We investigate the role of the initial screening order (ISO) in candidate screening. The ISO refers to the order in which the screener searches the candidate pool when selecting k candidates. Today, it is common for the ISO to be the product of an information access system, such as an online platform or a database query. The ISO has been largely overlooked in the literature, despite its impact on the optimality and fairness of the selected k candidates, especially under a human screener. We define two problem formulations describing the search behavior of the screener given an ISO: the best-k, where it selects the top k candidates; and the good-k, where it selects the first good-enough k candidates. To study the impact of the ISO, we introduce a human-like screener and compare it to its algorithmic counterpart, where the human-like screener is conceived to be inconsistent over time. Our analysis, in particular, shows that the ISO, under a human-like screener solving for the good-k problem, hinders individual fairness despite meeting group fairness, and hampers the optimality of the selected k candidates. This is due to position bias, where a candidate's evaluation is affected by its position within the ISO. We report extensive simulated experiments exploring the parameters of the best-k and good-k problems for both screeners. Our simulation framework is flexible enough to account for multiple candidate screening tasks, being an alternative to running real-world procedures.

Abstract Recent years have witnessed the rise of accurate but obscure classification models that hide the logic of their internal decision processes. Explaining the decision taken by a black-box classifier on a specific input instance is therefore of striking interest. We propose a local rule-based model-agnostic explanation method providing stable and actionable explanations. An explanation consists of a factual logic rule, stating the reasons for the black-box decision, and a set of actionable counterfactual logic rules, proactively suggesting the changes in the instance that lead to a different outcome. Explanations are computed from a decision tree that mimics the behavior of the black-box locally to the instance to explain. The decision tree is obtained through a bagging-like approach that favors stability and fidelity: first, an ensemble of decision trees is learned from neighborhoods of the instance under investigation; then, the ensemble is merged into a single decision tree. Neighbor instances are synthetically generated through a genetic algorithm whose fitness function is driven by the black-box behavior. Experiments show that the proposed method advances the state-of-the-art towards a comprehensive approach that successfully covers stability and actionability of factual and counterfactual explanations.

We describe the design, the architecture, and the evaluation of the Leasing Score Prediction (LSP) system - a credit scoring and credit rating system for the leasing sector deployed at the Italian association of leasing companies. Due to its challenging objectives, the design and complexity of the LSP system represent a unique contribution to the best practices in the field. We cover requirements by managers, users, and regulations about rigorous backtesting, statistical validation, calibration, explainability, robustness and uncertainty self-assessment. LSP relies on a machine learning model trained on a mixture of data distributions contributed by many associated leasing companies. We describe the technical solutions adopted and report on their performance evaluation, including the management of the data shifts due to the COVID-19 pandemic.

Decision Trees are accessible, interpretable, and well-performing classification models. A plethora of variants with increasing expressiveness has been proposed in the last forty years. We contrast the two families of univariate DTs, whose split functions partition data through axis-parallel hyperplanes, and multivariate DTs, whose splits instead partition data through oblique hyperplanes. The latter include the former, hence multivariate DTs are in principle more powerful. Surprisingly enough, however, univariate DTs consistently show comparable performances in the literature. We analyze the reasons behind this, both with synthetic and real-world benchmark datasets. Our research questions test whether the pre-processing phase of removing correlation among features in datasets has an impact on the relative performances of univariate vs multivariate DTs. We find that existing benchmark datasets are likely biased towards favoring univariate DTs.

The importance of achieving fairness in machine learning models cannot be overstated. Recent research has pointed out that fairness should be examined from a causal perspective, and several fairness notions based on the on Pearl’s causal framework have been proposed. In this paper, we construct a reweighting scheme of datasets to address causal fairness. Our approach aims at mitigating bias by considering the causal relationships among variables and incorporating them into the reweighting process. The proposed method adopts two neural networks, whose structures are intentionally used to reflect the structures of a causal graph and of an interventional graph. The two neural networks can approximate the causal model of the data, and the causal model of interventions. Furthermore, reweighting guided by a discriminator is applied to achieve various fairness notions. Experiments on real-world datasets show that our method can achieve causal fairness on the data while remaining close to the original data for downstream tasks.

We study the problem of estimating the total number of searches (volume) of queries in a specific domain, which were submitted to a search engine in a given time period. Our statistical model assumes that the distribution of searches follows a Zipf’s law, and that the observed sample volumes are biased accordingly to three possible scenarios. These assumptions are consistent with empirical data, with keyword research practices, and with approximate algorithms used to take counts of query frequencies. A few estimators of the parameters of the distribution are devised and experimented, based on the nature of the empirical/simulated data. For continuous data, we recommend using nonlinear least square regression (NLS) on the top-volume queries, where the bound on the volume is obtained from the well-known Clauset, Shalizi and Newman (CSN) estimation of power-law parameters. For binned data, we propose using a Chi-square minimization approach restricted to the top-volume queries, where the bound is obtained by the binned version of the CSN method. Estimations are then derived for the total number of queries and for the total volume of the population, including statistical error bounds. We apply the methods on the domain ofrecipes and cookingqueries searched in Italian in 2017. The observed volumes of sample queries are collected from Google Trends (continuous data) and SearchVolume (binned data). The estimated total number of queries and total volume are computed for the two cases, and the results are compared and discussed.

Segregation is the separation of social groups in the physical or in the online world. Segregation discovery consists of finding contexts of segregation. In the modern digital society, discovering segregation is challenging, due to the large amount and the variety of social data. We present a tool in support of segregation discovery from relational and graph data. The SCube system builds on attributed graph clustering and frequent itemset mining. It offers to the analyst a multi-dimensional segregation data cube for exploratory data analysis. The demonstration first guides the audience through the relevant social science concepts. Then, it focuses on scenarios around case studies of gender occupational segregation. Two real and large datasets about the boards of directors of Italian and Estonian companies will be explored in search of segregation contexts. The architecture of the SCube system and its computational efficiency challenges and solutions are discussed.

We study the problem of estimating the total volume of queries of a specific domain, which were submitted to the Google search engine in a given time period. Our statistical model assumes a Zipf's law distribution of the population in the reference domain, and a non-uniform or noisy sampling of queries. Parameters of the distribution are estimated using nonlinear least square regression. Estimations with errors are then derived for the total number of queries and for the total number of searches (volume). We apply the method on the recipes and cooking domain, where a sample of queries is collected by crawling popular Italian websites specialized on this domain. The relative volumes of queries in the sample are computed using Google Trends, and transformed to absolute frequencies after estimating a scaling factor. Our model estimates that the volume of Italian recipes and cooking queries submitted to Google in 2017 and with at least 10 monthly searches consists of 7.2B searches.

We introduce a framework for the data-driven analysis of social segregation of minority groups, and challenge it on a complex scenario. The framework builds on quantitative measures of segregation, called segregation indexes, proposed in the social science literature. The segregation discovery problem is introduced, which consists of searching sub-groups of population and minorities for which a segregation index is above a minimum threshold. A search algorithm is devised that solves the segregation problem by computing a multi-dimensional data cube that can be explored by the analyst. The machinery underlying the search algorithm relies on frequent itemset mining concepts and tools. The framework is challenged on a cases study in the context of company networks. We analyse segregation on the grounds of sex and age for directors in the boards of the Italian companies. The network includes 2.15M companies and 3.63M directors.

Attributed graphs model real networks by enriching their nodes with attributes accounting for properties. Several techniques have been proposed for partitioning these graphs into clusters that are homogeneous with respect to both semantic attributes and to the structure of the graph. However, time and space complexities of state of the art algorithms limit their scalability to medium-sized graphs. We propose SToC (for Semantic-Topological Clustering), a fast and scalable algorithm for partitioning large attributed graphs. The approach is robust, being compatible both with categorical and with quantitative attributes, and it is tailorable, allowing the user to weight the semantic and topological components. Further, the approach does not require the user to guess in advance the number of clusters. SToC relies on well known approximation techniques such as bottom-k sketches, traditional graph-theoretic concepts, and a new perspective on the composition of heterogeneous distance measures. Experimental results demonstrate its ability to efficiently compute high-quality partitions of large scale attributed graphs.

We present a framework for the analysis of corporate governance problems using network science and graph algorithms on ownership networks. In such networks, nodes model companies/shareholders and edges model shares owned. Inspired by the widespread pyramidal organization of corporate groups of companies, we model ownership networks as layered graphs, and exploit the layered structure to design feasible and efficient solutions to three key problems of corporate governance. The first one is the long-standing problem of computing direct and indirect ownership (integrated ownership problem). The other two problems are introduced here: computing direct and indirect dividends (dividend problem), and computing the group of companies controlled by a parent shareholder (corporate group problem). We conduct an extensive empirical analysis of the Italian ownership network, which, with its 3.9M nodes, is 30× the largest network studied so far.

We introduce a framework for a data-driven analysis of segregation of minority groups in social networks, and challenge it on a complex scenario. The framework builds on quantitative measures of segregation, called segregation indexes, proposed in the social science literature. The segregation discovery problem consists of searching sub-graphs and sub-groups for which a reference segregation index is above a minimum threshold. A search algorithm is devised that solves the segregation problem. The framework is challenged on the analysis of segregation of social groups in the boards of directors of the real and large network of Italian companies connected through shared directors. These keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.

The current availability of efficient algorithms for decision tree induction makes intricate post-processing techniques worth to be investigated both for efficiency and effectiveness. We study the simplification operator of subtree replacement, also known as grafting, originally implemented in the C4.5 system. We present a parametric bottom-up algorithm integrating grafting with the standard pruning operator, and analyze its complexity in terms of the number of nodes visited. Immediate instances of the parametric algorithm include extensions of error based, reduced error, minimum error, and pessimistic error pruning. Experimental results show that the computational cost of grafting is paid off by statistically significant smaller trees without accuracy loss.

With the support of the legally-grounded methodology of situation testing, we tackle the problems of discrimination discovery and prevention from a dataset of historical decisions by adopting a variant of k-NN classification. A tuple is labeled as discriminated if we can observe a significant difference of treatment among its neighbors belonging to a protected-by-law group and its neighbors not belonging to it. Discrimination discovery boils down to extracting a classification model from the labeled tuples. Discrimination prevention is tackled by changing the decision value for tuples labeled as discriminated before training a classifier. The approach of this paper overcomes legal weaknesses and technical limitations of existing proposals.

Concise representations of frequent itemsets sacrifice readability and direct interpretability by a data analyst of the concise patterns extracted. In this paper, we introduce an extension of itemsets, called regular, with an immediate semantics and interpretability, and a conciseness comparable to closed itemsets. Regular itemsets allow for specifying that an item may or may not be present; that any subset of an itemset may be present; and that any non-empty subset of an itemset may be present. We devise a procedure, called RegularMine, for mining a set of regular itemsets that is a concise representation of frequent itemsets. The procedure computes a covering, in terms of regular itemsets, of the frequent itemsets in the class of equivalence of a closed one. We report experimental results on several standard dense and sparse datasets that validate the proposed approach.

Discrimination discovery in databases consists in finding unfair practices against minorities which are hidden in a dataset of historical decisions. The DCUBE system implements the approach of [5], which is based on classification rule extraction and analysis, by centering the analysis phase around an Oracle database. The proposed demonstration guides the audience through the legal issues about discrimination hidden in data, and through several legally-grounded analyses to unveil discriminatory situations. The SIGMOD attendees will freely pose complex discrimination analysis queries over the database of extracted classification rules, once they are presented with the database relational schema, a few ad-hoc functions and procedures, and several snippets of SQL queries for discrimination discovery.

In the context of civil rights law, discrimination refers to unfair or unequal treatment of people based on membership to a category or a minority, without regard to individual merit. Discrimination in credit, mortgage, insurance, labor market, and education has been investigated by researchers in economics and human sciences. With the advent of automatic decision support systems, such as credit scoring systems, the ease of data collection opens several challenges to data analysts for the fight against discrimination. In this article, we introduce the problem of discovering discrimination through data mining in a dataset of historical decision records, taken by humans or by automatic systems. We formalize the processes of direct and indirect discrimination discovery by modelling protected-by-law groups and contexts where discrimination occurs in a classification rule based syntax. Basically, classification rules extracted from the dataset allow for unveiling contexts of unlawful discrimination, where the degree of burden over protected-by-law groups is formalized by an extension of the lift measure of a classification rule. In direct discrimination, the extracted rules can be directly mined in search of discriminatory contexts. In indirect discrimination, the mining process needs some background knowledge as a further input, for example, census data, that combined with the extracted rules might allow for unveiling contexts of discriminatory decisions. A strategy adopted for combining extracted classification rules with background knowledge is called an inference model. In this article, we propose two inference models and provide automatic procedures for their implementation. An empirical assessment of our results is provided on the German credit dataset and on the PKDD Discovery Challenge 1999 financial dataset.

Discrimination in social sense (e.g., against minorities and disadvantaged groups) is the subject of many laws worldwide, and it has been extensively studied in the social and economic sciences. We tackle the problem of determining, given a dataset of historical decision records, a precise measure of the degree of discrimination suffered by a given group (e.g., an etnic minority) in a given context (e.g., a geographic area) with respect to the decision (e.g. credit denial). In our approach, this problem is rephrased in a classification rule based setting, and a collection of quantitative measures of discrimination is introduced, on the basis of existing norms and regulations. The measures are defined as functions of the contingency table of a classification rule, and their statistical significance is assessed, relying on a large body of statistical inference methods for proportions. Based on this basic method, we are then able to address the more general problems of: (1) unveiling all discriminatory decision patterns hidden in the historical data, combining discrimination analysis with association rule mining, (2) unveiling discrimination in classifiers that learn over training data biased by discriminatory decisions, and (3) in the case of rule-based classifiers, sanitizing discriminatory rules by correcting their confidence. Our approach is validated on the German credit dataset and on the CPAR classifier.

In the context of civil rights law, discrimination refers to unfair or unequal treatment of people based on membership to a category or a minority, without regard to individual merit. Rules extracted from databases by data mining techniques, such as classification or association rules, when used for decision tasks such as benefit or credit approval, can be discriminatory in the above sense. In this paper, the notion of discriminatory classification rules is introduced and studied. Providing a guarantee of non-discrimination is shown to be a non trivial task. A naive approach, like taking away all discriminatory attributes, is shown to be not enough when other background knowledge is available. Our approach leads to a precise formulation of the redlining problem along with a formal result relating discriminatory rules with apparently safe ones by means of background knowledge. An empirical assessment of the results on the German credit dataset is also provided.

We present an analytic evaluation of the runtime behavior of the C4.5 algorithm which highlights some efficiency improvements. Based on the analytic evaluation, we have implemented a more efficient version of the algorithm, called EC4.5. It improves on C4.5 by adopting the best among three strategies for computing the information gain of continuous attributes. All the strategies adopt a binary search of the threshold in the whole training set starting from the local threshold computed at a node. The first strategy computes the local threshold using the algorithm of C4.5, which, in particular, sorts cases by means of the quicksort method. The second strategy also uses the algorithm of C4.5, but adopts a counting sort method. The third strategy calculates the local threshold using a main-memory version of the RainForest algorithm, which does not need sorting. Our implementation computes the same decision trees as C4.5 with a performance gain of up to five times.