Research On Sequence Recommendation With Attention Mechanism And Multimodality

With the rise of Internet and mobile and the explosive growth of data,people’s demand for the effectiveness and accuracy of information is increasing,and recommendation systems are born,and recommendation algorithms are the core of recommendation systems.Based on the user’s interaction records with various products on the web,recommendation algorithms model them and can predict the user’s next click behavior to a certain extent.In recent years,with the application of deep learning in personalized recommendation systems,the accuracy of prediction has also made great progress,but there are still the following problems that are difficult to solve: the recommendation model based on Recurrent Neural Network(RNN)is difficult to capture both long-term and short-term preferences of users,and cannot explicitly capture the interactions between items in the whole history.It is also difficult to determine the capturing dimension when capturing short-term preferences,and the user’s recent mis-clicks will make the recommendation inaccurate,while the capturing dimension is too high and the interest will drift.At the same time,in RNN-based models,the directly input features are often treated as independent individuals,such as audio,text,and images in micro-video recommendations,but features are intrinsically related to each other,and even different features have different importance in recommendations,and RNN-based recommendation models ignore the rich internal representation of each feature,which is not conducive to the improvement of recommendation performance.To address the above problems,the main research of this thesis includes.1.To address the problem that RNN-based recommendation models cannot capture both long-term and short-term preferences,we propose a method to divide long-term and short-term preferences into high and low orders,and expand the last click behavior into the last L-click behavior,which can effectively prevent the inaccuracy of the recommendation list due to misclicks.In order to prevent the interest drifting phenomenon caused by capturing the high dimension of short-term preferences,a recommendation model with a two-layer attention mechanism is proposed,which inputs the expanded L-click items into a self-attentive module,captures the internal information of them by matching them with itself,and assigns different weights to each click behavior,effectively preventing the interest drifting phenomenon caused by capturing the high dimension.The model was tested for comparison using ML-1M,Video Game and Digital Music datasets and showed significant improvements in both HR@50 and MRR metrics.The role of different parameters in the experiments is analyzed in detail through ablation experiments,which fully verifies the importance of the two-layer attention module in the recommendation algorithm.2.To address the problem of low performance caused by treating each input feature as an individual in the micro-video recommendation model and ignoring its internal relationship,an attention-based multimodal feature fusion method is proposed,which assigns different weights to each modality through the attention module and incorporates user features into the attention network to achieve personalized fusion.Also considering that the new features obtained by feature crossover are likely to hide the invisible representation of user preferences,an attention-based mechanism is also adopted to learn the crossover representation of features.In the experiments,the fusion method of directly splicing the features of each modality resulted in a 0.86% reduction in the Log Loss metric and a 1.16% improvement in the AUC metric,and the addition of the extracted crossover features resulted in a 1.4% reduction in the Log Loss metric and a 2.01% improvement in the AUC metric.The experimental results demonstrate that learning crossover features by attention helps to predict users’ preferences and has a significant effect on micro-video recommendation capability,which also verifies the importance of crossover features in recommendation systems.