Applied Machine Learning Day

Join us for an applied machine learning event in Microsoft Herzelia.

Agenda:

• 10:00 - 10:45 Keynote Speaker: Dr. Kira Radinsky (http://tx.technion.ac.il/~kirar/), Recognized by MIT Technology Review as one of the world's 35 top young innovators (http://www.technologyreview.com/lists/innovators-under-35/2013/inventor/kira-radinsky/) for accomplishments that are poised to have a dramatic impact on the world as we know it in the field of Computer Science: Predicting the future in Business

• 10:45 - 11:20 Dr. Danny Bickson, Co-Founder of GraphLab (http://graphlab.com/): Machine Learning in the Cloud with GraphLab

• 11:20 - 11:55 Dr. Ron Karidi, SparkBeyond (https://twitter.com/sparkbeyond): Art and Science in Feature Engineering

• 11:55 - 12:10 Coffe Break

• 12:10 - 12:45 Assaf Araki, Big Data Analytics Architect. Intel: Generic Recommendation package on Hadoop

• 12:45 - 13:20 Ari Yakir, Project Manager @ CogniTeam, Applied Machine Learning in Hebrew

• 13:20 - 14:45 Lunch break (on your own)

• 14:45 - 15:20 Dr. Noam Koenigstein & Dr. Royi Ronen, Microsoft: Sage and Xbox Recommendations

• 15:20 - 15:55 Dr. Ira Cohen, Chief Scientist, HP Software: Scaling the data scientist

• 15:55 - 16:30 Hani Neuvirth, IBM: Machine Learning Applied to Real-World Medical Data

Title: Predicting the future in Business (Dr. Kira Radinsky, SalesPredict)

Abstract:
In our highly data-driven environment, businesses are essentially becoming semi-autonomous agents, constantly competing for resources, customers and talent.

How can they adapt and remain competitive in an ever-changing market? What techniques could leverage real-time data to boost their customer acquisition process? How can they learn about their competition, when the competition is learning them? Can the effects of business decisions on the eco-system be anticipated? To answer those questions, I will present machine learning and text mining techniques, that mine data from heterogeneous sources, to predict in such multi-agent temporal environments.

Title: Machine Learning in the Cloud with GraphLab (Dr. Danny Bickson, GraphLab Inc.)

Abstract:

From social networks, to protein molecules and the web, graphs encode structure and context, enable advanced machine learning, and are rapidly becoming the future of big-data. In this talk we will present the next generation of GraphLab, an open-source platform and machine learning framework designed to process graphs with hundreds of billions of vertices and edges on hardware ranging from a single mac-mini to the cloud.

Title: Sage and Xbox Recommendations (Dr. Noam Koenigstein & Dr. Royi Ronen, Microsoft)

Abstract:

In this talk we will be introducing both project Sage and Xbox Recommendations

Sage is Microsoft's all-purpose recommender system, designed and deployed as an ultra-high scale cloud service. The Sage project focuses on both state of the art research well as high scale robust implementation. On the research front, we demonstrate new pre-processing and cleaning techniques, a novel probabilistic matrix factorization model for implicit one-class training data, and a relatively new evaluation framework for implicit recommenders. On the engineering front, we present a working service deployed on Microsoft's Azure cloud which provides easy-to-use interfaces to integrate on any website or online store.

The Xbox Recommendation team is responsible for serving recommendations to more than 50 million Xbox users worldwide. We provide personalized content in the games, movies, TV and music domains. Our recommendation engine is based on a novel Bayesian Matrix Factorization model designed specifically for the one-class problem. Unlike previous approaches, our model delineates the odds of a user disliking an item from simply not considering it. Relations between items and users are modelled as an unobserved random graph connecting users with items they might have encountered.

Title: Applied Machine Learning in Hebrew (Ari Yakir)

Abstract:

Using Machine Learning algorithms to infer word semantics or phonetic relations can be a challenging task. Doing so in Hebrew makes it that much harder. Many machine learning algorithms often outperform others on a wide context but when implemented in a specific domain perform terribly. In this short presentation I will cover the lessons learned from the implementation of Machine Learning in speech tel-rehabilitation and the various methods we used to solve problematic aspects of Hebrew as a language for machine learning.

Title: Scaling the data scientist (Dr. Ira Cohen, HP Software)

Abstract:

Organizations are starting to realize the potential of analyzing the huge amount of data they are collecting, reporting very large ROIs on advanced/predictive analytics projects (100s-1000s%). Data scientists are the experts that have the knowledge to bring such project to successful outcomes – but there is a real shortage in data scientists, and organizations are struggling to fulfill the promise of big data analytics. We describe our innovative approach to deal with the shortage: changes in process and development of an innovative development tool, both empowering the subject matter experts (SMEs) to do most of the role of the data scientist. We report on the success of the approach in HP-SW, where the number of advanced analytics capabilities being developed jumped from four per year to almost 20 in just two quarters without an increase in the number of data scientists in the organization.

Title: Generic Recommendation package on Hadoop (Assaf Araki, Intel)

Abstract:

Implementing holistic machine learning solution within a company is complex and build from several machine learning models. Building several predictions models mean reuse not only of the algorithms package but also reuse of the process. In this talk we will present a generic recommendation system solution build on top of Mahout ( machine learning on top of Hadoop ) which shorten dramatically the time to market for projects. We will present an hybrid solution from algorithm and platform perspectives and present two implementations.

Title: Art and Science in Feature Engineering (Dr. Ron Karidi, SparkBeyond)

Abstract:

Predictive models are eventually as good as their features; thus, we all had those moments when, after long struggling to squeeze higher precision/recall out of the model, we were inspired to introduce a new feature that we magically derived from other features and with which our model simply out-performed everything we tried before. Often we wish we could have some process to lead us to such breakthroughs; other times, we actually take pride in these sparks of creativity.

Similarly, all good models get to the point where they are suddenly not that good even after we retrain them; it is often when something fundamental has changed, and as data scientists we need to get back to the drawing canvas. Sometimes the underlying numeric interactions have somewhat changed, in other times it’s the semantic context (e.g. new titles, new categories, etc.) and often it’s just a data dictionary or a DB schema that were updated.

In this talk we discuss these challenges in the predictive model lifecycle, and the spectrum of tools we can throw at them. We propose a novel approach to find “good” predictive features by considering methodological search of the feature space, and demonstrate how it is used in actual examples to build and/or improve existing models.

Title: Machine Learning Applied to Real-World Medical Data (Hani Neuvirth, IBM)

Abstract:

Chronic diseases are an increasingly important factor in healthcare costs, as well as the patients' quality of life. Advanced analytic methods can nowadays utilize the vast amounts of data collected by the different stakeholders in order to provide a personalized treatment, identify risks and support advanced care management. In this talk I will describe how we leverage advance analytics of real-world evidence data of patients with chronic diseases, including diabetes and epilepsy, to improve patient care.