A Hybrid GAN-Based Approach to Solve Imbalanced Data Problem in Recommendation Systems

Abstract

With the advent of information technology, the amount of online data generation has been massive. Recommendation systems have become an effective tool in filtering information and solving the problem of information overload. Machine learning algorithms to build these recommendation systems require well-balanced data in terms of class distribution, but real-world datasets are mostly imbalanced in nature. Imbalanced data imposes a classifier to focus more on the majority class, neglecting other classes of interests and thus hindering the predictive performance of any classification model. There exist many traditional techniques for oversampling minority classes. Still, generative adversarial networks (GAN) have been showing excellent results in generating realistic synthetic tabular data that keeps the probability distribution of the original data intact. In this paper, we propose a hybrid GAN approach to solve the data imbalance problem to enhance recommendation systems' performance. We implemented conditional Wasserstein GAN with gradient penalty to generate tabular data containing both numerical and categorical values. We also augmented auxiliary classifier loss to enforce the model to explicitly generate data belonging to the minority class. We designed the discriminator architecture with the concept of PacGAN to receive m-packed samples as input instead of a single input. This inclusion of the PacGAN architecture eliminated the mode collapse problem in our proposed model. We did a two-fold evaluation of our model. Firstly based on the quality of the generated data and secondly on how different recommendation models perform using the generated data compared to original data.