Abstract:Multimedia Recommendation Systems (MRSs) have shown to be quite effective in learning consumer preferences and recommending the best multimedia products. Recent breakthroughs in adversarial machine learning have piqued the interest of researchers in the security of MRSs. It has been established that widely deployed MRSs are not resilient to detrimental perturbations applied to the learnt parameters, which can result in a significant loss in recommendation accuracy. Adversarial Multimedia Ranking (AMR) mitigates this problem by boosting Visual Bayesian Pairwise Ranking (VBPR) through adversarial learning. The quantitative gains in AMR’s performance on VBPR have led to its widespread implementation in numerous MRS models. However, in MRSs, this strategy overlooks the collaborative feature and is unable to effectively capture the smoothness of data distribution. We contend that modeling MRSs requires the collaborative feature, which displays the behavioral similarity between consumers and products. In this paper, we implement directed adversarial learning with the explicit introduction of the collaborative feature into the perturbation process. Technically, we propose the Adversarial Dynamic Collaborative Filtering (ADCF) for recommendation, which models visual characteristics and captures visual time dynamics. To reduce the influence of perturbation, we train the ADCF objective through minimax adversarial learning. Furthermore, we enhance the ADCF by directed adversarial learning. The objective is to restrict the direction of perturbation in the embedding space to other examples in the present embedding space. This enables us to integrate the collaborative feature into the learning process. Comprehensive evaluations using three Amazon datasets revealed that our technique outperformed baselines.

Robust multimedia recommender system based on dynamic collaborative filtering and directed adversarial learning