- 12 th (total score)
NAVER-AI-RUSH-1 / Image Classification
-
our solution : efficientNet + oct vNet Ensemble
-
17 th / 100 teams (first round)
-
the period actually participate : Aug 6 to Aug 13 (about 7 days)
NAVER-AI-RUSH-2 / Click-Through Rate (CTR) Prediction
-
our solution : A single CAT model, (only use 7 different features : read_len, read_cnt, total_cnt, read_prob, gender, age_range, hh)
-
11 th / 30 teams (final round)
-
the period actually participate : Aug 20 to Aug 28 (about 8 days)
-
tried : xDeepFM (feature, article_id, read_len, gender, age_range, hh, image_feature) - didn't go well..
-
tried : Embedding DNN network - didn't see it carefully..
-
tried : xgboost, lgbm for a lot of feature that we could make, (image_feature, catergory_id, cat_in, feature cross, means, std for row numeric values.. etc.. ) - trained well, got better score than our final submit score. But couldn't submit some memory issues, something wrong in features preprocess for test set..
-
Things we couldn't do because of not enough time :
- submit for image feature, category feature, means, std of numeric cols that we made at training
- Ensemble different models(xgboost, lgbm and CAT). even couldn't ensemble of result of k-fold of just single CAT, other model, respectively.
- train item2vec of read articles id (history).. ( i think if we did this and make article_id feature, then we could get very good feature for the model... we just thought but didn't try even though it's not that difficult)