A Recommender System using Holographic Associative Memory Matthew Rutledge-Taylor Institute of Cognitive Science Carleton University, 1125 Colonel By Drive Ottawa, Ontario, K1S 5B6 Canada mrtaylo2@connect.carleton.ca Andr´ e Vellino CISTI Research National Research Council Ottawa, Ontario K1A 0R6 andre.vellino@nrc.ca Abstract This paper describes a recommender system based on a cognitive model of associative memory that uses Holo- graphically Reduced Representations as the basis for its encoding of object associations. We compare such a recom- mender to a standard user-based collaborative filtering al- gorithm on two datasets - MovieLens and two bibliographic datasets for a digital library. Experimental results show .... 1 Introduction Most recommender systems have been built by cluster- ing similar items according to some characteristic of the item (content-based recommendation) or by measuring the similarity among ratings that users have given to items (col- laborative filtering, either user-based or model-based) or by combining the two (hybrid recommenders.) (see section 2 of [1] for a brief survey.) This paper describes a rather different recommendation technique based on a cognitive model of associative mem- ory. ¡give some brief summary of DHSM here¿CITE [?] and [2], The intent behind the series of experiments described in section 2 was to better understand the behavioural proper- ties of a recommender based on holographically reduced representations for the kind of very sparse bibliographic datasets that are found in a digital library. Our objective was not to discover a new recommendation technique that was more effective or efficient for typical recommender tasks. 1.1 Collaborative Filtering Typically, recommender systems operate on three kinds of entities, users, items and the preference ratings that users have given items. Given a set of ratings by users for certain items - whether they are obtained from users explicitly or implicity from, for example, browsing patterns - a collabo- rative filtering system will attempt to predict the rating of an item for a given user based on how other users previously rated the same item. Collaborative filtering, when applied to bibliographic items in a digital library, has several disadvantages. First the problem of data sparsity is much more pronounced in digital libraries than it is in recommenders for commodity products. Attempts at remedying this problem by using bib- liographic citations as a substitute for user ratings [3] is par- tial at best since bibliographic references, while an indicator of relevance, is not necessarily a favourable vote in the mind of the author. Secondly, bibliographic datasets have no standard benchmarks for testing recommender quality. Mention [4] 1.2 DHSM Dynamically Structured Holographic Memory (DSHM) is a computational model of human long-term memory [5, 6]. It makes use of holographic reduced representations (HRRs) to encode associations between concepts [2]. Each item is encoded by two large vectors of real valued numbers with a Euclidean length of 1.0. The environmental vector is static and is the system’s internal representation of the item and acts as the item’s name. The memory vector is dynamic and stores associations between the item and other items (and combinations of items) in the system. Associa- tions between items are formed when a set of items is given as input into the system. From a cognitive perspective this can be interpreted as the items co-occurring in a thought, verbal utterance or a perception. If the set of items is un- ordered, every item is associated with every sub-set of the other items in the set, up to a predefined maximum number of elements. These associations are recorded by binding the environmental vectors of the items in each sub-set to the memory vector of the given item. A binding is formed by recursively computing the circu-