A cross selling recommender system based on recurrent neural networks for online shopping

Recommender systems are considered to be capable of prediction of what is the next product to buy for a specific customer. For an effective cross-selling framework it is crucial to identify which customers are more suitable than others to target a product for the retail industry. This thesis proposes a hybrid model which implements recurrent neural network architectures with self-attention mechanism processing implicit feedback. Furthermore, the proposed design is capable of handling both sequential and non-sequential features, which correspond to purchase behavior and non-behavioral customer specific information respectively. This study represents an alternative solution to a well-known business problem: improving cross-selling effectiveness by estimating customers' likelihood for which products or services to buy next time. A recommender system which works on additional data configurations is the core concept of the framework. This study is verified with an online shopping dataset, and it is illustrated that the concatenation of relevant features provides additional information to the model, and it is obtained approximately 13% of improvement in evaluation metrics.