# Modeling User Behaviour in Research Paper Recommendation System

**Authors:** Arpita Chaudhuri, Debasis Samanta, Monalisa Sarma  
**arXiv:** [2107.07831](https://arxiv.org/abs/2107.07831)  
**Date:** 2021-07-16  
**Summary page:** https://www.summarizepaper.com/en/paper/2107.07831v1/

## Key points

- Shift towards modeling user intention in research paper recommendation systems
- Incremental dataset collection for user preferences like publication date and publishing authority
- User profiles based on topics provided by users, updated with new topics
- Deep learning-based system using implicit and explicit feedback for user profiles
- Representation of user profiles using Dynamic Normalized Tree of Concepts (DNTC)
- Content-based recommendation approach using weighted keyphrases from tagged papers
- Hybrid article recommendation system in social networks utilizing metadata like title, abstract, published journal, and year
- Importance of capturing user intention for accurate and personalized recommendations tailored to individual users' needs and interests

## Summary

, , , , 

In the realm of research paper recommendation systems, there has been a shift towards modeling user intention rather than solely focusing on user preferences. This shift is driven by the recognition that <kw>user intention</kw> is dynamic and plays a crucial role in enhancing the accuracy of recommendations. Various approaches have been proposed to capture <kw>user intention</kw> effectively. Dhanda and Verma introduced a recommender system that leverages incremental dataset collection to gather user preferences such as publication date and publishing authority from liked papers. These preferences are then utilized to recommend papers based on similarity matching with candidate papers. Hong et al. defined <kw>user profiles</kw> based on topics provided by users, which are updated with new topics. They used cosine similarity to identify related papers for recommendation. Hassan proposed a deep learning-based research paper recommendation system that creates <kw>user profiles</kw> using both implicit and explicit feedback. User preferences are extracted from preferred articles, while short-term interests are inferred from transaction logs. Recommendations are made using cosine similarity between candidate papers and those stored in the user profile. Alshaikh introduced a unique approach to representing <kw>user profiles</kw> using Dynamic Normalized Tree of Concepts (DNTC). The DNTC captures semantic relationships between concepts extracted from reading papers, reflecting users' short-term interests. Recommendations are generated based on similarities with these short-term interests. Ferrara et al. developed a content-based recommendation approach where weighted keyphrases from tagged papers were used to construct <kw>user profiles</kw>. Similarity matching was then employed to recommend papers aligned with these profiles. Sun et al. suggested a hybrid article recommendation system in social networks, utilizing metadata such as title, abstract, published journal, and year to define <kw>user profiles</kw>. Connectivity graphs were used to establish relations between users and articles, with recommendations generated through Random Walk Restart (RWR) method. Gautam et al. collected personal information explicitly from users' academic data to build their profiles using tags or keywords associated with their academic work. Overall, these diverse approaches highlight the importance of capturing <kw>user intention</kw> in research paper recommendation systems for more accurate and personalized recommendations tailored to individual users' needs and interests.