Simple low cost causal discovery using mutual information and domain knowledge

摘要: Bayesian networks (BNs) provide a means for representing, displaying, and making available in usable form the knowledge of experts given Weld. In this paper, we look at performance an expert constructed BN compared with other machine learning (ML) techniques predicting outcome (win, lose, or draw) matches played by Tottenham Hotspur Football Club. The period under study was 1995–1997 – start that period, based almost exclusively on subjective judgement. Our objective to determine retrospectively comparative accuracy some alternative ML models were built using data from two-year period. additional considered were: MC4, decision tree learner; Naive Data Driven (a whose structure node probability tables are learnt entirely data); K-nearest neighbour learner. results show is generally superior domain predictive accuracy. even more impressive BNs that, number key respects, assumptions place them disadvantage. For example, have assumed prediction ‘incorrect’ if predicts than one as equally most likely (whereas, fact, such would prove valuable somebody who could ‘each way’ bet outcome). Although has now long been irrelevant (since it contains variables relating players retired left club) here tend conWrm excellent potential when they reliable expert. ability accurate predictions without requiring much obvious bonus any where scarce. Moreover, relatively simple build its be used again similar types problems. © 2006 Elsevier B.V. All rights reserved.