Description
The Attentional Factorization Machine (AFM) is a sophisticated model designed to enhance the performance of factorization machines by learning the importance of feature interactions through attention networks. This repository offers a direct TensorFlow implementation of the AFM model, building upon the foundational concepts of Factorization Machines (FM).
Read More
The project is based on the research paper "Attentional Factorization Machines: Learning the Weight of Feature Interactions via Attention Networks" by Jun Xiao, Hao Ye, Xiangnan He, Hanwang Zhang, Fei Wu, and Tat-Seng Chua, presented at IJCAI 2017. The implementation supports both regression and classification tasks, optimizing for RMSE.
Users can leverage this code to replicate the research findings or apply AFM to their own datasets. The repository provides clear instructions for setting up the environment, which requires TensorFlow (version 1.0.1), NumPy, and Scikit-learn. Datasets should follow the input format of the LibFM toolkit, with an example using the MovieLens dataset for personalized tag recommendation.
The included quick example demonstrates how to train an FM model first for pretraining, and then use these weights to train the AFM model. This approach allows for a direct comparison of performance between standard FM and the enhanced AFM. The code also includes functionalities for evaluating the trained models on test datasets.
This implementation is particularly valuable for researchers and practitioners in the fields of recommender systems, machine learning, and deep learning who are looking to improve the accuracy and interpretability of their models by effectively capturing the nuances of feature interactions. The attention mechanism allows the model to dynamically weigh the importance of different feature pairs, leading to more precise predictions.
- CategoryAI Models & LLMs
- Pricing ModelOpen source
- DeploymentCloud, Self-hosted
- PlatformsWeb
Attentional Factorization Machine's Core Features
TensorFlow implementation of Attentional Factorization Machine (AFM)
Supports regression and classification tasks
Optimizes for RMSE
Learns the weight of feature interactions via attention networks
Includes implementation of Factorization Machines (FM) for pretraining
Uses one-hot encoding for feature vector conversion
Compatible with LibFM dataset format
Provides example usage with MovieLens dataset
Includes scripts for training and evaluation
Codebase includes AFM.py, FM.py, and LoadData.py
Supports model pretraining
Getting Started with Attentional Factorization Machine
Set up environment: Install TensorFlow (v1.0.1), NumPy, and Scikit-learn.
Prepare dataset: Format data according to LibFM toolkit specifications.
Train FM model: Execute `python FM.py` with specified dataset and parameters for pretraining.
Train AFM model: Execute `python AFM.py` using the pretrained FM weights.
Evaluate model: Run `python FM.py` or `python AFM.py` with the `--process evaluate` flag.
Attentional Factorization Machine's Use Cases
- Recommender Systems
- Tag Recommendation
- Click-Through Rate Prediction
- Feature Interaction Learning
- Machine Learning Model Development
