基于注意力因子分解机和深度神经网络的推荐算法

Recommendation algorithm has been a hot topic of research, which is a technology that uses historical user behavior data, combined with user characteristics and other information, and analyzes and processes them through algorithmic models to provide personalized recommendations to users. Recommendation algorithms can improve user satisfaction and usage experience, as well as increase the conversion rate and efficiency of e-commerce, advertising and other businesses. In this paper, we address the problem of identical feature weights in DeepFM algorithm, which leads to unimportant feature interactions wasting arithmetic power and hindering model performance improvement. With reference to the design of Attention Net in AFM model, DeepAFM (Attentional Factorization-Machine based on Deep Neural Network) algorithm is improved and proposed to assign weights to each intersection feature in the second-order part to solve the feature sparsity problem and effectively learn the inter-feature cross-relationships, avoid manual feature engineering, reduce arithmetic waste, and improve model efficiency and readiness. Experiments comparing the DeepAFM algorithm with six other baseline algorithms on two publicly available datasets demonstrate that the DeepAFM algorithm has better accuracy performance, and the algorithm is more efficient and interpretable.