Hello! We publish a report from the Avito Data Science Meetup: Personalization, which took place in our office. Participants discussed the modeling of user preferences in multimodal data and the clustering of volatile ads using the EM algorithm. Under the cut - videos, presentations, link to photo report.

Modeling user preferences in multimodal data. Hady W. Lauw, Maxim Tkachenko (Singapore Management University)

The key to good recommendations is modeling huge amounts of behavioral data that arise as a result of user interaction with online systems. These interactions are multimodal, that is, composed of various types of data, such as user ratings, reviews, photos, or their social interactions. This complicates the task. The speakers talk about data mining and machine learning methods for modeling user preferences in multimodal data and their use in creating a complete recommender system.

The report was in English, we translated it and added subtitles in Russian:

Presentation

Clustering volatile ads using the EM algorithm. Vasily Leksin (Avito)

Working on the development of recommendations on Avito, colleagues decided to cluster short-lived ads: this will help users see more relevant recommendations, and we will be less likely to retrain models and do it faster. Vasily presents an optimized EM algorithm capable of efficiently processing huge data arrays and talks about methods for assessing the quality of clustering and about the applied applications of the algorithm.

Presentation

Thanks to everyone who came to the meeting, watched the video. We posted the photo report from the meeting on Facebook . To be the first to know about technical events for Avito technicians, subscribe to our Timepad . And be sure to tell in the comments on what topics you would like to listen to the reports.

See you again!