1st Budapest RecSys & Personalization Meetup!
Details
Deep learning to the rescue - solving long standing problems of recommender systems
Balázs Hidasi - Head of Data Mining and Research @ Gravity RD
Recommender system research mainly focuses on a small fragment of recommendation tasks: personalization. While we are drowning in algorithms for personalized recommendations, they are not suitable for a wide range of practical recommendation scenarios.Standard algorithms are also not able to incorporate some key information, which greatly influences the users' decision on clicking the recommended item.
Meanwhile, the breakthroughs of the last decade in neural network research and the quick increasing of computational power resulted in the revival of deep neural networks and the field focusing on their training: deep learning.Deep learning methods have succeeded in complex tasks where other machine learning methods have failed, such as computer vision and natural language processing.Recently deep learning has begun to gain ground in recommender systems as well.
This talk briefly introduces deep learning and demonstrates how it can tackle long standing recsys problems and thus create a new generation of recommender algorithms.A concrete solution is also presented in detail: recurrent neural networks for item-to-session recommendations.
---------------------
Uncovering concepts and similarities at scale
Theodore Vasiloudis - Researcher @ SICS Swedish ICT
Finding similarities is one of the fundamental problems in machine learning.We use similarities between users to make recommendations, we use similarities between websites to do web searches etc.In this talk we will show how we can calculate similarities at scale using graph processing, and how we can then use the resulting similarity graphs to uncover higher-order concepts in multiple domains, including text, music, and molecular biology.
---------------------
Towards online learning for music recommendation
Róbert Pálovics - Associate researcher at the Informatics Laboratory of the Hungarian Academy of Sciences
While online learning algorithms have the capability of adapting very fast to new training data, they cannot iterate through the event stream several times, resulting in theoretically poor convergence properties.Online recommenders seem more restricted and one would expect inferior quality compared to their batch variants.Online methods however have the advantage of giving much more emphasis on recent events.In our music recommendation experiments, surprisingly, recommendation by online learning outperforms the batch methods.In our results, we show how to fuse the advantages of both online and batch machine learning worlds.We develop and evaluate methods that combine on-line analytics with real-time insights on current trends with batch processing that allows to explore in more details the trends that have emerged during the real-time phase.Our experiments are carried over the "30Mmusic" listening dataset.