Enhanced Recommendation Systems Using Deep Matrix Factorization with Sentiment Analysis for TripAdvisor
Dani Abizar Ahmad, Agung Toto Wibowo · 2025
This paper proposes an integration approach to deep matrix factorization (DeepMF) with user sentiment analysis to improve the recommendation accuracy of the TripAdvisor platform. Traditional recommendation systems often only utilize ratings as information obtained from user feedback. Sometimes, users can provide more in-depth opinions and assessments of an item through textual reviews. This indicates that traditional recommendation systems struggle to effectively combine quantitative information, such as ratings, with qualitative information, like user reviews. To overcome these limitations, we propose combining the Bidirectional Long Short-Term Memory (Bi-Lstm) model for sentiment analysis with DeepMF to utilize numerical ratings and textual feedback. The sentiment analysis part, which utilizes GloVe embedding and Bi-Lstm, achieves 93.41% accuracy in classifying the sentiment of the reviews. When combined with DeepMF using the optimal weighting scheme, the system achieved a 30% reduction in Mean Absolute Error (MAE) from 0.8391 to 0.5831 and a 33% reduction in Root Mean Square Error (RMSE) from 1.1060 to 0.7379 compared to regular DeepMF. These results demonstrate that incorporating sentiment analysis into deep matrix factorization in a structured manner can enhance recommendation accuracy while more effectively capturing the emotions and nuances of user preferences expressed in textual reviews. Our proposed approach can be adapted to other e-commerce datasets that emphasize the sentiment of user reviews. The ability of the recommendation system must also be improved by considering the results of sentiment analysis on e-commerce.