Combining Cmab with Matrix Factorization and Clustering For Enhanced Movie Recommendations

The contextual multi-armed bandit (CMAB) algorithm faces problems of a large-scale, sparse data matrix and many arms. This paper proposes a method that combines matrix factorization with a clustering algorithm to enhance the performance of the CMAB algorithm. For the MovieLens 1M dataset, data processing is conducted through matrix factorization to optimize the algorithm's input features, and a clustering method is used to decrease the number of arms, thereby improving the recommendation efficiency and performance of the CMAB algorithm. During the experiment of the proposed method, Funk Singular Value Decomposition (FunkSVD) is employed to fill the user-movie sparse matrix with predicted ratings, then the SSE of K-means clustering algorithm is utilized to determine the optimal number of clusters, in which 3952 movies is clustered into 96 groups, significantly reducing the number of arms and exploration cost of the algorithm. The hyperparameter of Linear Upper Confidence Bound (LinUCB) is also optimized by achieving the lowest cumulative regret value. By comparing the performance of LinUCB with Upper Confidence Bound (UCB) and Thompson Sampling (TS), LinUCB notably outperforms the traditional Multi-Armed Bandit (MAB) algorithms in cumulative regret values, regret bucket statistics (approximately 650 times/1000 runs of recommending high-quality movies), and convergence slope (0.0718).