Session-Based Recommendation by combining Probabilistic Models and LSTM

In this paper, we present the approach, we used as team ”DataLab HMU.GR”, for the ACM RecSys Challenge 2022 [1]. The challenge aims to predict the item that was purchased for a given sequence of item views (session). The full dataset, provided by Dressipi, consists of 1.1 million online retail sessions. Our proposed method, that solves the Session-Based Recommendation problem, relies on an efficient deterministic system based on a weighted combination of Probabilistic models and an LSTM neural network. Probabilistic models learn the transition probabilities between item-item interactions of each session, that are used to predict the purchase probability of an item in a new session. The LSTM neural network takes as input the context representation of the items in a session and a candidate item and predicts the purchase probability of the candidate item. The experimental results demonstrate the high performance and the computational efficiency of the probabilistic models. Our submission achieved the 13th rank and an overall score of 0.1963 in the final competition results. We release our source code at: https://github.com/cpanag79/recsys-Challenge-2022.