Hyper-parameter Optimization of Session-based Recommendation Systems

Abstract

This thesis proposes a comprehensive framework to recommend session-based algorithms and tune their hyperparameters. In the first part of this study, we present a comprehensive evaluation of the state-of-the-art deep learning approaches used in the session-based recommendation. Furthermore, we present an evaluation of neural-based models’ performance using the AutoML Neural Network Intelligence framework. In session-based recommendation, the system counts on the sequence of events made by a user within the same session to predict and endorse other more likely items to correlate with his preferences. Our extensive experiments investigate baseline techniques (e.g., nearest neighbors and pattern mining algorithms) and deep learning approaches (e.g., recurrent neural networks, graph neural networks, and attention-based networks). The first evaluation shows that advanced neural network models outperform the baseline techniques in most of the scenarios. However, we found that these models suffer more in long sessions when there is a drift in user interests, and there is not enough data to correctly model different items during training. The first study suggests that using hybrid models of different approaches combined with baseline algorithms could result in session-based recommendations based on dataset characteristics. The second evaluation shows that AutoML improved the performance of neural-based models by using a stateof- the-art algorithm such as Bayesian Optimization and hyperband (BOHB) to replace the random selection of hyper-parameters.