1. bookVolume 5 (2021): Issue 4 (October 2021)
Journal Details
License
Format
Journal
First Published
30 Mar 2017
Publication timeframe
4 times per year
Languages
English
access type Open Access

An Empirical Study on Knowledge Aggregation in Academic Virtual Community Based on Deep Learning

Published Online: 24 Sep 2021
Page range: 372 - 388
Received: 25 Apr 2021
Accepted: 15 Jul 2021
Journal Details
License
Format
Journal
First Published
30 Mar 2017
Publication timeframe
4 times per year
Languages
English
Abstract

Academic virtual community provides an environment for users to exchange knowledge, so it gathers a large amount of knowledge resources and presents a trend of rapid and disorderly growth. We learn how to organize the scattered and disordered knowledge of network community effectively and provide personalized service for users. We focus on analyzing the knowledge association among titles in an all-round way based on deep learning, so as to realize effective knowledge aggregation in academic virtual community. We take ResearchGate (RG) “online community” resources as an example and use Word2Vec model to realize deep knowledge aggregation. Then, principal component analysis (PCA) is used to verify its scientificity, and Wide & Deep learning model is used to verify its running effect. The empirical results show that the knowledge aggregation system of “online community” works well and has scientific rationality.

Keywords

Background

With the development of Web2.0 technology and computer technology such as big data and cloud computing, academic virtual community has changed the traditional way of work, life, and education. Users have transitioned from single form of knowledge acquisition to multi-dimensional knowledge release and knowledge exchange. As a subdivision product of social networking, academic social network has become an important platform for research personnel, professional personnel, and technical personnel to share knowledge and facilitate academic exchange. ResearchGate (RG) is chosen as an example for empirical study. RG was founded in Germany in 2008, is the professional network for scientists and researchers, collaborates with academic researchers in the world, makes progress in academic virtual community, and has >20 million members to share their own research work (About, 2008). Then, as the representative of RG, user-generated content (UGC) in academic virtual community is growing rapidly, and the knowledge accumulation is quite considerable. Subject categories and an index of mutual citation by authors are used to organize and match knowledge in RG. Knowledge organization and matching system are simple. In general, knowledge push and matching correlation are finished according to the interest and research direction chosen by users. This results in a situation wherein the accuracy of push information is low, and the matching of correlation degree is broad. Therefore, the knowledge integration of academic virtual community presents the features of less dimension and shallow level. The lacunae in effective guidance for users’ knowledge utilization in academic virtual community restrain knowledge exchange and knowledge innovation. “Compared with some relative concept,” we think “knowledge aggregation is a new direction of knowledge organization, and it contributes to the realization of knowledge service based on user demand.” Knowledge acquisition, knowledge recommendation, and knowledge discovery are the elements of knowledge aggregation in the big data environment (Li, 2016, p. 128). The concept of “knowledge aggregation” in the field of knowledge organization refers to the use of data mining, knowledge mapping, deep learning, artificial intelligence, and other technologies to describe the external and internal characteristics of knowledge elements, integrate the disordered and scattered knowledge elements through given organization rules, promote the association and reorganization of multi-source heterogeneous knowledge elements, and devote attention to the deep semantic analysis and integration of knowledge content.

The deep and effective knowledge aggregation system based on semantic association relies on artificial intelligence to provide efficient, intelligent, and semantically relevant computer processing technology. Deep learning is an emerging research direction in the AI field and realizes highly efficient treatment of input data by imitating the human brain. Deep learning is used to extract features from input data efficiently, and it can even extract more abstract features from data, thus enabling identification and getting more essential features of data. In this paper, deep learning model Word2Vec is used to annotate the knowledge structure of multi-dimensional space (50-dimensional vector representation), which can improve the quality of feature extraction and better understand the sentences in natural language processing (NLP). In the meantime, Wide & Deep learning is used to construct the classification model of titles with good memorizations and generalization. The goal of the paper is to design scientific approach to realize the deep knowledge aggregation in virtual community. The contribution made by this paper is that deep learning is used not only to overcome the semantic disclosure of knowledge and the lack of grammatical information in the network community, but also to express grammatical information more accurately. Therefore, it integrates the shallow features of knowledge into deep feature concepts and knowledge contents, which can express the concepts at the semantic level and reveal the multi-hidden contents. We emphasize that the semantic technology and its related algorithms based on neural network are used to reveal the knowledge association and multi-level complex knowledge interaction in academic virtual community, explore the knowledge disclosure and content aggregation of academic virtual community based on semantic association, and realize the deep aggregation standardization of knowledge resources and automatic processing from both theory and practice.

Related Works

This paper studies the means to realize the deep knowledge aggregation in virtual community. In this section, the existing knowledge aggregation methods and the existing deep learning models that are applied in this paper are reviewed.

Traditional Methods

Scholars start exploring the new direction of knowledge aggregation from different dimensions to implement the ways of knowledge aggregation in academic virtual community, such as ontology-based, metadata, associated data, social tag and cluster analysis, and knowmetrics. For example, Mentzas, Kafentzis, and Georgolios (2007) mentioned that the concept of Web-based knowledge service can propose an ontology-based framework of knowledge aggregation to satisfy the software company by expanding the knowledge management (KM) concept externally. Abel, Marenzi,, Nejdl, and Zerr. (2009) reported that the success of recent Web 2.0 platforms shows that knowledge information aggregation strategy is based on tags and metadata for sharing flexible/relevant learning resources of social networks. Tarko and Aligica (2011) designed “aggregation systems that rely either on meta-experts or on computer algorithms,” and they explored “the possible of using them for setting up virtual thinking tanks for foresight studies”. “By trans parenting biological evolution theory”, a periodic evolution model is put up “for knowledge push network based on social network under the action of multiple dynamic mechanisms. Then, the whole evolution process of knowledge push based on latent social network is observed and analyzed empirically by using network structure methods” (Yi., Mao, Deng, & Cao 2014, p. 50). Li et al. (2011) introduced the TWC(The Tetherless World Constellation) LOGD(Linked Open Government Data) Portal to use linked datasets for large-scaling distributed data integration, collaborate data manipulation, and trans parenting data consumption. Furthermore, Yin & Wang (2014) “propose a collapsed Gibbs Sampling algorithm for the Dirichlet Multinomial Mixture model for short text clustering (abbr. to GSDMM),” and they “find that GSDMM can infer the number of clusters automatically” and “can obtain the representative words of each cluster”. Mu, Goulermas, Korkontzelos, and Ananiadou (2016) “propose a novel descriptive clustering framework, refer to as CEDL”, and “it discovers an approximate cluster structure of documents in the common space” to improve the effect of knowledge aggregation in academic virtual community. Bi & Liu (2017) “put forward the approaches of digital literature resources ontology aggregation and service recommendation, based on HowNet and literature resources, using a clustering analysis, semantic similarity computation, collaborative filtering algorithm, and other methods”. Based on knowledge aggregation from different scholars, Grisci, Krause, and Dorn. (2021) pointed out, “we present relevance aggregation, an algorithm that combines the relevance computed from several samples as learned by a neural network and generates scores for each input feature. For poorly trained neural networks, relevance aggregation helped identify incorrect rules or machine bias”. The methods of knowledge aggregation could be deduced based on metadata, ontology-based, associated data, social tag and cluster analysis, and knowmetrics. We present meaning and characteristics focusing on each method (Table 1).

Comparison of Knowledge Aggregation Methods

Method Meaning Characteristics
Metadata Describe the property of data and realize the unified integration of heterogeneous knowledge resources Simple and easy to use, Strong standardization, Weak semantics
Ontology-based Formal description of concept system to improve the machine-readable and understandable data Strong standardization, Formalized and conceptualized, Semantic relevance
Associated data Naming network objects with uniform resource identifier and data publishing and resource association through HTTP protocol Reveal the semantic meaning and relationship of information to a certain extent
Social tag and cluster analysis Simple and easy to use, strong freedom, business collaboration Poor standardization, Loose structure and fuzzy semantics
Knowmetrics Large amount of data processing, multi-dimensional and visualization Weak semantics, Relying on auxiliary tools and methods

Therefore, scholars and research institutions focus on the network knowledge aggregation implementation. In practice, it has also been promoted, but there are still some deficiencies in the current research. From Table 1 in terms of research methods, deep aggregation of concepts and content at the semantic level of user source knowledge cannot be implemented, and the deep semantic information cannot be revealed. The sampling group is small and lacks the ability of transforming shallow features of knowledge into deep features. These deficiencies provide research space for the knowledge aggregation in academic virtual community. We use deep learning to overcome the above barriers.

Deep Learning

Deep learning is a subset of machine learning in which multi-layered neural networks—“learn” features from substantial amounts of data. Within each layer of the neural network, deep learning algorithms perform calculations, make predictions repeatedly, and gradually improve the accuracy of the outcome over time – all without human intervention (Mueller & Massaron, 2019). Deep learning is a nonlinear combination of multi-layer representation learning methods. Representation learning is “learning representations of the data that make it easier to extract useful information when building classifiers or other predictors” (Bengio, Courville, & Vincent, 2013). From the outset of the raw data, deep learning methods compute the representation at each layer into a representation at a higher and slightly more abstract level (LeCun, Bengio, & Hinton, 2015). Deep learning model Word2Vector and Wide & Deep learning are used in the empirical study.

Recently, deep learning Word2Vector model has achieved remarkable success in various text mining problems such as sentiment analysis, knowledge classification, spam filtering, document summarization, and web mining. Word2Vector Model is proposed on “Efficient Estimation of Word Representations in Vector Space” (Mikolov, Chen, Corrado, & Dean, 2013, p. 4). At the same time, Google Inc. provides that Google has developed an efficient tool to implement the algorithm, “an efficient implementation of the continuous” word vectors, which is called “Word2Vector” (Introduction, 2013). Many NLP applications are promoted and simplified by Word2Vector model in a critical way (Mikolov et al., 2013, p. 7). “Many current NLP systems and techniques treat words as atomic units - there is no notion of similarity between words, as these are represented as indices in a vocabulary”, and the choices of NLP are simple, robust, and it is observed that “simple models trained on huge amounts of data outperform complex systems trained on less data” (Mikolov et al., 2013, p. 1). Semantic relationships are the improvement of Word2Vector existing in NPL applications, including machine translation, information retrieval, and question answering systems. Semantic relationships are beneficial to invent the future of NPL applications (Mikolov et al., 2013, p. 5). Also, the future applications can develop high-quality Word2Vector significantly (Mikolov et al., 2013, p. 10). Some deep learning architectures are utilized in a question classification task in a highly inflectional language, namely Turkish, which uses an agglutinative-language–based word structure. They have built word embeddings using the Word2Vec method with a continuous bag-of-words (CBOW) and skip-gram models with different vector sizes on a large corpus composed of user questions (Yilmaz & Toklu, 2020, p. 2909).

Wide & Deep learning is proposed by Google Play, a commercial mobile app store with a tremendous number of users and apps, with the published dissertation “Wide & Deep learning for Recommender Systems”. Cheng et al. mentioned the generalized linear models and deep neural networks with embeddings and uses jointly trained wide linear model (Cheng et al., 2016). Wide & Deep learning, which combines the wide component and deep component, has shown good performance in recommendation systems. In order to improve the accuracy of online learning platform recommendation for learner learning resources and to alleviate the cold start problem, an online learning resource recommendation method based on Wide & Deep and Elmo model are proposed. Wide & Deep is used to deeply explore the deep features of learner characteristics and course content features under the condition of high-dimensional data sparseness. In addition, for the learner's text feature, it will use the ELMo language model to pre-train the feature vector to improve the recommendation accuracy (Liu, Zhang, & Liu., 2020, p. 1). “The Wide & Deep learning is characterized by the lack of a research and experimental result on regression analysis”. The paper experiments with the application of “Wide & Deep learning on regression analysis” and also presents a “new Wide & Deep structure named WDSI”. The paper also shows that “the WDSI outperforms a traditional machine-learning and deep-learning models in regression analysis” (Kim, Lee, & Kim., 2020, p. 8).

Briefly, the shallow features of knowledge are integrated and abstracted into deep feature concepts and content, and the semantic-level concept expression and multi-hidden content disclosure are carried out with Word2Vector Model to realize the deep aggregation of knowledge resources in academic virtual community. Wide & Deep Learning combining deep neural network with the linear model can be regarded as a content-based method. Furthermore, due to the phenomenon that the two parts’ learning speed is not the same, we need to ensure that the two parts are both well-trained. The application field of the Wide & Deep learning is also expanding.

A Design Science Approach to Building Knowledge Aggregation System Based on Deep Learning

The purpose of knowledge aggregation is to build a multi-dimensional, multi-perspective, multi-granularity knowledge association system to provide a strong guarantee for user-oriented knowledge services. Therefore, knowledge aggregation focuses on the knowledge organization based on content association, and excavates the deep learning expression and association of knowledge resources.

The problem of knowledge aggregation in academic virtual community has always been the attention of scholars’ research. We focus on the integration of deep learning into the realization of knowledge aggregation, study how to describe knowledge resources more accurately from multi-dimensional space, and integrate massive multi-source heterogeneous data to form an orderly knowledge aggregation system.

Building Knowledge Aggregation System Based on Deep Learning

This paper uses Word2Vec model and Wide & Deep learning model. In the meantime, the Word2Vec model can intelligently interpret the semantic relationship between words and sentences. Wide & Deep model can optimize generalization ability and memory ability at the same time. With the utilization of two models, the syntactic and semantic features in the text are extracted with high quality, and the hidden representation of information is revealed through the contextual semantic relationship as well as realizing the multi-dimensional aggregation and accurate classification of knowledge resources.

Word2Vector Model

Word2Vector uses deep learning of recurrent neural network to practice language model. Word2Vector has two model architectures for computing vector representations of words, such as CBOW model and continuous skip-gram model (Skip-gram). “The CBOW architecture predicts the current word based on” the output of W(t), and “the Skip-gram predicts surrounding words” based on the input of W(t) (Mikolov et al., 2013, p. 4–5). In Figure 1, the word representations are input, projection, hidden and output layers, and hierarchical SoftMax, and the word representations can delete less than the certain frequent words and prioritize the evaluation of vocabulary size to improve the efficiency of the softmax normalization (Mikolov et al., 2013, p. 2–3).

Figure 1

The CBOW architecture model and the Skip-gram architecture model (Mikolov et al., 2013, p.5).

Figure 2

Wide & Deep learning model (Cheng et al., 2016).

We utilize the Skip-gram architecture model in the empirical study. In order to improve the training efficiency of the model, the strategies of deleting hidden layer, negative sampling and hierachy softmax, deleting words less than a certain word frequency, and optimizing word weight are adopted in the process of word training. Through the model training, the words are mapped to the abstract high-dimensional vector space. Each word is represented by a dense semantic vector, and the semantic computation can be carried out between the vectors, which can be used in the fields of similar word computation, word clustering, and semantic computation analysis. In this paper, Word2Vec is used to calculate the word vector first. After calculating the vectorization representation of each word, the vector of the word is used to calculate corresponding words related to its semantics. We can capture the semantics of words according to the context of words. This paper innovatively uses the title vector for semantic analysis, and analyzes the relationship mapping and knowledge association between words and words at more levels. Word2Vector makes computers “improve the existing techniques for estimating the word vectors” (Mikolov et al., 2013, p. 3).

Wide & Deep Learning Model

Wide & Deep learning model starts designing to facilitate with recommend order, but the concept of generalization and memorization is preserved to learn. Wide & Deep learning consists of two sections. One is wide. The other is deep. “Wide learning is a generalized linear model, such as logistic regression or linear regression” (Bastani, Asgari, & Namavari, 2019, p. 212). Memorizations are valuable to a wide learning model in the additional interaction term. “Memorization is defined as learning the frequent interactions of features from the historical data” (Bastani et al., 2019, p. 210).

Deep learning refers to deep neural networks. Deep neural networks can generalize to unseen feature combinations and make the model achieve excellent generalization. Generalization refers to exploiting new feature combinations that have never or rarely occurred in historical data. Generalization has the adoptability of new samples to add new datasets on the original datasets by predicting reasonable output in the model training. Generalization has the better adoptability of new samples by predicting reasonable output in the model training. The model with excellent generalization can learn hidden rules in data. Hence, for creating a generalized model, it uses complex neural network with various hidden layers to predict the output, including heavy heterogeneity and mixed consequences.

Wide & Deep learning model can obtain both generalization and memorization at the same time. Theoretically, Wide & Deep learning model can achieve accurate classification and wider coverage.

Building Implementation Process of Knowledge Aggregation System

Based on the above theory, we construct the implementation process of knowledge aggregation system of academic virtual community based on deep learning (Figure 3).

Figure 3

The implementation process of the knowledge aggregation system in academic virtual community based on deep learning.

First, the Word2Vector model is used to train the big data that has been cleaned and participled, to calculate the word vector, to sum or average the vectors of all the words in the title, and to obtain the new title vector. Second, cosine distance is used to measure the semantic similarity of vectors; it analyzes the semantic mapping and knowledge association between words and vectors. Third, K-Means is used to cluster the semantic vectors of titles to find the knowledge content system after semantic aggregation.

Principal component analysis (PCA) is used to test the scientific nature and application of aggregation system by visual display. The operation effect of aggregation system is tested by Wide & Deep learning model, which can prioritize both generalization and memorization at the same time. That is, for any source, new publications according to its title can be automatically carried out clustering navigation. In this way, the whole knowledge aggregation content system can be prioritized by realizing user-oriented knowledge accurate push.

Empirical Study

“RG is the most popular academic social networking site (ASNS)” in the world and serves global scientific research users. After users register on RG, a profile page shows users’ information, such as brief biography, research items, and so forth. RG has networking skills with users, such as following users’ accounts, asking questions, and answering users’ questions. “Statistical information is available, such as data on readership, citations, recommendation counts for research items, and the numbers of questions and answers.” Notable user's scientific reputation shows the measurement of an RG score (Lee, Oh, Dong, Wang, & Burnett, 2019, p. 566–567). As a result, RG resource integration and information matching have the characteristics of generalization, but it is difficult to definitively improve the accuracy of information pushes. The purpose of this study is to explore the disordered knowledge in academic virtual community for effective semantic disclosure and knowledge link and to further realize the deep aggregation of knowledge content. Through big data download and the reveal of semantic vector, RG is selected as an example for the implementation of knowledge association and content depth aggregation. We refine the RG classification resources for user matching and push information precisely. All work is finished according to Figure 3.

Data Cleansing and Preparation

To verify the scientific and effective organizational structure of knowledge aggregation in network community based on deep learning, we take the RG (Find and share, 2008) website as an example; choose a “publications” model; search a keyword “Online Communities”; and use Python 3.6 to run 100,000 pieces of data, including authors, titles, abstract, etc. Then, the 100,000 pieces of data are cleaned and participled by Python. Similarly, meaningless words and punctuation marks are filtered to ensure the high purity of the corpus and a better effect of the word vector. Therefore, preprocessing could improve the effect of word vector when using corpus to train word vector.

Data Analysis

Word2Vector is used to reveal semantic vector. Through the training of Word2Vector, word vectors make full use of the contextual information of words and express good semantic features. The information of each dimension in the vector emphasizes the semantic information. We reveal the semantic information through the 50 dimensions of words. Semantic similarity is calculated by cosine distance between word vectors. Therefore, the word vector and semantic similarity based on Word2Vector can reveal the semantic vocabulary and the correlation of words from multi-dimensional space. Furthermore, it reveals the knowledge association between words and titles.

Calculation of the Word Vector Based on Word2Vector

“A neural network is constructed, where its input is a sequence of word embedding vectors generated from Convolutional Neural Networks (CNN)”. Afterward, the researchers mention different training models of word vectors. In the meantime, Thomas Mikolov states that “Word2Vec is a well-known word”. It maps words to a vector representation space and transforms the relationship between words and “surrounding words” into a dense vector. For similar words, the corresponding word vectors are similar. Using word vectors can solve many fields of machine learning, such as computer vision, speech recognition, and NLP (KLUNGPORNKUN & VATEEKUL, 2018, p. 121).

There are two kinds of methods obtaining a set of word vectors, which are statistical method and language-based model. The more famous method has neural network language model (NNLM), Skip-gram, CBOW, etc. We use Word2Vector to obtain the characteristics of the words, which depend on large-scale corpus to obtain accurate word vectors. In comparison, we collect titles of 100,000 retrieval data, clean and participle each title, concatenate all the titles after processing, and form a word sequence as the input data of the Word2Vector. The main parameters are that the word dimension takes 50 dimensions (i.e., 50 dimensions per word vector), the context window takes 5 words, the words appear at least 10 times, and the number of iterative model is 10 times. The result of the calculation is that there are 7988 words with word vector in 100,000 pieces of data.

Examples of word vectors are as follows (Table 2), showing only the first five dimensions, and the remaining dimensions are omitted:

Examples of Word Vectors

Word Dimension 1 Dimension 2 Dimension 3 Dimension 4 Dimension 5 Dimension 6–50
Online 0.034224428 0.43261373 0.259087235 −0.489322513 −0.007246093
Community −0.088755406 −0.020659156 −0.030128725 0.30018574 −0.380684167
Communication −0.122577295 0.214154214 0.150527656 0.252177447 −0.154901862
Communities −0.097516797 0.141274527 0.143667549 0.444990486 −0.327237487
Social −0.210871235 0.029080199 −0.112222128 −0.071832731 −0.195879236
Study −0.03604706 −0.101546846 −0.345023751 −0.303048939 −0.532510221

The purpose of clustering word vectors is to discover research topics (Table 3). The norm is used to select the representative words of each category after clustering. Table 3 takes some words with the largest norm (from big to small) as the representative words, which can best represent the research content of this category. The representative words and their norm of the clustering results are shown in the following table:

The Exhibition of Word Clustering Results

Line number Represent words Norm
1 Guinea 4.361
2 Lanka 4.2377
3 Sectional 4.2046
4 Congo 4.1324
5 Torres 4.1199
6 Strait 4.1101
7 Columbia 4.0069
8 Islander 3.9934
9 Leone 3.8981
10 Sierra 3.8625

Because of too fine word granularity, the category of research theme is hard to be refined accurately according to representative words. The clustering effect is not obvious based on word vector. Due to the composition of words, a title with richer semantic content covers more information and extracts the semantic relationship between words easily. Therefore, we further construct the title vector and use words with conceptual meaning in the title to analyze the semantic meaning of the title vector (Zhan & Dahal, 2017, p. 4). “Conceptual meaning words” are selected according to the norm length of the word vector in each category, and some words with the largest norm length are selected, namely representative words. So, “conceptual meaning title” is selected according to the norm of the title vector in each category, and some titles with the largest norm are selected.

Semantic Analysis of Title Vector

Word vectors can be calculated by addition and subtraction to calculate the similarity between sentences. Therefore, a new semantic vector is obtained by calculating the addition or subtraction of the word vector, which can express the meaning of the combination of words involved in the calculation. Based on word vector, a new semantic vector of title is obtained by calculating the sum and average of all word vectors in the title.

In this empirical study, we find that it is the uneven degree of heat and cold due to the inconsistency of the number of words in the training corpus. The norm of the calculated vector is quite different, and the norm of the words with high frequency is large. In comparison, the norm of the words with low frequency is small. If the vector is added or averaged directly, the calculated vector is easily affected by the large norm, and this phenomenon gives rise to what are known as “hot words.” Hot words often appear in subsequent application. Therefore, we mention the following calculation methods below to construct the title vector to eliminate the word hot degree for the effect of constructing title vector.

Norm Normalization of Word Vectors;

Assume a word vector is (x1, x2, ⋅ ⋅ ⋅, xm); the norm of this vector is norm=(i=1mxi2)12 {\rm{norm}} = {\left( {\sum\limits_{i = 1}^m {{\rm{x}}_{\rm{i}}^2} } \right)^{{1 \over 2}}} The word vector after norm normalization is (x1',x2',,xm') \left( {x_1^\prime,\,x_2^\prime, \cdots ,x_m^\prime} \right) in which xi'=xi/norm x_i^\prime = {x_i}/norm , i = 1,2, ⋯, m.

Sum the Vectors of all the Words in the Title;

Assume there are two-word vectors; the normalized vectors are (x11,x21,,xm1) \left( {x_1^1,\,x_2^1, \ldots ,x_m^1} \right) and (x12,x22,,xm2) \left( {x_1^2,\,x_2^2, \ldots ,x_m^2} \right) Hence, the sum of the word vector is (x11+x12,x21+x22,,xm1+xm2) \left( {x_1^1 + x_1^2,\,x_2^1 + x_2^2, \cdots ,x_m^1 + x_m^2} \right) and the sum of multiple vectors and so on.

Norm Normalization of Title Vectors;

The calculation results of (2) are normalized according to the method from (1).

We randomly choose six examples and present the first five-dimensions of title vector. The remaining dimensions are omitted (Table 4):

Title Vector Example

Title Dimension 1 Dimension 2 Dimension 3 Dimension 4 Dimension 5 Dimension 6–50
A practical guide to the development of an online course in adapted physical education −0.214557097 0.451087397 −0.089244092 −0.105165435 −0.064529825
Globalization populism conspiracism −0.100888215 −0.061636729 −0.163176164 0.05720861 −0.158507724
Cohort study evaluating pressure ulcer management in clinical practice in the UK following initial presentation in the community costs and outcomes −0.094149069 0.058877724 −0.01750199 −0.176174027 −0.150949746
A comparative analysis of metabarcoding and morphology-based identification of benthic communities across different regional seas −0.124478352 0.057873331 0.134045295 0.070697202 −0.062715003
Massive mimo antenna array deployment for airport in air-to-ground communications 0.080575986 0.146409381 0.097911457 −0.022636811 −0.045334639
Online brand community within the integrated marketing communication system when chocolate becomes seductive like a person −0.031164095 0.178977157 0.376543319 −0.280455168 −0.034272819
Measuring Semantic Similarity

Similarity at semantic level should consider the relations between words. Semantic similarity measurement relates to computing the similarity between terms or short text expressions that carry the same meaning or related information but are not lexicographically similar (Martinez-Gil & Montes, 2013, p. 399–400). Semantic similarity is one of the key research contents in artificial intelligence and NLP. Semantic similarity measurement can increase an efficient data matching and supply a precision service. Hence, semantic similarity measurement is widely used in NLP, information retrieval, and other fields.

According to the source of semantic information, semantic similarity measurement is divided into corpus-based and knowledge-based. Corpus is an enormous collection of text. Extracted text features in corpus are used for constructing feature vectors or statistical models. “Firstly, the continuous word vectors were trained from a textual corpus in advance by the neural network language model in deep learning. Then multiple semantic information and relationship information were extracted from corpus to augment original vectors and generate sense vectors for words. Hence, the semantic similarity between concepts can be measured by the similarity of sense vectors”(Lu, Cai, Che, & Lu, 2016, p. 311–312). We collect 100,000 pieces of data, construct a huge amount of textual corpus, extract text features, and construct feature vectors.

The distance between vectors can reflect the similarity of vectors. The common distance is Euclidean distance, cosine distance, and so on.

We use cosine distance to measure the similarity of the vectors. Calculation formulas are below:

Suppose there are two-word vectors, the normalized vectors are (x11,x21,,xm1) \left( {{\rm{x}}_1^1,\,{\rm{x}}_2^1, \cdots ,{\rm{x}}_{\rm{m}}^1} \right) and (x12,x22,,xm2) \left( {{\rm{x}}_1^2,\,{\rm{x}}_2^2, \cdots ,{\rm{x}}_{\rm{m}}^2} \right) Then, the cosine calculation formula is i=1mx11*x12 \sum\limits_{i = 1}^m {{\rm{x}}_1^1*\,{\rm{x}}_1^2} Because of the substantial number of words, the similarity calculated between each word (title) and all words (title) is too much. Based on big data processing software spark, the paper calculates the cosine similarity of each word (title) and all words (title) by method of matrix multiplication. Each line of the matrix represents a vector of words (titles) when building a matrix. The result of vector multiplication is cosine similarity after the norm of the vector needs to be normalized.

We calculate the similarity of each word and all words first. Then, the similarity is arranged from large to small, and the first several words are taken as the most similar words. In the same way, the title similarity can be calculated and analyzed.

Taking “health” as a similarity example between words and words, the 10 most relevant words are as follows (Table 5).

Word Similarity Example Exhibition

Rank Similar words of health Similarity
1 Healthcare 0.76234574
2 Workers 0.71603154
3 Care 0.69574523
4 Interventions 0.6920655
5 Mental 0.68639633
6 School-based 0.68160605
7 Preventive 0.68005763
8 Visiting 0.65810901
9 Medicare 0.65752329
10 Disparities 0.65668325

Taking “online brand community across cultures” as a similarity example between titles and titles, the 10 most relevant titles are shown in the following table (Table 6).

Title Similarity Example Exhibition

Rank Similar title of “online brand community across cultures” Similarity
1 Online brand communities loyal to the community or the brand 0.86801413
2 Brand evangelism among online brand community members 0.86336496
3 Online brand communities 0.85512465
4 Materiality of online brand community 0.85083147
5 Cultural differences in online community motivations exploring Korean automobile online brand communities(KAOBCs) and American automobile online brand communities(AAOBCs) 0.85036181
6 It's not a shoe it's a community – varumärkesupplevelserpå online brand communities 0.84567247
7 Online brand communities :there is more than one way to drive consumers’ online brand identity and interactivity 0.83576148
8 Nostalgia in online brand communities 0.83419945
9 Online brand community across cultures :a comparison between the US and Korea 0.83289393
10 Online brand community practices and the construction of brand legitimacy 0.8318584

The range of cosine is (−1, 1). The larger the number is, the larger the similarity is. In the meantime, 1 represents completely similarity, but −1 represents no similarity.

Semantic similarity can be calculated between title vectors. When the number between title semantic similarity is higher, the similarity of title is closer. Similarity calculation is the basis of clustering. We cluster similar titles into a class by K-Means++ method to implement similarity research topic clusters.

Based on word vector, we carry on the calculation of title vector for some purposes, such as annotating the semantic relationship between words and words more accurately; realizing the dense vector reflection of words concept, relation, attribute and so on; establishing the vector spatial mapping relation between elements; and obtaining the new title semantic vector. On this basis, we use K-Means++, “perhaps an even better initialization strategy (though not one implemented in the visualization)” (Visualizing K-means, 2014), to cluster the semantic vector of title and realize content aggregation based on title vector.

Knowledge Clustering

According to the idea of “clustering objects”, objects with the same feature are clustered into the same class. “Clustering is an example of unsupervised learning, in which we work with completely unlabeled data (or in which our data has labels, but we ignore them)” (Grus, 2019). Similarly, clustering refers to enabling an algorithm to recognize these clumps of points without help. The K-Means algorithm plays a vital role in clustering analysis. “The K-Means algorithm captures the insight that each point in a cluster should be near to the center of that cluster”. We start choosing k in various ways such that we want to find “the number of clusters in the data”. Then, centroids are the centers of those k clusters, “which are initialized in some fashion”. K-Means works best with the roughly same sized and shapes clusters in dataset. “Despite the fact that K-Means is guaranteed to converge, the final cluster configuration to which it is not coveraged in general unique and depends on the initial centroid locations” (Visualizing K-means, 2014).

K-Means++ is an improvement of the K-Means clustering algorithm in data mining. Based on the K-Means clustering algorithm, it made some improvements in choosing the k initial cluster center. The basic idea is that the k initial cluster centers should be as far away as possible. “The exact algorithm is as follows:”

“Choose one center uniformly at random among the data points”.

“For each data point x not chosen yet, compute D(x), the distance between x and the nearest center that has already been chosen”.

“Choose one new data point at random as a new center, using a weighted probability distribution where a point x is chosen with probability proportional to D(x)”.

“Repeat Steps 2 and 3 until k centers have been chosen”.

“Now that the initial centers have been chosen, proceed using standard K-Means clustering” (K-means++, 2020).

Taking the title vector as the input data, K-Means++ algorithm is used to cluster it. The number of clustering categories is set to 10, and the number of iterations of the model is 100. The clustering center vectors of 10 categories are obtained, as shown in Table 7.

Title Clustering Center Vector

Index Dimension 1 Dimension 2 Dimension 3 Dimension 4 Dimension 5 Dimension 49 Dimension 50
0 −0.0900 0.2218 0.0692 −0.0649 −0.1430 −0.0370 0.0318
1 −0.0552 0.2391 0.0992 −0.0776 −0.1130 −0.1805 0.0725
2 −0.0747 0.1618 0.0813 −0.0415 −0.1946 −0.0465 −0.0295
3 0.0175 0.2085 0.0767 −0.1440 −0.0342 −0.0199 −0.0804
4 0.0431 0.1801 0.1236 −0.1602 0.0378 −0.0791 −0.0500
5 −0.0503 0.1121 0.0518 0.0124 −0.1737 −0.0669 −0.1022
6 −0.0742 0.1920 0.0417 −0.1041 −0.0618 −0.1310 −0.2084
7 −0.1194 0.1587 0.1803 −0.2660 −0.1028 −0.0764 −0.0608
8 −0.0110 0.2546 0.1412 −0.0924 −0.1782 0.0270 −0.0244
9 −0.0151 0.1338 0.0822 −0.0194 −0.1065 0.0194 0.1752

The clustering center vector is the vector corresponding to the center point of each category. The clustering center vector is generally obtained by calculating the mean value of the vectors of all points belonging to the category. If the distance between a point and a clustering center vector is the shortest, the point belongs to this category. By clustering the title vector, we can get 10 clustering center vectors. Each title only belongs to one of the categories. That is to say, the distance between the title vectors corresponding to the clustering center vector is the shortest. The result of K-Means++ is demonstrated in Table 7. According to the title clustering center vector, it is initially clustered into 10 categories, and then further optimized the clustering results.

Knowledge Aggregation System

The knowledge content is clustered according to title semantic vector. After the clustering calculation is completed, the concept extraction and semantic association are carried out according to the high-frequency keywords in knowledge content. Because of the variety of expressions of the same or similar words, it is necessary for these to be examined and counted by people. Table 8 shows the high-frequency keywords in the brand community. The keyword is combined to synonyms, such as communities 830.0, community 727.0 as community 1557; and behavior 365.0, behaviour 101.0 as behavior 466. Table 8 clearly shows that under the “online community” research category with “brand” as the highest frequency keyword, it can be combined with the research category and representative title and can be determined as “brand community” for this category of research topics.

High-frequency Vocabulary of “Brand Community”

Word Online Community Brand Behavior Consumer Loyalty Participation Trust
Freq 4692 1557 895 466 460 459 452 447
Word Intention Knowledge Role Perspective Communication Sharing Interpersonal Study
Freq 423 423 214 417 409 375 366 339

After determining the research topic represented by a certain category, the first 50 titles with the largest norm are taken as the representative titles according to the norm of the unorganized title vector, and the research topics are analyzed again. Table 9 shows the first 10 records here and norm refers to the norm of title vector. The larger the norm is, the higher the correlation is between the title and the topic. At the same time, it indicates that the title is more representative of the research content of the topic. The categories of topics are ensured according to high-frequency keywords and representative titles in the same way.

“Brand Community” Representative Title Exhibition

Topic Line number Represent titles Norm
Brand community 1 Understanding consumer intention to participate in online travel community and effects on consumer intention to purchase travel online and WOM: An integration of innovation diffusion theory and TAM with trust 26.793
2 An empirical study on the relationship between online shopping mall characteristics and consumers repurchase intention – focused on Chinese consumers involved in overseas direct purchasing 25.6124
3 Is online consumers’ impulsive buying beneficial for e-commerce companies? An empirical investigation of online consumers’ past impulsive buying behaviors 23.8119
4 An empirical study of website personalization effect on user's intention to revisit e-commerce website through cognitive and hedonic experience: Proceedings of ICDMAI 2018 volume 2 20.7455
5 Digitalisation luxury fashion and “Chineseness”: The influence of the Chinese context for luxury brands and the online luxury consumers experience 20.1041
6 The effect of electronic word of mouth on brand image and purchase intention: An empirical study in the automobile industry in Iran 20.0951
7 A study on the effects of the attractiveness and credibility of online 1 personal media broadcasting B.J. on the viewing engagement perceived on media channel, interactivity, perceived enjoyment, and the user's responses. 19.3827
8 Predicting consumer purchase intention on fashion products in online retailer integration of self-determination theory and theory of planned behavior 18.9955
9 Extending expectancy violations theory to multiplayer online games: the structure and effects of expectations on attitude toward the advertising attitude toward the brand and purchase intent 18.4863
10 Negative online reviews of popular products: understanding the effects of review proportion and quality on consumers’ attitude and intention to buy 17.7826
Result

After data analysis, 10 categories are formed at first. But by observing the clustering results, it is found that there is high similarity between some categories. Some categories cover larger and more complete topics while some categories cover more professional and one-sided topics. So, it is difficult to refine each category of research topics and difficult to operate.

According to the possible hierarchical clustering relationship between these observations, we try to cluster five catalogues as the first class (large class) and cluster five catalogues as second class (small class) under each first-class catalogue. Thus, the knowledge aggregation system is formed. It is necessary to evaluate and optimize the clustering effect. After clustering, the center vector of each category is used as the vector of the category. The method in the optimization of clustering system is to compare semantic similarity between each category. If the similarity of two categories is higher than a certain threshold value, such as 0.8, it is thought that the two categories with high similarity need to be merged.

While computing the similarity of the first-class category, the similarity between most categories is low. The similarity between the catalogue “online health” and the catalogue “community intervention” is 0.805. Hence, comparing and analyzing “online health” and “community intervention” results in the finding that second-class category highly coincides. So, the two categories are merged.

After comparing the similarity of all the second-class categories under each first-class category in turn, it is found that the similarity between many categories is high. The reason is that these second-class categories belong to the same first-class category. Therefore, high-frequency keyword is directly adopted in the optimization of second-class categories.

“Online Community” Knowledge Aggregation System based on deep learning (Figure 4) consists of five first-class categories, which are semantic analysis, networking communication, online medical treatment, online health, and brand community. Therefore, we can construct the first-class categories system. After the first-class categories are determined, all samples under each first-class category are clustered by the same reason to get a more detailed category, namely the second-class categories. Five second-class are formed under the five first-class categories (Figure 4).

Figure 4

“Online community” knowledge aggregation system.

Testing Results

The “online community” knowledge system based on deep learning takes the world's most popular ASNS, RG, as an example. Downloading big data, revealing semantic vector between titles, clustering by knowledge association, and determining the category theme by high-frequency words have ensured the scientific nature of the aggregation system theoretically and experimentally, and realized knowledge association and the deep content aggregation. The empirical result is “online community” as a big category resource, which can be refined into five first-class categories and 25 second-class categories (Figure 4). To test the empirical results, the PCA model is used to visualize the knowledge aggregation system, and the Wide & Deep learning model is used to test the operation effect of the knowledge system.

Visual Display of Knowledge Aggregation System Based on PCA

We must first standardize, then perform PCA, extract the desired number of Principal Components (PCs), and finally use those PCs as input features to neural networks (known as neural networks with feature extractions). PCA is a useful and powerful tool to help us work with high-dimensional problems (many variables) and visualize data. Executing PCA is simple with scikit-learn to fit the model and apply the transformation for data analysis. PCA is sometimes referred as a feature extraction method based on many correlated input features to a predictive model (Yurko, 2020). “PCA is used widely in dimensionality reduction” (Dimensionality Reduction, n.d.), especially the visual exhibition of multi-dimensional data. If multidimensional data often encounters dimension greater than 3, it cannot be visualized. However, specifying the number of components is important to the PCs scores. The first two PCs are focused on visualization purposes according to the PC score. Based on the principle of PCA dimensionality reduction, we use PCA to visualize the scientific and reasonable nature of the “online community” knowledge content aggregation system based on deep learning. At the same time, this step can verify the reliability of the training samples based on Wide & Deep learning classification model.

The effect of classification model depends on the scientific characteristics and complexity of the model and the quality of training data. Traditional clustering model requires artificial tagging training data, but artificial tagging is slow and expensive. Therefore, Wide & Deep model is used for automatic generation of training data when it chooses training data. The specific methods are as follows:

Construct word vector of all titles (100,000 Pieces of Data).

Compute semantic similarity between title vector and cluster center vector (See Section 4.2.3).

The most relevant cluster center in ② is chosen, and the semantic similarity is greater than a certain threshold, such as 0.7. The title is used as a reliable training sample for the cluster center.

The training model predicts the title that is not be participated in the training.

To verify the scientific of the knowledge system, 15,358 training samples generated by the above methods in five first-class categories are compressed from 50-dimensional space to 2-dimensional space by PCA model in spark ml. After it is compressed to 2-dimensional space, 500 training samples are randomly chosen and covering 5 first-class categories. The scatter diagram is visualized by matplotlib.pyplot, as shown below.

In Figure 5, stars represent semantic analysis, squares represent network communication, pentagons represent online medical treatment, “+” represent online health, and triangles represent brand community. Due to the compression from 50-dimensional space to 2-dimensional space, a small spatial error is produced. It is displayed that the sample of five first-class categories is distinguished perceptibly.

Figure 5

“Online community” knowledge aggregation system visualization.

Some indicators such as AUC, recall rate, precision, etc. are applicable in the binary classification model when training with Wide & Deep model in TensorFlow. But in the multi-classification model, only the accuracy rate as the model calculation results is given. Therefore, this paper only lists the calculation results of the model accuracy.

The calculation results of the model accuracy are listed as follows:

The accuracy of the “online communities” category is 0.97112024. The accuracy rates of the first class categories are semantic analysis 0.90708065, network communication 0.8808989, online medical treatment 0.8619744, online health 0.8993711, and brand community 0.8224044. Hence, the accuracy of the broad category and the first-class category is remarkably high, and the knowledge aggregation system is scientific and reasonable. Reliability of training samples of Wide & Deep learning model is synchronically verified.

Operation Effect of Knowledge Aggregation System Based on Wide & Deep Learning Model
Wide & Deep Learning Model Application

The application of knowledge aggregation system is to build intelligent navigation system – that is, for any new date, according to its title, it can automatically cluster the corresponding research fields. Combined with the idea of Wide & Deep learning model, we integrate the model with knowledge aggregation system to construct the classification model of titles. When building the Wide & Deep learning model, the word of the title is used as the feature of the wide part to capture the literal meaning of the title; the semantic vector of the title is used as the deep part to capture the semantic meaning of the title.

The model figure is as follows below:

The words in title are represented from word 1 to word 10 in Figure 6. Each title gets at most 10 words. The choice principle is based on the frequency of each word in the title from high to low. If >10 words are present in title, the 10 words with the highest frequency are chosen. If <10 words are present in title, the default symbol (usually a question mark) is used to fill in 10 words. Chosen words from left to right are used as the input of the wide part of the model.

Figure 6

Title model based on Wide & Deep learning.

The semantic vector of title is represented from dense1 to dense50 in Figure 6, corresponding to the values of the 1st dimension to the 50th dimension of the semantic vector of the title in turn.

In Figure 6, multi-class represents predict category. When the first-class category is clustered, the category corresponds to five categories in the knowledge aggregation system. Similarly, when the second-class category is clustered, it corresponds to the second-class category under each first-class category – that is, five 5-category models.

Operation Effect Test of Knowledge Aggregation System

As the world's most popular research social network website, 20 million research users share their research results on the platform, and user-generated content is growing rapidly in RG. Therefore, the efficient knowledge aggregation system can quickly cluster multiple elements of latest information from users. We take “semantic analysis” as an example to verify the clustering effect for any new data(publications). The predict results of several publications are shown as follows:

As a result, the examples in Tables 10 and 11 show that the fusion of Wide & Deep Learning model and the “online community” knowledge content aggregation system based on deep learning can cluster publications quickly and accurately. In this way, RG online community publications in the knowledge base can be clustered into five first-class categories and 25 second-class categories.

Taking second-class category under “semantic analysis” as an example, the prediction results of randomly selected publications are as follows:

Title First-class Category Automatic Cluster Sample Exhibition

Title Prob. Category
Using stock prices as ground truth in sentiment analysis to generate profitable trading signals 0.907 Semantic analysis
Application of text mining techniques to the analysis of discourse in eWOM communications from a gender perspective 0.976 Semantic analysis
Creating an Arabic dialect text corpus by exploring Twitter, Facebook, and online newspapers 0.926 Semantic analysis
N-Gram representations for comment filtering 0.968 Semantic analysis

Title Second-class Category Automatic Cluster Sample Exhibition

Title Prob. Category
Utilizing text mining and Kansei engineering to support data-driven design automation at conceptual design stage 0.991 Text mining
What affects patients’ online decisions: An empirical study of online appointment service based on text mining: International conference ICSH 2018 Wuhan, China, July 1–3 2018 proceedings 0.991 Text mining
Optimizing cost for geo-distributed storage systems in online social networks 0.902 Social network analysis
Vulnerability of D2D communications from interconnected social networks 0.933 Social network analysis
Conclusions and Discussions

The research of knowledge aggregation in academic virtual community based on deep learning takes RG as an example, changes the single method of using Word2Vec to calculate word vector, converts word vector feature representation to higher layer title vector feature representation from raw data, creatively uses this model to calculate more semantic title vector, and learns the 50-dimensional vector representation of each title to label the semantic meaning of the title more accurately in order to find the complex knowledge association structure in high-dimensional data. The knowledge aggregation system is constructed empirically by big data. Furthermore, PCA and Wide & Deep learning model are used to verify its scientific and operation effects. The empirical results show that the “online community” knowledge aggregation system is scientific and reasonable, and the operation effect is good. Briefly, the process and method of knowledge aggregation system that is based on deep learning supply new angles and ideas in knowledge aggregation of academic virtual community.

This research has the deficiencies listed below that require to be addressed in the future research work, so as to ensure that the line of research elucidated in this paper is further optimized:

The deep learning model Word2Vector carries deficiencies in its model. The model cannot dispel ambiguity, and each word corresponds to a vector. It is difficult to learn effective feature vector representation for the features of words with low frequency. The Word2Vector model only selects the words in a certain window when considering semantic relationship of words in the context. The use of context information is limited, and the co-occurrence of word frequency information is not considered from the global perspective.

Unsupervised word vector training cannot make beneficial use of prior information. It is difficult to generate title vector accurately based on word vector. The title vector is obtained by the sum of word vectors. When there are too many words in title, the effect of sum may not be good.

The theme of aggregation classification is subjective. The theme of aggregation classification is clustered according to high-frequency keywords and representative title. The statistics of high-frequency keyword synonyms and single/plural word need to be finished by people.

Figure 1

The CBOW architecture model and the Skip-gram architecture model (Mikolov et al., 2013, p.5).
The CBOW architecture model and the Skip-gram architecture model (Mikolov et al., 2013, p.5).

Figure 2

Wide & Deep learning model (Cheng et al., 2016).
Wide & Deep learning model (Cheng et al., 2016).

Figure 3

The implementation process of the knowledge aggregation system in academic virtual community based on deep learning.
The implementation process of the knowledge aggregation system in academic virtual community based on deep learning.

Figure 4

“Online community” knowledge aggregation system.
“Online community” knowledge aggregation system.

Figure 5

“Online community” knowledge aggregation system visualization.
“Online community” knowledge aggregation system visualization.

Figure 6

Title model based on Wide & Deep learning.
Title model based on Wide & Deep learning.

Title Similarity Example Exhibition

Rank Similar title of “online brand community across cultures” Similarity
1 Online brand communities loyal to the community or the brand 0.86801413
2 Brand evangelism among online brand community members 0.86336496
3 Online brand communities 0.85512465
4 Materiality of online brand community 0.85083147
5 Cultural differences in online community motivations exploring Korean automobile online brand communities(KAOBCs) and American automobile online brand communities(AAOBCs) 0.85036181
6 It's not a shoe it's a community – varumärkesupplevelserpå online brand communities 0.84567247
7 Online brand communities :there is more than one way to drive consumers’ online brand identity and interactivity 0.83576148
8 Nostalgia in online brand communities 0.83419945
9 Online brand community across cultures :a comparison between the US and Korea 0.83289393
10 Online brand community practices and the construction of brand legitimacy 0.8318584

Comparison of Knowledge Aggregation Methods

Method Meaning Characteristics
Metadata Describe the property of data and realize the unified integration of heterogeneous knowledge resources Simple and easy to use, Strong standardization, Weak semantics
Ontology-based Formal description of concept system to improve the machine-readable and understandable data Strong standardization, Formalized and conceptualized, Semantic relevance
Associated data Naming network objects with uniform resource identifier and data publishing and resource association through HTTP protocol Reveal the semantic meaning and relationship of information to a certain extent
Social tag and cluster analysis Simple and easy to use, strong freedom, business collaboration Poor standardization, Loose structure and fuzzy semantics
Knowmetrics Large amount of data processing, multi-dimensional and visualization Weak semantics, Relying on auxiliary tools and methods

Title Vector Example

Title Dimension 1 Dimension 2 Dimension 3 Dimension 4 Dimension 5 Dimension 6–50
A practical guide to the development of an online course in adapted physical education −0.214557097 0.451087397 −0.089244092 −0.105165435 −0.064529825
Globalization populism conspiracism −0.100888215 −0.061636729 −0.163176164 0.05720861 −0.158507724
Cohort study evaluating pressure ulcer management in clinical practice in the UK following initial presentation in the community costs and outcomes −0.094149069 0.058877724 −0.01750199 −0.176174027 −0.150949746
A comparative analysis of metabarcoding and morphology-based identification of benthic communities across different regional seas −0.124478352 0.057873331 0.134045295 0.070697202 −0.062715003
Massive mimo antenna array deployment for airport in air-to-ground communications 0.080575986 0.146409381 0.097911457 −0.022636811 −0.045334639
Online brand community within the integrated marketing communication system when chocolate becomes seductive like a person −0.031164095 0.178977157 0.376543319 −0.280455168 −0.034272819

High-frequency Vocabulary of “Brand Community”

Word Online Community Brand Behavior Consumer Loyalty Participation Trust
Freq 4692 1557 895 466 460 459 452 447
Word Intention Knowledge Role Perspective Communication Sharing Interpersonal Study
Freq 423 423 214 417 409 375 366 339

Examples of Word Vectors

Word Dimension 1 Dimension 2 Dimension 3 Dimension 4 Dimension 5 Dimension 6–50
Online 0.034224428 0.43261373 0.259087235 −0.489322513 −0.007246093
Community −0.088755406 −0.020659156 −0.030128725 0.30018574 −0.380684167
Communication −0.122577295 0.214154214 0.150527656 0.252177447 −0.154901862
Communities −0.097516797 0.141274527 0.143667549 0.444990486 −0.327237487
Social −0.210871235 0.029080199 −0.112222128 −0.071832731 −0.195879236
Study −0.03604706 −0.101546846 −0.345023751 −0.303048939 −0.532510221

“Brand Community” Representative Title Exhibition

Topic Line number Represent titles Norm
Brand community 1 Understanding consumer intention to participate in online travel community and effects on consumer intention to purchase travel online and WOM: An integration of innovation diffusion theory and TAM with trust 26.793
2 An empirical study on the relationship between online shopping mall characteristics and consumers repurchase intention – focused on Chinese consumers involved in overseas direct purchasing 25.6124
3 Is online consumers’ impulsive buying beneficial for e-commerce companies? An empirical investigation of online consumers’ past impulsive buying behaviors 23.8119
4 An empirical study of website personalization effect on user's intention to revisit e-commerce website through cognitive and hedonic experience: Proceedings of ICDMAI 2018 volume 2 20.7455
5 Digitalisation luxury fashion and “Chineseness”: The influence of the Chinese context for luxury brands and the online luxury consumers experience 20.1041
6 The effect of electronic word of mouth on brand image and purchase intention: An empirical study in the automobile industry in Iran 20.0951
7 A study on the effects of the attractiveness and credibility of online 1 personal media broadcasting B.J. on the viewing engagement perceived on media channel, interactivity, perceived enjoyment, and the user's responses. 19.3827
8 Predicting consumer purchase intention on fashion products in online retailer integration of self-determination theory and theory of planned behavior 18.9955
9 Extending expectancy violations theory to multiplayer online games: the structure and effects of expectations on attitude toward the advertising attitude toward the brand and purchase intent 18.4863
10 Negative online reviews of popular products: understanding the effects of review proportion and quality on consumers’ attitude and intention to buy 17.7826

Title Second-class Category Automatic Cluster Sample Exhibition

Title Prob. Category
Utilizing text mining and Kansei engineering to support data-driven design automation at conceptual design stage 0.991 Text mining
What affects patients’ online decisions: An empirical study of online appointment service based on text mining: International conference ICSH 2018 Wuhan, China, July 1–3 2018 proceedings 0.991 Text mining
Optimizing cost for geo-distributed storage systems in online social networks 0.902 Social network analysis
Vulnerability of D2D communications from interconnected social networks 0.933 Social network analysis

The Exhibition of Word Clustering Results

Line number Represent words Norm
1 Guinea 4.361
2 Lanka 4.2377
3 Sectional 4.2046
4 Congo 4.1324
5 Torres 4.1199
6 Strait 4.1101
7 Columbia 4.0069
8 Islander 3.9934
9 Leone 3.8981
10 Sierra 3.8625

Title First-class Category Automatic Cluster Sample Exhibition

Title Prob. Category
Using stock prices as ground truth in sentiment analysis to generate profitable trading signals 0.907 Semantic analysis
Application of text mining techniques to the analysis of discourse in eWOM communications from a gender perspective 0.976 Semantic analysis
Creating an Arabic dialect text corpus by exploring Twitter, Facebook, and online newspapers 0.926 Semantic analysis
N-Gram representations for comment filtering 0.968 Semantic analysis

Word Similarity Example Exhibition

Rank Similar words of health Similarity
1 Healthcare 0.76234574
2 Workers 0.71603154
3 Care 0.69574523
4 Interventions 0.6920655
5 Mental 0.68639633
6 School-based 0.68160605
7 Preventive 0.68005763
8 Visiting 0.65810901
9 Medicare 0.65752329
10 Disparities 0.65668325

Title Clustering Center Vector

Index Dimension 1 Dimension 2 Dimension 3 Dimension 4 Dimension 5 Dimension 49 Dimension 50
0 −0.0900 0.2218 0.0692 −0.0649 −0.1430 −0.0370 0.0318
1 −0.0552 0.2391 0.0992 −0.0776 −0.1130 −0.1805 0.0725
2 −0.0747 0.1618 0.0813 −0.0415 −0.1946 −0.0465 −0.0295
3 0.0175 0.2085 0.0767 −0.1440 −0.0342 −0.0199 −0.0804
4 0.0431 0.1801 0.1236 −0.1602 0.0378 −0.0791 −0.0500
5 −0.0503 0.1121 0.0518 0.0124 −0.1737 −0.0669 −0.1022
6 −0.0742 0.1920 0.0417 −0.1041 −0.0618 −0.1310 −0.2084
7 −0.1194 0.1587 0.1803 −0.2660 −0.1028 −0.0764 −0.0608
8 −0.0110 0.2546 0.1412 −0.0924 −0.1782 0.0270 −0.0244
9 −0.0151 0.1338 0.0822 −0.0194 −0.1065 0.0194 0.1752

Abel, F., Marenzi, I., Nejdl, W., & Zerr, S. (2009). Sharing distributed resources in LearnWeb2.0. Lecture Notes in Computer Science, 5794, 154–159. doi: 10.1007/978-3-642-04636-0_16 AbelF. MarenziI. NejdlW. ZerrS. 2009 Sharing distributed resources in LearnWeb2.0 Lecture Notes in Computer Science 5794 154 159 10.1007/978-3-642-04636-0_16 Open DOISearch in Google Scholar

About. (2008). ResearchGate. Retrieved from https://www.research-gate.net/about About 2008 ResearchGate Retrieved from https://www.research-gate.net/about Search in Google Scholar

Bastani, K., Asgari, E., & Namavari, H. (2019). Wide and deep learning for peer-to-peer lending. Expert Systems with Applications, 134, 209–224. doi: 10.1016/j.eswa.2019.05.042 BastaniK. AsgariE. NamavariH. 2019 Wide and deep learning for peer-to-peer lending Expert Systems with Applications 134 209 224 10.1016/j.eswa.2019.05.042 Open DOISearch in Google Scholar

Bengio, Y., Courville, A., & Vincent, P. (2013). Representation learning: A review and new perspectives. IEEE Transactions on Pattern Analysis and Machine Intelligence, 35(8), 1798–1828. doi: 10.1109/TPAMI.2013.50 BengioY. CourvilleA. VincentP. 2013 Representation learning: A review and new perspectives IEEE Transactions on Pattern Analysis and Machine Intelligence 35 8 1798 1828 10.1109/TPAMI.2013.50 Open DOISearch in Google Scholar

Bi, Q., & Liu, J. (2017). Study on the method of aggregation and service recommendation of digital resource based on domain ontology. Journal of the China Society for Scientific and Technical Information, 36(5), 452–460. BiQ. LiuJ. 2017 Study on the method of aggregation and service recommendation of digital resource based on domain ontology Journal of the China Society for Scientific and Technical Information 36 5 452 460 Search in Google Scholar

Cheng, H., Koc, L., Harmsen, J., Shaked, T., Chandra, T., Aradhye, H., … Shah, H. (2016). Wide & deep learning for recommender systems. DLRS 2016: Proceedings of the 1st Workshop on Deep Learning for Recommender Systems. 7–10. doi: 10.1145/2988450.2988454 ChengH. KocL. HarmsenJ. ShakedT. ChandraT. AradhyeH. ShahH. 2016 Wide & deep learning for recommender systems DLRS 2016: Proceedings of the 1st Workshop on Deep Learning for Recommender Systems 7 10 10.1145/2988450.2988454 Open DOISearch in Google Scholar

Dimensionality reduction - rdd-based API. (n.d.). Apache Spark 3.1.1. Retrieved from https://spark.apache.org/docs/latest/mllib-dimensionality-reduction Dimensionality reduction - rdd-based API (n.d.). Apache Spark 3.1.1 Retrieved from https://spark.apache.org/docs/latest/mllib-dimensionality-reduction Search in Google Scholar

Find and share. (2008). ResearchGate. Retrieved from https://www.researchgate.net/ Find and share 2008 ResearchGate Retrieved from https://www.researchgate.net/ Search in Google Scholar

Grisci, B., Krause, M., & Dorn, M. (2021). Relevance aggregation for neural networks interpretability and knowledge discovery on tabular data. Information Sciences, 559, 111–129. doi: 10.1016/j.ins.2021.01.052 GrisciB. KrauseM. DornM. 2021 Relevance aggregation for neural networks interpretability and knowledge discovery on tabular data Information Sciences 559 111 129 10.1016/j.ins.2021.01.052 Open DOISearch in Google Scholar

Grus, J. (2019). Clustering. Data Science from Scratch (2nd ed.). Sebastopol, CA: O’Reilly Media, Inc. GrusJ. 2019 Clustering. Data Science from Scratch 2nd ed. Sebastopol, CA O’Reilly Media, Inc Search in Google Scholar

Introduction. (2013) Word2vector. Google code archive. Retrieved from https://code.google.com/archive/p/word2vec/ Introduction 2013 Word2vector. Google code archive Retrieved from https://code.google.com/archive/p/word2vec/ Search in Google Scholar

Kim, M., Lee, S., & Kim, J. (2020). A wide & deep learning sharing input data for regression analysis. IEEE International Conference on Big Data and Smart Computing (BigComp), 8–12. doi: 10.1109/BigComp48618.2020.0-108 KimM. LeeS. KimJ. 2020 A wide & deep learning sharing input data for regression analysis IEEE International Conference on Big Data and Smart Computing (BigComp) 8 12 10.1109/BigComp48618.2020.0-108 Open DOISearch in Google Scholar

Klungpornkun, M., & Vateekul, P. (2018). Hierarchical text categorization using level based neural networks of word embedding sequences with sharing layer information. [WJST]. Walailak Journal of Science and Technology, 16(2), 121–131. doi: 10.48048/wjst.2019.4145 KlungpornkunM. VateekulP. 2018 Hierarchical text categorization using level based neural networks of word embedding sequences with sharing layer information. [WJST] Walailak Journal of Science and Technology 16 2 121 131 10.48048/wjst.2019.4145 Open DOISearch in Google Scholar

K-means++. (November 26, 2020). Wikipedia. Retrieved from http://en.wikipedia.org/wiki/K-means%2B%2B K-means++ November 26 2020 Wikipedia Retrieved from http://en.wikipedia.org/wiki/K-means%2B%2B Search in Google Scholar

LeCun, Y., Bengio, Y., & Hinton, G. (2015). Deep learning. Nature, 521(7553), 436–444. doi: 10.1038/nature14539 LeCunY. BengioY. HintonG. 2015 Deep learning Nature 521 7553 436 444 10.1038/nature14539 Open DOISearch in Google Scholar

Lee, J., Oh, S., Dong, H., Wang, F., & Burnett, G. (2019). Motivations for self-archiving on an academic social networking site: A study on Researchgate. Journal of the Association for Information Science and Technology, 70(6), 563–574. doi: 10.1002/asi.24138 LeeJ. OhS. DongH. WangF. BurnettG. 2019 Motivations for self-archiving on an academic social networking site: A study on Researchgate Journal of the Association for Information Science and Technology 70 6 563 574 10.1002/asi.24138 Open DOISearch in Google Scholar

Li, D., Lebo, T., Erickson, J., DiFranzo, D., Williams, G., Li, X., . . . Hendler, J. (2011). TWC LOGD: A portal for linked open government data ecosystems. Journal of Web Semantics, 9(3), 325–333. doi: 10.1016/j.websem.2011.06.002 LiD. LeboT. EricksonJ. DiFranzoD. WilliamsG. LiX. HendlerJ. 2011 TWC LOGD: A portal for linked open government data ecosystems Journal of Web Semantics 9 3 325 333 10.1016/j.websem.2011.06.002 Open DOISearch in Google Scholar

Li, Y. (2016). A review on knowledge aggregation. Library and Information Service, 60(21), 128–136. LiY. 2016 A review on knowledge aggregation Library and Information Service 60 21 128 136 Search in Google Scholar

Liu, J., Zhang, H., & Liu, Z. (2020). Research on online learning resource recommendation method based on wide & deep and Elmo Model. Journal of Physics. Conference Series, 1437(1), 012015. Retrieved from doi: 10.1088/1742-6596/1437/1/012015 LiuJ. ZhangH. LiuZ. 2020 Research on online learning resource recommendation method based on wide & deep and Elmo Model Journal of Physics. Conference Series 1437 1 012015 Retrieved from 10.1088/1742-6596/1437/1/012015 Open DOISearch in Google Scholar

Lu, W., Cai, Y., Che, X., & Lu, Y. (2016). Joint semantic similarity assessment with raw corpus and structured ontology for semantic-oriented service discovery. Personal and Ubiquitous Computing, 20(3), 311–323. doi: 10.1007/s00779-016-0921-0 LuW. CaiY. CheX. LuY. 2016 Joint semantic similarity assessment with raw corpus and structured ontology for semantic-oriented service discovery Personal and Ubiquitous Computing 20 3 311 323 10.1007/s00779-016-0921-0 Open DOISearch in Google Scholar

Martinez-Gil, J., & Aldana-Montes, J. (2013). Similarity measurement using historical Google search patterns. Information Systems Frontiers, 1 Semantic 5(3), 399–410. Martinez-GilJ. Aldana-MontesJ. 2013 Similarity measurement using historical Google search patterns Information Systems Frontiers 1 Semantic 5 3 399 410 Search in Google Scholar

Mentzas, G. N., Kafentzis, K., & Georgolios, P. (2007). Knowledge services on the semantic web. Communications of the ACM, 50(10), 53–58. doi: 10.1145/1290958.1290962 MentzasG. N. KafentzisK. GeorgoliosP. 2007 Knowledge services on the semantic web Communications of the ACM 50 10 53 58 10.1145/1290958.1290962 Open DOISearch in Google Scholar

Mikolov, T., Chen, K., Corrado, G., & Dean, J. (2013). Efficient estimation of word representations in vector space, 1–12. Retrieved from https://arxiv.org/pdf/1301.3781.pdf MikolovT. ChenK. CorradoG. DeanJ. 2013 Efficient estimation of word representations in vector space 1 12 Retrieved from https://arxiv.org/pdf/1301.3781.pdf Search in Google Scholar

Mu, T., Goulermas, J., Korkontzelos, I., & Ananiadou, S. (2016). Descriptive document clustering via discriminant learning in a co-embedded space of multilevel similarities. Journal of the Association for Information Science and Technology. 67, 106–133. MuT. GoulermasJ. KorkontzelosI. AnaniadouS. 2016 Descriptive document clustering via discriminant learning in a co-embedded space of multilevel similarities Journal of the Association for Information Science and Technology 67 106 133 Search in Google Scholar

Mueller, J., & Massaron, L. (2019). Deep Learning for dummies (1st ed). For Dummies: Hoboken, NJ, United States. MuellerJ. MassaronL. 2019 Deep Learning for dummies 1st ed For Dummies Hoboken, NJ, United States Search in Google Scholar

Tarko, V., & Aligica, P. (2011). From “broad studies” to internet-based “Expert knowledge aggregation.” notes on the methodology and technology of knowledge integration. Futures, 43(9), 986–995. doi: 10.1016/j.futures.2011.06.008 TarkoV. AligicaP. 2011 From “broad studies” to internet-based “Expert knowledge aggregation.” notes on the methodology and technology of knowledge integration Futures 43 9 986 995 10.1016/j.futures.2011.06.008 Open DOISearch in Google Scholar

Visualizing k-means clustering. (2014, January 19). Naftali Harris. Retrieved from https://www.naftaliharris.com/blog/visualizing-k-means-clustering/ Visualizing k-means clustering 2014 January 19 Naftali Harris Retrieved from https://www.naftaliharris.com/blog/visualizing-k-means-clustering/ Search in Google Scholar

Yi, M., Mao, J., Deng, W., & Cao, G. (2014). Evolution of knowledge push network based on social network in social tagging system. Journal of Library Science in China, 40(02), 50–66. YiM. MaoJ. DengW. CaoG. 2014 Evolution of knowledge push network based on social network in social tagging system Journal of Library Science in China 40 02 50 66 Search in Google Scholar

Yilmaz, S., & Toklu, S. (2020). A deep learning analysis on question classification task using word2vec representations. Neural Computing & Applications, 32(7), 2909–2928. doi: 10.1007/s00521-020-04725-w YilmazS. TokluS. 2020 A deep learning analysis on question classification task using word2vec representations Neural Computing & Applications 32 7 2909 2928 10.1007/s00521-020-04725-w Open DOISearch in Google Scholar

Yin, J., & Wang, J. (2014) A dirichlet multinomial mixture model-based approach for short text clustering. Association for Computing Machinery Digital Library, 233–242. YinJ. WangJ. 2014 A dirichlet multinomial mixture model-based approach for short text clustering Association for Computing Machinery Digital Library 233 242 Search in Google Scholar

Yurko, J., Dr (Presenter). (2020, October 9). Principal components analysis (PCA). Lecture presented at University of Pittsburgh, Pittsburgh, PA, United States. YurkoJ. Dr (Presenter). 2020 October 9 Principal components analysis (PCA) Lecture presented at University of Pittsburgh Pittsburgh, PA, United States Search in Google Scholar

Zhan, J., & Dahal, B. (2017). Using deep learning for short text understanding. Journal of Big Data, 4(1), 1–15. doi: 10.1186/s40537-017-0095-2 ZhanJ. DahalB. 2017 Using deep learning for short text understanding Journal of Big Data 4 1 1 15 10.1186/s40537-017-0095-2 Open DOISearch in Google Scholar

Recommended articles from Trend MD

Plan your remote conference with Sciendo