Do you have a Pop face? Here is a Pop song. Using profile pictures to mitigate the cold-start problem in Music Recommender Systems Eugenio Tacchini Ramon Morros Veronica Vilaplana Enrique Sañoso Università Cattolica di Universitat Politècnica de Universitat Politècnica de Universitat Politècnica de Piacenza Catalunya Catalunya Catalunya eugenio.tacchini@unicatt.it ramon.morros@upc.edu veronica.vilaplana@upc.edu enriquesv19@gmail.com ABSTRACT We focus just on the new user problem in the music domain, When a new user registers to a recommender system service, the assuming we do not have any information about user preferences system does not know her taste and cannot propose meaningful and relying exclusively on her profile picture. To our knowledge, suggestions (cold-start problem). This preliminary work attempts there is no work that has attempted to use profile pictures to guess to mitigate the cold-start problem using the profile picture of the preferences. This method is not proposed as a substitution for user as a sole information, following the intuition that a existing approaches; it can be used when the RS does not have correspondence may exist between the pictures that people use to information about a user and can be combined with other methods represent themselves and their taste. We proved that, at least in in order to increase the accuracy in a cold-start situation. the small music community we used for our experiments, our method can improve the precision of both a classifier and a Top-N 3. METHODOLOGY music recommender system in a cold-start condition. 3.1 Background, goals and method Our dataset comes from Last.FM: through the Last.fm APIs we Categories and Subject Descriptors can retrieve users' profile pictures and listening logs. Last.FM also H.3.3 [Information Search and Retrieval] allows users to create and join Groups; a group is a place where people talk about a topic, in particular, there are groups related to Keywords music genres e.g. the Pop group or the Jazz group. We assumed Top-N recommendations, cold-start, evaluation, pictures. that a user joins a group if she is interested in that specific genre. Our experiments have been performed with users belonging to 1. INTRODUCTION three quite different genre groups: Pop, Black Metal and Jazz (P, One of the limits of Collaborative Filtering (CF) Recommender M and J). The three genres traditionally have a different audience: Systems (RSs) is the "cold start problem": when a new user their fans tend to have not only a different music taste but also registers to a RS, the system does not know her taste and cannot different styles so we thought that a dataset coming from M, P and propose meaningful suggestions until the user provides some J was a good starting point to test our intuition. We retrieved for feedback. Nowadays it is very common to provide a profile each user her profile picture and listening logs, we built a picture when we register to a Web site (including web-based playcount matrix M(n,m) (n=3,000 users, m=48,868 artists) and RSs), furthermore in many cases users register using Social for each user-artist pair we stored the number of times that user Network accounts, which allows in turn to access their profile listened to that artist. The users dataset, together with some details pictures. This work analyzes users’ profile pictures to provide and examples, has been released here [5]. Starting from M(n,m) hints about their musical taste and thus provide better we built a preferences binary matrix P(n,m) that represents, for recommendations since the registration, without additional input. each user, the artists she liked. P was computed using an approach similar to the one used in [6]; we assumed a user liked an artist if she listened to the artist more than five times. 2. RELATED WORK The cold-start problem hides two different subproblems: the new We used the dataset for two different goals. The first one was a user problem (users need to rate some items before getting classification problem: given the profile picture of a user (without meaningful suggestions) and the new item problem (items need to other information) and the information related to all the other be rated by some users before being suggested). For the new user users (profile pictures and groups they belong to), can we predict problem, a solution is to fill the missing ratings with default to which group the user belongs to (M, P or J)? To guess the values such as, for each item, the average rating received by other group a user belongs to, we used a k-nearest neighbors (kNN) users [1]. Other approaches involve the segmentation of the users approach. The nearest neighbors of a user in this context were the in homogeneous classes and the suggestion of items suitable for a ones having the most similar profile pictures (see section 3.2). The specific class. Some classification criteria in literature are: data prediction was based on the groups the picture-neighbors coming from questionnaires and demographic data [2,3]. An belonged to: if most of the picture-neighbors of user Ux belonged alternative approach [4] relies on the detection of communities to M, we predicted Ux belonged to M as well. through Social Networks analysis: for a new user, the RS suggests The second goal was related to a Top-N recommendation problem the items typically liked by the community she belongs to. in a cold-start situation: a user Ux has just subscribed to a RS service and we want to suggest N artists. If we do not know anything about her taste we will end up suggesting random artists Copyright is held by the author(s). RecSys 2015 Poster Proceedings, or the most popular artists. Given the profile picture of Ux and the September 16-20, 2015, Austria, Vienna profile pictures and preferences (P) of all the other users, can we provide to Ux a meaningful Top-N artists recommendation list? To those artists. Fig. 1 (bottom) shows the results: the random exploit the information from the user's picture, we mimicked a CF approach performs very poorly (precision 0.003). On our user-based technique using, as a similarity measure between two FaceBasedRecommender method, as expected for a kNN users, the similarity between their profile pictures. Given a user approach, the precision increases as k increases and at some point Ux, we selected her k nearest picture-neighbors, we computed the starts decreasing until it reaches the value of SuggestPopular. At k list of the N most appreciated artists by the picture-neighborhood = 250 reaches its maximum: 0.2883, overcoming the precision and we suggested them to Ux. provided by SuggestPopular by 10.01%. This preliminary work shows that profile pictures can be used to 3.2 Image analysis mitigate the cold-start problem. Our hypothesis have been tested Visual inspection of several profile pictures shows that the content with both a classification and a Top-N recommendation of the pictures is very heterogeneous. Some people use pictures of experiment. As future work, we will explore ways to explicitly their faces, while other use images of objects, places, logos, model correlations between musical taste and pictures using cartoons, etc. However, we expect that users with similar musical KCCA [8]. Another line of research will be improving the tastes select profile images that are related in some way. description of the image content by combining color and texture To compute image similarities, images can be described using information using Bag of Features [9] based on color SIFT different types of visual information such as color, texture, shape descriptors. Also, we will make experiments with more users, of objects or similar characteristics. The MPEG-7 Visual Standard from different genre groups. [7] specifies several content-based descriptors which can be used to efficiently identify, filter or browse images or video. The experiments have been performed using several MPEG7 image descriptors [7] (Dominant Colors, Color Layout, Color Structure and Edge Histogram). For space reasons we only present results obtained with Color Structure (CS), the best performing one. CS captures information about both color content and spatial arrangement of this color content. It is a histogram counting the number of times a color is present in a windowed neighborhood, as this window progresses over the image rows and columns. This enables it to distinguish, for example, between an image in which pixels of each color are distributed uniformly and an image in which the same colors occur in the same proportions, but are located in distinct blocks. The matching function used to compare the CS of two images is the L1 metric. We then convert distance into similarity multiplying distance values by -1. 4. RESULTS AND FUTURE WORK To evaluate our method we used a leave-one-out approach. In the Figure 1: Classification and recommendation precision classification experiment, for each of the n users, we alternatively hid the group she belonged to and we tried to predict it according 5. REFERENCES to her picture-neighbors. Fig. 1 (top) shows the results of the [1] Ricci, F. et al. (2011). Recommender systems handbook. experiment: the precision at various levels of k (number of New York: Springer. (Chapter 4 and 8) neighbors). Our method (FaceBasedClassifier) already overcomes [2] Park, S.T. et al. Pairwise preference regression for cold-start RandomClassifier at k=1 and reaches the maximum value at recommendation. Proc of the 3rd ACM conference on k=281 where the precision is 0.467 (46.70% correct predictions); Recommender system. 2009. since M, P and J were composed by the same number of users, we assume, for RandomClassifier, a precision of 0.333; therefore our [3] Lika, B. et al. (2014). Facing the cold start problem in method overcomes RandomClassifier, at k=281, by 40.24%. recommender systems. Expert Systems with Applications, 41(4), 2065-2073. In the recommendation experiment, for each of the n users, we [4] Sahebi, S. et al. "Community-based recommendations: a alternatively hid her preferences and we tried to predict them solution to the cold start problem.", RSWEB. 2011. using her picture-neighbors. We used, as an accuracy metric, the precision, defined as number of true positive divided by the sum [5] Dataset and some more details: http://ds.dreamhosters.com/ of true positive plus false positive, where a true positive here is an [6] Tacchini, E. (2012), Serendipitous Mentorship in Music artist correctly guessed. Precision is a typical metric used for off- Recommender Systems. (Ph.D. Thesis). line evaluation of a Top-N task ([1]). We experimented with [7] B. S. Manjunath et al. Introduction to MPEG-7, Multimedia N=20 and we tested the precision at various level of k (number of Content Description Interface, J. Wiley and Sons, Ltd., 2002. neighbors), the final precision is the average of the users' precision and we compared it with the precision provided by [8] D. R. Hardoon et al. Canonical Correlation Analysis: An SuggestRandom and SuggestPopular. The SuggestRandom Overview with Application to Learning Methods. Neural method suggested, for each user, a different set of artists, Computation, 16(12), 2004. randomly extracted among all the m available. The [9] G. Csurka et al. Visual categorization with bags of SuggestPopular method computed in advance the 20 most keypoints. In Workshop on Statistical Learning in Computer appreciated artists in the community of 3,000 users and suggested Vision, ECCV, volume 1, page 22. Citeseer, 2004.