סמינר מחלקתי - המחלקה להנדסת תעשייה

Targeted Bayesian Network Learning

Aviv Gruber - Ph.D. student

Abstract:

In this talk, I will present the Targeted Bayesian Network Learning (TBNL) method, the central theme of my PhD research. The TBNL is a Bayesian network learning method, designed to approximate the marginal probability distribution of a predefined target variable, conditioned on other related attributes within the domain. Unlike canonical methods, of which objectives are to best approximate the joint probability distribution of the domain subject to some constraints, the TBNL considers the objective as the target variable, already at the learning stage, increasing the learning efficiency.  On the other hand, while other generative target-oriented BN classifiers (such as the naive-Bayes and the Tree Augmented Network), have quite limited flexibility with regard to their structure, the TBNL puts together a discriminative approach along with principles from Information Theory, making it a flexible target-oriented learner.

I will present the method and show via a Pareto front over a complexity-accuracy plane how in classification problems, the proposed method effectively manages the tradeoff between the classification error and the model complexity.

I shall also present two real-world examples of applications: in one, the telecom domain, the method has been used for effectively detecting suspect and anomalies based on the behavioural patterns. In the second example, a freight-car fleet management simulation model, based on a European operator, the TBNL was used for predicting failures, in order to optimize the preventive maintenance plan in a condition-based maintenance fashion.

העבודה נעשתה בהנחיית פרופ' עירד בן גל .

ההרצאה תתקיים ביום ג', 19.06.12, בשעה 14:00 בחדר 206, בנין וולפסון הנדסה, הפקולטה להנדסה, אוניברסיטת תל-אביב.