Citation Context

...el that does not require an external similarity measure and that is trained directly on the data of interest, namely historical playlists. In particular, we extend the Logistic Markov Embedding (LME) =-=[3]-=- approach to probabilistic sequence modeling to incorporate social tags, unifying song and tag embeddings in a single Euclidean space. This provides a probabilistically well-founded and constructive w...

Citation Context

...at generate trajectories of musical preferences by jointly embedding listeners and the songs they play in a single metric space. In order to do this, we extend existing probabilistic playlists models =-=[1, 2]-=- by adding time dynamics, allowing users and songs to change position on a multi-year scale. By examining these models, we can draw conclusions about the behavior of listeners and musical artists Perm...

Citation Context

...eds to be sufficiently rich and diverse, but also reflect real playlists “in the wild”. In this paper, we used two separate sources. The first, the Yes.com archive, is corpus collected by Chen et al. =-=[7]-=-. These playlists and related tag data were respectively crawled from Yes.com and Last.fm. Chen et al. harvested data between December 2010 and May 2011, yielding 75,262 songs and 2,840,553 transition...

Citation Context

...ee whether modeling intransitivity helps or not. Tangentially related, pairwise comparison also arises as a subroutine of multi-class classification problems [23, 4]. The goal here is to assign one or more best classes to each instance given the pairwise comparisons among all the classes. It differs from ours, as our focus is about any individual comparison for prediction. The idea of learning multi-dimensional representations in a semantically meaningful latent space has also become a popular and effective method in many applications, including language modeling [30, 34], playlist generation [10, 31, 11], co-occurrence data modeling [17], recommendation system [19, 41] and image/social media tagging [39, 40]. Also related is the work [1], in which the authors use matrix factorization to predict scores of professional basketball games. The idea of using different feature functions for offense and defense is analogous to the model we propose. The major difference lies in three aspects. First, it does not explicitly studies the intransitive behavior. Second, the input differs, as our models takes simple binary win or lose results and theirs needs detailed scores. Lastly, we go beyond the specifi...

Citation Context

...dding problems that go beyond the Gaussian model of rating prediction. For example, when the embedding is used to model probability distributions over discrete objects like sequences (e.g., playlists =-=[18]-=-, words in a sentence[10], purchases [16]) or complex preferences (e.g., as extension to [15]), computation time and memory for storing data and model become a bottleneck. In this paper, we explore tr...

Citation Context

... features, meta-data, or social tags [9, 15, 25]. However, even though long-term user preferences intuitively play a key role in personalizing the recommendations, nearly all the playlist generation approaches reviewed in the recent surveys in [7, 18] only consider the user’s short-term listening history. In this work, we therefore aim to explore the value of long-term preferences for personalizing the next-track recommendation process. More precisely, instead of basing the recommendations solely on the sequence of seed tracks from the users’ recent listening histories, as done, e.g., in [7], [12], or [15], we incorporate multi-dimensional signals extracted from the users’ long-term preferences and their social network into the recommendation process. In particular, we look at a user’s long-term preferences to identify and recommend (a) tracks that the user liked in the past (track repetition), (b) tracks of artists that the user liked in the past (favorite artists), (c) tracks that are semantically similar to past liked ones (topic similarity), and (d) tracks that are often played together with those that the user likes (track co-occurrence). In addition, we look for (e) tracks and ar...

Citation Context

.../2507157.2507215. Figure 1: A screenshot of ihou.com, where a user is singing. music recommendation techniques. Methods such as neighborhood based methods[14], matrix factorization[9] and Markov model=-=[5]-=- are commonly adopted. These technical achievements further secured the success of many music station websites, such as Pandora1 and douban.fm2. In summary, most of the existing works try to predict t...

Citation Context

... known items can be compensated for. This is called the cold-start problem [20]. Content-based approaches which are designed to solve the cold-start problem can suffer from lack of diversity when recommended items are selected simply by similarity. This is often called top-N recommendation. It is well known that unexpectedness, surprise or serendipity play an important role in the music recommendation and discovery [4]. Compared to other strategies, focusing on transitions can naturally provide these qualities. There have been approaches that primarily focus on the transitions of tracks [13], [3], [15]. They assumed the Markov property of hidden states or embeddings of tracks. Using the Markov property, it is assumed that future events only depend on the current one and does not depend on the past. This has been successfully used for sequence modelling for instance in speech [19] too. In music computing, playlist datasets [16], [14], [3] collaboratively created for reference by DJs and listeners were used for training and evaluation of sequence modelling approaches. Although these datasets consist of a large number of tracks, e.g. 101k playlists in [16], the lack of audio data fundame...

Citation Context

...while balancing exploration (of new policies) and exploitation (of known good policies) (Li et al. 2010). However, learning a policy from scratch, by exploring and exploiting feedback obtained from real users, is a dangerous game: by the time a good policy is learned, all users would have abandoned the system. One is therefore forced to collectively leverage other available data sources to bootstrap a good initial policy. Conversely, much work has focused on learning from explicitly labelled data (Koren, Bell, and Volinsky 2009; Mnih and Salakhutdinov 2007), or user-provided example policies (Chen et al. 2012a). However, these approaches do not directly optimise online performance, and it is currently not clear how to bridge the gap between these approaches. Copyright c© 2016, Association for the Advancement of Artificial Intelligence (www.aaai.org). All rights reserved. In this paper we formalise the task of learning a behaviour policy from heterogeneous data sources as a policy transfer learning problem. Given a set of exploration data, and additional heterogeneous data sources (e.g., user-provided example policies), how can the additional data be used to guide the policy learned from exploratio...

Citation Context

...-art collaborative filtering models for implicit feedback with additional tagging activity. As future work, we want to investigate the automated generation of music playlists. As already suggested in =-=[2, 7, 21, 29]-=-, user-generated data can be successfully exploited for this task. We will explore how our extensions to state-ofthe-art collaborative filtering models can be translated into sequential recommendation...