Although deep learning has demonstrated astonishing performance in many applications, there are still concerns on their dependability. One desirable property of deep learning applications with societal impact is fairness (i.e., non-discrimination). Unfortunately, discrimination might be intrinsically embedded into the models due to discrimination in the training data. As a countermeasure, fairness testing systemically identifies discriminative samples, which can be used to retrain the model and improve its fairness. Existing fairness testing approaches however have two major limitations. First, they only work well on traditional machine learning models and have poor performance (e.g., effectiveness and efficiency) on deep learning models. Second, they only work on simple tabular data and are not applicable for domains such as text. In this work, we bridge the gap by proposing a scalable and effective approach for systematically searching for discriminative samples while extending fairness testing to address a challenging domain, i.e., text classification. Compared with state-of-the-art methods, our approach only employs lightweight procedures like gradient computation and clustering, which makes it significantly more scalable. Experimental results show that on average, our approach explores the search space more effectively (9.62 and 2.38 times more than the state-of-art methods respectively on tabular and text datasets) and generates much more individual discriminatory instances (24.95 and 2.68 times) within reasonable time. The retrained models reduce discrimination by 57.2% and 60.2% respectively on average.
翻译:虽然深层次的学习在许多应用中表现出惊人的成绩,但仍然令人对其可靠性感到关切。深层次学习应用对社会有影响的可取属性之一是公平(即不歧视)。不幸的是,由于培训数据中的歧视,歧视可能内在地嵌入模型中。作为一种反措施,公平测试系统化地识别歧视性样本,可以用来对模型进行再培训并提高其公平性。虽然现有的公平测试方法有两大局限性。首先,它们仅仅在传统机器学习模型上运作良好,在深层次学习模型上表现不佳(例如,效力和效率)。第二,它们只处理简单的表格数据,不适用于文本等领域。在这项工作中,我们提出一种可扩展和有效的方法,系统搜索歧视性样本,同时将公平测试扩大到一个具有挑战性的领域,即文本分类。与最先进的方法相比,我们的方法只是使用较轻的模型,如梯度计算和组合,因此其业绩(例如,效力和效率和效率)较差。实验结果显示,我们的方法仅涉及简单的表格数据,对搜索空间进行更有效的探索,不适用于文本等。(462和2-3倍地分别对平均时间段和单个数据进行了更精确的重新计算)。