4th Place Solution for Mercari Price Suggestion Challenge on Kaggle

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.


Highlights of our method are as follows:

Model Architecture


Please find the slide of our solution here.

About this project

This is the 4th text mining competition I have attend on Kaggle. The other three are:

In these previous competitions, I took the general ML based methods, i.e., data cleaning, feature engineering (see the solutions of CrowdFlower and 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.