HomeNewsListPh.D candidate Jie Ding from Harvard University visited on December 30, 2015

Ph.D candidate Jie Ding from Harvard University visited on December 30, 2015

Ph.D candidate Jie Ding from Harvard University visited our lab on December 30, 2015. He gave a talk on "Bridge criterion: A new model selection criterion beyond AIC and BIC" in Room 9-206, Rohm Building at 3PM.

Jie Ding is currently a PhD candidate at Harvard John A. Paulson School of Engineering and Applied Sciences, supervised by Prof. Vahid Tarokh. He has broad interests in combinatorics, statistical inference, machine learning, and information theory. Specifically, his recent work includes key pre-distribution for wireless sensor networks, sequence design for coded aperture imaging systems, cyclic difference set constructions, capacity of Markov random fields, model selection criteria, manifold learning, causal inference. He and his colleagues are currently leading an exciting two-year project on high-dimensional time series analysis and prediction.

 

Title:

Bridge criterion: A new model selection criterion beyond AIC and BIC

Abstract:

We introduce a new criterion to determine the order/dimension of an autoregressive model fitted to time series data. The proposed technique is shown to give a consistent and asymptotically efficient order estimation under mild assumptions. It has the benefits of the two state-of-the-art model selection techniques, the Akaike information criterion and the Bayesian information criterion. When the true model dimension is relatively large compared with the sample size, the Akaike information criterion is known to be efficient, and the new criterion behaves in a similar manner. When the true dimension is finite and small compared with the sample size, the Bayesian information criterion is known to be consistent, and so is the new criterion. Thus, the new criterion, referred to as the bridge criterion, automatically builds a bridge between the two classical criteria. In fact, it is a purely data driven criterion that can be understood as an optimal infinite sequence of hypothesis tests on the model dimension. In practice, where the observed time series is given without any prior information about the dimension, bridge criterion is more flexible and robust compared with classical approaches. A possible future work is to study to what extent the superiority of bridge criterion can be extended to various other probabilistic models.