Skip to content

Latest commit

 

History

History
30 lines (21 loc) · 1.33 KB

README.md

File metadata and controls

30 lines (21 loc) · 1.33 KB
Raw

MeLi Data Challenge 2019 | Deep Learning

Author: Mariano Leonel Acosta | Leaderboard #17 - 0.89764

https://ml-challenge.mercadolibre.com/final_results

I developed a predictive system for product classification with 1588 different categories. Using Natural Language Processing (NLP) combined with Deep Learning, I was able to analyze over two million product descriptions from Mercado Libre and predict new cases with a balanced accuracy of 89,76%.

The final model consists of a Neural Network ensemble, a combination of Long Short Term Memory RNNs (LSTM) and Convolutional Nets (CNN). Each sub-system was trained independently on different subset of the dataset. Then, to make the final prediction, each output is combined using weighted sums.

Implementation

In order to try this project on your own, first you need to download the dataset (using Bash):

$wget https://meli-data-challenge.s3.amazonaws.com/train.csv.gz 
$wget https://meli-data-challenge.s3.amazonaws.com/test.csv 
$wget https://meli-data-challenge.s3.amazonaws.com/sample_submission.csv 
$gunzip resources/train.csv.gz

(Alternately, the resources can be downloaded manually from HERE)

Next, simply run the main.py script.

The following libraries are required:

  • Numpy
  • Pandas
  • SciKit Learn
  • Tensorflow
  • Keras