Abstract

Abstract

Abstract
There have been several research works proposing new Explainable AI (XAI) methods designed to generate model explanations having specific properties, or desiderata, such as fidelity, robustness, or human-interpretability. However, explanations are seldom evaluated based on their true practical impact on decision-making tasks. Without that assessment, explanations might be chosen that, in fact, hurt the overall performance of the combined system of ML model + end-users. This study aims to bridge this gap by proposing XAI Test, an application-grounded evaluation methodology tailored to isolate the impact of providing the end-user with different levels of information. We conducted an experiment following XAI Test to evaluate three popular XAI methods - LIME, SHAP, and TreeInterpreter - on a real-world fraud detection task, with real data, a deployed ML model, and fraud analysts. During the experiment, we gradually increased the information provided to the fraud analysts in three stages: Data Only, i.e., just transaction data without access to model score nor explanations, Data + ML Model Score, and Data + ML Model Score + Explanations. Using strong statistical analysis, we show that, in general, these popular explainers have a worse impact than desired. Some of the conclusion highlights include: i) showing Data Only results in the highest decision accuracy and the slowest decision time among all variants tested, ii) all the explainers improve accuracy over the Data + ML Model Score variant but still result in lower accuracy when compared with Data Only; iii) LIME was the least preferred by users, probably due to its substantially lower variability of explanations from case to case. © 2021 ACM.

Abstract
Several algorithms have been proposed for offline multi-label classification. However, applications in areas such as traffic monitoring, social networks, and sensors produce data continuously, the so called data streams, posing challenges to batch multi-label learning. With the lack of stationarity in the distribution of data streams, new algorithms are needed to online adapt to such changes (concept drift). Also, in realistic applications, changes occur in scenarios with infinitely delayed labels, where the true classes of the arrival instances are never available. We propose an online unsupervised incremental method based on self-organizing maps for multi-label stream classification in scenarios with infinitely delayed labels. We consider the existence of an initial set of labeled instances to train a self-organizing map for each label. The learned models are then used and adapted in an evolving stream to classify new instances, considering that their classes will never be available. We adapt to incremental concept drifts by online updating the weight vectors of winner neurons and the dataset label cardinality. Predictions are obtained using the Bayes rule and the outputs of each neuron, adapting the prior probabilities and conditional probabilities of the classes in the stream. Experiments using synthetic and real datasets show that our method is highly competitive with several ones from the literature, in both stationary and concept drift scenarios. © 2021 ACM.

Abstract
Social networks are becoming larger and more complex as new ways of collecting social interaction data arise (namely from online social networks, mobile devices sensors, ...). These networks are often large-scale and of high dimensionality. Therefore, dealing with such networks became a challenging task. An intuitive way to deal with this complexity is to resort to tensors. In this context, the application of tensor decomposition has proven its usefulness in modelling and mining these networks: it has not only been applied for exploratory analysis (thus allowing the discovery of interaction patterns), but also for more demanding and elaborated tasks such as community detection and link prediction. In this work, we provide an overview of the methods based on tensor decomposition for the purpose of analysing time-evolving social networks from various perspectives: from community detection, link prediction and anomaly/event detection to network summarization and visualization. In more detail, we discuss the ideas exploited to carry out each social network analysis task as well as its limitations in order to give a complete coverage of the topic.

Abstract
The usage of non-traditional data in credit scoring, from microfinance institutions, is very useful when trying to address the problem, very common in emerging markets, of the lack of a verifiable customers' credit history. In this context, this paper relies on data acquired from smartphones in a loan classification problem. We conduct a set of experiments concerning feature selection, strategies to deal with imbalanced datasets and algorithm choice, to define a baseline model. This model is, then, compared to others adding network features to the original ones. For that comparison, we generate a network that links a given user to its phone book contacts which are users of a given mobile application, taking into account the ethics and privacy concerns involved, and use some feature extraction techniques, such as the introduction of centrality measures and the definition of node embeddings, in order to capture certain aspects of the network's topology. Several node embedding algorithms are tested, but only Node2Vec proves to be significantly better than the baseline model, applying Friedman's post hoc tests. This node embedding algorithm outperforms all the other, representing a relative improvement, in comparison with the baseline model, of 5.74% on the mean accuracy, 7.13% on the area under the Receiver Operating Characteristic curve and 30.83% on the Kolmogorov-Smirnov statistic scores. This method, therefore, proves to be very promising when trying to discriminate between "good" and "bad" customers, in credit scoring classification problems.