A Recommender system based on the Immune Network Proceedings CEC2002, pp 807-813, Honolulu, USA, 2002 Steve Cayzer and Uwe Aickelin Hewlett-packardlaBoratoriesFiltonRoadBs126qzBristol,Uk,stevecayzer@hp.com School of Computer Science, University of Nottingham, NG8 1BB UK, uxa @cs. nott. ac uk Abstract-The immune system is a complex biological system continue to spread through the population and potentially has with a highly distributed, adaptive and self-organising nature. much explanatory power. The idiotypic network has been This paper presents an artificial immune system(AIS)that formalised by a number of theoretical immunologists[15] exploits some of these characteristics and is applied to the task There are many more features of the immune system, evolution and in particular the immune system have not been including adaptation, immunological memory and protection designed for classical optimisation. However, for this problem, against auto-Immune attack. Since these are not directly we are not interested in finding a single optimum. Rather we relevant to this work, they will not be reviewed here ntend to identify a sub-set of good matches on which recommendations can be based. It is our hypothesis that an AIs Overview of collaborative Filtering be an ideal candidate to achieve this: Antigen .antibody t In this paper, we are using an AIS as a CF technique. CFis term for a broad range of algorithms that use similarity interaction for matching and antibody - antibody interaction for measures to obtain recommendations. The best-known diversity.computatioNalresultsarepresentedinsupportofthisexampleisprobablythe"peoplewhoboughtthisalso bought"feature of the internet company Amazon [2] However, any problem domain where users are required to . INTRODUCTION rate items is amenable to CF techniques. Commercial applications are usually called recommender systems [16].A Over the last few years, a novel computational intelligence canonical example is movie recommendation chnique, inspired by biology, has emerged: the artificial In traditional CF, the items to be recommended are treated immune system(AlS). This section introduces the AIs and That is, yo commendations are based problems. In essence, the immune system is used here as content of the item. The preferences of a user, usually a set of inspiration to create an unsupervised machine-learning votes on an item, comprise a user profile, and these profiles lgorithm. The immune system metaphor will be explore are compared to build a neighbourhood. The key decisions to involving a brief overview of the basic immunological be made are theories that are relevant to our work. We also introduce the Data encoding: Perhaps the most obvious representation basic concepts of collaborative filtering(CF) for a user profile is a string of numbers, where the length the number of items, and the position is the item identifier Overview of the Immune System Each number represents the 'vote for an item. Votes are a detailed overview of the immune system can be found sometimes binary(e.g. did you visit this web page? but can many textbooks [14]. Briefly, the purpose of the immune also be integers in a range(say [0, 5])or rational numbers stem is to protect the body against infection and includes a Similarity Measure: The most common method to cor set of mechanisms collectively termed humoral immunity. two users is a correlation-based measure like Pearson or This refers to a population of circulating white blood cells Spearman, which gives two neighbours a matching score called B-lymphocytes, and the antibodies they create The features that are particularly relevant to our research between vectors, and probabilistic methods are alternative are matching, diversity and distributed control. Matching The canonical example is the k Nearest Neighbo Diversity refers to the fact that, in order to achieve optimal algorithm, which uses a matching method to select k antigen space coverage, antibody diversity must be reviewers with high similarity measures. The votes from hese reviewers, suitably weighted, are used to make encouraged [11]. Distributed control means that there is no predictions and recommendations central controller, rather, the immune system is governed by local interactions between cells and antibodies Many improvements on this method are possible [10]. For The idiotypic network hypothesis [13](disputed by some example, the user profiles are usually extremely sparse immunologists )builds on the recognition that antibodies can because many items are not rated. This means that similarity match other antibodies as well as antigens. Hence,an antibody may be matched by other antibodies, which in turn dimensionality,) and difficult to calculate due to the small may be matched by yet other antibodies. This activation can overlap. Default votes are sometimes used for items a userA Recommender System based on the Immune Network Proceedings CEC2002, pp 807-813, Honolulu, USA, 2002. Steve Cayzer1 and Uwe Aickelin2 1Hewlett-Packard Laboratories, Filton Road, BS12 6QZ Bristol, UK, steve_cayzer@hp.com 2 School of Computer Science, University of Nottingham, NG8 1BB UK, uxa@cs.nott.ac.uk Abstract-The immune system is a complex biological system with a highly distributed, adaptive and self-organising nature. This paper presents an artificial immune system (AIS) that exploits some of these characteristics and is applied to the task of film recommendation by collaborative filtering (CF). Natural evolution and in particular the immune system have not been designed for classical optimisation. However, for this problem, we are not interested in finding a single optimum. Rather we intend to identify a sub-set of good matches on which recommendations can be based. It is our hypothesis that an AIS built on two central aspects of the biological immune system will be an ideal candidate to achieve this: Antigen - antibody interaction for matching and antibody - antibody interaction for diversity. Computational results are presented in support of this conjecture and compared to those found by other CF techniques. I. INTRODUCTION Over the last few years, a novel computational intelligence technique, inspired by biology, has emerged: the artificial immune system (AIS). This section introduces the AIS and shows how it can be used for solving computational problems. In essence, the immune system is used here as inspiration to create an unsupervised machine-learning algorithm. The immune system metaphor will be explored, involving a brief overview of the basic immunological theories that are relevant to our work. We also introduce the basic concepts of collaborative filtering (CF). Overview of the Immune System A detailed overview of the immune system can be found in many textbooks [14]. Briefly, the purpose of the immune system is to protect the body against infection and includes a set of mechanisms collectively termed humoral immunity. This refers to a population of circulating white blood cells called B-lymphocytes, and the antibodies they create. The features that are particularly relevant to our research are matching, diversity and distributed control. Matching refers to the binding between antibodies and antigens. Diversity refers to the fact that, in order to achieve optimal antigen space coverage, antibody diversity must be encouraged [11]. Distributed control means that there is no central controller, rather, the immune system is governed by local interactions between cells and antibodies. The idiotypic network hypothesis [13] (disputed by some immunologists) builds on the recognition that antibodies can match other antibodies as well as antigens. Hence, an antibody may be matched by other antibodies, which in turn may be matched by yet other antibodies. This activation can continue to spread through the population and potentially has much explanatory power. The idiotypic network has been formalised by a number of theoretical immunologists [15]. There are many more features of the immune system, including adaptation, immunological memory and protection against auto-immune attack. Since these are not directly relevant to this work, they will not be reviewed here. Overview of Collaborative Filtering In this paper, we are using an AIS as a CF technique. CF is the term for a broad range of algorithms that use similarity measures to obtain recommendations. The best-known example is probably the “people who bought this also bought” feature of the internet company Amazon [2]. However, any problem domain where users are required to rate items is amenable to CF techniques. Commercial applications are usually called recommender systems [16]. A canonical example is movie recommendation. In traditional CF, the items to be recommended are treated as ‘black boxes’. That is, your recommendations are based purely on the votes of your neighbours, and not on the content of the item. The preferences of a user, usually a set of votes on an item, comprise a user profile, and these profiles are compared to build a neighbourhood. The key decisions to be made are: Data encoding: Perhaps the most obvious representation for a user profile is a string of numbers, where the length is the number of items, and the position is the item identifier. Each number represents the 'vote' for an item. Votes are sometimes binary (e.g. did you visit this web page?) but can also be integers in a range (say [0,5]) or rational numbers. Similarity Measure: The most common method to compare two users is a correlation-based measure like Pearson or Spearman, which gives two neighbours a matching score between -1 and 1. Vector based, e.g. cosine of the angle between vectors, and probabilistic methods are alternative approaches. The canonical example is the k Nearest Neighbour algorithm, which uses a matching method to select k reviewers with high similarity measures. The votes from these reviewers, suitably weighted, are used to make predictions and recommendations. Many improvements on this method are possible [10]. For example, the user profiles are usually extremely sparse because many items are not rated. This means that similarity measurements are both inefficient (the so-called ‘curse of dimensionality’) and difficult to calculate due to the small overlap. Default votes are sometimes used for items a user