Research On Recommendation System Based On Deep Neural Factorization Machine

In the Internet era of information explosion,various industries generate hundreds of millions of data every day.On the one hand,in order to improve their own competitiveness,companies need to continuously optimize their management to provide users with high-quality services,meanwhile quickly and accurately recommend products that are currently or potentially interesting to users,thereby digging into profit margins.On the other hand,in the face of rapidly changing products and services,users also want to be able to quickly locate the products they need.However,it will consume immeasurable manpower and material resources by manually selecting and processing data to accurately recommend products for millions of active users.Therefore,designing a recommendation system with excellent performance and sound interpretability is of great significance for enterprises to improve product competitiveness and expand market space.Most of the existing factor models are not interpretable,and the interactions between features are unpredictable.For enterprises,a good recommendation system must not only have higher prediction accuracy,but also more importantly how to make the model interpretable.The advantage of an interpretable model is that it can also provide the hidden results of the network's internal feature interaction to a certain extent while giving prediction results,thereby making the ”black box” mechanism inside the network visible to the enterprise and helping the enterprise develop more reliable strategies.Therefore,this paper designs an interpretable neural factor model based on explicit interactions between domains.The feature interaction mode of the model enables the domain representation vectors to maintain practical significance in the network,thereby further introducing the self-attention mechanism to provide models with interpretability.On this basis,the model also improves the prediction performance to a certain extent.In addition,the performance improvement of the existing factor-based integrated models based on individual models is very limited,and the interaction between individual models cannot be fully utilized.Therefore,a new integrated model is proposed in this paper,and the correlation between individual models is strengthened through feature fusion.The prediction performance can be further improved compared with existing integrated models.The innovations and contributions of this paper are as follows:· The interaction mode occurs at the domain level,which can keep the domain's dimensions unchanged,and provides a structural basis for introducing the attention mechanism.In addition,this paper also takes the lead in adopting a subdomain separation method,which uses convolution layers that do not share parameters to perform feature extraction,which significantly improves the representation ability.· The self-attention mechanism is introduced into a higher-order neural factor model,and a weight matrix is learned by introducing the self-attention mechanism in each subdomain,which provides a sound interpretability for the model's prediction results.· A dual-flow network structure combining explicit and implicit interaction is designed.The implicit interaction network is a selfencoding network with vector-level feature interaction.the fusion of the two components occurs in the earlier feature extraction layer,which strengthens the interaction between the two components and makes the performance of the dual-flow model more robust.· click-through rate prediction and movie rating prediction tasks ontwo large datasets,Criteo and Movie Lens,are designed,and conducte a lot of comparative experiments with multiple advanced and classic benchmark models.In addition,the paper expatiates the interpretability of the proposed model by visualizing the attention matrix.In the era of big data on the Internet,the recommendation system is of great significance for companies to enhance their own competitiveness,increase user stickiness,and increase profit margins.In this paper,the recommendation system is studied in depth from three aspects of explicit interaction,interpretability and multi-stream network,and a neural factor decomposition model based on explicit interaction between domains is proposed.Experiments on two large data sets demonstrate the effectiveness of the model proposed in this paper in real scenarios.