ELEKTROTEHNI ˇ SKI VESTNIK 78(4): 177–180, 2011 ENGLISH EDITION Emotional properties of latent factors in an image recommender system Marko Tkalˇ ciˇ c, Andrej Koˇ sir, ˇ Stefan Dobravec, Jurij Tasiˇ c Univerza v Ljubljani, Fakulteta za elektrotehniko, Trˇ zaˇ ska 25, 1000 Ljubljana, Slovenija E-poˇ sta: marko.tkalcic@fe.uni-lj.si Abstract. In this paper we analyze the relations between the latent factors with high variance description and affective parameters in an image recommender system. Using the matrix factorization approach we identify the main two factors in the user-item rating database. We exploit the affective metadata related to each item to identify relations between the main factors and the affective metadata. Results show that the first latent factor is strongly related with the valence and dominance while the arousal does not appear to be related. The second factor, however, shows no relation with the affective parameters. Keywords: recommender systems, affective computing, matrix factorization, latent factors 1I NTRODUCTION Recommender systems for multimedia content (e.g. films, images, music, books, etc.) are systems that exploit the knowledge about an observed user’s pref- erences (the user profile) and the knowledge about the multimedia item properties (the item profile) to filter a limited set of multimedia content suited to the user’s tastes. Recommender systems have outgrown the labo- ratory environment and found their place in commercial applications. Amazon, for example, has been using their solution to recommend shopping items on their online store [5]. Research work has been following mainly two paths: (i) the algorithmic path, where the existing algorithms have been improved, or new algorithms have been proposed, and (ii) the descriptor path, where new features have been sought in order to increase the variance. The first algorithms were very simple content-based recommenders (CBR) [6] and collaborative filtering recommenders (CF) [1]. However, during the Netflix competition (http://www.netflixprize.com/), the matrix factorization approach turned out to outperform other approaches in recommender systems [2], [3]. So far, the matrix factorization approach is regarded as the best approach when the training dataset is big enough. On the other hand, researchers were looking out for new features to include in the user and item profiles that would carry more information for predicting relevant content to recommend. The first features used were generic metadata like the genre, actors, director, etc. [1], [7]. As the algorithms were improving, they were able to get the maximum information from these metadata Received August 26, 2011 Accepted September 25, 2011 and researchers started exploring new kinds of features for improving the performance of recommender systems. In our previous work we identified emotional metadata to describe considerable variance in user’s data thus improving the performance of the CBR system [10]. The introduction of the matrix factorization approach in recommender system replaced the former human understandable features (e.g. genre, etc.) with latent features, which are not necessarily human understand- able. The first two latent factors in the Netflix dataset turned out to be explained as intellectual-shalllow and masculine-feminine, respectively [3], [2]. This line of reasoning, the interpretation of the properties of the main latent factors, is interesting for further research because it opens different perspectives on users’ preferences and their modeling. 1.1 Problem statement The goal of this paper is to explore the properties of the main latent factors of a recommender system dataset in terms of the emotions that the content items induce in end users. The problem adressed in this paper arises from two presumptions: (i) the matrix factorization algo- rithm identifies the main latent factors that describe the variance in the user-item rating matrix and (ii) the items’ affective parameters (the parameters that describe the emotion that an item induces in the user) vary along the main factors’ axes. Thus we aim at using the explanatory factor analysis to identify the affective properties that characterize the items at the ends of the main latent factors, as we depicted in Fig. 1. We denote the groups of items as the groups G 1.1 ,G 1.2 ,G 2.1 and G 2.2 . The result is the visualization and interpretaion of these properties.