Abstract
We study BIC-like model selection criteria and in particular, their refinements that include a constant term involving the Fisher information matrix. We observe that for complex Bayesian network models, the constant term is a negative number with a very large absolute value that dominates the other terms for small and moderate sample sizes. We show that including the constant term degrades model selection accuracy dramatically compared to the standard BIC criterion where the term is omitted. On the other hand, we demonstrate that exact formulas such as Bayes factors or the normalized maximum likelihood (NML), or their approximations that are not based on Taylor expansions, perform well. A conclusion is that in lack of an exact formula, one should use either BIC, which is a very rough approximation, or a very close approximation but not an approximation that is truncated after the constant term.
Notes
- 1.
Our earlier paper on this topic appeared as an invited paper at the ITA-2013 workshop. Hence, no prior peer-reviewed publication of this material exists beyond the basic Monte Carlo approximation proposed in [13]. In particular, this is the first study where the lower-order terms of information criteria are discussed in conjunction with a model class for which model selection criteria are being intensively developed.
- 2.
An interesting line of future research will be to zoom in into the differences in model complexity within classes of networks with a fixed number of parameters by the techniques we use here.
- 3.
Unlike in the numerical studies in the previous section, here we want to take into account the fine-grained differences between FII values between different Bayesian network models with a fixed number of parameters.
References
Clarke, B.S., Barron, A.R.: Jeffreys prior is asymptotically least favorable under entropy risk. J. Stat. Plan. Infer. 41(1), 37–61 (1994)
Grünwald, P.D.: The Minimum Description Length Principle. MIT Press, Cambridge (2007)
Han, C., Carlin, B.P.: Markov chain Monte Carlo methods for computing Bayes factors. J. Am. Stat. Assoc. 96(455), 1122–1132 (2001)
Jeffreys, H.: An invariant form for the prior probability in estimation problems. J. Roy. Stat. Soc. A 186(1007), 453–461 (1946)
Kass, R.E., Raftery, A.E.: Bayes factors. J. Am. Stat. Assoc. 90(430), 773–795 (1995)
Kontkanen, P., Myllymäki, P.: A linear-time algorithm for computing the multinomial stochastic complexity. Inf. Process. Lett. 103(6), 227–233 (2007)
Kontkanen, P., Myllymäki, P., Silander, T., Tirri, H., Grünwald, P.: On predictive distributions and Bayesian networks. Stat. Comput. 10, 39–54 (2000)
Krichevsky, R., Trofimov, V.: The performance of universal coding. IEEE Trans. Inf. Theor. 27(2), 199–207 (1981)
Navarro, D.: A note on the applied use of MDL approximations. Neural Comput. 16(9), 1763–1768 (2004)
Rasmussen, C. E., Ghahramani, Z.: Occam’s razor. In: Leen, T., Dietterich, T., Tresp, V. (eds.) Advances in Neural Information Processing Systems, pp. 294–300 (2001)
Rissanen, J.: Fisher information and stochastic complexity. IEEE Trans. Inf. Theor. 42(1), 40–47 (1996)
Rissanen, J.: Information and Complexity in Statistical Modeling. Springer, New York (2007)
Roos, T.: Monte Carlo estimation of minimax regret with an application to MDL model selection. In: Proceedings of the IEEE Information Theory Workshop, pp. 284–288. IEEE Press (2008)
Roos, T., Rissanen, J.: On sequentially normalized maximum likelihood models. In: Rissanen, J., Liski, E., Tabus, I., Myllymäki, P., Kontoyiannis, I., Heikkonen, J. (eds.) Proceedings of the Workshop on Information Theoretic Methods in Science and Engineering (WITMSE 2008), Tampere, Finland (2008)
Roos, T., Zou, Y.: Keep it simple stupid – on the effect of lower-order terms in BIC-like criteria. In: Information Theory and Applications Workshop (ITA), pp. 1–7. IEEE Press (2013)
Schwarz, G.: Estimating the dimension of a model. Ann. Stat. 6, 461–464 (1978)
Shtarkov, Y.M.: Universal sequential coding of single messages. Probl. Inf. Transm. 23(3), 3–17 (1987)
Silander, T., Roos, T., Kontkanen, P., Myllymäki, P.: Factorized normalized maximum likelihood criterion for learning Bayesian network structures. In: Jaeger, M., Nielsen, T. D. (eds.) Proceedings of the 4th European Workshop on Probabilistic Graphical Models (PGM 2008), pp. 257–272 (2008)
Silander, T., Roos, T., Myllymäki, P.: Learning locally minimax optimal Bayesian networks. Int. J. Approx. Reason. 51(5), 544–557 (2010)
Xie, Q., Barron, A.R.: Asymptotic minimax regret for data compression, gambling, and prediction. IEEE Trans. Inf. Theor. 46(2), 431–445 (2000)
Author information
Authors and Affiliations
Corresponding author
Editor information
Editors and Affiliations
Rights and permissions
Copyright information
© 2015 Springer International Publishing Switzerland
About this paper
Cite this paper
Zou, Y., Roos, T. (2015). On Model Selection, Bayesian Networks, and the Fisher Information Integral. In: Suzuki, J., Ueno, M. (eds) Advanced Methodologies for Bayesian Networks. AMBN 2015. Lecture Notes in Computer Science(), vol 9505. Springer, Cham. https://doi.org/10.1007/978-3-319-28379-1_9
Download citation
DOI: https://doi.org/10.1007/978-3-319-28379-1_9
Published:
Publisher Name: Springer, Cham
Print ISBN: 978-3-319-28378-4
Online ISBN: 978-3-319-28379-1
eBook Packages: Computer ScienceComputer Science (R0)