Weighted Hybrid recommendation system using autoencoders

No Thumbnail Available
Narayana, Sushmitha Chennaiahnapalya
Issue Date
Msc in Artificial Intelligence
Dublin Business School
Deciphering user-item interaction data effectively has become critical to improving user experience and driving business growth in the exponentially expanding digital environment. This research is primarily concerned with exploring the possibility of enhancing recommendation systems' accuracy through the use of a weighted hybrid model, By extracting latent features through autoencoders and seamlessly integrating them with item-based filtering, our system aims to capture and predict user preferences with improved precision. In our experiments, the base autoencoder model exhibited an RMSE score of 0.4980, while the base item-based approach registered an RMSE score of 0.2813. Conclusively, the final weighted hybrid model yielded a score of 0.2666913, underscoring the efficacy of combining different models. The results showcased that the model exhibited an RMSE score of 0.2666913 and the corresponding weights were w1(weight for item-based)=0.1 and w2(weight for autoencoder)=0.9. The report initially reviews the current state of recommendation systems and autoencoders, followed by a comprehensive understanding of the autoencoder mathematical framework.