Publication

Abstract : Collaborative filtering, while a powerful tool in movie recommendation systems, encounters substantial challenges such as sparsity, scalability, diversity, and interpretability, which influence recommendation fidelity. Although the sparsity issue can be tackled through traditional model-based collaborative filtering algorithms like matrix factorization, these models often fail to capture the full depth of user-movie interactions due to their reliance on a simple dot product for rating prediction. Recently, research has leaned towards the application of deep learning to harness the complex and non-linear relationships between users and movies. However, these deep learning methods, despite their non-linear attributes, are susceptible to high variance and overfitting, potentially compromising their capacity for generalization. In the present study, an ensemble of neural networks has been implemented to diminish variance and generalization error by amalgamating the predictions from multiple models. This approach enhances overall performance and tackles the inherent limitations of deep learning approaches. A novel ensemble-based deep collaborative filtering model (Deep CF), in concert with a unique optimizer (AdaMVRGO), has been introduced to address sparsity and to exploit the non-linear, complex relationships between users and movies. It also aims to minimize the variance and generalization error of the neural network. The proposed architecture, termed Deep CF-AdaMVRGO, employs an ensemble of multi-layer perceptrons (MLP) to augment prediction accuracy. A pioneering optimizer, the adaptive moment variance reduced gradient optimization (AdaMVRGO), has been developed, drawing upon the ADAM and SVRG optimizers. This optimizer eliminates noise by calculating the first and second moments of predicted ratings using a variance-reduced gradient, similar to the SVRG algorithm, thereby expediting algorithm convergence. It is employed to fine-tune the parameters of the MLP and decrease the reconstruction error. Deep CF-AdaMVRGO has been evaluated against six existing models using the RMSE metric on the M-1M and M-10M datasets. The simulation results demonstrated that the proposed framework outperformed state-of-the-art deep learning-based collaborative filtering approaches on both datasets in terms of lower RMSE values. Further, the performance of the proposed AdaMVRGO optimizer within the ensemble framework was compared to existing optimizers such as Adagrad, RMSProp, ADAM, and SVRG on M-100K, M-1M, and M-10M datasets using RMSE, MAE, and MSE metrics. Experimental results affirmed that AdaMVRGO converged more rapidly to the optimum compared to other optimizers.