Learning-to-rank

This Repository built a lambdarank algorithm for anime recommendations with gradient boosting algorithm. The code style has been configured to use Black, and the maximum line length has been set to 120 characters.

Dataset

I used Anime Recommendation LTR in the Kaggle dataset.

Requirements

By default, hydra-core==1.3.0 was added to the requirements given by the competition. For pytorch, refer to the link at https://pytorch.org/get-started/previous-versions/ and reinstall it with the right version of pytorch for your environment.

You can install a library where you can run the file by typing:

$ conda env create --file environment.yaml

Run code

CatBoost parameter Setting

params:
  task_type: CPU
  loss_function: StochasticRank:metric=NDCG
  eval_metric: NDCG
  learning_rate: 0.05
  l2_leaf_reg: 0.02
  bagging_temperature: 3
  min_data_in_leaf: 57
  od_type: Iter
  iterations: 20000
  allow_writing_files: False

Running the learning code shell allows learning

$ sh scripts/run.sh

Benchmark

Results

LightGBM Recommendation

Anime Name	Already Liked	Predicted Score
Ningen Doubutsuen	Majo no Takkyuubin	3.062
Bakugan Battle Brawlers	Tenkuu no Shiro Laputa	2.981
Chain Chronicle: Haecceitas no Hikari	Pumpkin Scissors	2.981
Kure-nai	Omoide Poroporo	2.821
Break Blade Picture Drama	Heisei Tanuki Gassen Ponpoko	2.819
Coral no Tanken	Tonari no Totoro	2.805
None	Zetsuai 1989	2.795
Rokujouma no Shinryakusha!? (TV)	Monster	2.795
Platonic Chain: Web	xxxHOLiC Kei	2.793
Seikimatsu Occult Gakuin Specials	Shounen Onmyouji	2.793

CatBoost Recommendation

Anime Name	Already Liked	Predicted Score
Gad Guard	Majo no Takkyuubin	4.797
Fuyu no Hi	Tenkuu no Shiro Laputa	4.566
Hatsukoi Limited.: Gentei Shoujo	Pumpkin Scissors	4.566
Slam Dunk Movie	Omoide Poroporo	4.541
Koukaku Kidoutai Nyuumon Arise	Heisei Tanuki Gassen Ponpoko	4.522
Joshiochi!: 2-kai kara Onnanoko ga... Futtekita!?	Tonari no Totoro	4.266
Ai Yori Aoshi	Zetsuai 1989	4.266
Shin Mitsubachi Maya no Bouken	Monster	4.141
Denshinbashira no Okaasan	xxxHOLiC Kei	4.141
Schoolgirl Strikers: Animation Channel	Shounen Onmyouji	4.126

What not worked

Stochastic Rank method took a lot of training time.

ds-wook/learning-to-rank