outbrain-click-prediction
Data
Kaggle Outbrain click prediction challenge: https://www.kaggle.com/c/outbrain-click-prediction/data
The problem is about predicting which display ads are going to be clicked. There are two major challenges. First, how to handle the massive data sets? page_views.csv alone is 80G with 2bil rows. Second, there are 9 data sets without clear explanation on relationship. Figuring out how to connect dots between those data sets is another major challenge.
Field Aware Factorization Machine(FFM)
A state-of-the-art technology in recommendation was used. To be more specific, libffm was customized to include linear weights in order to account for features that interaction with other features is not so meaningful.
FFM is a recommendation system that incorporates contextual data along with user and item data. In basic matrix factorization, only a single set of latent factors exist per feature, however there exists multiple sets of latent factors in FFM. One set per field. (Features are grouped by fields they belong to.)
Structure
Generate interaction feature
- pick_topk_ad_characteristics.py
- main.cpp
- normalize_interaction.py
Transform data into ffm format
- gen_ffm_format.py
Train & Predict
- libffm (customized to incorporate linear weights)
Make submission file
- make_submission.py
Quick-start
$ g++ -std=c++14 main.cpp util/data.h util/data.cpp util/helpers.h util/helpers.cpp -o main -lboost_iostreams -lpthread
Move to libffm directory:
$ make
Come back to parent directory:
$ python run.py