Глава
Encode Me If You Can: Learning Universal User Representations via Event Sequence Autoencoding
В книге
Рассматривается способ улучшения производительности рекомендательных систем при помощи предварительного выделения групп пользователей с похожим поведением. Для разбиения пользователей на группы используются распределенная версия алгоритма k-средних и алгоритм canopy для определения начальных центроидов.
The first and the second edition of the FCA4AI Workshop showed that many researchers working in Artificial Intelligence are indeed interested by a well-founded method for classi- fication and mining such as Formal Concept Analysis (see http://www.fca4ai.hse.ru/). The first edition of FCA4AI was co-located with ECAI 2012 in Montpellier and published as http://ceur-ws.org/Vol-939/ while the second edition was co-located with IJCAI 2013 in Beijing and published as http://ceur-ws.org/Vol-1058/. Based on that, we decided to continue the series and we took the chance to organize a new edition of the workshop in Prague at the ECAI 2014 Conference. This year, the workshop has again attracted many different researchers working on actual and important topics, e.g. recommendation, linked data, classification, biclustering, parallelization, and various applications. This shows the diversity and the richness of the relations between FCA and AI. Moreover, this is a good sign for the future and especially for young researchers that are at the moment working in this area or who will do.
Recommender systems are becoming an inseparable part of many modern Internet web sites and web shops. The quality of recommendations made may significantly influence the browsing experience of the user and revenues made by web site owners. Developers can choose between a variety of recommender algorithms; unfortunately no general scheme exists for evaluation of their recall and precision. In this chapter, the authors propose a method based on cross-validation for diagnosing the strengths and weaknesses of recommender algorithms. The method not only splits initial data into a training and test subsets, but also splits the attribute set into a hidden and visible part. Experiments were performed on a user-based and item-based recommender algorithm. These algorithms were applied to the MovieLens dataset, and the authors found classical user-based methods perform better in terms of recall and precision.
Concept discovery is a Knowledge Discovery in Databases (KDD) research field that uses human-centered techniques such as Formal Concept Analysis (FCA), Biclustering, Triclustering, Conceptual Graphs etc. for gaining insight into the underlying conceptual structure of the data. Traditional machine learning techniques are mainly focusing on structured data whereas most data available resides in unstructured, often textual, form. Compared to traditional data mining techniques, human-centered instruments actively engage the domain expert in the discovery process. This volume contains the contributions to CDUD 2011, the International Workshop on Concept Discovery in Unstructured Data (CDUD) held in Moscow. The main goal of this workshop was to provide a forum for researchers and developers of data mining instruments working on issues with analyzing unstructured data. We are proud that we could welcome 13 valuable contributions to this volume. The majority of the accepted papers described innovative research on data discovery in unstructured texts. Authors worked on issues such as transforming unstructured into structured information by amongst others extracting keywords and opinion words from texts with Natural Language Processing methods. Multiple authors who participated in the workshop used methods from the conceptual structures field including Formal Concept Analysis and Conceptual Graphs. Applications include but are not limited to text mining police reports, sociological definitions, movie reviews, etc.