Item Vector Representation in Classified Website Based on Content and User Behavior

With the potential increase in the amount of online information, recommender systems and search engines became the most convenient and effective way to overcome such data overload and thus are one of the critical problems for major Internet companies. However, despite sharing the common task of understanding user intents, each company has to find a unique solution due to a significant disparity in content that needs to be ranked and recommended. This paper focuses on the specific challenge of recommendations design faced by Avito, one of the world largest Classifieds websites, where users have interests in somewhat distant areas, and listings can either change over time or are closed shortly after publication. Based on an idea of semantic vector representations using deep neural networks we describe the concept of \textit{item embeddings}, which utilizes information about user preferences, past user-item interactions, and advertisement content. This approach is evaluated on a large offline dataset of user interactions. Finally, a comparison between variuos modifications of \textit{item embeddings} model will be made.