The intensification of international technological innovation competition and the evolution of scientific research paradigms have led to a continuous expansion of scientific literature, making information analysis increasingly complex and diversified. To address the challenges of accurately assessing topic evolution within the context of vast literature big data, traditional methods of expert evaluation or visualization analysis based on scientific knowledge networks are inadequate. From the perspectives of artificial intelligence and big data, this paper proposes a universal method for automated and intelligent discrimination and prediction of research topic evolution hotness. This method involves integrating content and structural features of keywords to track the evolution of keyword frequency strength over time in research topic networks characterized by keywords. This study conducts a case analysis in the field of information science. The results demonstrate that the prediction of keyword strength is improved after integrating content and structural features, which has significant reference value for tasks such as future research topic evolution trend discrimination, research direction, and policy planning.
With the intensification of international technological innovation competition and the evolution of the fourth paradigm of scientific research driven by big data development, the growing volume of scientific literature, shifting scholarly interests, and the emergence of new research topics pose significant challenges to traditional methods of research topic analysis[1,2,3]. How to comprehensively and finely reveal the research topics and their characteristic dimensions: content, structure, and strength[4,5]. Utilizing the powerful representation and feature learning capabilities of deep representation learning algorithms such as word embedding and graph embedding[6], it is possible to model the complex nonlinear relationships between entities represented by keywords, the smallest units of knowledge[7,8]. Tracking changes in topic strength, as indicated by keyword frequency over time, can reflect the evolving trends of topics[9,10,11].
Therefore, this paper utilizes a multiplex-graph representation learning method combined with interactions in topic keyword content and structure to assess changes in topic strength, achieving prediction of topic evolution hotness. And select the field of "Information Science" for case analysis, aiming to address the following two scientific questions: 1. How to reveal the evolution process of topics at the micro level, thereby tracking the evolutionary trends of research topics? 2. How to effectively and comprehensively model and integrate the multi-dimensional Step1:Data Preparation Field-specific literature is selected from databases such as Web of Science and Scopus, with titles, keywords, abstracts, and references extracted as basic data. After cleaning and filtering the data, an original dataset is constructed, which specifically includes the keyword citation relationship dataset , the keyword frequency dataset , and the integrated dataset containing titles, abstracts, and keywords. -- LDaantagsueatg:eW: OEnSglish Data cleaning - Collection: SCI & SSCI - Query : ‘Information Science’ Hump naming - Year : 2010 – 2023 - Amount: 58,119 - Fields : TI, AB, DE, CR, PY, DOI Step3: Evolution prediction
MAE
MSE
Evaluation and testing Forecasting task MAE MSE 1.93 13.41 1.91 12.63 1.90 13.26 1.88 12.24
Null-data validation data process
Word Frequency
Ai 16 BigData 84 Gcn 26 Cat 9 ... ...
Bert 33
Gpt 124 Example of result
Data extraction Year: 2010 . 2022 2023
Datasets for 2010-2023 + + + + Representation fusion experimental group
MLP
forecast Word frequency
Distance calculation Fusion of representations features of research topic evolution representations on knowledge networks? After tracking these multi-dimensional evolutionary features, will the assessment of topic evolution trends become more accurate?
This study aims to predict the strength of keyword frequency, using changes in keyword frequency strength to reflect variation in topic evolution hotness.
The research process is divided into 3 steps: Step 1 involves retrieving and cleaning the source data to obtain all the data needed for subsequent experiment. Step 2 involves obtaining content, structure, and strength representations of keywords. Step 3 utilizes deep learning models to integrate multi-dimensional data representations and conduct prediction of topic evolution hotness with the integrated experimental data. Figure 1 shows the detailed research process. Step2:Multidimensional data representation
Dataset of titles, abstracts, and keywords Dataset of keyword citation relationships
Dataset of keyword frequencies
Year:2010 . 2022 2023 YeGar:lo20V1e0, .N2u0m22py2,02P3andas and oYtehaer:r2m01o0d.el2s0a22nd20t2o3olkits Representation of content Representation of reference structure Representation of strength GCN GAT
Hotness Citation Semantic
Represent data MAE
Back-Propagation
To prepare for multi-dimensional feature integration, this study will perform feature extraction on keyword data across three representational dimensions: content, structure, and strength of keywords.
（1）Content feature extraction In the process of extracting keyword content features, this study opts to use the static word embedding method to capture the semantic relationships of keywords in their global context[12], facilitating the embedding of keywords, as it offers greater stability and requires less computational resources[13]. The principle is as shown in 2-1 and 2The generated word vector is then used to calculate the cosine similarity between words using formula 2-3. This process results in obtaining the semantic distance matrix keyword content feature extraction.
= = ( , ), = , for = [ 1, 2, , ], B = [ 1, 2, , ] = ∑ =1 × √∑ =1 2 × √∑ =1 2 （2）Structure feature extraction Some scholars have proposed using a "keywordcitation-keyword"
method to construct keyword
citation networks[14], meaning when literature 1
cites literature 2, there exists a "Keyword-Cartesian
product mapping" citation relationship between the
keywords of the two literatures. The specific principle
is shown in Figure 2. Based on this theory, this study
constructs a keyword citation network with citation
frequency as edge weight, resulting in a keyword
citation matrix = = ( , ), = .
(2-1)
(
| − ̂ | = =1 − ̂ 2 ∑ 2, where is the number of times word appears in the context of word , is the total number of words appearing in the context of word , and is the probability of word appeare in the context of word .
= ∑
= | =
(
Graph Attention Network(GAT)based on the attention mechanism, can effectively capture complex semantic dependencies between keywords, while
Convolutional Network(GCN) can efficiently process the strcture information of the graph structure itself by aggregating the features of neighboring nodes. Therefore, this study chooses to use GAT and GCN graph neural network models to capture the relationships between nodes in graph-structured data from content and structure perspectives, respectively, and employs an
regression model to integrate multi-dimensional features of keywords for strength prediction.
This study constructs an ablation experiment group, as shown in Table 1, for predicting the hotness of topic evolution. And details of the model settings are shown in Figure 1. It's worth noting that set an initial identity matrix allows the neural network to gradually adjust and optimize feature representations during the learning process.
Therefore, after obtaining the content matrix and the structure matrix
, GAT and GCN are used to perform convolution operations on these two matrices on a predefined 50-dimensional identity matrix. After obtaining content representation and structure representation of keywords, the three representation data are concatenated directly for integration, and prediction is made based on MLP[15].
Above, represents the ith element of , and is the number of elements.
The detailed process of data acquisition can be found in the appendix under . .
When extracting features from experimental data, this study choose to work with four sets of yearly data: 2017-2019, 2018-2020, 2019-2021, and 2020-2022. The first three sets are used as the training groups, and the last set as the test group. Specifically, the data from 2019 to 2022 serve as the basis for operations (all operations on yearly data will follow this fouryear standard). Taking 2019 as an example, semantic content matrix , citation structure matrix , and frequency strength matrix are constructed for that year's keyword data.
The objective is to determine the optimal parameters
for various model groups in order to ensure accurate
predictions. This study has set four hyperparameters:
learning rate (1e-2, 1e-3, 1e-4), number of training
epochs (10, 50, 100, 200), hidden layer dimensions
(10, 30, 50), and stopping steps (
The prediction results are evaluated using two indicators: MAE and MSE, with the evaluation results listed in Table 3. From the table, it can be observed that using only the strength representation for predicting topic hotness, the MAE and MSE between the predicted and actual values are 1.93484 and 13.41658, respectively. However, after integrating the content representation or structure representation , the values of MAE and MSE both decrease. The best result for predicting topic hotness are achieved by integrating all three types of representations, resulting in the lowest values of MAE and MSE.
The results indicate that predicting the evolution hotness of research topics by integrating multidimensional features such as content and structure through multiplex-graph representation learning is more accurate than traditional prediction methods.
This study proposes a novel approach based on multiplex-graph representation learning to predict the evolution of research topics. And the contributions are follows: First, in feature modeling, GCN and GAT graph neural network models are used to perform convolution operations on content and structure features on unit matrices of specified dimensions, adaptively aligning data across different dimensions and time windows to ensure comparability. Second, this study integrates semantic content features, citation structure features, and frequency strength features of keywords for research topic hotness prediction, showcasing the interaction between knowledge structures and cognitive structures from a multidimensional perspective, offering a deeper insight into predicting research topic evolution hotness. Third, after integrating content and structure features, a domain case analysis is conducted, and the result indicates that combining these two types of features indeed makes the prediction of research topic evolution hotness more accurate.
Owing to the desire to directly validate whether integrating multiple representations of topic evolution enhances the accuracy of topic evolution analysis, this paper choose to predict the future frequency of topic keywords, which has certain limitations. Subsequent tasks such as research topic trend discrimination, research direction, and policy planning can be developed based on the effective analysis results of this study.
This work was funded by the National Natural Science Fund of China (No. 71874129), the Open-end Fund of Information Engineering Lab of ISTIC and the Independent Innovation Foundation of Wuhan University of Technology (No. 233103002). keywords in building an initial reading list of research papers in scientific paper retrieval and recommender systems." Information Processing & Management 53.3 (2017): 577-594. [10] Yoon, Young Seog, et al. "Exploring the dynamic knowledge structure of studies on the Internet of things: Keyword analysis." ETRI Journal 40.6 (2018): 745-758. [11] Ohniwa, Ryosuke L., and Aiko Hibino. "Generating process of emerging topics in the life sciences." Scientometrics 121.3 (2019): 1549-1561. [12] Pennington, Jeffrey, Richard Socher, and Christopher D. Manning. "Glove: Global vectors for word representation." Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP). 2014. [13] Wang, Yuxuan, et al. "From static to dynamic word representations: a survey." International Journal of Machine Learning and Cybernetics 11 (2020): 1611-1630. [14] Q. Chen, J. Wang, and W. Lu. "Discovering Domain Vocabularies Based on Citation Co-word Network" Data Analysis and Knowledge Discovery 3.6 (2019): 57-65. (in Chinese) [15] Liu, Weijia, et al. "Category-universal witness discovery with attention mechanism in social network." Information Processing & Management 59.4 (2022): 102947. [16] Yan, Yuwei, et al. "Data mining of customer choice behavior in internet of things within relationship network." International Journal of Information Management 50 (2020): 566-574. [17] Gandhudi, Manoranjan, et al. "Causal aware parameterized quantum stochastic gradient descent for analyzing marketing advertisements and sales forecasting." Information Processing & Management 60.5 (2023): 103473.
The resources of this article can be downloaded at https://github.com/Hipkevin/EEKE-hotness.
This study uses "Information Science" as a case study topic, selecting the SCI and SSCI core databases in WOS. Conducting literature searches in the "Information Science & Library Science" field with the search query "Document Types: Article or Review Article; Languages: English," ultimately selecting literature from 2010 to 2023, totaling 58,119 articles, as experimental data.