A hybrid of deep auto-encoder and feature embedding model for an e-commerce recommender system

Abstract

Recommender systems aim to predict user interests and suggest products that are likely to be of interest. These systems are widely used across various platforms, including online shopping, streaming services, and music stores, to provide personalized suggestions. Traditional machine learning-based models, such as collaborative filtering and content-based algorithms, often face challenges like low accuracy, data sparsity, and the cold start problem. The cold start problem occurs when a system lacks sufficient data to make accurate recommendations for new users or items. This study specifically focuses on addressing the visitor cold start problem, where the system does not have prior information about the new user’s preferences or behavior, making personalized recommendations difficult. To address this issue, a model was developed using deep auto-encoders integrated with feature embedding (DAE-FE), designed to improve item prediction accuracy for new users in an e-commerce recommender system. The model introduces an embedding layer after the dropout layer in the deep neural network, which automatically captures user data points such as time and location. These data points help in constructing a user profile necessary for prediction. This feature not only improves the accuracy of item predictions but also speeds up the process by filling in missing data for new users, allowing the system to proceed directly to prediction. An experimental research design was employed to compare the performance of the developed model with previous models that relied solely on provided datasets. In the experiment, user location and time of login were used as independent variables, while model accuracy served as the dependent variable. The model was trained and tested using the MovieLens 100k dataset, which was adapted to meet the requirements of the DAE-FE model. The hybrid model achieved a mean squared error of 0.0241 and a root mean squared error of 0.1443, indicating minimal deviation from the actual values. As a result, the model attained approximately 96% accuracy in predicting recommendations for cold start users. Overall, the model demonstrated strong performance and appears to be a promising solution for the cold start problem in ecommerce systems. The research found that incorporating more side information from users and items on the dataset during the model's training will yield more accuracy in item prediction.