A Cross-Domain Recommender System for Literary Books Using Multi-Head Self-Attention Interaction and Knowledge Transfer Learning

Introduction

With the growing number of book resources, readers face increasing difficulty in finding books that align with their interests. To address this challenge, book recommendation technology has emerged (Moore et al., 2022). This technology utilizes user information, book information, and historical user behavior to predict books of potential interest, allowing users to quickly and accurately obtain relevant books and saving them significant time (Naghiaei et al., 2022; Zdravko, 2020). In addition, high-quality book recommendation methods can help book sales platforms improve their precision recommendation capabilities, thereby enhancing their core competitiveness (Lin, 2022; Shen et al., 2020).

Currently, literary books account for the highest proportion within the book industry. With the increasing sophistication of intelligence in the book industry, there is a growing emphasis on personalized recommendation of literary books. This, in turn, has become a significant area of research. However, the field of literary book recommendation faces several challenges (Behera & Nain, 2022; Choi et al., 2023; Herce et al., 2022):

• 1.

  The significant growth of data in the literary book domain has resulted in an enormous amount of information. This has led to the problem of data sparsity, which means the total number of books read by readers is only a small fraction of the available books. This makes it difficult to calculate the nearest neighbor of readers or books, leading to a sharp decline in the recommendation quality of the recommendation system.

• 2.

  The cold start problem refers to the challenge of recommending books to new readers or new books. This is due to the system having insufficient data on these users or items, hindering accurate recommendations.

• 3.

  Traditional book retrieval systems based on search engine technology struggle to properly reflect current user preferences. This limitation makes it difficult to achieve “truly personalized” recommendations, resulting in lower accuracy and recall of search results.

Cross-domain recommendation (CDR) aims to utilize information like user preferences and project features from other domains to help improve the accuracy of recommendations in the target domain (Zhou et al., 2023). This approach enables a more comprehensive modeling of users or projects in the target domain, effectively alleviating data sparsity and user cold start issues (Banik et al., 2023; Cao et al., 2022). Literature shows a significant correlation between users and projects across different fields (Liu et al., 2023). For example, users in different fields often share similar hobbies, such as those who love horror movies tend to purchase horror books, while those who love lyrical music tend to purchase romantic books.

To address the issues in the field of literary book recommendation, a CDR model is proposed, integrating multi-head self-attention interaction and knowledge transfer learning. The innovation of the proposed method lies in: