# Attentional Factorization Machines: Learning the Weight of Feature Interactions via Attention Networks

**Authors:** Jun Xiao, Hao Ye, Xiangnan He, Hanwang Zhang, Fei Wu, Tat-Seng Chua  
**arXiv:** [1708.04617](https://arxiv.org/abs/1708.04617)  
**Date:** 2017-08-15  
**Summary page:** https://www.summarizepaper.com/en/paper/1708.04617v1/

## Key points

- Attentional Factorization Machines (AFM) is a supervised learning approach that improves upon the limitations of Factorization Machines (FMs)
- FMs assign all feature interactions with the same weight, leading to noise and decreased performance
- AFM introduces a neural attention network to learn the importance of each feature interaction from data
- AFM enhances the predictive power of FM by discriminating between useful and useless interactions
- AFM outperforms FM with an 8.6% relative improvement in experiments on real-world datasets
- AFM surpasses state-of-the-art deep learning methods such as Wide&Deep and DeepCross
- AFM achieves these results with a simpler structure and fewer model parameters
- The implementation of AFM is publicly available at https://github.com/hexiangnan/attentional_factorization_machine

## Summary

Attentional Factorization Machines (AFM) is a supervised learning approach that improves upon the limitations of Factorization Machines (FMs). FMs are effective in modeling feature interactions but assign all interactions with the same weight, which can lead to noise and decreased performance. To address this issue, AFM introduces a neural attention network to learn the importance of each feature interaction from data. This allows AFM to enhance the predictive power of FM by discriminating between useful and useless interactions. Extensive experiments on two real-world datasets demonstrate that AFM outperforms FM with an 8.6% relative improvement and surpasses state-of-the-art deep learning methods such as Wide&Deep and DeepCross. Additionally, AFM achieves these results with a simpler structure and fewer model parameters. The implementation of AFM is publicly available at https://github.com/hexiangnan/attentional_factorization_machine. Overall, AFM presents a promising approach for improving the performance of FM through consideration of varying importance of different feature interactions via a neural attention network.