# Joint Optimization of Tree-based Index and Deep Model for Recommender Systems

**Authors:** Han Zhu, Daqing Chang, Ziru Xu, Pengye Zhang, Xiang Li, Jie He, Han Li, Jian Xu, Kun Gai  
**arXiv:** [1902.07565](https://arxiv.org/abs/1902.07565)  
**Date:** 2019-02-19  
**Summary page:** https://www.summarizepaper.com/en/paper/1902.07565v2/

## Key points

- Large-scale industrial recommender systems face computational challenges due to the large dataset size
- Efficient index structure is crucial for timely recommendations
- Tree-based Deep Model (TDM) improves recommendation accuracy by utilizing a tree index
- TDM achieves logarithmic computational complexity relative to dataset size
- Quality of tree index and user-node preference prediction model are important for recommendation accuracy
- Interdependence between learning the tree index and preference model
- Proposed method jointly learns index structure and user preference prediction model
- Hierarchical user preference representation leveraging the tree index hierarchy is introduced
- Experimental evaluations on real-world datasets show improved recommendation accuracy
- Online A/B test results confirm effectiveness in production environments
- Research focuses on addressing computational challenges in large-scale recommender systems through optimization of both index structure and preference model simultaneously

## Summary

Large-scale industrial recommender systems often face computational challenges due to the large size of the dataset. To address this issue and provide timely recommendations to users, an efficient index structure is crucial. The Tree-based Deep Model (TDM) introduced in previous work has significantly improved recommendation accuracy by utilizing a tree index. By organizing items in a hierarchical tree structure and training a user-node preference prediction model that follows a max-heap property, TDM achieves logarithmic computational complexity relative to the dataset size. This allows for the use of advanced models in candidate retrieval and recommendation. In tree-based recommendation methods, both the quality of the tree index and the user-node preference prediction model play vital roles in determining recommendation accuracy. However, there is an interdependence between learning the tree index and the preference model. Therefore, this paper aims to develop a method that jointly learns the index structure and user preference prediction model. The proposed joint optimization framework involves learning both components under a unified performance measure. Additionally, a novel hierarchical user preference representation is introduced, leveraging the hierarchy of the tree index. Experimental evaluations conducted on two large-scale real-world datasets demonstrate that the proposed method significantly improves recommendation accuracy. Furthermore, online A/B test results conducted on a display advertising platform confirm its effectiveness in production environments. Overall, this research focuses on addressing computational challenges in large-scale recommender systems by optimizing both the tree-based index structure and user preference prediction model simultaneously. The proposed method not only enhances recommendation accuracy but also demonstrates its practicality through successful implementation in real-world scenarios.