TalkingData AdTracking Fraud Detection Challenge

models and scores

model definition can be found in scripts/model_lib.py

• model1 LGBM with 83 (76 numerical, 7 categorical) features.

• model2 keras with 27(18 numerical, 9 categorical) features, You can see network structure in model.png

feature engineering and scripts

Most of these features have already been discussed on the kaggle forum.

• counting features

  • mk_feat_count.py
  • mk_feat_count_time.py
  • mk_feat_countRatio.py

• cumulative count

  • mk_feat_cumcount.py
  • mk_feat_recumcount.py
  • mk_feat_cumratio.py

• time to next click

  • mk_feat_nextClick_leak_day.py
  • mk_feat_nextClick_filter.py

• time bucket count.(make multiple time intervals, and count the number of buckets which the IP exists)

  • mk_feat_rangecount.py
  • mk_feat_rangecount_minute.py

• variance

  • mk_feat_var.py

• common IP

  • mk_feat_common_ip.py

• unique count

  • mk_feat_uniq_count2.py

• target encoding: woe

  • mk_feat_woe_all_prev.py
  • mk_feat_woe_bound.py

Features will be calculated once and saved to disk.

Importance from LGBM is found in importance.txt.

Requirements

I used following environment

Hardware:

• Memory: 256GB RAM, 256GB SWAP
• CPU: 20 core, 2.10GHz
• GPU: 1080Ti

Python3 packages:

• numpy==1.14.2
• pandas==0.22.0
• lightgbm==2.1.0
• keras==2.1.5

How to run

At first, put sample_submission.csv test.csv test_supplement.csv train.csv to input directory.

Then run shell scripts as follows,

$ cd scripts/

$ ./run_mk_feats.sh

$ ./run_mk_model1.sh

$ ./run_mk_model2.sh

Output prediction files will be in csv directory.

It took about one day for feature extraction(run_mk_feats.sh).

It needs large memory(~256GB) to build model1(run_mk_model1.sh), sorry.

GPU is required to build model2(run_mk_model2.sh)