Research Of Sequential Recommendation Algorithm Based On Multi-level Behavior Modeling And Diversified Preference Mining

Recommender systems learn user preferences from their historical behavior,providing personalized information and services to users,and are one of the key technologies to alleviate information overload.Sequential recommendation is an important field of recommendation system.Compared with traditional recommendation methods,sequential recommendation takes into account the sequential dependencies between user’s behaviors,that is,the user’s current behavior may be based on his previous behaviors.By modeling the sequential dependencies,the sequential recommendation model can better mine the changing preferences of users and improves recommendation accuracy.In spite of plenty of achievements in this field,existing methods still face two main challenges: 1)Existing methods only consider user interactions on the item level,making it difficult for these methods to effectively model the sequential dependencies due to the sparsity of behavior sequence data;2)Users often interact with items for various reasons so that the model cannot determine which aspect of the preference is the main cause of the behavior if a coupled approach is used to model users’ multi-faceted preferences.This would result in a lack of interpretability in the recommendation results.Specifically,the main contributions of this paper include:1)To address the first challenge,we propose a sequential recommendation algorithm based on multi-level behavior modeling.Through statistical analysis experiments,we find that user behavior sequences at the category level,compared to item-level behavior sequences,are more conducive to modeling sequential dependencies.Therefore,this algorithm introduces additional item category information and performs multi-level modeling of user behavior to better mine user preferences.Specifically,the algorithm first uses two self-attention based encoders to model user behavior sequences from both the item and item category perspectives,learning user preferences on the item and category levels.Since category information is coarse-grained,there is many meaningless repetitive behaviors in category level sequence.To address this,we design a category sequence entropy feature that determines the importance of each behavior for users’ preference learning based on the category transition’s apparency indicated by the feature.Finally,we design an auxiliary loss function that supervises the learning of preferences on both levels based on the next item and its category.The NDCG and Recall are used as evaluation metrics to conduct comparison experiments with baseline models on three real-world e-commerce datasets,and the results show that the algorithm has better performance.2)To address the second challenge,we propose a sequential recommendation algorithm based on preference decoupling.In the context of telecom package recommendations,this algorithm decouples user preferences for flux and voice to achieve personalized recommendations for users.Specifically,we first construct two preference representations to represent users’ preferences for flux and voice in plan separately,and extract data reflecting only one aspect of user preference from the dataset based on the collider effect in causal inference to train their corresponding preference representations.Then,we design an attention mechanism that evaluates users’ current preferences for data and voice based on their short-term consumption tendencies and accordingly fuses the two preference representations.Finally,the model recommends packages to users based on their shortterm consumption tendencies and the fused preference representation.We test our model on a realworld service provider plan subscription dataset,and the results show that the recommended results generated by the proposed model have good interpretability.