Attention based Long-Short Term Memory Model for Product Recommendations with Multiple Timesteps

摘要

A recommender system is an engine that helps users find information on products and services. These products and services can be anything from books, digital content, music, videos, essentials, and more. Shuffling through numerous pages on the web may be hectic, that is where a recommender system may come handy in suggesting them products or services based on their past browsing history. The key motivation of this research is to improve the accuracy of existing benchmark model (Long-Short Term Memory with Multi Period - LSTM_MP) based on the previous purchase history. In the previous studies, limitations are seen in certain areas for e.g., usage of very less amount of data for evaluations due to which changing preferences of customers have not been identified. Also, they fail to accurately predict products properly. To overcome such shortfalls, researcher has come up with ALSTMM which is attention based Long Short-Term Memory Model which tells network where attention should be paid in input sequence of items corresponding to output sequence. The proposed model uses enormous and diversified real transactional data from Amazon, identifies changing preferences of customers over period and recognizes more patterns based on past purchase transactions. Our research focuses on suggesting a multi-time stamp product recommender system that is efficient enough to study past purchase patterns and accurately predict the next purchase order. With the help of RNN (Recurrent Neural Networks) for time series data analysis, the model segments recommendation periods into numerous time stamps which can then be used by the system to recommend products. Several experiments have been performed using datasets from Amazon and Instacart, which reveal that this model is a substantially improved version of the older systems and excel in accuracy and diversity when compared to CF-based models. Based on the results shown at the end of this study, the proposed model is found to be best for multi-time-based product recommendation purposes. In terms of applications, this model can effectively contribute towards reducing the shopping times and manual effort.