Time series anomaly detection (TSAD) finds many applications such as monitoring environmental sensors, industry KPIs, patient biomarkers, etc. A two-fold challenge for TSAD is a versatile and unsupervised model that can detect various different types of time series anomalies (spikes, discontinuities, trend shifts, etc.) without any labeled data. Modern neural networks have outstanding ability in modeling complex time series. Self-supervised models in particular tackle unsupervised TSAD by transforming the input via various augmentations to create pseudo anomalies for training. However, their performance is sensitive to the choice of augmentation, which is hard to choose in practice, while there exists no effort in the literature on data augmentation tuning for TSAD without labels. Our work aims to fill this gap. We introduce TSAP for TSA “on autoPilot”, which can (self-)tune augmentation hyperparameters end-to-end. It stands on two key components: a differentiable augmentation architecture and an unsupervised validation loss to effectively assess the alignment between augmentation type and anomaly type. Case studies show TSAP’s ability to effectively select the (discrete) augmentation type and associated (continuous) hyperparameters. In turn, it outperforms established baselines, including SOTA self-supervised models, on diverse TSAD tasks exhibiting different anomaly types.

Novel joint techniques capture both the microscopic context and the mesoscopic structure of networks by leveraging two previously separated fields of research: node representation learning (NRL) and community detection (CD). However, several limitations exist in the literature. First, a comprehensive comparison between these joint NRL-CD techniques is non-existent. Second, baseline techniques, datasets, evaluation metrics, and classification algorithms differ significantly between each method. Thirdly, the literature lacks a synchronized experimental approach, thus rendering comparison between these methods strenuous. To overcome these limitations, we present a uni-fied experimental setup mutually comparing six joint NRL-CD techniques and comparing them with corresponding NRL/CD baselines in three different settings: non-overlapping and over-lapping CD and node classification. Our results show that joint methods underperform on the node classification task but achieve relatively solid results for overlapping community detection. Our research contribution is two-fold: first, we show specific weaknesses of selected joint techniques in different tasks and data sets; and second, we suggest a more thorough experimental setup to benchmark joint techniques with simpler NRL and CD techniques.

In many real-life applications it is crucial to be able to, given a collection of link states of a network in a certain time period, accurately predict the link state of the network at a future time. This is known as dynamic link prediction, which compared to its static counterpart is more complex, as capturing the temporal characteristics is a non-trivial task. This explains while still majority of today's research in network representation learning focuses on static setting ignoring temporal information. In this work, we focus on one such case and aim at extending node2vec, representation learning method successfully applied for static link prediction, to a dynamic setup. This extended method is applied and validated on several real-life networks with different properties. Results show that taking into account dynamic aspect outperforms static approach. Additionally, based on the network properties, recommendations are given for the node2vec parameters.

When content consumers explicitly judge content positively, we consider them to be engaged. Unfortunately, explicit user evaluations are difficult to collect, as they require user effort. Therefore, we propose to use device interactions as implicit feedback to detect engagement. We assess the usefulness of swipe interactions on tablets for predicting engagement and make the comparison with using traditional features based on time spent. We gathered two unique datasets of more than 250,000 swipes, 100,000 unique article visits, and over 35,000 explicitly judged news articles by modifying two commonly used tablet apps of two newspapers. We tracked all device interactions of 407 experiment participants during one month of habitual news reading. We employed a behavioral metric as a proxy for engagement, because our analysis needed to be scalable to many users, and scanning behavior required us to allow users to indicate engagement quickly. We point out the importance of taking into account content ordering, report the most predictive features, zoom in on briefly read content and on the most frequently read articles. Our findings demonstrate that fine-grained tablet interactions are useful indicators of engagement for newsreaders on tablets. The best features successfully combine both time-based aspects and swipe interactions.

Most of the recent studies on churn prediction in telco utilize social networks built on top of the call (and/or SMS) graphs to derive informative features. However, extracting features from large graphs, especially structural features, is an intricate process both from a methodological and computational perspective. Due to the former, feature extraction in the current literature has mainly been addressed in an ad-hoc and hand-crafted manner. Due to the latter, the full potential of the structural information is unexploited. In this work, we incorporate both interaction and structural information by devising two different ways of enriching original graphs with interaction information, delineated by the well-known RFM model. We circumvent the process of extensive manual feature engineering by enriching the networks and improving the scalability of the renowned node2vec approach to learn node representations. The obtained results demonstrate that our enriched network outperforms baseline RFM-based methods.

Relational learning in networked data has been shown to be effective in a number of studies. Relational learners, composed of relational classifiers and collective inference methods, enable the inference of nodes in a network given the existence and strength of links to other nodes. These methods have been adapted to predict customer churn in telecommunication companies showing that incorporating them may give more accurate predictions. In this research, the performance of a variety of relational learners is compared by applying them to a number of CDR datasets originating from the telecommunication industry, with the goal to rank them as a whole and investigate the effects of relational classifiers and collective inference methods separately. Our results show that collective inference methods do not improve the performance of relational classifiers and the best performing relational classifier is the network-only link-based classifier, which builds a logistic model using link-based measures for the nodes in the network.

Past research has shown that both social and local features are informative for customer churn, however some studies have found that combining both kinds of data into a single model is ineffective. People who churn based on their neighbors' behavior are a distinct subset of customers from those who churn for personal reasons. However, for an effective retention campaign, it is desired to identify both groups of likely churners, attempt to explain the factors that lead to churn in both cases, and still determine the customers most likely to churn so they can be contacted. The goal of this research is to evaluate different techniques for combining features and models based on customer attributes and customer social networks to identify the best approaches to deal with this problem.

Fraud is a social process that occurs over time. We introduce a new approach, called AFRAID, which utilizes active inference to better detect fraud in time-varying social networks. That is, classify nodes as fraudulent vs. non-fraudulent. In active inference on social networks, a set of unlabeled nodes is given to an oracle (in our case one or more fraud inspectors) to label. These labels are used to seed the inference process on previously trained classifier(s). The challenge in active inference is to select a small set of unlabeled nodes that would lead to the highest classification performance. Since fraud is highly adaptive and dynamic, selecting such nodes is even more challenging than in other settings. We apply our approach to a real-life fraud data set obtained from the Belgian Social Security Institution to detect social security fraud. In this setting, fraud is defined as the intentional failing of companies to pay tax contributions to the government. Thus, the social network is composed of companies and the links between companies indicate shared resources. Our approach, AFRAID, outperforms the approaches that do not utilize active inference by up to 15% in terms of precision.

The selection of classifiers which are profitable is becoming more and more important in real-life situations such as customer churn management campaigns in the telecommunication sector. In previous works, the expected maximum profit (EMP) metric has been proposed, which explicitly takes the cost of offer and the customer lifetime value (CLV) of retained customers into account. It thus permits the selection of the most profitable classifier, which better aligns with business requirements of end-users and stake holders. However, modelers are currently limited to applying this metric in the evaluation step. Hence, we expand on the previous body of work and introduce a classifier that incorporates the EMP metric in the construction of a classification model. Our technique, called ProfLogit, explicitly takes profit maximization concerns into account during the training step, rather than the evaluation step. The technique is based on a logistic regression model which is trained using a genetic algorithm (GA). By means of an empirical benchmark study applied to real-life data sets, we show that ProfLogit generates substantial profit improvements compared to the classic logistic model for many data sets. In addition, profit-maximized coefficient estimates differ considerably in magnitude from the maximum likelihood estimates.

Process mining encompasses the research area which is concerned with knowledge discovery from event logs. One common process mining task focuses on conformance checking, comparing discovered or designed process models with actual real-life behavior as captured in event logs in order to assess the “goodness” of the process model. This paper introduces a novel conformance checking method to measure how well a process model performs in terms of precision and generalization with respect to the actual executions of a process as recorded in an event log. Our approach differs from related work in the sense that we apply the concept of so-called weighted artificial negative events toward conformance checking, leading to more robust results, especially when dealing with less complete event logs that only contain a subset of all possible process execution behavior. In addition, our technique offers a novel way to estimate a process model's ability to generalize. Existing literature has focused mainly on the fitness (recall) and precision (appropriateness) of process models, whereas generalization has been much more difficult to estimate. The described algorithms are implemented in a number of ProM plugins, and a Petri net conformance checking tool was developed to inspect process model conformance in a visual manner.

As social networks offer a vast amount of additional information to enrich standard learning algorithms, the most challenging part is extracting relevant information from networked data. Fraudulent behavior is imperceptibly concealed both in local and relational data, making it even harder to define useful input for prediction models. Starting from expert knowledge, this paper succeeds to efficiently incorporate social network effects to detect fraud for the Belgian governmental social security institution, and to improve the performance of traditional non-relational fraud prediction tasks. As there are many types of social security fraud, this paper concentrates on payment fraud, predicting which companies intentionally disobey their payment duties to the government. We introduce a new fraudulent structure, the so-called spider constructions, which can easily be translated in terms of social networks and included in the learning algorithms. Focusing on the egonet of each company, the proposed method can handle large scale networks. In order to face the skewed class distribution, the SMOTE approach is applied to rebalance the data. The models were trained on different timestamps and evaluated on varying time windows. Using techniques as Random Forest, logistic regression and Naive Bayes, this paper shows that the combined relational model improves the AUC score and the precision of the predictions in comparison to the base scenario where only local variables are used.

Process mining encompasses the research area which is concerned with knowledge discovery from information system event logs. Within the process mining research area, two prominent tasks can be discerned. First of all, process discovery deals with the automatic construction of a process model out of an event log. Secondly, conformance checking focuses on the assessment of the quality of a discovered or designed process model in respect to the actual behavior as captured in event logs. Hereto, multiple techniques and metrics have been developed and described in the literature. However, the process mining domain still lacks a comprehensive framework for assessing the goodness of a process model from a quantitative perspective. In this study, we describe the architecture of an extensible framework within ProM, allowing for the consistent, comparative and repeatable calculation of conformance metrics. For the development and assessment of both process discovery as well as conformance techniques, such a framework is considered greatly valuable.

The interest for data mining techniques has increased tremendously during the past decades, and numerous classification techniques have been applied in a wide range of business applications. Hence, the need for adequate performance measures has become more important than ever. In this paper, a cost-benefit analysis framework is formalized in order to define performance measures which are aligned with the main objectives of the end users, i.e., profit maximization. A new performance measure is defined, the expected maximum profit criterion. This general framework is then applied to the customer churn problem with its particular cost-benefit structure. The advantage of this approach is that it assists companies with selecting the classifier which maximizes the profit. Moreover, it aids with the practical implementation in the sense that it provides guidance about the fraction of the customer base to be included in the retention campaign.

Process discovery is the learning task that entails the construction of process models from event logs of information systems. Typically, these event logs are large data sets that contain the process executions by registering what activity has taken place at a certain moment in time. By far the most arduous challenge for process discovery algorithms consists of tackling the problem of accurate and comprehensible knowledge discovery from highly flexible environments. Event logs from such flexible systems often contain a large variety of process executions which makes the application of process mining most interesting. However, simply applying existing process discovery techniques will often yield highly incomprehensible process models because of their inaccuracy and complexity. With respect to resolving this problem, trace clustering is one very interesting approach since it allows to split up an existing event log so as to facilitate the knowledge discovery process. In this paper, we propose a novel trace clustering technique that significantly differs from previous approaches. Above all, it starts from the observation that currently available techniques suffer from a large divergence between the clustering bias and the evaluation bias. By employing an active learning inspired approach, this bias divergence is solved. In an assessment using four complex, real-life event logs, it is shown that our technique significantly outperforms currently available trace clustering techniques.

Within process mining research, one of the most important fields of study is process discovery, which can be defined as the extraction of control-flow models from audit trails or information system event logs. The evaluation of discovered process models is an essential but difficult task for any process discovery analysis. With this paper, we propose a novel approach for evaluating discovered process models based on artificially generated negative events. This approach allows for the definition of a behavioral F-measure for discovered process models, which is the main contribution of this paper.

Support vector machines (SVMs) are currently state-of-the-art for the classification task and, generally speaking, exhibit good predictive performance due to their ability to model nonlinearities. However, their strength is also their main weakness, as the generated nonlinear models are typically regarded as incomprehensible black-box models. In this paper, we propose a new Active Learning-Based Approach (ALBA) to extract comprehensible rules from opaque SVM models. Through rule extraction, some insight is provided into the logics of the SVM model. ALBA extracts rules from the trained SVM model by explicitly making use of key concepts of the SVM: the support vectors, and the observation that these are typically close to the decision boundary. Active learning implies the focus on apparent problem areas, which for rule induction techniques are the regions close to the SVM decision boundary where most of the noise is found. By generating extra data close to these support vectors that are provided with a class label by the trained SVM model, rule induction techniques are better able to discover suitable discrimination rules. This performance increase, both in terms of predictive accuracy as comprehensibility, is confirmed in our experiments where we apply ALBA on several publicly available data sets.

We address the problem of credit scoring as a classification and feature subset selection problem. Based on the current framework of sophisticated feature selection methods, we identify features that contain the most relevant information to distinguish good loan payers from bad loan payers. The feature selection methods are validated on several real-world datasets with different types of classifiers. We show the advantages following from using the subspace approach to classification. We discuss many practical issues related to the applicability of feature selection methods. We show and discuss some difficulties that used to be insufficiently emphasized in standard feature selection literature. © 2005 Wiley Periodicals, Inc. Int J Int Syst 20: 985–999, 2005.

In this paper, we evaluate and contrast two types of fuzzy classifiers for credit scoring. The first classifier uses evolutionary optimization and boosting for learning fuzzy classification rules. The second classifier is a fuzzy neural network that employs a fuzzy variant of the classic backpropagation learning algorithm. The experiments are carried out on a real life credit scoring data set. It is shown that, for the case at hand, the boosted genetic fuzzy classifier performs better than both the neurofuzzy classifier and the well-known C4.5(rules) decision tree(rules) induction algorithm. However, the better performance of the genetic fuzzy classifier is offset by the fact that it infers approximate fuzzy rules which are less comprehensible for humans than the descriptive fuzzy rules inferred by the neurofuzzy classifier. © 2002 Wiley Periodicals, Inc.

We study the problem of repeat-purchase modeling in a direct marketing setting using Belgian data. More specifically, we investigate the detection and qualification of the most relevant explanatory variables for predicting purchase incidence. The analysis is based on a wrapped form of input selection using a sensitivity based pruning heuristic to guide a greedy, stepwise, and backward traversal of the input space. For this purpose, we make use of a powerful and promising least squares support vector machine (LS-SVM) classifier formulation. This study extends beyond the standard recency frequency monetary (RFM) modeling semantics in two ways: (1) by including alternative operationalizations of the RFM variables, and (2) by adding several other (non-RFM) predictors. Results indicate that elimination of redundant/irrelevant inputs allows significant reduction of model complexity. The empirical findings also highlight the importance of frequency and monetary variables, while the recency variable category seems to be of somewhat lesser importance to the case at hand. Results also point to the added value of including non-RFM variables for improving customer profiling. More specifically, customer/company interaction, measured using indicators of information requests and complaints, and merchandise returns provide additional predictive power to purchase incidence modeling for database marketing. © 2001 John Wiley & Sons, Inc.