TD-DNN: A Time Decay-Based Deep Neural Network for Recommendation System

Abstract

In recent years, commercial platforms have embraced recommendation algorithms to provide customers with personalized recommendations. Collaborative Filtering is the most widely used technique of recommendation systems, whose accuracy is primarily reliant on the computed similarity by a similarity measure. Data sparsity is one problem that affects the performance of the similarity measures. In addition, most recommendation algorithms do not remove noisy data from datasets while recommending the items, reducing the accuracy of the recommendation. Furthermore, existing recommendation algorithms only consider historical ratings when recommending the items to users, but users’ tastes may change over time. To address these issues, this research presents a Deep Neural Network based on Time Decay (TD-DNN). In the data preprocessing phase of the model, noisy ratings are detected from the dataset and corrected using the Matrix Factorization approach. A power decay function is applied to the preprocessed input to provide more weightage to the recent ratings. This non-noisy weighted matrix is fed into the Deep Learning model, consisting of an input layer, a Multi-Layer Perceptron, and an output layer to generate predicted ratings. The model’s performance is tested on three benchmark datasets, and experimental results confirm that TD-DNN outperforms other existing approaches.