<![CDATA[E-commerce has become one of the most widely used digital applications globally, enabling personalized product discovery and purchasing. To support these services, recommender systems are essential, offering item suggestions based on user preferences. Most recommender systems rely on machine learning algorithms to estimate user-item relevance scores, often utilizing product ratings. However, a persistent challenge in this domain is the issue of data sparsity, where only a small fraction of users provides explicit ratings, leading to reduced accuracy in recommendation results. In this study, we introduce a novel hybrid recommendation algorithm, named AMIKOM-RECSYS, designed to address the sparsity problem and enhance rating prediction. Our model integrates three main components included a Large Language Model (LLM) using ChatGPT, a Transformer-based encoder (BERT), and Probabilistic Matrix Factorization (PMF). The LLM generates descriptive information about movies based on specific prompts, which is then passed to BERT to encode the content into meaningful 2D vector representations. These enriched embeddings are subsequently utilized by the PMF algorithm to predict missing user-item ratings. We evaluate the proposed model on two benchmark datasets, ML-1M and ML-10M using Root Mean Squared Error (RMSE) as the evaluation metric. The AMIKOM-RECSYS model achieved RMSE values of 0.8681 on ML-1M and 0.7791 on ML-10M under a 50:50 data split, outperforming several baseline models including CNN-PMF, LSTM-PMF, and Attention-PMF. These results highlight the effectiveness of integrating LLM and Transformer-based contextual understanding into matrix factorization frameworks. In future work, we plan to extend this framework by incorporating other matrix factorization techniques such as Singular Value Decomposition (SVD) and integrating additional sources of user information, including social media activity, to further improve recommendation performance.]]>
Copyrights © 2025
