Recommender systems workshop
Joseph A. Konstan (University of Minnesota)
Recommender Systems: Beyond Machine Learning
Recommender systems help users find items of interest and help websites and marketers select items to promote. Today's recommender
systems incorporate sophisticated technology to model user preferences,
model item properties, and leverage the experiences of a large community
of users in the service of better recommendations. Yet all too often better
recommendations--at least by traditional measures of accuracy and precision--fail to meet the goal of improving user experience. This talk will take a look at successes and failures in moving beyond basic machine
learning approaches to recommender systems to emphasize factors tied to user behavior and experience. Along the way, we will explore a generalizable approach to combining human-centered evaluation with
data mining and machine learning techniques.
Joseph A. Konstan is Distinguished McKnight University Professor and Distinguished University Teaching Professor in the Department of Computer Science and Engineering at the University of Minnesota. His research addresses a variety of human-computer interaction issues, including personalization (particularly through recommender systems), eliciting on-line participation, and designing computer systems to improve public health. He is probably best known for his work in collaborative filtering recommenders (the GroupLens project, work which won the ACM Software Systems Award and Seoul Test of Time Award). Dr. Konstan received his Ph.D. from the University of California, Berkeley in 1993. He is a Fellow of the ACM, IEEE, and AAAS, and a member of the CHI Academy. Konstan is co-Chair of the ACM Publications Board, served as President of ACM SIGCHI and as a member of the ACM Council.
Shiri Kremer Davidson (IBM)
Recommendations for raising your social eminence in the enterprise
Oren Barkan (Microsoft)
One Class Matrix Factorization
Oren Somekh (Yahoo)
Adaptive Online Hyper-Parameters Tuning for Ad Event-Prediction Models
Yahoo's native advertising (also known as Gemini native) is one of its fastest growing businesses, reaching a run-rate of several hundred Millions USD in the past year. Driving the Gemini native models that are used to predict both, click probability (pCTR) and conversion probability (pCONV), is Offset - a feature enhanced collaborative-filtering (CF) based event prediction algorithm. Offset is a one-pass algorithm that updates its model for every new batch of logged data using a stochastic gradient descent (SGD) based approach. As most learning algorithms, Offset includes several hyper-parameters that can be tuned to provide best performance for a given system conditions. Since the marketplace environment is very dynamic and influenced by seasonality and other temporal factors, having a fixed single set of hyper-parameters (or configuration) for the learning algorithm is sub-optimal.
In this work we present an online hyper-parameters tuning algorithm, which takes advantage of the system parallel map-reduce based architecture, and strives to adapt the hyper-parameters set to provide the best performance at a specific time interval. Online evaluation via bucket testing of the tuning algorithm showed a significant 4.3% revenue lift overall traffic, and a staggering 8.3% lift over Yahoo Home-Page section traffic. Since then, the tuning algorithm was pushed into production, tuning both click- and conversion-prediction models, and is generating a hefty estimated revenue lift of 5% yearly for Yahoo Gemini native.
The proposed tuning mechanism can be easily generalized to fit any learning algorithm that continuously learns on incoming streaming data, in order to adapt its hyper-parameters to temporal changes.
Oren Somekh received his BSc, MSc, and PhD in Electrical Engineering from the Technion in 1989, 1991, and 2005, respectively. During 1991-1996 he served in the IDF Signals Corps. During 1998-2002 he was the VP R&D and later CTO of Surf Communication Solutions Ltd. During 2005-2008 he has been a visiting research fellow at the EE Departments of NJIT and Princeton University. Dr. Somekh Joined Yahoo Labs in 2009 and serves there as a senior Research Scientist since. He was a recipient of the European Community Marie-Curie Outgoing International Fellowship, a co-recipient of the first IEEE Information Theory Society ISIT best student paper award, a co-recipient of Yahoo 2015 Master Inventor award, and holds patents in the fields of Communications and Internet technologies. His current research interests are scientific aspects of Internet technologies, such as Recommendation Systems and Computational Advertising.
14:00 – 14:45
Evgenia Wasserman Pritsker (Haifa University)
Assessing the Contribution of Twitter's Textual Information to Graph-based Recommendation
14:45 - 15:00
15:00 - 15:45
Moran Gavish (Outbrain)
15:45 – 16:30
The effect of temporal trends on popularity
Abstract: The rich get richer principle, manifested by the Preferential Attachment (PA) mechanism, stipulates that popular nodes are bound to be more attractive than less popular nodes. However, it overlooks the transient nature of popularity, which is often governed by trends. Here, we suggest a quantitative methodology to detect trends in empirical data, and show their effects in 11 real-world datasets. We show that the numbers of links accumulated in any two time-steps is correlated due to their correlation with the overall number of garnered links, but once one controls for this effect, a clear picture of trends emerges—recent time-steps determine the number of accumulated links much more than earlier time-steps. As expected, the real-world networks differ in the rate by which earlier time steps lose their relevance, and this rate determines the networks’ susceptibility to short-lived trends. To formally quantify a network’s trendiness factor we use a natural generalization of PA, which we call Trending Preferential Attachment (PA). We show that TPA predicts the temporal popularity of nodes better than PA. We find that the real-world networks we studied were moderately to highly trendy. Importantly, a network’s trendiness factor remains constant throughout the network’s lifetime, and therefore constitutes a fundamental network property effecting node's popularity.
Osnat (Ossi) Received her B.Sc. in Computer Science and M.Sc. in Electrical Engineering from the Technion, Israel Institute of Technology, Haifa, Israel. She received her Ph.D. in Computer Science from the Hebrew University, Jerusalem, Israel, in 2004.
Ossi has worked at leading companies such as Intel and IBM research labs in Haifa, Israel.
She served as faculty and the head of the Internet and Networks Lab in The Academic College of Tel-Aviv Yaffo from 2008-2015. Starting 2015 She is faculty at The Information and Knowledge Management Dept., University of Haifa.
Her recent research focuses on social media data mining and recommender systems, and social networks.