Content-based recommender systems using network graphs

다국어 초록 (Multilingual Abstract)

The development of the internet and web technologies has expanded the range of activities and items such as entertainment, e-commerce, and education. Recommender systems are defined as techniques used to predict the rating of a social entity or items. These items can be music, videos, books, movies, etc. The main goal of recommender systems is to provide suggestions to online users to them make better choices from the many alternatives available over the Web. A better recommender system is directed more toward personalized recommendations by taking into consideration the available digital footprint of the user and information about a product, such as specifications, feedback from the users, comparison with other products, and so on, before making recommendations.
Content-based filtering, one of the most successful recommendation techniques, is based on correlations between content. Content-based filtering uses item information, represented as attributes, to calculate the similarities between items. In this study, I propose a novel content-based filtering method that uses a multiattribute network to effectively reflect several attributes when calculating correlations as a basis for recommendations to users. In network analysis, I measure the similarities between directly and indirectly linked items. Moreover, the proposed method uses centrality and clustering techniques to evaluate the mutual relationships among items as well as to determine the structural patterns of these interactions. This mechanism ensures that a variety of items are recommended to a user, which improves performance. I used MovieLens data to compare the proposed approach with existing approaches and found that the approach outperformed existing methods in terms of accuracy and robustness. The proposed method can address problems of sparsity and over-specialization that frequently affect recommender systems. Furthermore, the proposed method depends only on ratings data obtained from a user’s own past information; consequently, it is not affected by the cold start problem.
A recommender system is important not only to improve the intrinsic mechanism of an algorithm but also to build an appropriate methodology by analyzing the characteristics and environment of the recommended domain. In this dissertation, I propose a methodology for recommendation of document containing citations. An academic paper is a representative example of a document with citation information. Researchers typically need to filter several academic papers to find those relevant to their research. This filtering is cumbersome and time-consuming because the number of published academic papers is growing exponentially. Some researchers have focused on developing better recommender systems for academic papers by using citation analysis and content analysis. Most traditional content analysis is implemented by using a keyword matching process, and thus it cannot consider the semantic context of items. Further, citation analysis-based techniques rely on the number of links directly citing or being cited in a single-level network. Consequently, it may be difficult to recommend the appropriate papers when the paper of interest lacks sufficient citation information. To address these problems, I propose a recommender system that combines citation analysis and network analysis. The proposed method is based on multilevel citation networks that compare all the indirectly linked papers to the paper of interest by inspecting the structural and semantic relationships among them. Thus, the proposed method tends to recommend informative and useful papers related to both the research topic and the academic theory. The comparison results based on real data showed that the proposed method outperformed the Google Scholar and SCOPUS algorithms. The proposed methods in this dissertation are capable of handling important issues that can be encountered in recommender systems. I believe that all of the proposed methodologies in this dissertation improve existing problems and can be applied to actual problems.

목차 (Table of Contents)

• 1. Introduction 14
• 1.1 Recommender systems 14
• 1.2 Various techniques of recommender systems 15
• 1.2.1 Information filtering 16
• 1.2.2 Association rule and demographic 19
• 1.2.3 Hybrid modeling 21
• 1.3 Issues in recommender systems 22
• 1.3.1 Data collection 22
• 1.3.2 Sparsity 23
• 1.3.3 Over-specialization 25
• 1.3.4 Scalability 25
• 1.4 Outline of dissertation 26
• 2. Content-Based Filtering for Recommendation Systems Using Multiattribute Networks 28
• 2.1 Introduction 28
• 2.2 Related work 31
• 2.2.1 Multiattribute recommender systems 31
• 2.2.1.1 Pure CBF 31
• 2.2.1.2 Feature weighting 32
• 2.2.2 Semantic analysis 33
• 2.2.3 Network analysis 34
• 2.3 Proposed method 36
• 2.3.1 Data acquisition 36
• 2.3.2 Item similarity computation 37
• 2.3.3 Generating an MN 39
• 2.3.4 Item clustering 40
• 2.3.5 Recommendation 43
• 2.3.5.1 Centrality of items in the MN 43
• 2.3.5.2 Ego-focused centrality 46
• 2.3.6 Extensions 50
• 2.4 Experimental Evaluation 52
• 2.4.1 Experimental data and evaluation measure 52
• 2.4.2 Experimental results 56
• 2.5 Conclusions 58
• 3. Academic Paper Recommender System Using Multilevel Simultaneous Citation Networks 62
• 3.1 Introduction 62
• 3.2 Related work 65
• 3.2.1 Content-based filtering and collaborative filtering 65
• 3.2.2 Information retrieval techniques 68
• 3.2.3 Citation networks 69
• 3.3 Propsed method 71
• 3.3.1 Overview of proposed recommendation method 71
• 3.3.2 Generating multilevel citation networks 73
• 3.3.3 Selection of candidate papers 75
• 3.3.4 Determination of recommend papers 81
• 3.4 Experimental evaluation 85
• 3.4.1 Experimental data and evaluation measure 85
• 3.4.2 Experimental results 89
• 3.4.2.1 Correlation between centralities 89
• 3.4.2.2 Experimental results 90
• 3.5 Conclusions 98
• 4. Summary and Future Studies 102
• Reference 105