Improving User and Item Representation for Recommender Systems with Textual Data Öffentlichkeit

Abstract

Conversational Recommender Systems (CRS) aim to provide personalized and contextualized recommendations through natural language conversations with users.

This dissertation capitalizes on the recent developments in conversational interfaces to advance the field of Recommender Systems in several directions.

Specifically, I address the user and item representation and recommendation explainability problems in recommender systems.

First, I investigate several approaches for improving user representation.

One approach is to map conversational users to reviewers with more data, using semantic similarity between the conversation and the texts of reviews. Another approach is to extract the items the user has mentioned, the sentiment the user has expressed towards them, and what specifically the user said about them, and use that information to represent the user.

Second, I investigate improving user and item representations by leveraging textual information. I develop methods to incorporate textual features such as item descriptions into the user-item interaction graph, which introduce additional semantic and behavioral information unavailable from the purely topological structure of the interaction graph.

I also investigate whether the knowledge learned by LLMs during pretraining can be leveraged to improve the user and item representations by generating new textual features about the users and items.

Third, I investigate ways to improve explainability. One approach to enhance the explainability and transparency of CRS is to generate justifications. However, existing methods, such as rule-based and template-based methods, have limitations. In this work, I develop an extractive method using a corpus of reviews to identify relevant information for generating concise and coherent justifications.

The research questions I am tackling are:

How to infer and represent user preferences during a conversation with the system?

How to better represent users and items using structured and unstructured knowledge for improving the quality of recommendations?

How to improve the explainability of conversational recommendations?

This thesis improves the effectiveness of conversational recommender systems and advances the state-of-the-art in the field by introducing novel approaches for user and item representation for improving conversational recommendation systems.

Table of Contents

1 Introduction 4

1.1 Research Questions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

1.2 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

2 Related Work 7

2.1 Recommender Systems Overview . . . . . . . . . . . . . . . . . . . . . 7

2.1.1 Content-Based Filtering . . . . . . . . . . . . . . . . . . . . . . 7

2.1.2 Collaborative Filtering . . . . . . . . . . . . . . . . . . . . . . . 7

2.1.3 Hybrid Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

2.1.4 Neural Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

2.1.5 Graph-based Methods . . . . . . . . . . . . . . . . . . . . . . . 9

2.1.6 LLM-based Methods . . . . . . . . . . . . . . . . . . . . . . . . 10

2.2 Conversational Recommender Systems . . . . . . . . . . . . . . . . . . 11

2.2.1 Types of Conversational Recommender Systems . . . . . . . . . 11

2.2.2 Conversational User Representation for Recommendations . . . 12

2.3 Explainable Recommendations . . . . . . . . . . . . . . . . . . . . . . . 12

2.3.1 Review-based Methods . . . . . . . . . . . . . . . . . . . . . . . 13

2.3.2 Deep Learning Methods . . . . . . . . . . . . . . . . . . . . . . 13

2.3.3 Context-Aware and Conversational Methods . . . . . . . . . . . 13

2.3.4 LLM-Based Methods . . . . . . . . . . . . . . . . . . . . . . . . 14

2.3.5 Evaluation and Impact . . . . . . . . . . . . . . . . . . . . . . . 14

2.4 Knowledge Injection for Recommender Systems . . . . . . . . . . . . . 14

2.5 Data Augmentation Using Foundational Language Models . . . . . . . 15

2.5.1 Conversational Recommender Systems . . . . . . . . . . . . . . 15

2.5.2 Traditional Recommender Systems . . . . . . . . . . . . . . . . 16

2.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

3 User Preference Representation in Conversational Recommendations 18

3.1 Method Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

3.1.1 Collaborative Filtering . . . . . . . . . . . . . . . . . . . . . . . 19

3.1.2 Domain Adaptation . . . . . . . . . . . . . . . . . . . . . . . . . 19

3.1.3 Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

3.2 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

3.2.1 MovieSent: Sentiment Elicitation Dataset . . . . . . . . . . . . 21

3.2.2 Reviews dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

3.3 Metrics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

3.4 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22

3.4.1 Baselines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22

3.4.2 ConvExtr . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22

3.4.3 Evaluation Procedure . . . . . . . . . . . . . . . . . . . . . . . . 23

3.5 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

3.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

3.7 Limitations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

4 Knowledge Injection for Better Recommendations 25

4.1 Method Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

4.1.1 Training LightGCN . . . . . . . . . . . . . . . . . . . . . . . . . 26

4.1.2 User and Item Representation . . . . . . . . . . . . . . . . . . . 27

4.2 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

4.3 Metrics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

4.4 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

4.4.1 Baselines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

4.4.2 TextGCN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29

4.5 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30

4.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32

4.7 Limitations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32

5 Enhancing Recommender Systems with LLM-Driven User and Item Insights 33

5.1 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35

5.1.1 Model Training . . . . . . . . . . . . . . . . . . . . . . . . . . . 36

5.1.2 Implementation Details . . . . . . . . . . . . . . . . . . . . . . . 36

5.1.3 Alternative Backbones . . . . . . . . . . . . . . . . . . . . . . . 37

5.2 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38

5.2.1 Source . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38

5.2.2 Preprocessing . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38

5.3 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39

5.3.1 Feature Generation . . . . . . . . . . . . . . . . . . . . . . . . . 39

5.4 Results and Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40

5.4.1 User Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42

5.4.2 Item Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42

5.4.3 LightGCN vs MMSSL . . . . . . . . . . . . . . . . . . . . . . . 43

5.4.4 Feature Interactions . . . . . . . . . . . . . . . . . . . . . . . . 43

5.4.5 Feature Importances . . . . . . . . . . . . . . . . . . . . . . . . 44

5.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44

5.6 Limitations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45

6 Using Textual Data for Explanation Construction 46

6.1 Method Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47

6.1.1 Elementary Discourse Units (step 1) . . . . . . . . . . . . . . . 47

6.1.2 Generate and Select Candidates (step 4) . . . . . . . . . . . . . 48

6.2 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48

6.2.1 Conversation Corpus . . . . . . . . . . . . . . . . . . . . . . . . 48

6.2.2 Reviews Corpus . . . . . . . . . . . . . . . . . . . . . . . . . . . 48

6.3 Metrics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49

6.4 Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49

6.4.1 Procedure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49

6.4.2 Baselines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49

6.5 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50

6.6 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51

6.7 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51

6.8 Limitations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52

7 Summary and Discussion 53

7.1 RQ1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53

7.2 RQ2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53

7.2.1 RQ2(a,b) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53

7.2.2 RQ2(c) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54

7.3 RQ3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54

7.4 Ethical Considerations . . . . . . . . . . . . . . . . . . . . . . . . . . . 55

7.4.1 Bias and Fairness . . . . . . . . . . . . . . . . . . . . . . . . . . 55

7.4.2 Privacy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55

7.4.3 Transparency . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56

7.4.4 Factuality in CONJURE . . . . . . . . . . . . . . . . . . . . . . 57

7.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57

A Prompts for Feature Generation 70

About this Dissertation

Rights statement

• Permission granted by the author to include this thesis or dissertation in this repository. All rights reserved by the author. Please contact the author for information regarding the reproduction and use of this thesis or dissertation.

School	Laney Graduate School
Department	Computer Science and Informatics
Degree	Ph.D.
Submission	Dissertation
Language	English
Research Field	Computer Science
Stichwort	recommender systems
Committee Chair / Thesis Advisor	Eugene Agichtein, Emory University
Committee Members	Filip Radlinski, Google Joyce Ho, Emory University Jinho Choi, Emory University

Zuletzt geändert

Primary PDF

Thumbnail	Title	Date Uploaded	Actions
	Improving User and Item Representation for Recommender Systems with Textual Data ()	2024-08-08 23:19:04 -0400	Press to Select an action Download

Supplemental Files

Thumbnail	Title	Date Uploaded	Actions