https://github.com/chenglongchen/tensorflow-xnn
4th Place Solution for Mercari Price Suggestion Competition on Kaggle using DeepFM variant.
https://github.com/chenglongchen/tensorflow-xnn
ctr ctr-prediction deep-ctr deepfm factorization-machines fm kaggle-competition snapshot-ensemble
Last synced: about 2 months ago
JSON representation
4th Place Solution for Mercari Price Suggestion Competition on Kaggle using DeepFM variant.
- Host: GitHub
- URL: https://github.com/chenglongchen/tensorflow-xnn
- Owner: ChenglongChen
- License: mit
- Created: 2018-02-22T12:41:33.000Z (about 7 years ago)
- Default Branch: master
- Last Pushed: 2018-07-24T14:05:05.000Z (almost 7 years ago)
- Last Synced: 2025-03-24T08:34:47.757Z (about 2 months ago)
- Topics: ctr, ctr-prediction, deep-ctr, deepfm, factorization-machines, fm, kaggle-competition, snapshot-ensemble
- Language: Python
- Homepage: https://www.kaggle.com/c/mercari-price-suggestion-challenge
- Size: 1.07 MB
- Stars: 283
- Watchers: 12
- Forks: 76
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
# tensorflow-XNN
4th Place Solution for [Mercari Price Suggestion Challenge on Kaggle](https://www.kaggle.com/c/mercari-price-suggestion-challenge)
## The Challenge
Build a model to suggest the price of product on Mercari. The model is required to train (including all the preprocessing, feature extraction and model training steps) and inference within 1 hour, using only 4 cores cpu, 16GB RAM, 1GB disk. Data include unstructured text (product title & description) and structured ones, e.g., product category and shipping flag etc.
## Summary
Highlights of our method are as follows:
* very minimum preprocessing with focus on end-to-end learning with multi-field inputs, e.g., textual and categorical;
* hybrid NN consists of four major compoments, i.e., embed, encode, attend and predict. FastText and NN based FM are used as building block;
* purely bagging of NNs of the same architecture via snapshot ensemble;
* efficiency is achieved via various approaches, e.g., lazynadam optimization, fasttext encoding and average pooling, snapshot ensemble, etc.
### Model Architecture
Please find the slide of our solution [here](./doc/Mercari_Price_Suggesion_Competition_ChenglongChen_4th_Place.pdf).
## About this project
This is the 4th text mining competition I have attend on Kaggle. The other three are:
* [CrowdFlower Search Results Relevance Competition](https://www.kaggle.com/c/crowdflower-search-relevance), 1st Place
* [Home Depot Product Search Relevance Competition](https://www.kaggle.com/c/home-depot-product-search-relevance), 3rd Place
* [The Hunt for Prohibited Content Competition](http://www.kaggle.com/c/avito-prohibited-content), 4th Place
In these previous competitions, I took the general ML based methods, i.e., data cleaning, feature engineering (see the solutions of [CrowdFlower](https://github.com/ChenglongChen/Kaggle_CrowdFlower) and [HomeDepot](https://github.com/ChenglongChen/Kaggle_HomeDepot) for how many features have been engineered), VW/XGBoost training, and massive ensembling.
Since I have been working on CTR & KBQA based on deeplearning and embedding models for some time, I decided to give this competition a shot. With data of this competition, I have experimented with various ideas such as NN based FM and snapshot ensemble.