Project topics and materials
Get Complete Project Material File(s) Now! »
Recommender Systems Function
In one hand, recommender systems are expected to suggest useful items to users by supporting them in various decision-making processes, such as what items to buy, what news to read, or which restaurants to try. On the other hand, electronic retailers and online service providers which use RS hope to increase their profits. For example, a travel intermediary wants to sell more hotel rooms by the increase the number of tourists to some destinations, while tourists wish to find a suitable hotel and interesting attractions when visiting a destination. Recommender systems aim to come up to these expectations. Thus they have to play a variety of functions. We list below some of them.
– Increase the conversion rate is probably the most important function of any RS, particularly for commercial RS where it amounts to raise the number of items sold.
The conversion rate is the percentage of users who accept the recommendations and consume some items. This is the main function for which someone would want to exploit a recommender system.
– Increase satisfaction and fidelity of users through their experiences with the system. By relevant recommendations, and perhaps a properly designed humancomputer interaction, the users could better evaluate the system and enjoy using it. This in turn will increase the system usage and the likelihood that the recommendations will be accepted. The first consequence of the users’ enjoyment to use a system is their fidelity. The latter comes into more effect for a Web site with the recognition of users when they are visited the site and treated as a valuable visitor. This is fully carried out by RS with gradually refined suggestions as they can leverage the information acquired from the users in previous interactions.
– Sell more diverse items is another major function of a RS. Indeed while giving to users the possibility to get some hard-to-find items, it allows to bring to front all the items, not just the most popular ones, and then to reduce the long tail [129, 35]. This could be difficult without a RS.
Classification of Recommender Systems
Recommendation systems are usually classified on the basis of their approaches to estimate the interests of users [116, 127]. The authors usually agree on two broad classes.
– Content-based filtering recommend similar items to the ones a user was interested in the past. Their basic process consist in matching up the attributes of an item and the description of a user profile in which preferences and interests are stored.
– As for collaborative filtering (CF) systems, they recommend items to users based on the interests the other users expressed for those items. In practice, these interests may be expressed by the ratings users give to items (e.g., 1-5 stars) or the purchases they made. CF focuses on the relationship between users and items. It is the most widely used prediction technique in recommendation.
The major difference between collaborative filtering and content-based approaches is that CF mainly uses the interest of users on items to make predictions and recom- mendations, while content-based recommender systems rely on the features of users and items for predictions [127].
A third type of recommendation are hybrid systems which combine both above approaches in some manner. They allow to overcome their limitations, as we will present shortly.
Content-based Recommender Systems
Content-based recommendation systems (CbRS) share in common a means of rep- resenting the items that may be recommended, a way for creating a user’s profile that contains information about its tastes, preferences and needs. Users can construct such a profile explicitly, but it can also be learned from the user’s interaction with the system [105, 83]. CbRS rely on matching methods in order to determine which items may interest the most the user and recommend them. Let us notice here that, in some applications, it can be appropriate to recommend an item that the user has already seen (i.e., purchased or rated) while in some others it is not interesting. Thus some items may be discarded from a recommendation list depending on the application. For example, a system should continue to recommend items that wear out (e.g., a razor blade) or are expended (e.g., print cartridge), while there is little value in recommending a movie that a user watched. Furthermore, because the representation of items differs from a context to another, matching functions change according to that. In this section, we present alternative rep- resentations of items and some matching functions depending on each of these represen- tations. We then discuss the strengths and weaknesses of content-based recommendation systems.
Item Representation
Items can be represented by a set of features and its description stored in structured data. For example, in a movie recommendation application, each movie can be described by the actors and directors involved in it, its genre and release year. In this case, many machine learning algorithms can learn a user profile from the ones of the items which interested him before, or a menu interface can be created to allow a user to easily create his profile. Section 1.3.1.2 discusses some approaches to learn a user profile from structured data.
In most content-based filtering systems, all item features are not available in or- ganized form. Some information is extracted from free text (e.g., web pages, news articles or product descriptions). Thus unlike structured data, they are not attributes with well-defined values. In this case, term-based profiles may be used to represent the users [119, 142]. Each user is then represented by a vector of terms (i.e., words) selected from the free text describing the items he saw. In other words, each term is viewed as an attribute to which we can assign a weight depending on its popularity for the user. Moreover the root forms of terms are typically used thanks to some stemming processes [107, 101, 100].
Table of contents :
Contents
List of Figures
List of Tables
General Introduction
I A survey of recommender systems
1 Introduction to Recommender Systems
1.1 Formal Definition of Recommendation Task
1.2 Recommender Systems Function
1.3 Classification of Recommender Systems
1.3.1 Content-based Recommender Systems
1.3.1.1 Item Representation
1.3.1.2 User Profile
1.3.1.3 Advantages and limitations
1.3.2 Collaborative Filtering Systems
1.3.2.1 Memory-based CF
1.3.2.2 Model-based CF
1.3.2.3 Advantages and limitations
1.3.3 Hybrid Systems
1.4 Conclusion
2 Challenges and Evaluation of Recommender Systems
2.1 Challenges of Recommender Systems
2.1.1 Data Sparsity
2.1.2 Scalability
2.1.3 Diversity vs. accuracy
2.1.4 User interface
2.1.5 Vulnerability to attacks
2.1.6 Some other Challenges
2.2 Evaluating Recommender Systems
2.2.1 Evaluation approaches
2.2.1.1 Offline evaluation
2.2.1.2 User studies
2.2.1.3 Online Evaluation
2.2.2 Evaluation measures
2.2.2.1 Prediction Accuracy
2.2.2.2 Coverage
2.2.2.3 Diversity
2.2.2.4 Novelty
2.3 Conclusion
II The contributions of the thesis
3 Using Cluster-based Biases for Dynamic Recommendations
3.1 Preliminaries
3.2 The Dynamicity Problem
3.3 Related Work
3.3.1 Distributed MF techniques
3.3.2 Online-updating approaches
3.4 Making Dynamic Recommendations
3.4.1 Biased MF
3.4.2 The interest of cluster-based local biases
3.4.3 The CBMF model
3.4.4 Integration of incoming ratings
3.5 Complexity analysis
3.6 Experimental Results
3.6.1 Implementation and experimental setup
3.6.2 Datasets
3.6.3 Initial quality
3.6.4 Large training sets improve the quality of the model
3.6.5 Quantifying the need for online integration
3.6.6 Robustness over time of our online integration model
3.6.7 Quality vs. Performance tradeoff for online integration
3.6.8 Benefit of refactorization
3.7 Conclusion
4 Making Social and Popularity-based Tag Recommendations
4.1 Preliminaries
4.1.1 Similarity propagation
4.1.2 Extended neighborhood opinion
4.2 Social and Popularity-based Tag Recommendation
4.2.1 Score model and tag relevance
4.2.2 The FasTag Algorithm
4.2.3 Handling the Network Partitioning
4.3 Related Work
4.4 Experimentation
4.4.1 Datasets
4.4.2 Evaluation Measures and Methodology
4.4.3 Effectiveness of FasTag
4.4.4 Comparison with the result of ECML PKDD challenge
4.4.5 Scalability of FasTag
4.5 Conclusion
5 Optimizing Tag Recommendation List Size
5.1 Preliminaries
5.1.1 Factor Models for Tag Recommendation
5.1.2 FolkRank – A Topic-Specific Ranking
5.1.3 Recommending the Most Popular Tags
5.1.4 Social and popularity-based Recommender
5.2 Adjusted Recommendation list size
5.2.1 Linear combination models
5.2.2 The blsC algorithm
5.2.2.1 Simple relevance measure
5.2.2.2 Refining the relevance measure
5.3 Experimentation
5.3.1 Datasets
5.3.2 Evaluation Measures and Methodology
5.3.3 Experimental Results
5.3.3.1 Effectiveness of our proposal
5.3.3.2 Giving up some recall for more precision
5.3.3.3 Distribution of optimal list sizes
5.4 Conclusion
Research Perspectives
Self References
Bibliography
