Explanation Generation For Black-box Recommendation Systems

In recent years,there have been numerous model approaches related to explanation generation for recommendation system.Among them,the post-hoc explanation generation for recommendation method has received increasingly widespread attention because it can generate explanations for any complex black-box recommendation system.Post-hoc explanation generation for recommendation is mainly divided into two types: template-based and natural language-based.Template-based methods complete explanation generation by filling in keywords in pre-defined templates.Although the quality of the explanations obtained through this method is relatively stable,this approach lacks scalability and has low sentence diversity,making it difficult to meet the many requirements in real-world situations.Natural language-based methods learn new sentences autonomously from data and generate explanations,which makes them highly scalable and significantly improves the diversity of generated explanations,and therefore has received more and more attention.However,this paper found that existing natural language-based methods,although capable of generating effective explanations in some cases,still have many problems.Some models can generate fluent explanations,but these explanations are often empty,lack personalization,and lack relevance to the input recommendation results.Furthermore,through preliminary experiments,this paper found that these problems are mainly caused by the following three issues:(1)There is a significant difference between the vector representations of input users and items and the representation of internal text in the language model,which makes it difficult for the language model toutilize the representations of input users and items,and it is necessary to find ways to narrow the gap between these two representations.(2)When generating explanations,the model often ignores the input recommendation results and overly relies on the prior knowledge learned by the language model internally,which leads to poor controllability of the generated explanations by the input recommendation results and low relevance between the generated explanations and the given recommendation results.(3)Due to the lack of feature descriptions for users and items in the data,the model finds it difficult to generate explanations containing explicit user preferences and item attributes.Although the explanation factors in the latent variables corresponding to users and items can provide feature descriptions of users and items,these explanation factors are highly coupled and cannot be effectively utilized by the decoder.To solve the above problems,this paper has made the following main efforts:1.Explanation Generation for Recommendation based on Multi-Granularity Variational Autoencoder.To address the problem of large differences between input and output representations,this paper proposes a multi-granularity conditional VAE.The text generation model based on multi-granularity VAE is conducive to the mutual conversion of representations in different representation spaces,thus reducing the impact of representation differences between input and output.2.User-Item Conditional Enhancement based on Contrastive Learning.For the conditional signal in the VAE,this paper uses contrastive learning to enhance its representation,so that it can accurately express the different preferences of different users and the different attributes of different items,thus guiding the decoder to generate explanations accurately and promoting the model to generate personalized explanations based on different users and items.3.Identification of Explanation Factors based on Decoupling of Latent Factors.Although the latent variables obtained by encoding the input users and itemsusing VAE contain explanation factors,these factors are highly coupled,making it difficult for the decoder to effectively utilize them.To address this problem,this paper proposes to identify potential item attributes and user preferences using decoupling of latent variables.Each chapter of this paper includes sufficient experiments and related analyses to effectively validate the effectiveness and rationality of the proposed method,which further promotes the development of related research.