Next item prediction using neural networks with embedding initialized weights

Session-based recommendation systems became a very part of humankind's daily life, as a result of the increasing transaction volume of e-commerce and e-marketing fields. Companies that can analyze the trails left by their current users on their systems and know their customers better than other firms, can become one step ahead of their rivals. In this concept, representations of items become a key point when related deep learning models are investigated closer. Since the relationship between each item within a session will have a direct effect on the task that involves predicting the next item in that session, extracting these relationships among each item needs to be handled in an effective manner. Using auto encoder systems to achieve the task of revealing hidden features between items and representing these relationships in a more meaningful way, will result in both boosting current state-of-art models' performance and offering new session-based methods that can overperform the current state-of-art models. In this paper, state-of-the-art graph neural networks SR-GNN's and TA-GNN's weights are initialized with item embeddings that are obtained from autoencoder with RBM layers, and the performance of the models are compared with random weight initialization. According to the proposed weight initialization, using pre-trained item embeddings will increase the performance of the recommender model. Thanks to the pre-trained item embeddings, the hidden relationship between items modeled in a better way, and the introduced model has overperformed the current state-of-art techniques.