关键词: 推荐算法, 自编码器, 注意力机制, 协同过滤

Abstract: With the arrival of an era of big data, the data of Internet is growing at an explosive rate, and users are inundated with quite a few choices, and this phenomenon is known as information overload. As an indispensable decision-making tool, recommender systems can effectively alleviate the information overload, and have been widely applied in various scene. However, the data collected in recommender systems are often sparse, leading to a higher susceptibility of the algorithm to overfitting, which has become one of the key challenges to designing high quality personalized recommendation algorithm. Moreover, the majority of recommendation algorithms overlook the distribution of users' attention towards item characteristics, making it difficult to mine comprehensive and accurate preference information and suggest satisfactory items.
In order to effectively extract user preference information and improve the performance of recommendation results, an autoencoder recommendation algorithm fused with attention mechanism is proposed. To improve the generalization ability and memory ability of the classical encoder, the proposed algorithm first designs the corresponding feature extraction modules for the low-order and high-order features contained in the data, which are named low-order feature extraction module and high-order feature extraction module. Then, the algorithm fuses the low-order feature and high-order feature to obtain the final vector representing users' preference information through the designed attention mechanism. Finally, a decoder is used for calculating the preference rating on items of user, and generate the recommendation result based on preference ratings.
To validate the effectiveness of the proposed algorithm, we conduct an ablation study on ML-100K dataset. The experimental results demonstrate that both the low-order feature extraction module and high-order feature extraction module contribute to mining user preferences, and the feature fusion of low-order feature and high-order feature based on attention mechanism can obtain more precise preference information. Furthermore, we compare our algorithm with ItemPop, CDAE, CFGAN, and Wide&Deep, which are the classic and the state of the art models. The experimental results on ML-100K, ML-1M and Yahoo Music three datasets show that the proposed algorithm significantly improves Precision, Recall, F1 value, and normalized discounted cumulative gain (NDCG), respectively.
The proposed algorithm is applicable to Internet recommendation scenarios, which can fully mine users' preference information in data, provide users with high-quality recommendation results, improve users' satisfaction and increase product transaction volume. However, the model in this paper primarily focuses on the interaction between users and items, neglecting contextual scene information. Thus, future research can consider incorporating more user information, item information, and contextual scene information into the modeling process to further improve the performance of the proposed model. Additionally, as user preferences evolve over time and in response to environmental changes, integrating time and environmental factors should be considered a pivotal research focus in the future.

Key words: recommender system, autoencoder, attention mechanisms, collaborative filtering