Using generative adversarial networks for improving classifification effectiveness in credit card fraud detection
Abstract:
In the last years, the number of frauds in credit card-based online payments has grown dramatically, pushing banks and e-commerce organizations to implement automatic fraud detection systems, performing data mining on huge transaction logs. Machine learning seems to be one of the most promising solutions for spotting illicit transactions, by distinguishing fraudulent and non-fraudulent instances through the use of supervised binary classification systems properly trained from pre-screened sample datasets. However, in such a specific application domain, datasets available for training are strongly imbalanced, with the class of interest considerably less represented than the other. This significantly reduces the effectiveness of binary classifiers, undesirably biasing the results toward the prevailing class, while we are interested in the minority class. Oversampling the minority class has been adopted to alleviate this problem, but this method still has some drawbacks. Generative Adversarial Networks are general, flexible, and powerful generative deep learning models that have achieved success in producing convincingly real-looking images. We trained a GAN to output mimicked minority class examples, which were then merged with training data into an augmented training set so that the effectiveness of a classifier can be improved. Experiments show that a classifier trained on the augmented set outperforms the same classifier trained on the original data, especially as far the sensitivity is concerned, resulting in an effective fraud detection mechanism.
Generative Adversarial Networks (GANs) are deep learning technologies, building up multiple layers of abstraction to learn hierarchies of concepts, that have achieved considerable success in generating convincing examples, especially reallooking images. They are very general, though, and can be applied to several settings. GANs are composed of two models, a generative one and a discriminative one, which compete against one another, playing a zero-sum minimax game [17]. In the usual setting, where the two adversaries are multilayer perceptrons, some of the issues arising with training GANs are reduced.
To address the imbalanced dataset problems in supervised classification-based credit card fraud detection, we build an augmented training set.
Generative adversarial networks:
1.A GAN consists of two feed-forward neural networks, a Generator G and a Discriminator D competing against each other, with G producing new candidates and its adversary D evaluating their quality.
2.The main idea in GANs is to refine a generative model by making it confront an adversary, a discriminative model that has the goal of separating the generated examples from real ones.
3.The main idea in GANs is to refine a generative model by making it confront an adversary, a discriminative model tha has the goal of separating the generated examples from real ones. The generator takes random noise z as input, transforms it through a function and produces examples, while the discriminator learns to determine whether an example has been produced by the generator (see Fig. 1).
4.The training goal for the generator is tricking the discriminator into believing that generated examples are real. The discriminator is trained by minimizing its prediction error, whereas the generator is trained on the basis of maximizing the prediction error by the discriminator. This results in a competition between generator and discriminator that can be formalized as a minimax game:
where pD is the data distribution, pZ is the prior distribution of the generative network, and θG (resp., θD) are the parameters of the generator (resp., discriminator) network.
5.The first is the discriminator within the GAN,whose purpose is to distinguish the synthetic examples produced by the GAN generator G from the real examples; the second is the classifier C whose task is to identify fraudulent examples from non-fraudulent ones. 、
6.As regards the type of activation function in each layer, alternatives explored included the logistic sigmoid, defined as：
7.Weights and biases are adapted to the input on the basis of the desired output by means of training, when such parameters are progressively modified (on the basis of the learning rate) so that performance improves.
8.While we have presented our framework in the context of credit card fraud detection, it is quite general and can readily be extended to other application domains characterized by significant class imbalance rates. We are actively pursuing that endeavor. We have seen how, as it could be expected, injecting synthetic examples in a training set causes an increase in false positives. Clearly, this can be a limiting factor in settings where the cost of a false positive is relatively high. However, the use of ensemble methods can remedy that and it is worth investigating. The proposed mechanism is intrinsically dependent on the availability of labeled instances of fraudulent transactions. In an unsupervised setting, it would be difficult to apply our framework. It should be taken into account that, in the specific context of credit card fraud, customer complaints are a valuable source of labeled data. Finally, while our method can handle frauds which are similar, in essence, to malevolent transactions seen before, it can be expected to be largely ineffective in spotting frauds that are completely novel, where there is no information to generalize upon.
To validate our framework, we performed experiments on publicly available credit card fraud detection data, where the minority class is severely underrepresented. Our framework achieved an improved sensitivity at the cost of a limited increase in false positives. Another interesting point is computing performance, since the potentially costly portion of the elaboration, training the GAN, is done on a small subset of the training data. Directions for future study are manifold. We are planning to devise a strategy to reduce the decrease in specificity to a minimum.
论文的一点总结和体会：
1.这篇论文是CCFA类会议，主要用了有监督的方法——对抗生成网络（GAN）来提高行用卡欺诈检测的分类故障，论文中介绍了由于样本正例和负例类别不平衡等问题，用GAN训练多层感知器输出模拟的少数族类样本，将这些样本与训练数据合并到一个增强的训练集中，从而提高分类的有效性。GAN将每一层学到的知识视为原始输入的表示层，可以理想地与抽象层合成能力相关联。
2.生成网络的目的是学习训练数据的概率分布，通过将z映射到这种分布，以生成尽可能接近真实数据实例的新的人工候选对象。另一方面，对抗性鉴别器网络的目标是通过惩罚生成器生成人工候选实例的活动，正确区分真实数据和人工候选。显然，生成器网络通过生成尽可能逼真的合成实例来欺骗鉴别器网络，以提高对手的错误率，从而导致猫和老鼠游戏，在该游戏中，两个竞争者都会提高自己的能力，直到达到平衡，在平衡中生成示例与真实的不可区分。
3.鉴别器通过最小化其预测误差来训练，而发生器则通过鉴别器最大化误差来训练。
4.当一个组分对于另一个组分出现严重不平衡时，整个GAN失效。
5.最后，作者通过实验得出结论：虽然我们已经在信用卡欺诈检测的背景下提出了我们的框架，但它是非常普遍的，可以很容易地扩展到其他应用领域，这些领域具有显著的类不平衡率。我们正在积极地努力。我们已经看到了，正如我们所预期的，在训练集中注入合成的例子是如何导致假阳性的增加的。显然，在假阳性的成本相对较高的情况下，这可能是一个限制因素。然而，使用集成方法可以弥补这一不足，值得研究。提议的机制本质上取决于标记的欺诈交易实例的可用性。在无人监督的环境中，很难应用我们的框架。应考虑到，在信用卡欺诈的特定背景下，客户投诉是标记数据的宝贵来源。最后，虽然我们的方法可以处理类似于以前看到的恶意交易的欺诈行为，但在发现完全新颖的欺诈行为方面，它在很大程度上是无效的，因为没有可归纳的信息。
为了验证我们的框架，我们对公开可用的信用卡欺诈检测数据进行了实验，其中少数类严重缺乏代表性。我们的框架以有限增加误报为代价，提高了敏感性。另一个有趣的点是计算性能，因为细化的潜在成本部分，即对GAN的培训，是在培训数据的一小部分上完成的。未来的研究方向是多方面的。我们计划制定一个策略，将特异性降低到最低限度。