Yandex_ML_track_2018

Result: 29th place, mean NDCG x 100000 = 85352

Problem:

Rank replies by relevance according to the context. Context usually consists of 3 replicas with several replies. Each reply has relevance and confidence - I use it's product as a target variable.

Solution:

Use pretrained fastText model to represent each sequence as concatenated replicas embeddings and match it all with target. Let's create KFold cross-validation with 10 folds, train separate simple lightgbm regressor on every iteration and then calculate mean of all models predictions to make result more "stable".

Pipeline:

• Install requirements.txt pip3 install -r requirements.txt
• Change paths in config.py and then run python3 prep.py
• Download fastText and build it using make.
• Download fastText model trained on wikipedia and common crawl.
• Fill get_vectors.txt with needed paths and copy it to the fastText folder.
• Run bash get_vectors.txt
• Make numpy array from processed dataset python3 prep_fasttext_data.py
• Train model python3 fasttext_lgbm.py