Kohei Ozaki / 小嵜 耕平

Kohei Ozaki
Kohei Ozaki (a.k.a. @smly) is a Software Engineer at Ubie. My specialties are around data and machine learning. I love to discover knowledge and hidden values in data.
保険/金融/広告ほか様々な事業でデータ分析や研究開発などの業務を 10 年以上経験してきました。 現在はUbie株式会社でソフトウェアエンジニアをしています。またTURING株式会社でフェローをしています。

Data Science Competitions

Kaggle Office
kaggle book.
I am a Grandmaster of Kaggle competitions (an honor for top competitors with outstanding data science skills) with 19 gold medals in Kaggle competitions. My highest Kaggle rank is 4th out of 530,000+ data scientists as of April 2016.
I also wrote a book on deep learning for kagglers.

Programming Contests

ICPC 2008 Aizu.
ACM/ICPC is an annual multi-tiered competitive programming competition among the universities of the world. I joined this contest with my friends at Tokyo University of Science and advanced to the Asia Regional Contest.

Hobby Projects

AI Mahjong Parlour.

Mahjong AI Salon: Mahjong AI is a very interesting research topic, but it's hard to evaluate the strength of the Mahjong AI agent you have developed. Mahjong AI Salon is a platform where you can play your Mahjong AI agent with other agents.

This application uses TypeScript and React to implement the front-end and is deployed on Firebase. The API server is deployed on Cloud Run. A VM launched by Cloud Function will evaluate the submitted Docker containers.

Interview / Media


Google Scholar
  • Shuhei Yokoo, Kohei Ozaki, Edgar Simo-Serra, Satoshi Iizuka, "Two-stage discriminative re-ranking for large-scale landmark retrieval", 2020. [arxiv]
  • Yusuke Niitani, Toru Ogawa, Shuji Suzuki, Takuya Akiba, Tommi Kerola, Kohei Ozaki, Shotaro Sano, "Team PFDet's Methods for Open Images Challenge 2019", 2019. [arxiv]
  • Kohei Ozaki, Shuhei Yokoo, "Large-scale landmark retrieval and recognition under a noisy and diverse dataset ", 2019. [arxiv][github]
  • Kohei Ozaki, Masashi Shimbo, Mamoru Komachi, Yuji Matsumoto, "Mutual k-Nearest Neighbor Graph Construction in Graph-based Semi-Supervised Classification ", Tra nsactions of the Japanese Society for Artificial Intelligence, 2013.
  • Kohei Ozaki, Masashi Shimbo, Mamoru Komachi, Yuji Matsumoto, "Using the mutual k-nearest neighbor graphs for semi-supervised classification on natural language data", Proceedings of the fifteenth conference on computational natural language learning, 2011.
  • Asad Habib, Masayuki Asahara, Yuji Matsumoto, Kohei Ozaki, "JaPak IEOU: Japan-Pakistan's Input English Output Urdu a case sensitive standard input system for Perso-Arabic script clients", IEEE 2010 International Conference on Information and Emerging Technologies, 2010.


  • 2019-09-19, "情報科学特論2019", 首都大学東京 (集中講義・演習、首都大学東京)
  • 2018-07-24, "ビッグデータ解析" (ゲスト担当分、大阪大学大学院情報科学研究科)
  • 2015-10-15, "競技データマイニングにおけるテクニックと動向", 鹿島研究室公開セミナー, 京都大学

Talks / Presentations

  • 2022-12-08 "機械学習コンテストの登り方", Vision Engineering Workshop (ViEW) 2022 ビジョン技術の実利用ワークショップ
  • 2019-06-16 "Large-scale Landmark Retrieval/Recognition under a Noisy and Diverse Dataset", CVPR'19 WS, Las Vegas
  • 2018-05-19 "データ分析コンテストの解答から学ぶ", PyCon mini Osaka, 大阪
  • 2018-03-03 "データ分析コンテストの 勝者解答から学ぶ", ステアラボ人工知能シンポジウム 2018, 千葉工業大学
  • 2018-06-18 "Local vs Global Descriptor? Relevance Scoring with using Both", CVPR'18 WS, Salt Lake City
  • 2017-10-20 "データ分析コンテストの技術と最近の進展", 第14回ステアラボ人工知能セミナー, 千葉工業大学
  • 2017-07-29 "画像処理コンペティションにおける技術とワザ", Kaggle Student Meetup by Sansan
  • 2016-09-09 "Python とデータ分析コンテストの実践", FIT2016 第15回情報科学技術フォーラム, 富山大学
  • 2016-07-04 "実験の再現性と効率化の話(Docker と Serialization 周辺)", ビッグデータ基盤技術勉強会, 筑波大学
  • 2016-03-05 "Workflow, Serialization & Docker for Kaggle", Kaggle Tokyo Meetup #1, 東京
  • 2016-02-01 "国際チームにおけるモデリング手法とチームガイドライン", 情報処理学会誌 (2016年2月号) 小特集
  • 2015-10-27 "Techniques (Tricks) for Data Mining Competition", BrainPad 社内講演
  • 2015-10-23 "データマイニングコンテストにおけるテクニック", PyData.Tokyo Meetup #6
  • 2015-08-09 "3-Stage Ensemble and Feature Engineering for MOOC Dropout Prediction", KDD'15 workshop. (発表のみ)
  • 2013-11-30 "InfluxDB & LevelDB Inside-out", モニタリングカジュアル#5, 京都
  • 2012-06-09 "Distributed Graph System & Related Topics", TokyoWebMining18