Hybrid Approach for Improvement of Recommendation System with Latent Features and Improvement of Sparsity Using Inference Rules

Abstract: Streaming services today really struggle with providing personalized recommendations at a large scale, which calls for some pretty advanced modeling techniques. This paper presents a new recommendation framework designed specifically for Netflix, combining collaborative, content-based, and time forecasting methods. We used Apriori association rule mining to find hidden patterns in genres and metadata, built a knowledge graph to make things easier to explain, and applied k-Means clustering to learn similarities. For predictions, we went with a multi-layer perceptron (MLP) deep neural network, and on top of that, we included a Prophet time-series model to forecast content trends. When we tested our approach, we ended up with a Mean Absolute Error (MAE) of 38.23 and a Mean Squared Error (MSE) of 2144.39 for regression tasks, while the k-Means part scored a silhouette score of 0.4170. Overall, this hybrid setup showed better robustness and interpretability than traditional systems, making it a solid option for suggesting content in large video-on-demand platforms.

Keywords: Hybrid Recommendation Systems, Association Rule Mining, Knowledge Graphs, Deep Learning, k - Means Clustering, Time-Series Forecasting, Lecture Notes in Computer Science.