king0980692/otto_rec

Otto Recommendation

Quick Start

Original Data format is jsonl format, but jsonl take tool much time to process with, so just use the processed data by kaggle's user.

0. Prepare

Prepare Kaggle cli tool and its credentials

1. check if ~/.kaggle/kaggle.json exist
2. pip install kaggle

Create some necessary folder

mkdir data exp

1. Train/Valid data

Get the data

kaggle datasets download -d cdeotte/otto-validation -p data/
unzip data/otto-validation.zip -d data/split_chunked_parquet/

This data contains:

• train
  • processed training data in chunked parquet format in train_parquet folder
• valid
  • processed validation data in chunked parquet format in test_parquet folder

2. Train/Test data

Get the data

kaggle datasets download -d columbia2131/otto-chunk-data-inparquet-format -p data/
unzip data/otto-chunk-data-inparquet-format.zip -d data/chunked_parquet/

This data contains:

• train
  • processed training data in chunked parquet format in train_parquet folder
• test
  • processed test data in chunked parquet format in test_parquet folder

3. Baseline score

Co-visitation

Local Score:

Overall Recall	0.56134
clicks_recall	0.52557
carts_recall	0.40675
orders_recall	0.64460

run

bash scripts/run_covisitation.sh

Word2vec

Local Score:

=============

Overall Recall	0.51324
clicks_recall	0.42695
carts_recall	0.36441
orders_recall	0.60204
=============

run

bash scripts/run_w2v.sh