A Distinctive Ensemble Approach for Unbiased Movie Recommendations using Variants of Graph Convolution Networks

In recent years, entertainment platforms like Netflix, Amazon Prime etc. have given people personalization control over what they want to watch. These services use recommendation systems to suggest shows and movies based on what users like. Recommendation algorithms use complex computations to understand user’s preferences and provide personalized recommendations based on one’s liking. This empirical study advances the field of movie recommendation by proposing a novel Graph Neural Networks (GNN) ensemble framework for forecasting similar movie preferences, utilizing the Movielens-1M dataset for model training and evaluation. stratified sample from the original dataset is used to ensure the inclusion of all classes equally thereby reducing the variance. Given the limited number of features to be predicted, we opted for a linear Support Vector Machine(SVM), as a meta-classifier, to aggregate the ensemble’s results obtained from three models namely Matrix factorization(MF), Ultra Graph Convolution Network(UltraGCN) and Light Graph Convolution Network(LightGCN). The proposed approach achieved an impressive F1 score of 0.2275 and Normalised Discounted Cummulative Gain(NDCG) of 0.3777, surpassing the individual model scores within the ensemble by 13.5% and 6.1% respectively. The achieved performance is above par with the existing models.