Turing Test's Solution for Home Depot Product Search Relevance Competition on Kaggle
Submission | CV RMSE | Public LB RMSE | Private LB RMSE | Position |
---|---|---|---|---|
Simplified Single Model from Igor and Kostia (10 features) | 0.44792 | 0.45072 | 0.44949 | 31 |
Best Single Model from Igor and Kostia | 0.43787 | 0.44017 | 0.43895 | 11 |
Best Single Model from Chenglong | 0.43832 | 0.43996 | 0.43811 | 9 |
Best Ensemble Model from Igor and Kostia | - | 0.43819 | 0.43704 | 8 |
Best Ensemble Model from Chenglong | 0.43550 | 0.43555 | 0.43368 | 6 |
Best Final Ensemble Model | - | 0.43433 | 0.43271 | 3 |
See ./Doc/Kaggle_HomeDepot_Turing_Test.pdf
for documentation.
Before proceeding, one should place all the data from the competition website into folder ./Data
.
Note that in the following, all the commands and scripts are executed and run in directory ./Code/Chenglong
.
We used Python 3.5.1 and modules comes with Anaconda 2.4.1 (64-bit). In addition, we also used the following libraries and modules:
- gensim 0.12.4
- hyperopt 0.0.3.dev
- keras 0.3.2
- matplotlib-venn 0.11.3
- python-Levenshtein 0.12.0
- regex 2.4.85
- xgboost 0.4
We used the following packages installed via install.packages()
:
- data.table
- Rtsne
We used the following thirdparty packages:
We used pre-trained Word2Vec models listed in this Github repo. In specific:
We used glove-gensim to convert GloVe vectors into Word2Vec format for easy usage with Gensim. After that, put all the models in the corresponding directory (see config.py
for detail).
We also used the following external data:
- Color data from this Kaggle forum post, i.e.,
./Data/dict/color_data.py
in this repo. - Google spelling correction dictionary from this Kaggle forum post, i.e.,
google_spelling_checker_dict.py
in this repo. - Home-made word replacement dictionary, i.e.,
./Data/dict/word_replacer.csv
in this repo. - NLTK corpora and taggers data downloaded using
nltk.download()
, specifically:stopwords.zip
,wordnet.zip
andmaxent_treebank_pos_tagger.zip
.
To generate data and features, one should run python run_data.py
. While we have tried our best to make things as parallelism and efficient as possible, this part might still take 1 ~ 2 days to finish, depending on the computational power. So be patient :)
Note that various text processing are useful for introducing diversity into ensemble. As a matter of fact, one feature set (i.e., basic20160313
) from our final solution is generated before the Fixing Typos post, i.e., not using the Google spelling correction dictionary. Such version of features can be generated by turning off the GOOGLE_CORRECTING_QUERY
flag in config.py
.
After team merging with Igor&Kostia, we have rebuilt everything from scratch, and most of our models used different subsets of Igor&Kostia's features. For this reason, you should also need to generate their features. Since Igor&Kostia's features are in .csv
dataframe format, we provide a converter turing_test_converter.py
to convert them to the format we use, i.e., .pkl
.
In step 3, we have generated a few thousands of features. However, only part of them will be used to build our model. For example, we don't need those features that have very little predictive power (e.g., have very small correlation with the target relevance.) Thus we need to do some feature selection.
In our solution, feature selection is enabled via the following two successive steps.
This approach is implemented as get_feature_conf_*.py
. The general idea is to include or exclude specific features via regex
operations of the feature names. For example,
- one can specify the features that he want to include via the
MANDATORY_FEATS
variable, despite of its correlation with the target - one can also specify the features that he want to exclude via the
COMMENT_OUT_FEATS
variable, despite of its correlation with the target (MANDATORY_FEATS
has higher priority thanCOMMENT_OUT_FEATS
.)
The output of this is a feature conf file. For example, after running the following command:
python get_feature_conf_nonlinear.py -d 10 -o feature_conf_nonlinear_201605010058.py
we will get a new feature conf ./conf/feature_conf_nonlinear_201605010058.py
which contains a feature dictionary specifying the features to be included in the following step.
One can play around with MANDATORY_FEATS
and COMMENT_OUT_FEATS
to generate different feature subset. We have included in ./conf
a few other feature confs from our final submission. Among them, feature_conf_nonlinear_201604210409.py
is used to build the best single model.
With the above generated feature conf, one can combine all the features into a feature matrix via the following command:
python feature_combiner.py -l 1 -c feature_conf_nonlinear_201604210409 -n basic_nonlinear_201604210409 -t 0.05
The -t 0.05
above is used to enable the correlation base feature selection. In this case, it means: drop any feature that has a correlation coef lower than 0.05
with the target relevance.
TODO(Chenglong): Explore other feature selection strategies, e.g., greedy forward feature selection (FFS) and greedy backward feature selection (BFS).
In our solution, a task
is an object composite of a specific feature
(e.g., basic_nonlinear_201604210409
) and a specific learner
(XGBoostRegressor
from xgboost). The definitions for task
, feature
and learner
are in task.py
.
Take the following command for example.
python task.py -m single -f basic_nonlinear_201604210409 -l reg_xgb_tree -e 100
- It runs a
task
withfeature
basic_nonlinear_201604210409
andlearner
reg_xgb_tree
. - The
task
is optimized with hyperopt for100
evals for searching the best parameters forlearner
reg_xgb_tree
. - The
task
performs both CV and final refit. CV in this case has two purposes: 1) guide hyperopt to find the best parameters, and 2) generate predictions for each CV fold for further (2nd and 3rd level) stacking. - For all the available learners and the corresponding parameter searching space, please see
model_param_space.py
.
During the competition, we have run various tasks (i.e., various features and various learners) to generate a diverse 1st level model library. Please see ./Log/level1_models
for all the tasks we have included in our final submission.
After generating the feature
basic_nonlinear_201604210409
(see step 4 how to generate this), run the following command to generate the best single model:
python task.py -m single -f basic_nonlinear_201604210409 -l reg_xgb_tree_best_single_model -e 1
This should generate a submission with local CV RMSE around 0.438 ~ 0.439.
After building some diverse 1st level models, run the following command to generate the best ensemble model:
python run_stacking_ridge.py -l 2 -d 0 -t 10 -c 1 -L reg_ensemble -o
This should generate a submission with local CV RMSE around 0.436.
Before proceeding, one should specify correct paths in file config_IgorKostia.py
and place all the data from the competition website into folder specified by variable DATA_DIR
. To reproduce our Ensemble_B
from Step IK5 one should place the used feature sets into folder specified by variable FEATURESETS_DIR
. Note that in the following, all the commands and scripts are executed and run in directory ./Code/Igor&Kostia
.
We used Python 2.7.11 on Windows platform and modules comes with Anaconda 2.4.0 (64-bit), including:
- scikit-learn 0.17.1
- numpy 1.10.1
- pandas 0.17.0
- re 2.2.1
- matplotlib 1.4.3
- scipy 0.16.0
In addition, we also used the following libraries and modules:
- NLTK 3.1 (use
nltk.download()
command) - gensim 0.12.2
- xgboost 0.4
Some descriptive analysis and final model blending was also done in Excel 2007 and Excel 2010.
We do all text preprocessing before any feature generation and save the results to files. It helped us save a few computing days since the same preprocessing steps are necessary to generate different features.
- Run
text_processing.py
. - Run
text_processing_wo_google.py
.
The necessary replacement data is loaded automatically from files homedepot_functions.py
and google_dict.py
.
We need to run consequently the following files:
feature_extraction1.py
.grams_and_terms_features.py
.dld_features.py
.word2vec.py
.
To generate features without using the Google dictionary, we also need to run:
feature_extraction1_wo_google.py
.word2vec_without_google_dict.py
.
As a result, we will have a few csv files with the necessary features for model building.
- Run
generate_feature_importances.py
.
One part of the ensemble Ensemble_A
is generated from the following code:
generate_models.py
.generate_model_wo_google.py
.generate_ensemble_output_from_models.py
.
To get the other part Ensemble_B
, we need to run these files:
ensemble_script_imitation_version.py
(It just reproduces the selection of random features generated fromensemble_script_random_version.py
. You do not need to runensemble_script_random_version.py
again).model_selecting.py
.
These two parts can be generated in parallel. Our final submission from Igor&Kostia was then prodused in Excel as:
Output
=0.75 Ensemble_A
+ 0.25 Ensemble_B
So, we had two ensembles prepared using different methodologies. We observed that our ensembles behave differently in different parts of the datasets (part1
: id<=163700
, part2
: 163700 < id <= 221473
,
part_3
: id > 221473
. Since we observed regular patterns in the data as well, we thought that one of the ensembles might be especially prone to overfitting in some parts. So, while blending our ensembles for final submissions, we made different bets assuming that in some parts one of the models would behave much worse in private than in public.
Our two final submission were produced in Excel with the weights from the table below (the weight for Chenglong's and Igor&Kostia's parts add up to 1). Both these submissions scored the same 0.43271
on the private leaderboard.
Weight Chenglong for part1 and part2 |
Weight Chenglong for part3 |
Public LB RMSE | Private LB RMSE | |
---|---|---|---|---|
Submission 1 | 0.75 | 0.8 | 0.43443 | 0.43271 |
Submission 2 | 0.6 | 0.3 | 0.43433 | 0.43271 |