Explainable Sequential Recommendation With Time-aware Gated Graph Neural Network

The progress of human society depends on the exchange of information.As the core technology to filter information,the recommendation system plays an important role in information entry.With the richness of information types and the diversity of information display methods,a variety of information recommendation forms have also been derived.The sequential recommendation algorithm is a stable and efficient recommendation algorithm,which can effectively model user's recent preferences while integrating user's historical preferences.This paper focuses on sequential recommendation and model interpretability,aiming at the recommendation scenarios of academic papers,and carries out the following studies:(1)Time-GGNN,a time-aware gating graph neural network unit,is proposed to model user behavior time interval information.Time interval information of user behavior is the time interval between two clicks,which is an important information reflecting the correlation between user behaviors.In addition,compared with the Recurrent neural network,the gate graph neural network can better capture the complex transition in the user behavior sequence.On this basis,this paper proposes a graph neural network unit with time interval information.(2)A sequential recommendation algorithm TGGNN4 REC is proposed based on Time-GGNN.After the user behavior sequence is transformed from chain structure to graph structure,the user behavior graph is modeled by Time-GGNN to obtain the vector representation of each behavior,and then the weight of each behavior in the graph of the next click prediction is calculated by the attention mechanism.Finally,the global vector of the sequence graph is calculated by using this weight and each behavior vector.The global vector and the last behavior vector is used as the prediction vector of the next click.Compared with other sequence recommendation algorithms,the performance on two open datasets is improved,and the performance of different length sequences is also improved.(3)In order to explain the process of sequential recommendation by neural network of sequential gate graph,the explainable recommendation algorithm TGGNN4 ER is proposed.Firstly,we use the sequential recommendation algorithm of Time-GGNN to model the user behavior sequential map as the user attribute,and then use the method of combining convolution neural network and attention mechanism to model the comment information of the item as the project preference,in which the comment vector of the attention mechanism to the item is used compared with the user vector,the user's weight of each comment statement is calculated as the interpretation of the recommendation results.Finally,the user preference and item preference are predicted by the factor machine.In the two open datasets,the performance of recommendation is better than that of other interpretable recommendation algorithms.Using "academic headlines" user behavior data and real user feedback to evaluate the interpretation effect is better than other explainable recommendation algorithms.The above work is an important part of Jilin province's key scientific and technological research and development project "research,development and application of fast knowledge sharing system in the era of big data and mobile Internet".The recommended algorithm proposed in this paper has been applied to the "academic headlines" APP(https://www.acheadline.com/)supported by the project.