Recently, a paper that Max, Eva and Günther authored together with Stefan Brandl and Markus Schedl of the Johannes Kepler Universität Linz was accepted for publication at the ISMIR 2022 conference. The paper is titled Verse Versus Chorus: Structure-aware Feature Extraction for Lyrics-based Genre Recognition, and investigates how the predictive power of lyrical features is impacted by which part of a song they are extracted from.

We look forward to presenting our work at ISMIR 2022 in December!

Together with Christine Bauer (Utrecht University) and Alan Said (University of Gothenburg), Eva will co-organize the 2nd Workshop on Perspectives on the Evaluation of Recommender Systems at this year's ACM Recommender Systems Conference.

Our manuscript on HOT (Height Optimized Tries) has now been published in ACM Transactions on Database Systems (ranked A*). In this article, we present an extended version of our previous SIGMOD paper on HOT - a fast and space-efficient in-memory index structure.

This week our our submission "HSP Datasets: Insights on Song Popularity Prediction" has been accepted for the IJSC special issue on IEEE ISM 2021.

Our IJSC journal paper is an extension of our paper "Novel Datasets for Evaluating Song Popularity Prediction Tasks" accepted at the ISM 2021. We extended our publication by providing further experiments and results. This extension is in three directions. First, we now include experiments utilizing Essentia’s mel-band features contained in our dataset. Second, we now additionally include experiments using the Essentia features computed on the short audio samples. Lastly, we extended the number of models used per data source and feature set. We now include a k-nearest neighbor approach for spectral-based features and extend the number of models utilized for predictions based on Mel-spectrograms. Doing so increased the total number of experiments where we report results from 96 to 330.

This also results in a fully reworked Results and Discussion section. We now show how to utilize our results (and subsets of them) to answer a diverse set of research questions regarding song popularity prediction. Based on our baselines, we discuss the best model per feature type. Further, we elaborate on the best hit song prediction approach among the included baseline models. In addition, we shed light on the question if all included measures of popularity are equally hard to predict and finally, we discuss if features extracted from the short audio samples have equal predictive power to those gathered from AcousticBrainz.

Last week, a dataset paper authored by DBIS members has been accepted at the ISM 2021:

The paper "Novel Datasets for Evaluating Song Popularity Prediction Tasks" authored by Michael Vötter, Maximilian Mayerl, Günther Specht, and Eva Zangerle has been accepted at the 23rd IEEE International Symposium on Multimedia (ISM 2021). In this paper, we present two novel datasets that can be utilized for hit song prediction tasks. For further details on the datasets, see: https://doi.org/10.5281/zenodo.5383858

This week, we welcome a new DBIS member: Andreas Peintner. Andreas will work as a PhD student in the field of recommender systems, particularly working on our HumRec project. 

It's a busy week for DBIS: on Wednesday and Thursday, the results of the PAN challenges are presented at the CLEF conference (Conference and Labs of the Evaluation Forum). Maximilian and Eva have co-organized the Style Change Detection task as part of PAN (together with Martin Potthast and Benno Stein), Eva will present the task and the participant's results and approaches on Thursday. Find an overview and participant's papers as part of the proceedings on ceur-ws.

On Saturday, the Perspectives on the Evaluation of Recommender Systems Workshop (co-located with ACM Recommender Systems 2021) will take place. Eva is co-organizing this workshop together with Christine Bauer (Utrecht University) and Alan Said (Gothenburg University). Find the program, teaser videos for the papers and further information on the workshop website. We published the workshop proceedings on ceur-ws.

Benjamin's paper "Small-Scale Cross-Language AuthorshipAttribution on Social Media Comments" was published at the Low Resource Machine Translation Workshop of this years' Machine Translation Summit 2021. In the paper, Benjamin analyzes different methods to cross the language gap between documents of multilingual authors.

In their last board meeting, the Austrian Science Fund (FWF) approved Eva's stand-alone project proposal titled "Humans and Recommender Systems: Towards a Mutual Understanding" (grant awarded: 599k EUR). In the next three years, we will focus on music recommendations and strive to enhance the understanding of human decision-making underlying the choice of certain music in a given situational context. Furthermore, we aim to advance the users' understanding for the decisions that lead to the recommendation of certain (sequences of) tracks. We believe that an increased understanding and communication between users and the system can contribute to improved user models and, thus, recommendation performance. A previously largely unexplored aspect will be the development of techniques for sequential recommendation strongly targeted at giving explanations for these sequences and considering user feedback.

The project consortium features, besides Eva as principal investigator, Markus Schedl (Johannes Kepler University Linz), Peter Knees (Vienna University of Technology), Marcel Zentner (University of Innsbruck, Department of Psychology), and Michael Huber (University of Music and Performing Arts Vienna).

Our new study published in EPJ Data Science shows that music recommendations for fans of beyond-mainstream music, such as hard rock and ambient, may receive be less accurate recommendations than for fans of mainstream music, such as pop. Together with Dominik Kowald, Peter Müllner and Elisabeth Lex (TU Graz and Know-Center, Austria), Christine Bauer (University of Utrecht, The Netherlands), and Markus Schedl (Johannes-Kepler-Universität Linz, Austria), Eva has investigated the effects of popularity bias in music recommender systems for listeners of non-mainstream-listeners. Our findings provide valuable insights for developing improved user models and recommendation approaches to better serve beyond-mainstream music listeners. 

Our findings have also been covered by press, among others, by the Austrian newspaper derStandard, Rolling Stone Italia, El Pais (Spain), de Volkskrant (The Netherlands), BioMedCentral, or a blog post by TU Graz, Austria. For a more complete list of news coverage, visit Altmetric.
 

Eva will be co-organizing the PERSPECTIVES Workshop: Perspectives on the Evaluation of Recommender Systems at the ACM Recommender Systems Conference 2021 in Amsterdam, The Netherlands. Together with Christine Bauer and Alan Said. In this workshop, we aim to capture the current state of evaluation of recommender systems, and gauge whether there is, or should be, a different target that RS evaluation should strive for. 

Our article titled "Leveraging Affective Hashtags for Ranking Music Recommendations" recently appeared in IEEE Transactions on Affective Computing (impact factor 7.512). Together with Yi-Hsuan Yang (Academia Sinica, Taiwan) and Chih-Ming Chen and Ming-Feng Tsai (both National Chengchi University, Taiwan), Eva extracted affective contextual information from hashtags that music listeners use to describe music on Twitter. The gathered information is modelled as a graph and via state-of-the-art network embedding methods, we learn latent feature representations of users, tracks and hashtags. Based on these representations, we propose eight ranking methods for personalized music recommendations.
 

DBIS is presenting two workshop papers at the MediaEval 2020 on Monday.

At 14:00, we will present our solution to Music mood and theme detiction, and at 16:00 we talk about our solution for detecting conspiracies regarding 5g and Corona.

The registration is free, so feel free to drop in and join the discussions.

This week, we finished two submissions to two challenges organized by the MediaEval benchmarking initiative.

In FakeNews: Corona virus and 5G conspiracy, the task is to classify Twitter messages into three categories, and determine whether they contain assumptions correlating 5G with the outbreak of the Corona virus, or not.

In Emotions and Themes in Music, we predict user-defined tags associated to music tracks by analyzing low-level audio features.

Stay tuned for the results of our approaches in these challenges, as well as our working notes.

Last week, two conference papers co-authored by DBIS member Eva Zangerle have been accepted at conferences:

The paper "Personality Bias of Music Recommendation Algorithms" together with Alessandro Melchiorre and Markus Schedl of JKU Linz has been accepted as a short paper at the 14th ACM Conference on Recommender Systems (acceptance rate for short papers: 20%). The paper "Pandemics, music, and collective sentiment: evidence from the outbreak of COVID-19" together with Meijun Liu and Xiao Hu (both University of Hong Kong) and Alessandro Melchiorre and Markus Schedl (both JKU Linz) has been accepted at the International Society for Music Information Retrieval Conference.