sgyan/kdd_cup_2012_track2

Predict the CTR through all data(about 150 million), the auc of test set(about 20 million) is 0.7874.

The Data Mining Project

PREDICT THE CLICK-THROUGH RATE(CTR) OF ADS GIVEN THE QUERY AND USER INFORMATION

Data: KDD CUP 2012 Track2

Predict the CTR through all data(about 150 million), the auc of test set(about 20 million) is 0.7874.

Running Steps

Pre-processing

(1) python feature_statistic.py

(2) python features_min_max.py // for dense features

(3) python construct_mapping_fn.py // for sparse features

(4) python count_features_fn.py // for sparse features

(5) python tokens_vector.py

(6) python sum_idf.py

(7) python construct_tokens_vectors.py

(8) python shuffle_big_file.py or python shuffle_file_enough_memory.py

(9) python combine_data.py train/test

(10) python divide_data_to_train_valid.py

Training

(11) python deep_fm_combined.py or python deep_fm_enough_memory.py

changing the parameters in file to deal with training and test data, and training more models

(12) python auc.py

more details: see the documents in directory "doc"

Model

Modified the deepfm which is implemented by Weichen Shen,[email protected] in DeepCTR

more details: see the contents in directory "deepctr"

Try

(1) The words in tokens with tf less than other 75% words are thrown away: did not improve the auc.

(2) The cold start of item in test data is set to the mean of all records in training data: a small improvement in auc.

(3) The value of dense feature with large number is being log() before being normalized to one: did not improve the auc.

(4) The way that may be improve auc is: make the feature selection a little more detailed.

Dependency

pip install joblib