Information Market-Based Decision Fusion

Improved classification performance has practical real-world benefits ranging from improved effectiveness in detecting diseases to increased efficiency in identifying firms that are committing financial fraud. Multiclassifier combination (MCC) aims to improve classification performance by combining the decisions of multiple individual classifiers. In this paper, we present information market-based fusion (IMF), a novel multiclassifier combiner method for decision fusion that is based on information markets. In IMF, the individual classifiers are implemented as participants in an information market where they place bets on different object classes. The reciprocals of the market odds that minimize the difference between the total betting amount and the potential payouts for different classes represent the MCC probability estimates of each class being the true object class. By using a market-based approach, IMF can adjust to changes in base-classifier performance without requiring offline training data or a static ensemble composition. Experimental results show that when the true classes of objects are only revealed for objects classified as positive, for low positive ratios, IMF outperforms three benchmarks combiner methods, majority, average, and weighted average; for high positive ratios, IMF outperforms majority and performs on par with average and weighted average. When the true classes of all objects are revealed, IMF outperforms weighted average and majority and marginally outperforms average.