A Factorization Based Solution to the Implicit Feedback Problem 

Get Complete Project Material File(s) Now! »

Points-Of-Interest and Social Networks

A point-of-interest6 is a uniquely identified specific site generally associated to a specific category of activities (e.g. museum, restaurant, university etc.). Similarly to checkins (defined below) in LBSNs a point-of-interest is generally also associated with some content which corresponds to the set of all comments, pictures, opinions that users have uploaded during the checkins they made. For instance in table 1.1 the point-of-interest Eiffel Tower is associated with some of the categories it belongs to. However in many practical cases the categories or other point-of-interest descriptions are not disclosed for various reasons (e.g. privacy, confidentiality etc.). This is why in our approaches we have assumed that we only know the locations, i.e. the pairs (latitude; longitude) for all points-of-interest.
On the other hand the checkins correspond to the visits made by users in points-of-interest. Therefore checkins are always associated at least with a POI and a location (i.e. a pair {longitude; latitude}) and a date (e.g. a timestamp). These information are required to deal with geographical and temporal dimensions. Checkins can also be associated to different content. However taking into account of these metadata requires more complex input models, which eventually increases the training duration and the computational complexity of the recommendation model. We propose on figure 1.2 an overview of the standard structure of a LBSN. On this figure we distinguish three main layers: first the map and the points-of-interest (i.e. the geographical layer or the physical layer), then the users (i.e. the social layer) and finally the content shared online by the users (i.e. the content layer). The information contained by these layers come with specific constraints and characteristics that are likely to influence the final quality of the models and that require to be taken into account. For instance it has been demonstrated that the geographical layer content has the most significant impact on the recommendation final quality [Lian et al. 2014]. As a result it is necessary to consider how to exploit all the layers in order to increase the efficiency of our recommendation models. Unfortunately the information contained by the content layer are often not disclosed for privacy purpose. Moreover it does not exist any universal method to manage these data. For these reasons we will not exploit the content layer information directly in this thesis.

A Recommender Systems Overview

The most general goal of a recommender system (also noted RS) is to suggest to online users items to consume or to select [Adomavicius and Tuzhilin 2005]. Most of the time these items are expected to be new or at least that she could not find on her own. Furthermore these items are also expected to match the user preferences and so to contribute positively to the user experience. This is why these systems propose to the user a personalized exploration of a large space of possible choices. Differently from pure information retrieval systems where the user navigates into this possible space of choices by expressing an explicit query, the RS is not aware ”a priori” of what the user really wants or prefers. In other words in a standard recommendation scenario the RS have to infer the implicit personal information needs of the users. As a consequence, because there is no explicit query, the RS can only exploit all past interactions of the user with the system to generate recommendations. This is why recommender systems collect and analyse all past user’s preferences in order to predict future preferences. Moreover, from a business point of view the RS goal is to transform a standard user into a consumer. The business value here is to increase the conversion rate of the POIs owner [Ricci, Rokach, Shapira, and Kantor 2010]. This task is completed especially by enhancing the loyalty of the user to the system, which is done by improving her browsing experience. Thus the recommender systems business purpose is to improve the quality of the users’ interaction with the system, and to help the user through a large space of possible relevant items to select and consume.

READ  Cognitive Loads Theory and Theory of Instructional Dialogue

POI Recommendation

Many recommendation services are provided together in most of LBSN, such as user recommendation, activity recommendation, or POI recommendation. POI recommendation is one of the most challenging problems that received attention both in the academic community (with international conferences dedicated specifically to this problem such as ACM RecSys2) and the industry community as well due to its business exploitation. Our aim in this section is to present a general overview of existing models and approaches proposed in literature. We start to define our problem in subsection 2.2.1. Then we provide details about distinct POI recommendation problems in subsection 2.2.2. Subsections 2.2.3 details hybrid methods. Then we explore graph-based approaches in subsection 2.2.4. Finally we investigate matrix factorization approaches in subsection 2.2.5. Notice that we present a comprehensive summary of all existing approaches proposed for POI recommendation in subsection 2.3.1.

Table of contents :

1 Introduction 
1.1 Research Motivation
1.2 Points-Of-Interest and Social Networks
1.3 General Objectives
1.4 Research Goals
1.5 Contributions
1.6 General Definitions
1.7 Structure of the Thesis
2 A Survey on Points-Of-Interest Recommender Systems 
2.1 A Recommender Systems Overview
2.1.1 Background
2.1.2 Algorithms Classification
2.1.3 Challenges
2.1.4 Evaluation
2.2 POI Recommendation
2.2.1 Problem Definition
2.2.2 Different POI Recommendation Problems
2.2.3 Hybrid Collaborative Filtering Models
2.2.4 Graph Based Approaches
2.2.5 Matrix Factorization Models
2.3 Overview of Important Models
2.3.1 Existing Methods
2.3.2 Models of this Thesis
3 An Efficient Matrix Factorization Model for POI Recommendation
3.1 Introduction
3.2 Related Matrix Factorization Models
3.3 Geographical Influence for Factorization Models
3.3.1 Weighted Matrix Factorization
3.3.2 Modelling Geographical Influence
3.4 GeoMF with Temporal Dependencies: GeoMF-TD
3.5 Experiments
3.5.1 Dataset and Experimental Setup
3.5.2 Evaluation Metrics
3.5.3 Results and Discussions
3.6 Conclusions
4 A Factorization Based Solution to the Implicit Feedback Problem 
4.1 Introduction
4.2 Existing Implicit Feedback Approaches
4.3 A Factorization Model for Implicit Feedback
4.4 GeoSPF: Modeling Geographical and Social Influences
4.4.1 General Idea
4.4.2 Geographical Accessibility
4.4.3 AGRA: Accessibility Graph
4.4.4 GeoSPF: An Implicit Social Factorization
4.4.5 Inference
4.5 Experimental Evaluation
4.5.1 Data Sets and Metrics Description
4.5.2 Comparison with competitor models
4.6 Conclusion
5 ALGeoSPF: A Clustering Based Factorization Model for Large Scale POI Recommendation 
5.1 Introduction
5.1.1 Contributions
5.1.2 Road Map
5.2 POI Recommendation at Large Scale
5.3 ALGeoSPF: Local-Global Spatial Influence Modeling
5.3.1 General Idea
5.3.2 Super-POIs
5.3.3 Mobility Behaviors
5.3.4 Final Objective
5.4 Hierachical SuperPOIs Layers
5.4.1 Geographical Clustering Algorithm
5.4.2 Personalized Class Selection
5.5 Experimental Evaluation
5.5.1 Datasets and Metrics Description
5.5.2 Comparison with competitor models
5.6 Conclusion
6 Conclusion 
6.1 Summary
6.2 Outlook
A Résumé en français 
A.1 Introduction
A.2 Axes de recherche
A.3 Contributions
A.4 GeoMF-TD : Un modèle de factorisation de matrices pour la recommandation de POI
A.4.1 Factorisation de matrices géographique
A.4.2 GeoMF avec dépendances temporelles
A.4.3 Résultats expérimentaux
A.4.4 Conclusions
A.5 GeoSPF : influences sociales implicites
A.5.1 Factorisation de Poisson et feedback implicite
A.5.2 Modèle d’influence sociale
A.5.3 Résultats expérimentaux
A.5.4 Conclusion
A.6 Passage à l’échelle avec ALGeoSPF
A.6.1 Idée générale
A.6.2 Hiérarchie de superPOI
A.6.3 Résultats expérimentaux
A.6.4 Conclusion
A.7 Conclusion générale . .


Related Posts