We look at a specific aspect of model interpretability: models often need to be constrained in size for them to be considered interpretable. But smaller models also tend to have high bias. This suggests a trade-off between interpretability and accuracy. Our work addresses this by: (a) showing that learning a training distribution (often different from the test distribution) can often increase accuracy of small models, and therefore may be used as a strategy to compensate for small sizes, and (b) providing a model-agnostic algorithm to learn such training distributions. We pose the distribution learning problem as one of optimizing parameters for an Infinite Beta Mixture Model based on a Dirichlet Process, so that the held-out accuracy of a model trained on a sample from this distribution is maximized. To make computation tractable, we project the training data onto one dimension: prediction uncertainty scores as provided by a highly accurate oracle model. A Bayesian Optimizer is used for learning the parameters. Empirical results using multiple real world datasets, various oracles and interpretable models with different notions of model sizes, are presented. We observe significant relative improvements in the F1-score in most cases, occasionally seeing improvements greater than 100% over baselines. Additionally we show that the proposed algorithm provides the following benefits: (a) its a framework which allows for flexibility in implementation, (b) it can be used across feature spaces, e.g., the text classification accuracy of a Decision Tree using character n-grams is shown to improve when using a Gated Recurrent Unit as an oracle, which uses a sequence of characters as its input, (c) it can be used to train models that have a non-differentiable training loss, e.g., Decision Trees, and (d) reasonable defaults exist for most parameters of the algorithm, which makes it convenient to use.
翻译:我们查看了模型可解释性的一个具体方面:模型通常需要限制其大小才能被视为可以解释。但较小的模型也往往具有高度偏差。这表明在解释性和准确性之间有一个权衡。我们的工作通过以下方式解决这个问题:(a) 显示学习培训分布(通常不同于测试分布)可以提高小模型的准确性,因此可以用作补偿小模型的策略,并(b) 提供模型-不可知的算法来学习这种培训分布。我们提出分配空间学习问题,作为基于 Dirichlet 进程的Infinite Beta 特性模型优化参数的一个问题。这表示在解释性和准确性之间有一个权衡。我们的工作通过下列方法可以实现这一点:(a) 显示培训分布(通常不同于测试分布分布分布), 我们把培训数据投放到一个维度上: 预测精确度以非常精确的模型提供的不确定性分数, 并且(b) 用于学习默认参数。 (b) 使用多个真实的世界数据集、 各种或标志和解释性模型的精度的精度, 具有不同模型的精度的精度的精度的精度 。(ericreal d) 将显示一个比其精度的精度的精度的精度的精度的精度, 在使用一个不细度框架的精度框架中, 我们观察到的精度能度1 显示的精度上显示的精度可以显示的精度的精度的精度的精度的精度的精度, 用于在使用一个不细的精度的精度, 。