Andreas Weigend, who served as Amazon.com's Chief Scientist until January 2004, will discuss sources of data in e-business and characteristics of click and purchase data. He will present some work on the discovery of probabilistic models over customer attributes, and on dynamic probabilistic relational models that predict customer intentions and modalities from click streams in real time. This work illustrates the interactive and iterative process of both offline data analysis and online experiments, consisting of the steps: Design, Measurement, Characterization, Prediction and Evaluation, and Action.
While experiments on the web can give rapid short-term feedback, creating a framework for modeling and predicting long-term customer behavior in response to actions is a promising research problem. Such a framework should, for example, answer strategic questions like: What is the long-term impact of displaying at an e-business website sponsored links that take customers to competitors? The talk ends with the discussion of a recent platform of active learning, Amazon.com's two-panel Goldbox that launched last month.
Throughout the talk, we will see that three ingredients are necessary for research that leads to actionable outcomes: (i) Measurement (data and infrastructure), (ii) Methodology (statistical modeling and machine learning), and (iii) Domain knowledge and insights from behavioral economics and human decision making.