Frank W. Takes

Rachel G. de Jong, Mark P. J. van der Loo, Frank W. Takes

With the introduction of large-scale network data, including population-scale social networks, techniques for privacy-aware sharing of network data become increasingly important. While existing $k$-anonymity approaches can model different attacker scenarios, they typically assume that attacker knowledge exactly matches the published network structure. We argue that exact knowledge is often unrealistic and introduce $ϕ$-$k$-anonymity, a fuzzy variant of $k$-anonymity in which parameter $ϕ$ captures the level of uncertainty in attacker knowledge. Across a benchmark of $39$ real-world networks, a realistic level of uncertainty ($ϕ=5\%$) renders, on average, $64\%$ of previously unique nodes anonymous. To further enhance anonymity, we apply anonymization algorithms under a $5\%$ edge modification budget. While full anonymization is often unattainable under exact $k$-anonymity, with low uncertainty ($ϕ=10\%$) our newly proposed Greedy algorithm anonymizes over $99\%$ of the nodes. Uncertainty also enables effective anonymization in otherwise difficult to anonymize dense synthetic graphs. Additionally, data utility in terms of structural properties and performance on network analysis tasks is well preserved, with most metrics changing less than $5\%$. Overall, our findings suggest that modest uncertainty assumptions yield high levels of anonymity and utility, motivating further research on uncertainty-aware privacy guarantees for network data.

António Pereira Barata, Frank W. Takes, H. Jaap van den Herik et al.

When dealing with sensitive data in automated data-driven decision-making, an important concern is to learn predictors with high performance towards a class label, whilst minimising for the discrimination towards any sensitive attribute, like gender or race, induced from biased data. A few hybrid tree optimisation criteria exist that combine classification performance and fairness. Although the threshold-free ROC-AUC is the standard for measuring traditional classification model performance, current fair tree classification methods mainly optimise for a fixed threshold on both the classification task as well as the fairness metric. In this paper, we propose a compound splitting criterion which combines threshold-free (i.e., strong) demographic parity with ROC-AUC termed SCAFF -- Splitting Criterion AUC for Fairness -- and easily extends to bagged and boosted tree frameworks. Our method simultaneously leverages multiple sensitive attributes of which the values may be multicategorical or intersectional, and is tunable with respect to the unavoidable performance-fairness trade-off. In our experiments, we demonstrate how SCAFF generates models with performance and fairness with respect to binary, multicategorical, and multiple sensitive attributes.

Q. A. Meertens, C. G. H. Diks, H. J. van den Herik et al.

In this paper, we study the accuracy of values aggregated over classes predicted by a classification algorithm. The problem is that the resulting aggregates (e.g., sums of a variable) are known to be biased. The bias can be large even for highly accurate classification algorithms, in particular when dealing with class-imbalanced data. To correct this bias, the algorithm's classification error rates have to be estimated. In this estimation, two issues arise when applying existing bias correction methods. First, inaccuracies in estimating classification error rates have to be taken into account. Second, impermissible estimates, such as a negative estimate for a positive value, have to be dismissed. We show that both issues are relevant in applications where the true labels are known only for a small set of data points. We propose a novel bias correction method using Bayesian inference. The novelty of our method is that it imposes constraints on the model parameters. We show that our method solves the problem of biased classification-based aggregates as well as the two issues above, in the general setting of multi-class classification. In the empirical evaluation, using a binary classifier on a real-world dataset of company tax returns, we show that our method outperforms existing methods in terms of mean squared error.

Francesco Bortolussi, Zeger Hoogeboom, Frank W. Takes

Cryptocurrencies such as Bitcoin and Ethereum have recently gained a lot of popularity, not only as a digital form of currency but also as an investment vehicle. Online marketplaces and exchanges allow users across the world to convert between dozens of different cryptocurrencies and regular currencies such as euros or dollars. Due to the novelty of this concept, the volatility of these markets and the differences in maturity and usage of particular marketplaces, currency pairs may appear at multiple marketplaces but at different trading prices. This paper proposes a novel algorithmic approach to take advantage of these mispricings and capitalize upon the pricing differences that exist between exchanges and currency pairs. To do so, we model each combination of a currency and a market as one node in a graph. A directed link between two nodes indicates that a conversion between these two currency/market pairs is possible. The weight of the link relates to the exchange rate of executing this particular currency exchange. To leverage the mispricings, we seek for cycles in the graph such that upon multiplying the weights of the links in the cycle, a value greater than 1 is found and thus a profit can be made. Our goal is to do this efficiently, without exhaustively enumerating all possible cycles in the graph. Therefore, we convert our data and address the problem in terms of finding minimum weight triangles in graphs with integer weights, for which efficient algorithms can be utilized. We experiment with parameter settings (heuristics) related to the conversion of exchange rate data into integer weight values. We show that our approach improves upon a reasonable baseline algorithm in terms of computation time. Furthermore, using a real-world dataset, we demonstrate how the obtained minimal weight cycles indeed unveil a number of currency exchange cycles that result in a net profit.