4 Idiots' Approach for Click-through Rate Prediction

Our team consists of:

Name              Kaggle ID         Affiliation
====================================================================
Yu-Chin Juan      guestwalk         National Taiwan University (NTU)
Wei-Sheng Chin    mandora           National Taiwan University (NTU)
Yong Zhuang       yolicat           National Taiwan University (NTU)
Michael Jahrer    Michael Jahrer    Opera Solutions

Our final model is an ensemble of NTU's model and Michael's model. Michael's model is based on his work in Opera Solutions, so he cannot release his part. Therefore, in the codes and documents we only present NTU's model.

This README introduces how to run our code up. For the introduction to our approach, please see

http://www.csie.ntu.edu.tw/~r01922136/slides/kaggle-avazu.pdf

The model we use for this competition is called `field-aware factorization machines.' We have released a package for this model at:

http://www.csie.ntu.edu.tw/~r01922136/libffm

System Requirement

• 64-bit Unix-like operating system

• Python 3

• g++ (with C++11 and OpenMP support)

• pandas (required if you want to run the bag' part. See Step-by-step' below.)

Step-by-step

Our solution is an ensemble of 20 models. It is organized into the following three parts:

name       public score     private score     description        
===========================================================================
base             0.3832            0.3813     2 basic models

bag              0.3826            0.3807     2 models using bag features.

ensemble         0.3817            0.3797     an ensemble of the above 4 
                                              models and 16 new small models

Because the bag' part consumes a huge amount of memory (more than 64GB), and the ensemble' part takes a long time to run, this instruction guides you to run our `base' part first. If you want reproduce our best result, please run the commands in the final step on a suitable machine.

1. First, please use the following command to run a tiny example up

   $ ./run.sh x

2. Create a symbolic link to the training dataset

   $ ln -sf <training_set_path> tr.r0.csv

3. Add a dummy label to the test set

   $ ./add_dummy_label.py <test_set_path> va.r0.csv

4. Checksum

   $ md5sum tr.r0.csv va.r0.csv f5d49ff28f41dc993b9ecb2372abb033 tr.r0.csv 6edd380a5897bc16b61c5a626062f7b3 va.r0.csv

5. Reproduce our base submission

   $ ./run.sh 0

   Note: base.r0.prd is the submission file

6. (optional) Reproduce our best submission

   $ ./run_all.sh x

   If success, then run

   $ ./run_all.sh 0

   Note: The algorithm in the `bag' part is non-deterministic. That is, the result can be slightly different when you run it two or more times.

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

If you want to trace these codes, please be prepared that it will take some efforts. We do not have enough time to polish the codes here to improve the readability. Sorry about it.