AG 1, AG 2, AG 3, AG 4, AG 5, RG1, SWS, MMCI   Info

Public Audience

English

60 Minutes

Saarbrücken

When using machine learning methods to build predictive models from data, one often assumes that all data encountered at training and test time come from a single, homogeneous data source. However, in many scenarios available data are shaped by regional effects (spatial variation) or change dynamically (temporal variation). In this talk, I will discuss machine learning models that are optimized for these scenarios. Varying coefficient models can represent spatial or temporal variation in data by letting model parameters vary across space or time. Inference in varying coefficient models identifies spatial or temporal trends in the data and extrapolates these trends to novel locations and points in time. However, existing inference algorithms are computationally expensive; applications of these models have therefore been limited to small data sets. I present results for a novel inference algorithm that can reduce computational complexity by orders of magnitude for certain varying coefficient models. Empirical studies in seismic

risk analysis and real estate price prediction confirm the effectiveness of the proposed approach. Temporally varying data also poses challenges for model evaluation, because the predictive performance of a model will change over time as the distribution of data changes. I discuss methods for evaluating models efficiently in this scenario, specifically the problem of evaluating the performance of document ranking models in information retrieval.

Connie Balzert

--email hidden

passcode not visible

logged in users only