# Learning Recommendations from User Actions in the Item-poor Insurance Domain

**Authors:** Simone Borg Bruun, Maria Maistro, Christina Lioma  
**arXiv:** [2211.15360](https://arxiv.org/abs/2211.15360)  
**Date:** 2022-11-28  
**Summary page:** https://www.summarizepaper.com/en/paper/2211.15360v1/

## Key points

- Bruun et al. address the challenge of generating automatic recommendations in data-sparse domains like insurance purchasing.
- They propose a novel recurrent neural network recommendation model that leverages past user sessions to learn recommendations.
- The model considers various types of user actions within sessions and captures dependencies across sessions to improve recommendation accuracy.
- The approach is unique as it focuses on the target action of purchasing insurance, which typically occurs outside the input sessions.
- Evaluation on a real-world dataset from an insurance vendor shows that the proposed model outperforms existing baselines.
- The dataset includes information on user actions categorized by section (e-commerce, claims reporting, information), object (items and services), and type (click, start, act).
- Key findings from the dataset include most actions occurring on personal account pages and e-commerce sections, demographic features of users, and actions leading up to purchase events.

## Summary

In their study, Bruun et al. address the challenge of generating automatic recommendations in data-sparse domains such as insurance purchasing. Traditional personalized recommendation systems may not be as effective due to limited user feedback on products in this domain. To overcome this limitation, the researchers propose a novel recurrent neural network recommendation model that leverages past user sessions to learn recommendations. The model takes into account various types of user actions within sessions and captures dependencies across sessions to improve recommendation accuracy. The proposed approach is unique in that it considers the target action of purchasing insurance, which typically occurs outside the input sessions. Evaluation on a real-world dataset from an insurance vendor shows that the proposed model outperforms existing baselines. Additionally, Bruun et al. provide insights into the dataset used in their study, including information on user actions categorized by section (e-commerce, claims reporting, information), object (items and services), and type (click, start, act). Some key findings from the dataset include
- Most actions occur on personal account pages and e-commerce sections as existing customers interact with services before making new purchases. - The dataset also includes demographic features of users and captures actions leading up to purchase events. Overall, Bruun et al. 's work contributes a novel session-based recommendation model tailored for the insurance domain and provides valuable insights into addressing data scarcity challenges in recommendation systems. Furthermore, they make their dataset publicly available for further research in this area.