Geoffray Bonnin

Automated Generation of Music Playlists: Survey and Experiments

Most of the time when we listen to music on the radio or on our portable devices, the order in which the tracks are played is governed by so-called playlists. These playlists are basically sequences of tracks that traditionally are designed manually and whose organization is based on some underlying logic or theme. With the digitalization of music and the availability of various types of additional track-related information on the Web, new opportunities have emerged on how to automate the playlist creation process. Correspondingly, a number of proposals for automated playlist generation have been made in the literature during the past decade. These approaches vary both with respect to which kind of data they rely on and which types of algorithms they use. In this article, we review the literature on automated playlist generation and categorize the existing approaches. Furthermore, we discuss the evaluation designs that are used today in research to assess the quality of the generated playlists. Finally, we report the results of a comparative evaluation of typical playlist generation schemes based on historical data. Our results show that track and artist popularity can play a dominant role and that additional measures are required to better characterize and compare the quality of automatically generated playlists.

What Recommenders Recommend – An Analysis of Accuracy, Popularity, and Sales Diversity Effects

In academic studies, the evaluation of recommender system (RS) algorithms is often limited to offline experimental designs based on historical data sets and metrics from the fields of Machine Learning or Information Retrieval. In real-world settings, however, other business-oriented metrics such as click-through-rates, customer retention or effects on the sales spectrum might be the true evaluation criteria for RS effectiveness. In this paper, we compare different RS algorithms with respect to their tendency of focusing on certain parts of the product spectrum. Our first analysis on different data sets shows that some algorithms – while able to generate highly accurate predictions – concentrate their top 10 recommendations on a very small fraction of the product catalog or have a strong bias to recommending only relatively popular items than others. We see our work as a further step toward multiple-metric offline evaluation and to help service providers make better-informed decisions when looking for a recommendation strategy that is in line with the overall goals of the recommendation service.

A low-order markov model integrating long-distance histories for collaborative recommender systems

Recommender systems provide users with pertinent resources according to their context and their profiles, by applying statistical and knowledge discovery techniques. This paper describes a new approach of generating suitable recommendations based on the active users navigation stream, by considering long and short-distance resources in the history with a tractable model. The Skipping Based Recommender we propose uses Markov models inspired from the ones used in language modeling while integrating skipping techniques to handle noise during navigation. Weighting schemes are also used to alleviate the importance of distant resources. This recommender has also the characteristic to be anytime. It has been tested on a browsing dataset extracted from Intranet logs provided by a French bank. Results show that the use of exponential decay weighting schemes when taking into account non contiguous resources to compute recommendations enhances the accuracy. Moreover, the skipping variant we propose provides a high accuracy while being less complex than state of the art variants.

Biases in Automated Music Playlist Generation: A Comparison of Next-Track Recommending Techniques

Playlist generation is a special form of music recommendation where the problem is to create a sequence of tracks to be played next, given a number of seed tracks. In academia, the evaluation of playlisting techniques is often done by assessing with the help of information retrieval measures if an algorithm is capable of selecting those tracks that also a human would pick next. Such approaches however cannot capture other factors, e.g., the homogeneity of the tracks that can determine the quality perception of playlists. In this work, we report the results of a multi-metric comparison of different academic approaches and a commercial playlisting service. Our results show that all tested techniques generate playlists with certain biases, e.g., towards very popular tracks, and often create playlists continuations that are quite different from those that are created by real users.

History Dependent Recommender Systems Based on Partial Matching

This paper focuses on the utilization of the history of navigation within recommender systems. It aims at designing a collaborative recommender based on Markov models relying on partial matching in order to ensure high accuracy, coverage, robustness, low complexity while being anytime. Indeed, contrary to state of the art, this model does not simply match the context of the active user to the context of other users but partial matching is performed: the history of navigation is divided into several sub-histories on which matching is performed, allowing the matching constraints to be weakened. The resulting model leads to an improvement in terms of accuracy compared to state of the art models.

Skipping-Based Collaborative Recommendations inspired from Statistical Language Modeling

Due to the almost unlimited resource space on the Web, efficient search engines and recommender systems have become a key element for users to find resources corresponding to their needs. Recommender systems aims at helping users in this task by providing them some pertinent resources according to their context and their profiles, by applying various techniques such as statistical and knowledge discovery algorithms. One of the most successful approaches is Collaborative Filtering, which consists in considering user ratings to provide recommendations, without considering the content of the resources; however the ratings are the only criterion taken into account to provide the recommendations, although including some other criterion should enhance their accuracy. One such criterion is the context, which can be geographical, meteorological, social, etc. In this chapter we focus on the temporal context, more specifically on the order in which the resources were consulted. The appropriateness of considering the order is domain dependent: for instance, it seems of little help in domains such as online moviestores, in which user transactions are barely sequential; however it is especially appropriate for domains such as Web navigation, which has a sequential structure. We propose to follow this direction for this domain, the challenge being to find a low enough complexity sequential model while providing a better accuracy. We first put forward similarities between Web navigation and natural language, and propose to adapt statistical language models to Web navigation to compute recommendations. Second, we propose a new model inspired from the n-gram skipping model. This model has several advantages: (1) It has both a low time and a low space complexity while providing a full coverage, (2) it is able to handle parallel navigations and noise, (3) it is able to perform recommendations in an anytime framework, (4) weighting schemes are used to alleviate the importance of distant resources. Third, we provide a comparison of this SLM inspired model to the state of the art in terms of features, complexity, accuracy and robustness and present experimental results. Tests are performed on a browsing dataset extracted from Intranet logs provided by a French bank. Results show that the use of exponential decay weighting schemes when taking into account non contiguous resources highly improves the accuracy, and that the anytime configuration is able to provide a satisfying trade-off between an even lower computation time and a good accuracy while conserving a good coverage.

Comparing Audio Features and Playlist Statistics for Music Classification

In recent years, a number of approaches have been developed for the automatic recognition of music genres, but also more specific categories (styles, moods, personal preferences, etc.). Among the different sources for building classification models, features extracted from the audio signal play an important role in the literature. Although such features can be extracted from any digitised music piece independently of the availability of other information sources, their extraction can require considerable computational costs and the audio alone does not always contain enough information for the identification of the distinctive properties of a musical category. In this work we consider playlists that are created and shared by music listeners as another interesting source for feature extraction and music categorisation. The main idea is that the tracks of a playlist are often from the same artist or belong to the same category, e.g. they have the same genre or style, which allows us to exploit their co-occurrences for classification tasks. In the paper, we evaluate strategies for better genre and style classification based on the analysis of larger collections of user-provided playlists and compare them to a recent classification technique from the literature. Our first results indicate that an already comparably simple playlist-based classifiers can in some cases outperform an advanced audio-based classification technique.

Empfehlungssysteme, automatische Erzeugung von Wiedergabelisten und Musikdatenbanken

Die passende Musik fur einen gewunschten Anwendungszweck auszuwahlen, etwa fur eine Wiedergabeliste fur Hintergrundmusik oder als Untermalung in einem Werbespot, ist aufgrund von verschiedensten Anforderungen und der schieren Menge an verfugbaren Stucken ein aufwandiger Prozess. Es existieren zahlreiche Kriterien, beispielsweise Metadaten, aber auch die Beschaffenheit der Musik selbst, anhand derer ein Stuck charakterisiert werden kann. Mithilfe von Empfehlungssystemen – speziellen Algorithmen, die Elemente anhand festgelegter Kriterien auswahlen konnen – lasst sich dieser Prozess vereinfachen und teilweise automatisieren. Ihre Daten beziehen solche Systeme oft aus sogenannten Musikdatenbanken, die Informationen uber Musikstucke aggregieren und kategorisieren, und damit die Moglichkeit bieten, Titel nach verschiedenen Kriterien zu finden, dem Anwendungszweck gemas auszuwahlen und oft auch direkt zu erwerben oder abzuspielen. In diesem Kapitel wird das Problem der automatisierten Erstellung von Wiedergabelisten charakterisiert sowie algorithmische Ansatze im Uberblick vorgestellt. Anschliesend wird eine Ubersicht uber aktuelle Online-Musikdatenbanken gegeben.