日本データベース学会の皆様， （重複して受け取られた場合にはご容赦ください） 京都大学の加藤です． 9月19日（火）10:30-11:30にWebDB Forum 2017（@お茶の水女子大学）にて予定されている Maarten de Rijke教授によるiDB招待講演「Learning from Bandit Feedback」 のご案内を申し上げます． WebDB Forum 2017へ参加登録されている方はもちろん， 登録されていない方も当日参加を受け付けておりますので， 是非ご参加くださいますようよろしくお願い申し上げます． Maarten de Rijke教授は情報検索の分野で幅広く活躍されており （トピックモデル，クエリ推薦，情報推薦，対話的検索，オンライン評価，ランキング学習， クリックモデル，情報要約，パーソナライズ検索，エンティティ検索など）， 今現在最もProductiveな情報検索チームを率いています． https://staff.fnwi.uva.nl/m.derijke/ 講演内容はBandit Feedback（実際に解きたい問題の正解ラベルは明示的に与えられない状況， e.g. 検索結果へのクリックは，検索結果をランキングするという問題に対して正解ラベルを与えない） からの学習です．情報検索のコンテキストでの話にはなると思いますが，機械学習に大きく関わる話になると思います． 特に，実サービスで得られたログデータからシステム改善を行っている方々にとっては興味深い話ではないかと思います． どうぞよろしくお願い申し上げます． ======================================= iDB招待講演 講演タイトル: Learning from Bandit Feedback 講演者: Maarten de Rijke (University of Amsterdam) 日時: 9月19日（火） 10:30-11:30 場所: お茶の水女子大学 ### 講演概要 ### Log data is one of the most ubiquitous forms of data available, as it can be recorded from a variety of online systems (e.g., search engines, recommender systems, online stores) at little cost. Interaction logs typically contain a record of the input to the system (e.g., features describing the user and the query), the action that was taken by the system (e.g., a result page that was generated) and the feedback (e.g., clicks on a URL, or dwell time). This feedback, however, provides only partial information -- ``contextual-bandit feedback'' -- limited to the particular action taken by the system. The feedback for all the other actions (e.g., other URLs or verticals) the system could have taken is typically not known. This makes learning from log data fundamentally different from traditional supervised learning, where ``correct'' predictions together with a loss function provide full-information feedback, in the sense that the system has access to the quality of each action in each context during training. The abundant availability of log data motivates the problem of learning from bandit feedback. In the talk I will discuss recent advances. I will briefly touch on online learning from bandit feedback but will spend most of the talk on batch learning from bandit feedback, highlighting recent advances in the choice of estimators and model classes. I will illustrate these advances with examples from search, ad placement, and classification. This talk is based on joint work with Artem Grotov, Katja Hofmann, Thorsten Joachims, Adith Saminathan, Anne Schuth, Masrour Zoghi. ### 講演略歴 ### Maarten de Rijke is full professor of Information Retrieval in the Informatics Institute at the University of Amsterdam. He holds MSc degrees in Philosophy and Mathematics (both cum laude), and a PhD in Theoretical Computer Science. He worked as a postdoc at CWI, before becoming a Warwick Research Fellow at the University of Warwick, UK. He joined the University of Amsterdam in 1998, and was appointed full professor in 2004. He is a member of the Royal Dutch Academy of Arts and Sciences (KNAW) and a recipient of a Pioneer Personal Innovation grant, the Bloomberg Data Science Research Award, the Criteo Faculty Research Award, the Microsoft PhD Research Fellowship Award, and the Yahoo Faculty and Research Engagement Program Award. De Rijke leads the Information and Language Processing Systems group, one of the world’s leading academic research groups in information retrieval. His research focus is on intelligent information access, with projects on self-learning search engines, on semantic search, and on the interface between information retrieval and artificial intelligence. A Pionier personal innovational research incentives grant laureate (comparable to an advanced ERC grant), De Rijke has helped to generate over 60MEuro in project funding. With an h-index of 63 he has published over 700 papers, published or edited over a dozen books, is editor-in-chief of ACM Transactions on Information Systems, co-editor-in-chief of Foundations and Trends in Information Retrieval and of Springer’s Information Retrieval book series, (associate) editor for various journals and book series, and a current and former coordinator of retrieval evaluation tracks at TREC, CLEF and INEX. Recently, he was co-chair for SIGIR 2013, general chair for ECIR 2014 and WSDM 2017, co-chair “web search systems and applications” for WWW 2015, short paper co-chair for SIGIR 2015, and program co-chair for information retrieval for CIKM 2015. He is also general co-chair of ICTIR 2017. He is the director of Amsterdam Data Science. He’s a former director of the Intelligent Systems Lab (ISLA), of the Center for Creation, Content and Technology (CCCT), and of the University of Amsterdam’s Ad de Jonge Center for Intelligence and Security Studies. ======================================= -- +++++++++++++++++++++++++++++++++++++++ Makoto P. Kato, Program-specific Assistant Professor Graduate School of Informatics, Kyoto University Yoshida Honmachi, Sakyo, Kyoto 606-8501, Japan Email: kato@dl.kuis.kyoto-u.ac.jp Tel : +81-75-753-5702 HP : http://www.mpkato.net/ +++++++++++++++++++++++++++++++++++++++