NIPS 2014 Workshop
December 12, 2014
From online news to online shopping to scholarly research, we are inundated with a torrent of information on a daily basis. With our limited time, money and attention, we often struggle to extract actionable knowledge from this deluge of data. A common approach for addressing this challenge is personalization, where results are automatically filtered to match the tastes and preferences of individual users.
This workshop aims to bring together researchers from industry and academia in order to describe recent advances and discuss future research directions pertaining to the personalization of digital systems, broadly construed. We aim to highlight new and emerging research opportunities for the machine learning community that arise from the evolving needs for personalization.
09.00-09.45 Ralf Herbrich (Amazon)
09.45-10.00 Poster Spotlights
10.00-10.30 Coffee Break
10.30-11.15 Polo Chau (Georgia Tech) [slides]
11.15-12.30 Poster Session
12.30-15.00 Lunch Break
15.00-15.45 Emma Brunskill (CMU) [slides]
15.45-16.30 Karthik Raman (Cornell) [slides]
16.30-17.00 Coffee Break
17.00-17.45 Jason Weston (Facebook)
17.45-18.30 Panel Discussion + Closing
Emma Brunskill (Carnegie Mellon University)
Partially Personalized Policies
Interactive machine learning systems have the opportunity to provide better education, healthcare and marketing. Such systems often interact many times with the same user, and across many users, offering the chance to customize the experience per person. In this talk I will frame doing so as an instance of transfer learning across related stochastic decision processes, which lead to partially personalized policies. This approach leads to formal benefits in terms of performance and sample complexity. I also wil briefly describe one of our current applications efforts for this work.
Emma Brunskill is an assistant professor of computer science at CMU, where she is also affiliated with the machine learning department. Her work has been recognized with a NSF Career award, a Microsoft Research Faculty Fellow award, and has received best paper nominations at CHI and Educational Data Mining.
Polo Chau (Georgia Tech)
Human-in-the-loop Graph Sensemaking: Bridging HCI and Data Mining
Massive datasets now arise in virtually all domains. Yet, making sense of these data remains a fundamental challenge. At the Polo Club of Data Science, we are innovating at the intersection of Data Mining and Human-Computer Interaction (HCI), to combine the best from both worlds to create novel tools for making sense of graphs with billions of nodes and edges. I will describe some of our latest works and ideas:
Duen Horng (Polo) Chau is an Assistant Professor at Georgia Tech's School of Computational Science and Engineering, and an Associate Director of the MS Analytics program. Polo holds a PhD in Machine Learning and a Masters in human-computer interaction (HCI). His PhD thesis won Carnegie Mellon's Computer Science Dissertation Award, Honorable Mention. Polo’s research lab bridges data mining and HCI to solves large-scale, real world problems. They develop scalable, interactive, and interpretable tools for big data analytics. Their patented Polonium malware detection technology protects 120 million people worldwide. Their auction fraud detection research was widely covered by media. Their fake review detection research received the Best Student Paper Award at the 2014 SIAM Data Mining Conference. Polo received a Yahoo Faculty Research and Engagement Award, a Raytheon Faculty Fellowship, and LexisNexis Dean's Excellence Award. He is the only two-time Symantec fellow and an award-winning designer. Polo designed Carnegie Mellon's ID card.
Ralf Herbrich (Amazon)
Learning to Rank at Scale
In large-scale e-commerce platforms, purchase logs provide a rich source of information with respect to customer preferences. The signal contained in these logs can be used in several ways to improve the quality of product recommendations. However, a variety of deployment constraints limits the range of viable options: (1) the dimensionality and amount of available data is so large that in-memory learning approaches might not be possible, (2) preference might drift over time making it necessary to retrain the ranking model regularly over freshly collected data, and (3) ranking in real time becomes necessary whenever the query context becomes so complex that predictions can no longer be cached; such a constraint requires both to minimize prediction latency and maximize the overall throughput. In this talk, we will discuss these practical challenges in learning to rank. Sparse models will be the key to reduce prediction latency, while one-pass stochastic optimization will minimize training time and also limiting the memory footprint. Extensive experiments show that one-pass learning is further capable of preserving most of the prediction accuracy. We will present an analysis of the relative advantages and disadvantages of sparsification methods. We will present results from public benchmarks in text document retrieval, where our approach leads to state-of-the-art results, in spite of its simplicity and of being tailored in the first place to the large scale. The presented work is joint work with Antonino Freno, Rodolphe Jenatton, Martin Saveski and Cédric Archambeau.
Ralf Herbrich is Director of Machine Learning at Amazon. From 2011 to 2012, he worked at Facebook leading the Unified Ranking and Allocation team building large-scale machine learning infrastructure for learning user-action-rate predictors that enabled unified value experiences across Facebook products. From 2009 to 2011, he was Director of Microsoft's Future Social Experiences (FUSE) Lab UK working on the development of computational intelligence technologies on large online data collections. From 2006 to 2010, Ralf was co-leading the Applied Games and Online Services and Advertising group at Microsoft Research Cambridge which engaged in research at the intersection of machine learning and computer games and in the areas of online services, search and online advertising combining insights from machine learning, information retrieval, game theory, artificial intelligence and social network analysis. Ralf joined Microsoft Research in 2000 as a Postdoctoral researcher and Research Fellow of the Darwin College Cambridge. Prior to joining Microsoft, he obtained both a diploma degree in Computer Science in 1997 and a Ph.D. degree in Statistics in 2000 from Technical University of Berlin. Ralf has published over 80 papers and holds over 30 patents. His research interests include Bayesian inference and decision making, kernel methods, statistical learning theory, distributed systems and programming languages. Ralf is one of the inventors of the Drivatars™ system in the Forza Motorsport series as well as the TrueSkill™ ranking and matchmaking system in Xbox 360 Live. He also co-invented the adPredictor click-prediction technology launched in 2009 in Bing's online advertising system.
Karthik Raman (Cornell University)
"By the User, For the User, With the Learning System": Learning from User Interactions
Online information systems like search engines and recommender systems have used machine learning algorithms to learn and adapt quickly so as to increase their performance. However, to improve fidelity, robustness and cost-effectiveness of these complex systems, we need to look beyond conventional learning techniques which rely on expert-labeled data. In this talk, I will present principled learning algorithms that learn continuously from and with the users in an interactive manner. I will demonstrate that these algorithms perform well in end-to-end evaluation studies with live users, while also admitting theoretical guarantees. I will first describe how these algorithms can be used to overcome noise and biases present in user feedback. Second, I will show how we can learn the dependencies across different items (e.g. documents of a ranking), by explicitly modeling the joint utility of a set of items as a submodular set function. Third, I will describe how we can learn to reconcile the conflicting preferences of a diverse user population, to obtain socially beneficial solutions.
Karthik Raman is a final year PhD student at Cornell University. Motivated by applications such as search, recommendation and educational assessment, his research aims to tackle learning problems with a human-in-the-loop. His work on understanding the role of diversity in complex search tasks won a best paper award at SIGIR. He is supported by a Google PhD Fellowship and a Yahoo! Key Scientific Challenge Award.
Jason Weston (Facebook)
Hashtags, Clicks and Likes: Supervision for Content-based Post Recommendation [EMNLP paper]
We study the problem of understanding the content of short textual posts popular in social networks, with the aim of using that understanding to improve personalization and recommendation. There are a variety of weakly supervised signals for such a goal, including hashtags provided by the users themselves, as well as various kinds of click, comment and like interactions. We show how employing word embedding neural network models can effectively utilize this supervision.
This is joint work with Keith Adams, Sumit Chopra, Misha Denil, Emily Denton, Rob Fergus, Marc’Aurelio Ranzato, Ledell Wu and Ming Yang.
Jason Weston is a research scientist at Facebook, NY since February 2014. He earned his PhD in machine learning at Royal Holloway, University of London and at AT&T Research in Red Bank, NJ (advisors: Alex Gammerman, Volodya Vovk and Vladimir Vapnik) in 2000. From 2000 to 2002, he was a researcher at Biowulf technologies, New York. From 2002 to 2003 he was a research scientist at the Max Planck Institute for Biological Cybernetics, Tuebingen, Germany. From 2003 to 2009 he was a research staff member at NEC Labs America, Princeton. From 2009 to 2014 he was a research scientist at Google, NY. His interests lie in statistical machine learning and its application to text and images. Jason has published over 90 papers, including best paper awards at ICML and ECML. He was also part of the YouTube team that won a National Academy of Television Arts & Sciences Emmy Award for Technology and Engineering for Personalized
Recommendation Engines for Video Discovery.
Christophe Dupuy, Francis Bach and Christophe Diot. Review Prediction Using Topic Models.
Vijay Kamble, Nadia Fawaz and Fernando Silveira. Sequential Relevance Maximization with Binary Feedback.
Huitian Lei, Ambuj Tewari and Susan Murphy. An Actor-Critic Contextual Bandit Algorithm for Personalized Interventions using Mobile Devices.
Xiujun Li, Chenlei Guo, Wei Chu, Ye-Yi Wang and Jude Shavlik. Deep Learning Powered In-Session Contextual Ranking using Clickthrough Data.
Ulrich Paquet, Noam Koenigstein and Ole Winther. A Scalable Bayesian Alternative to Density Estimation with a Bilinear Softmax Function.
Maja Rudolph, San Gultekin, John Paisley and Shih-Fu Chang. Probabilistic Canonical Tensor Decomposition for Predicting User Preference.
Svitlana Volkova and David Yarowsky. Improving Gender Prediction of Social Media Users via Weighted Annotator Rationales.
Xiaoting Zhao and Peter Frazier. A Markov Decision Process Analysis of the Cold Start Problem in Bayesian Information Filtering.
We welcome the following types of papers:
We encourage submissions from a wide range of disciplines, from machine learning to HCI to the social sciences. Topics of interest include (but are not limited to):
Submissions should be 4-8 pages long, and adhere to the NIPS format. Please make the author information visible.
Submissions will be accepted online here.
Khalid El-Arini (Facebook)
Yisong Yue (Caltech)
Dilan Görür (Microsoft)