Project topics and materials
Get Complete Project Material File(s) Now! »
Recommendation Approaches
A recommendation approach, also referred to as recommendation method or recommendation algorithm, is expected to predict the utilities of items for a set of target users and generate appropriate recommendations. A large range of approaches have been proposed to oer accurate recommendations and can be classied according to multiple criteria including the recommendation problem they address, i.e., rating prediction or top-N recommendation, or the type of feedback they employ, i.e., explicit or implicit feedback (Section 2.1.1). The most common classication found in the literature refers to how the information related to users and items is exploited for the recommendation task and establishes the following categories:
• Content-Based Filtering (CBF) approaches. These approaches make explicit use of domain knowledge related to users or items.
• Collaborative Filtering (CF) approaches. Recommendations are only based on the user behavior and on previous interactions, without further explicit information about users or items.
• Hybrid approaches. These approaches combine the two previous strategies using both user and item information, and user interactions.
• Context-aware approaches. These approaches leverage contextual information about users and items in order to propose appropriate recommendations.
Content-Based Filtering Approaches
Content-Based Filtering (CBF) approaches [Pazzani and Billsus, 2007] use content information about users and items in order to generate recommendations. This information can take various forms such as features, textual descriptions, and tags. In order to recommend relevant items, the main idea is to match user proles and item proles based on user preferences for item attributes. Therefore, users receive suggestions about items that are similar to the ones they previously interacted with. Deploying a CBF approach requires extracting relevant information about the content of items, building item proles and user proles, and ltering items according to the similarity between proles.
From Context-Aware to Context-Driven Recommender Systems
Recent work [Pagano et al., 2016] highlighted the emergence of the contextual turn, creating the need for Context-Driven RS (CDRS) for which context is critical rather than additional. CDRS aim to contextualize recommendations, i.e., tailor recommendations to the user intent and situation, rather than personalize, i.e., tailor recommendations to the individual. The main assumption is that users have more in common with other users in the same situation than with the past version of themselves. Recommendations are based on what is going around the user, i.e., the user’s situation, and on what the user is trying to accomplish, i.e., the user’s intent. CDRS are therefore able to generate recommendations without past information about users, i.e., in the cold-start setting. Several families of RS can be seen as a special case of CDRS. We mention for example session-based RS [Quadrana et al., 2018] that denes the context as a series of interactions carried out within a session.
Table of contents :
List of Figures
List of Tables
List of Abbreviations
I Introduction and Background
1 Introduction
1.1 Context in Real-World Recommender Systems
1.2 Contributions
1.3 Organization of the Thesis
1.3.1 Publications
2 Recommender Systems
2.1 The Recommendation Problem
2.1.1 Problem Formulation
2.1.2 Types of Feedback
2.1.3 Challenges and Limitations
2.1.4 Data Representation and Notations
2.2 Evaluation of Recommender Systems
2.2.1 Evaluation Methods
2.2.1.1 Oine Evaluation
2.2.1.2 User Studies
2.2.1.3 Online Evaluation
2.2.2 Evaluation Criteria
2.2.3 Evaluation Metrics
2.3 Recommendation Approaches
2.4 Content-Based Filtering Approaches
2.4.1 Example of a Content-Based Filtering Approach
2.4.2 Advantages and Disadvantages
2.4.3 Related Recommendation Approaches
2.5 Collaborative Filtering Approaches
2.5.1 Memory-Based Approaches
2.5.1.1 User-Based Collaborative Filtering
2.5.1.2 Item-Based Collaborative Filtering
2.5.1.3 Extensions, Complexity, Advantages and Disadvantages
2.5.2 Matrix Factorization Approaches
2.5.2.1 Matrix Factorization Framework
2.5.2.2 Singular Value Decomposition
2.5.2.3 Minimizing Squared Loss and Other Loss Functions
2.5.2.4 Dealing with Implicit Feedback
2.5.2.5 Probabilistic Models
2.6 Hybrid Approaches
2.7 Context-Aware Approaches
2.7.1 Paradigms for Incorporating Context
2.7.2 From Context-Aware to Context-Driven Recommender Systems
2.8 Conclusion
II Partially Observable Context in Hotel Recommendation
3 The Hotel Recommendation Problem
3.1 Introduction
3.2 Scope of Our Work
3.3 Comparison with Recommendation in Other Domains
3.4 Challenges and Limitations
3.5 Related Work
3.6 Context in the Hotel Domain
3.7 Conclusion
4 Leveraging Explicit Context
4.1 Introduction
4.2 In uence of the Physical Context
4.3 Inuence of the Social Context
4.3.1 Collaborative Topic Modeling
4.3.2 Handling Positive and Negative Reviews
4.4 Inuence of the Modal Context
4.5 Overview of the System
4.6 Experimental Results
4.6.1 Contribution of the Physical Context
4.6.2 Contribution of the Social and Modal Contexts
4.6.3 User Segmentation and Performance
4.7 Conclusion
5 Leveraging Implicit Context
5.1 Introduction
5.2 Event Recommendation
5.3 Problem Formulation
5.4 Data Collection and Analysis
5.4.1 Event Dataset
5.4.2 Booking Dataset
5.5 Proposed Framework
5.5.1 Overview
5.5.2 Notations and Denitions
5.5.3 Modules
5.5.3.1 Building All-Inclusive Proles Based on Location and Time
5.5.3.2 Measuring Events’ Similarities
5.5.3.3 Building Limited Proles Based on Cohesiveness
5.5.3.4 Learning Preferences for Hotels and Events
5.5.3.5 Recommending Hotels and Events
5.6 Experimental Results
5.6.1 Qualitative Evaluation Through Concrete Examples
5.6.2 Quantitative Evaluation Through Oine Experiments
5.7 Discussion
5.8 Conclusion
6 Transferring Context Knowledge Across Domains
6.1 Introduction
6.2 Cross-Domain Recommendation
6.3 Proposed Approach
6.3.1 Mapping Items from Both Domains
6.3.2 Mapping Users from Both Domains
6.3.3 Merging Preferences from Both Domains
6.4 Experimental Results
6.5 Conclusion
III From Unobservable Context to Online Adaptive Recommendation
7 The Online Adaptive Recommendation Problem
7.1 Introduction
7.2 Time Dimension in Recommender Systems
7.3 Adaptive Data Stream Mining
7.4 Online Adaptive Recommendation
7.4.1 Memory Module
7.4.2 Learning Module
7.4.2.1 Incremental Memory-Based Approaches
7.4.2.2 Incremental Matrix Factorization and Other Model-Based
Approaches
7.4.3 Retrieval Module
7.4.4 Evaluation Module
7.5 Conclusion
8 Dynamic Local Models
8.1 Introduction
8.2 Local Models for Recommendation
8.3 Proposed Approach
8.4 Experimental Results
8.5 Conclusion
9 Adaptive Incremental Matrix Factorization
9.1 Introduction
9.2 Learning Rate Schedules for Matrix Factorization
9.3 Proposed Approach
9.4 Experimental Results
9.4.1 Performance of AdaIMF on Synthetic Datasets
9.4.2 Performance of AdaIMF on Real-World Datasets
9.5 Conclusion
10 Adaptive Collaborative Topic Modeling
10.1 Introduction
10.2 Related Work
10.3 Proposed Approach for Online Topic Modeling
10.4 Proposed Approach for Online Recommendation
10.5 Experimental Results
10.5.1 Performance of AWILDA for Online Topic Modeling
10.5.1.1 Results for Topic Drift Detection
10.5.1.2 Results for Document Modeling
10.5.2 Performance of CoAWILDA for Online Recommendation
10.6 Conclusion
IV Concluding Remarks
11 Conclusions and Future Work
11.1 Summary and Conclusions
11.2 Future Work
A Resume en francais
A.1 La notion de contexte dans les systemes de recommandation du monde reel
A.2 Contributions
A.3 Organisation de la these
A.3.1 Publications
Bibliography
