Theory of electoral college framework-based multi-classifier ensembling and its applications to subjective pattern recognition

本文的研究将应用于一类我们称之为主观模式识别的模式识别应用，以人脸识别和基于内容的信息检索为典型例子。我们对大脑如何处理这类信息知之甚少；事实上，判断相似/不相似的标准是主观的——通常需要人力来验证机器得出的任何结论。尽管每年都有新的算法被开发出来，但大部分研究遵循的是一般的“试错”程序：给定一个数据集，对于已知的算法，我们尝试各种参数；如果对精度不满意，可以尝试新的算法；然后对于新算法，我们尝试不同的参数；等等......。本文提出的研究将导致一种新的主观模式识别算法，该算法具有可预测的高稳定性，从而保证在高水平上执行（即准确性）。

The proposed research will be applied to a class of pattern recognition applications, which we call subjective pattern recognition, with face identification and content-based information retrieval as typical examples. We know very little about how our brains process this type of information; indeed, the standards for judging the similarity/dissimilarity are subjective -- manpower is usually required to verify any conclusions made by machine. Although new algorithms are developed every year, much of the research follows a general "trial-and-error" procedure: Given a data set, for a known algorithm, we try various parameters; if we are not satisfied with the accuracy, we try a new algorithm; and then for the new algorithm, we try various parameters; and so on. The proposed research will lead to a new type of algorithm for subjective pattern recognition, which has predictable high stability and therefore is guaranteed to perform (i.e. accuracy) at a high level. The research proposed is closely related to the Electoral College (EC). The EC voting format has been used for many years in political elections: a nation is partitioned into regions; the winner of each region is determined by a majority of its voting population; the final winner is selected according to the weighted sum of each candidate's winning regions based on the winner-take-all principle. It has also been used in many areas of scientific research. This research attempts to develop a model, our so-called EC framework-based multi- classifier ensembling, to improve the regular EC format in that general decision making approaches rather than simple vote counting will be used in each region for local decision making, and advanced ensembling technology rather than simple winner-take-all rule will be used for combining local decisions into final ones. On the theoretical front, our model will be the first that integrates classifier ensembling techniques with the EC framework; it will also be the first that illustrates the stabilities and applicabilities of the EC framework- based multi-classifier ensembles. The proposed research will establish guidelines for adopting our model in pattern recognition applications, particularly in subjective pattern recognition.