# DCNv3: Towards Next Generation Deep Cross Network for CTR Prediction

**Authors:** Honghao Li, Yiwen Zhang, Yi Zhang, Hanwei Li, Lei Sang  
**arXiv:** [2407.13349](https://arxiv.org/abs/2407.13349)  
**Date:** 2024-07-18  
**Summary page:** https://www.summarizepaper.com/en/paper/2407.13349v1/

## Key points

- Deep & Cross Network (DCN) and its derivative models strike a balance between efficiency and performance in CTR prediction
- Implicit feature interactions through DNN lack interpretability
- Explicit feature interaction methods underperform compared to DNN approaches
- Difficulty filtering noise while enhancing feature interactions
- Fusion methods lack appropriate supervision signals for different techniques
- Prioritize interpretability in feature interaction modeling
- Elevate the order of feature interactions for genuine Deep Crossing
- Implement Self-Mask operation to filter noise and reduce parameter complexity by half
- Utilize Tri-BCE method in fusion layer for tailored supervision signals
- Extensive experiments validate effectiveness, efficiency, and interpretability of DCNv3 and SDCNv3 in CTR prediction

## Summary

, , , , 

In the realm of CTR prediction, Deep & Cross Network and its derivative models have emerged as a pivotal paradigm, striking a delicate balance between computational efficiency and predictive performance. However, these models encounter four significant limitations that hinder their full potential. Firstly, many models claim to capture high-order feature interactions but do so implicitly through deep neural networks (DNN), leading to a lack of interpretability in predictions and compromising trustworthiness. Secondly, existing explicit feature interaction methods tend to underperform compared to implicit DNN approaches, casting doubt on their necessity in the modeling process. Thirdly, numerous models struggle with effectively filtering out noise while enhancing the order of feature interactions, posing a challenge in achieving optimal predictive accuracy. Lastly, fusion methods employed by most models often fail to provide appropriate supervision signals for different interaction techniques, further complicating the modeling process. To address these limitations head-on, this paper introduces the next generation Deep Cross Network (DCNv3) and Shallow & Deep Cross Network (SDCNv3). These innovative models prioritize interpretability in feature interaction modeling while significantly elevating the order of feature interactions for genuine Deep Crossing rather than superficial cross-networking. A Self-Mask operation is implemented to efficiently filter noise and reduce parameter complexity within the cross network by half. Additionally, a novel loss weight calculation method named Tri-BCE is utilized in the fusion layer to offer tailored supervision signals for enhanced model performance. Extensive experiments conducted across six datasets underscore the effectiveness, efficiency, and interpretability of DCNv3 and SDCNv3. The comprehensive findings validate the efficacy of these next-generation models in overcoming existing limitations and advancing CTR prediction capabilities. For those interested in replicating or further exploring this research endeavor, detailed code implementations, running logs, and hyperparameter configurations are readily accessible at https://anonymous.4open.science/r/DCNv3-E352. The authors behind this groundbreaking work - Honghao Li, Yiwen Zhang, Yi Zhang, Hanwei Li, and Lei Sang - have made significant strides towards propelling CTR prediction into new realms of accuracy and reliability with their pioneering contributions through DCNv3 and SDCNv3 methodologies.