LEARNING USER PREFERENCES USING EVOLUTION Supiya Lijin and PeterS. Bentley Department of Computer Science University College London, Gower Street, London WCIE 6BT S Ujin @cs. ucl. ac uk P, Bentley @cs. ucl. ac uk ABSTRACT opropriate similarities is often difficult. To overcome these problems, current systems(such as MovieLens)use Recommender systems are new types of internet-based stochastic and heuristic-based models to speed up and software tools, designed to help users find their way improve the quality of profile matching. This work takes through today's complex on-line shops and entertainment such ideas one step further, by applying an evolutionary websites. This paper describes a new recommender algorithm to the problem of profile matching system, which employs a genetic algorithm to learn In this research. the movieLens dataset personalpreferencesofusersandprovidetailored(http://www.movielens.umn.edu),wasusedforinitial suggestions experiments. The evolutionary recommender system uses 22 features from this data set: movie rating, age, gender 1 INTRODUCTION occupation and 18 movie genre frequencies: action, As the name suggests, recommender systems' task is to documentary, drama, fantasy, film-noir, horror, musical commend or suggest items or products to the customer mystery, romance, sci-fi, thriller, war, western based on his/her preferences. These systems are often used by E-commerce websites as marketing tools to increase 2. 1. Profile Generator revenue by presenting products that the customer is likely to buy. An internet site using a recommender system can Before recommendations can be made, the movie data is exploit knowledge of customers'likes and dislikes to build processed into separate profiles, one for each person, an understanding of their individual needs and thereby defining that person's movie preferences. We define increase customer loyalty [1, 2 profile, i) to mean the profile for user on movie item i, fine-tune a profile-matching algorithm within a collection of profile, i) for all the items i that j has see 9 This paper focuses on the use of evolutionary search to e fig. 1. The profile of j, profile() is therefore recommender system, tailoring it to the preferences of individual users. This enables the recommender system to I Rating 2 22 18 Genre make more accurate predictions of users' likes and dislikes. and hence better recommendations to users Figure 1: profile, i)- profile for user j with rating on movie item i, if i The paper is organised as follows: section 2 outlines has a rating of 5 related work, and section 3 describes the recommender Once profiles are built, the process of recommendation system and genetic algorithm. Section 4 provides can begin. Given an active user A, a set or neighbourhood xperimental results and analysis. Finally section 5 of profiles similar to profile(A)must be found 2.2. Neighbourhood Selection 2. SYSTEM OVERVIEW The success of a collaborative filtering system is highly The system described in this paper is based around dependent upon the effectiveness of the algorithm in collaborative filtering approach, building up profiles of finding the set or neighbourhood of profiles that are most users and then using an algorithm to find profiles similar to that of the active user. It is vital that, for a to the current user. (In this paper, we refer to the particular neighbourhood method, only the best or closest user as the active user, A). Selected data from m profiles are chosen and used to generate new profiles are then used to build recommendations. Because recommendations for the user. There is little tolerance fo profiles contain many attributes, many of which have inaccurate or irrelevant predictions incomplete data [4 the task of finding The neighbourhood selection algorithm consists of threeLEARNING USER PREFERENCES USING EVOLUTION Supiya Ujjin and Peter J. Bentley Department of Computer Science University College London, Gower Street, London WC1E 6BT S.Ujjin@cs.ucl.ac.uk P.Bentley@cs.ucl.ac.uk ABSTRACT Recommender systems are new types of internet-based software tools, designed to help users find their way through today’s complex on-line shops and entertainment websites. This paper describes a new recommender system, which employs a genetic algorithm to learn personal preferences of users and provide tailored suggestions. 1. INTRODUCTION As the name suggests, recommender systems’ task is to recommend or suggest items or products to the customer based on his/her preferences. These systems are often used by E-commerce websites as marketing tools to increase revenue by presenting products that the customer is likely to buy. An internet site using a recommender system can exploit knowledge of customers' likes and dislikes to build an understanding of their individual needs and thereby increase customer loyalty [1, 2]. This paper focuses on the use of evolutionary search to fine-tune a profile-matching algorithm within a recommender system, tailoring it to the preferences of individual users. This enables the recommender system to make more accurate predictions of users' likes and dislikes, and hence better recommendations to users. The paper is organised as follows: section 2 outlines related work, and section 3 describes the recommender system and genetic algorithm. Section 4 provides experimental results and analysis. Finally section 5 concludes. 2. SYSTEM OVERVIEW The system described in this paper is based around a collaborative filtering approach, building up profiles of users and then using an algorithm to find profiles similar to the current user. (In this paper, we refer to the current user as the active user, A). Selected data from those profiles are then used to build recommendations. Because profiles contain many attributes, many of which have sparse or incomplete data [4], the task of finding appropriate similarities is often difficult. To overcome these problems, current systems (such as MovieLens) use stochastic and heuristic-based models to speed up and improve the quality of profile matching. This work takes such ideas one step further, by applying an evolutionary algorithm to the problem of profile matching. In this research, the MovieLens dataset (http://www.movielens.umn.edu), was used for initial experiments. The evolutionary recommender system uses 22 features from this data set: movie rating, age, gender, occupation and 18 movie genre frequencies: action, adventure, animation, children, comedy, crime, documentary, drama, fantasy, film-noir, horror, musical, mystery, romance, sci-fi, thriller, war, western. 2.1. Profile Generator Before recommendations can be made, the movie data is processed into separate profiles, one for each person, defining that person’s movie preferences. We define profile(j,i) to mean the profile for user j on movie item i, see fig. 1. The profile of j, profile(j) is therefore a collection of profile(j,i) for all the items i that j has seen. 1 Rating 2 Age 3 Gender 4 Occupation ..22 18 Genre frequencies 5 23 0 45 000000100010000000 Figure 1: profile(j,i) - profile for user j with rating on movie item i, if i has a rating of 5. Once profiles are built, the process of recommendation can begin. Given an active user A, a set or neighbourhood of profiles similar to profile(A) must be found. 2.2. Neighbourhood Selection The success of a collaborative filtering system is highly dependent upon the effectiveness of the algorithm in finding the set or neighbourhood of profiles that are most similar to that of the active user. It is vital that, for a particular neighbourhood method, only the best or closest profiles are chosen and used to generate new recommendations for the user. There is little tolerance for inaccurate or irrelevant predictions. The neighbourhood selection algorithm consists of three