Comparison and Optimization of the Robustness of Recommendation Models in the Face of Noised Data in E-Commerce

Abstract

Recommendation systems play a crucial role in e-commerce, but their performance is often degraded by noisy data such as accidental clicks, erroneous implicit interactions, and ambiguous user behavior. This study compares the robustness of three recommendation approaches: collaborative filtering, denoising autoencoders, and the DeepFM hybrid model, using the OTTO RecSys dataset containing 2.5 million interactions from e-commerce platforms. An experimental noise injection protocol (0%, 10%, 20%, and 30%) was applied to evaluate model stability in perturbed environments. Performance was measured using the Recall@10, Accuracy@10, F1@10, RMSE, and MAE metrics. The results show a progressive improvement in performance between the models. Collaborative filtering achieves an F1@10 of 0.136 with an RMSE of 0.35. The autoencoder improves performance with an F1@10 of 0.179 and an RMSE of 0.28. DeepFM shows the best results with an F1@ of 0.221, a Recall@10 of 42%, an Accuracy@10 of 15%, and an RMSE of 0.22. The robustness analysis also reveals that DeepFM is the most noise-resistant model. At 30% noise, its performance degradation remains limited to 15.8%, compared to 26.8% for the autoencoder and 37.5% for collaborative filtering. Statistical tests (ANOVA, p<0.05) confirm the significance of the differences observed between the models. These results demonstrate that hybrid architectures based on deep learning, particularly DeepFM, offer better generalization capabilities and greater robustness in noisy e-commerce environments. This study thus confirms the value of integrating robust hybrid models into decision support and commercial personalization systems.