Search

Federated Learning is an emerging privacy-preserving distributed machine learning approach to building a shared model by performing distributed training locally on participating devices (clients) and aggregating the local models into a global one. As this approach prevents data collection and aggregation, it helps in reducing associated privacy risks to a great extent. However, the data samples across all participating clients are usually not independent and identically distributed (non-i.i.d.), and Out of Distribution (OOD) generalization for the learned models can be poor. Besides this challenge, federated learning also remains vulnerable to various attacks on security wherein a few malicious participating entities work towards inserting backdoors, degrading the generated aggregated model as well as inferring the data owned by participating entities. In this work, we propose an approach for learning invariant (causal) features common to all participating clients in a federated learning setup and analyse empirically how it enhances the Out of Distribution (OOD) accuracy as well as the privacy of the final learned model. Although Federated Learning allows for participants to contribute their local data without revealing it, it faces issues in data security and in accurately paying participants for quality data contributions. In this report, we also propose an EOS Blockchain design and workflow to establish data security, a novel validation error based metric upon which we qualify gradient uploads for payment, and implement a small example of our Blockchain Causal Federated Learning model to analyze its performance with respect to robustness, privacy and fairness in incentivization., L’apprentissage fédéré est une approche émergente d’apprentissage automatique distribué préservant la confidentialité pour créer un modèle partagé en effectuant une formation distribuée localement sur les appareils participants (clients) et en agrégeant les modèles locaux en un modèle global. Comme cette approche empêche la collecte et l’agrégation de données, elle contribue à réduire dans une large mesure les risques associés à la vie privée. Cependant, les échantillons de données de tous les clients participants sont généralement pas indépendante et distribuée de manière identique (non-i.i.d.), et la généralisation hors distribution (OOD) pour les modèles appris peut être médiocre. Outre ce défi, l’apprentissage fédéré reste également vulnérable à diverses attaques contre la sécurité dans lesquelles quelques entités participantes malveillantes s’efforcent d’insérer des portes dérobées, dégradant le modèle agrégé généré ainsi que d’inférer les données détenues par les entités participantes. Dans cet article, nous proposons une approche pour l’apprentissage des caractéristiques invariantes (causales) communes à tous les clients participants dans une configuration d’apprentissage fédérée et analysons empiriquement comment elle améliore la précision hors distribution (OOD) ainsi que la confidentialité du modèle appris final. Bien que l’apprentissage fédéré permette aux participants de contribuer leurs données locales sans les révéler, il se heurte à des problèmes de sécurité des données et de paiement précis des participants pour des contributions de données de qualité. Dans ce rapport, nous proposons également une conception et un flux de travail EOS Blockchain pour établir la sécurité des données, une nouvelle métrique basée sur les erreurs de validation sur laquelle nous qualifions les téléchargements de gradient pour le paiement, et implémentons un petit exemple de notre modèle d’apprentissage fédéré blockchain pour analyser ses performances.

Estimation of causal effects using machine learning methods has become an active research field in econometrics. In this paper, we study the finite sample performance of meta-learners for estimation of heterogeneous treatment effects under the usage of sample-splitting and cross-fitting to reduce the overfitting bias. In both synthetic and semi-synthetic simulations we find that the performance of the meta-learners in finite samples greatly depends on the estimation procedure. The results imply that sample-splitting and cross-fitting are beneficial in large samples for bias reduction and efficiency of the meta-learners, respectively, whereas full-sample estimation is preferable in small samples. Furthermore, we derive practical recommendations for application of specific meta-learners in empirical studies depending on particular data characteristics such as treatment shares and sample size., 60 pages, 17 figures, 17 tables

We investigate the impact of the presence of university dropouts on the academic success of first-time students. Our identification strategy relies on quasi-random variation in the proportion of returning dropouts. The estimated average zero effect of dropouts on first- time students' success masks treatment heterogeneity and non-linearities. First, we find negative effects on the academic success of their new peers from dropouts re-enrolling in the same subject and, conversely, positive effects of dropouts changing subjects. Second, using causal machine learning methods, we find that the effects vary nonlinearly with different treatment intensities and prevailing treatment levels.

We provide new insights regarding the finding that Medicaid increased emergency department (ED) use from the Oregon experiment. We find meaningful heterogeneous impacts of Medicaid on ED use using causal machine learning methods. The treatment effect distribution is widely dispersed, and the average effect is not representative of most individualized treatment effects. A small group—about 14% of participants—in the right tail of the distribution drives the overall effect. We identify priority groups with economically significant increases in ED usage based on demographics and prior utilization. Intensive margin effects are an important driver of increases in ED utilization.

Estimating individualized treatment effects (ITEs) from observational data is crucial for decision-making. In order to obtain unbiased ITE estimates, a common assumption is that all confounders are observed. However, in practice, it is unlikely that we observe these confounders directly. Instead, we often observe noisy measurements of true confounders, which can serve as valid proxies. In this paper, we address the problem of estimating ITE in the longitudinal setting where we observe noisy proxies instead of true confounders. To this end, we develop the Deconfounding Temporal Autoencoder (DTA), a novel method that leverages observed noisy proxies to learn a hidden embedding that reflects the true hidden confounders. In particular, the DTA combines a long short-term memory autoencoder with a causal regularization penalty that renders the potential outcomes and treatment assignment conditionally independent given the learned hidden embedding. Once the hidden embedding is learned via DTA, state-of-the-art outcome models can be used to control for it and obtain unbiased estimates of ITE. Using synthetic and real-world medical data, we demonstrate the effectiveness of our DTA by improving over state-of-the-art benchmarks by a substantial margin. ISSN:2640-3498

Binary treatments are often ex-post aggregates of multiple treatments or can be disaggregated into multiple treatment versions. Thus, effects can be heterogeneous due to either effect or treatment heterogeneity. We propose a decomposition method that uncovers masked heterogeneity, avoids spurious discoveries, and evaluates treatment assignment quality. The estimation and inference procedure based on double/debiased machine learning allows for high-dimensional confounding, many treatments and extreme propensity scores. Our applications suggest that heterogeneous effects of smoking on birthweight are partially due to different smoking intensities and that gender gaps in Job Corps effectiveness are largely explained by differences in vocational training.

We implement a generalized random forest (Athey et al., 2019) to a differencein-difference setting to identify substantial heterogeneity in earnings losses across displaced workers. Using administrative data from Austria over three decades we document that a quarter of workers face cumulative 11-year losses higher than 2 times their pre-displacement annual income, while almost 10% of individuals experience gains. Our methodology allows us to consider many competing theories of earnings losses. We find that the displacement firm's wage premia and the availability of well paying jobs in the local labor market are the two most important factors. This implies that earnings losses can be understood by mean reversion in firm wage premia and losses in match quality, rather than by a destruction of firm-specific human capital. We further show that 94% of the cyclicality of earnings losses is explained by compositional changes of displaced workers over the business cycle.

Ce numéro de Regards économiques analyse la question de la formation des demandeurs d’emploi en Belgique, du côté néerlandophone en particulier. Ces formations sont-elles efficaces pour améliorer l’insertion professionnelle des chômeurs et si oui, pour quels demandeurs d’emploi en particulier ? Pour mener cette étude, Bart Cockx, spécialiste des questions du marché de l’emploi, a travaillé en collaboration avec le VDAB (le service public de l’emploi flamand) et le professeur Michael Lechner (Université de Saint-Gall en Suisse). Ils ont utilisé les techniques d’analyse statistique des Big Data. Ils les ont appliquées à un grand ensemble de données administratives anonymisées sur des demandeurs d’emploi résidant en Flandre. Ces techniques permettent de prévoir le gain d’emploi qu’un chômeur peut espérer de sa participation à une formation donnée. Les trois résultats majeurs présentés dans cette étude sont les suivants : 1. Quelles sont les formations les plus efficaces ? Parmi les types de formation étudiés (sessions d’orientation professionnelle, formation professionnelle de type court et formations de durée moyenne), ce sont les formations professionnelles courtes qui sont les plus efficaces; elles augmentent la durée moyenne passée en emploi des participants de 3,5 mois (soit + 21%) par rapport à l’absence de participation. 2. Pour quel profil de demandeur d’emploi les formations sont les plus efficaces ? Les formations ont le plus d’impact sur le retour à l’emploi des immigrés récents dont la connaissance du néerlandais est limitée. 3. Comment améliorer l’efficacité des formations ? Si le VDAB attribuait les programmes de formation aux demandeurs d’emploi selon l’algorithme informatique développé par les chercheurs, on obtiendrait une augmentation de 20 % de la durée passée en emploi grâce aux formations. Cette étude démontre que les outils informatiques et statistiques peuvent apporter un soutien dans l’attribution des politiques d’aide à l’emploi pour les chômeurs, mais également que les gains d’emploi peuvent être conséquents si on alloue les formations à celles et ceux qui en tirent le plus de bénéfices.

Online dating emerged as a key platform for human mating. Previous research focused on socio-demographic characteristics to explain human mating in online dating environments, neglecting the commonly recognized relevance of sport. This research investigates the effect of sport activity on human mating by exploiting a unique data set from an online dating platform. Thereby, we leverage recent advances in the causal machine learning literature to estimate the causal effect of sport frequency on the contact chances. We find that for male users, doing sport on a weekly basis increases the probability to receive a first message from a woman by 50%, relatively to not doing sport at all. For female users, we do not find evidence for such an effect. In addition, for male users the effect increases with higher income.

Limited runway capacity is a common problem faced by most airports worldwide. The two important factors that affect runway throughput are the wake-vortex separation and Runway Occupancy Time (ROT). Therefore, to improve runway throughput, Wake Turbulence Re-categorisation program (RECAT) was introduced to reduce the minimum separation distance required between successive aircraft on final approach. As a result, the constraining impact of ROT on runway throughput has now become significant. The objective of this paper is to identify data-driven intervention to reduce the ROT of landing aircraft. Specifically, we propose a data-driven approach to estimate the causal effect of landing parameters on ROT. We propose categorisation of each landing parameter into groups using Gaussian process models and employ Generalised Random Forest (GRF) to estimate the average treatment effect and the standard deviation of each landing parameters. Experimental results show that a few procedural changes to current landing procedure may reduce ROT. The results establish that slowing down the aircraft speed in the final approach phase leads to shorter ROT. In the final approach phase, ROTs of aircraft which are at least 10 knots slower than the average aircraft speed are on an average 2.63 seconds shorter. Furthermore, aircraft that are at least 10 knots faster than the average aircraft have on average 4 seconds longer ROTs. The second finding of this work is that flexible glide-slope angles should be introduced for the different aircraft types to achieve better ROT performance. Therefore, our findings also validate the industry need for Ground-Based Augmented System landing system which provides landing guidance with flexible glide-slopes. Civil Aviation Authority of Singapore (CAAS) Nanyang Technological University Accepted version This research has been supported by the Civil Aviation Authority of Singapore (CAAS) and the Air Traffic Management Research Institute (ATMRI), Singapore.

We investigate heterogenous employment effects of Flemish training programmes. Based on administrative individual data, we analyse programme effects at various aggregation levels using Modified Causal Forests (MCF), a causal machine learning estimator for multiple programmes. While all programmes have positive effects after the lock-in period, we find substantial heterogeneity across programmes and types of unemployed. Simulations show that assigning unemployed to programmes that maximise individual gains as identified in our estimation can considerably improve effectiveness. Simplified rules, such as one giving priority to unemployed with low employability, mostly recent migrants, lead to about half of the gains obtained by more sophisticated rules.

Based on administrative data of unemployed in Belgium, we estimate the labour market effects of three training programmes at various aggregation levels using Modified Causal Forests, a causal machine learning estimator. While all programmes have positive effects after the lock-in period, we find substantial heterogeneity across programmes and unemployed. Simulations show that “black-box” rules that reassign unemployed to programmes that maximise estimated individual gains can considerably improve effectiveness: up to 20% more (less) time spent in (un)employment within a 30 months window. A shallow policy tree delivers a simple rule that realizes about 70% of this gain.

Based on administrative data of unemployed in Belgium, we estimate the labour market effects of three training programmes at various aggregation levels using Modified Causal Forests, a causal machine learning estimator. While all programmes have positive effects after the lock-in period, we find substantial heterogeneity across programmes and unemployed. Simulations show that “black-box” rules that reassign unemployed to programmes that maximise estimated individual gains can considerably improve effectiveness: up to 20% more (less) time spent in (un)employment within a 30 months window. A shallow policy tree delivers a simple rule that realizes about 70% of this gain.

Based on administrative data of unemployed in Belgium, we estimate the labour market effects of three training programmes at various aggregation levels using Modified Causal Forests, a causal machine learning estimator. While all programmes have positive effects after the lock-in period, we find substantial heterogeneity across programmes and unemployed. Simulations show that “black-box” rules that reassign unemployed to programmes that maximise estimated individual gains can considerably improve effectiveness: up to 20% more (less) time spent in (un)employment within a 30 months window. A shallow policy tree delivers a simple rule that realizes about 70% of this gain.

Based on administrative data of unemployed in Belgium, we estimate the labour market effects of three training programmes at various aggregation levels using Modified Causal Forests, a causal machine learning estimator. While all programmes have positive effects after the lock-in period, we find substantial heterogeneity across programmes and unemployed. Simulations show that “black-box” rules that reassign unemployed to programmes that maximise estimated individual gains can considerably improve effectiveness: up to 20% more (less) time spent in (un)employment within a 30 months window. A shallow policy tree delivers a simple rule that realizes about 70% of this gain.

Matching-type estimators using the propensity score are the major workhorse in active labour market policy evaluation. This work investigates if machine learning algorithms for estimating the propensity score lead to more credible estimation of average treatment effects on the treated using a radius matching framework. Considering two popular methods, the results are ambiguous: We find that using LASSO based logit models to estimate the propensity score delivers more credible results than conventional methods in small and medium sized high dimensional datasets. However, the usage of Random Forests to estimate the propensity score may lead to a deterioration of the performance in situations with a low treatment share. The application reveals a positive effect of the training programme on days in employment for long-term unemployed. While the choice of the 'first stage' is highly relevant for settings with low number of observations and few treated, machine learning and conventional estimation becomes more similar in larger samples and higher treatment shares. Propensity Score Matching-Schätzer sind ein zentrales Tool, um Wirkungen aktiver Arbeitsmarktpolitik mit Individualdaten zu analysieren. Diese Studie untersucht, inwieweit Machine-Learning Algorithmen zur Schätzung des Propensity Score zu einer verbesserten Schätzung von durchschnittlichen Maßnahmeteilnahmeeffekten für die Teilnehmenden führt, wenn Radius-Matching verwendet wird. Es werden zwei gängige Methoden verwendet und die Analyse kommt zu keinem eindeutigen Ergebnis: Die Anwendung des LASSO auf der Basis von Logit-Modellen zur Schätzung des Propensity Score liefert überzeugendere Befunde als konventionelle Schätzmethoden, wenn kleine oder mittelgroße Datensätze verwendet werden. Wenn hingegen Random Forests zur Schätzung des Propensity Score verwendet werden, kann unter Verwendung von Stichproben mit einem geringen Anteil von Teilnehmenden das Gegenteil der Fall sein. Den Schätzergebnisse zufolge, führen die untersuchten Schulungsmaßnahmen dazu, dass die Tage in Beschäftigung nach Förderbeginn für die langzeitarbeitslosen Schulungsteilnehmenden höher ausfallen als ohne diese Förderung. Während bei Stichproben mit einer geringen Anzahl von Beobachtungen und einem kleinen Anteil von Maßnahmeteilnehmenden die Wahl des Schätzers für den Propensity Score für die Ergebnisse hoch relevant ist, sind die Resultate von Machine-Learning- und konventionellen Schätzungen bei einer hohen Anzahl von Beobachtungen und einem relativen hohen Anteil von Maßnahmeteilnehmenden ähnlicher.

Based on administrative data of unemployed in Belgium, we estimate the labour market effects of three training programmes at various aggregation levels using Modified Causal Forests, a causal machine learning estimator. While all programmes have positive effects after the lock-in period, we find substantial heterogeneity across programmes and unemployed. Simulations show that “black-box” rules that reassign unemployed to programmes that maximise estimated individual gains can considerably improve effectiveness: up to 20% more (less) time spent in (un)employment within a 30 months window. A shallow policy tree delivers a simple rule that realizes about 70% of this gain.

We propose an econometric procedure based mainly on the generalized random forests method. Not only does this process estimate the quantile treatment effect nonparametrically, but our procedure yields a measure of variable importance in terms of heterogeneity among control variables. We also apply the proposed procedure to reinvestigate the distributional effect of 401(k) participation on net financial assets, and the quantile earnings effect of participating in a job training program.

Uncovering the heterogeneity of causal effects of policies and business decisions at various levels of granularity provides substantial value to decision makers. This paper develops new estimation and inference procedures for multiple treatment models in a selection-on-observables frame-work by modifying the Causal Forest approach suggested by Wager and Athey (2018). The new estimators have desirable theoretical and computational properties for various aggregation levels of the causal effects. An Empirical Monte Carlo study shows that they may outperform previously suggested estimators. Inference tends to be accurate for effects relating to larger groups and conservative for effects relating to fine levels of granularity. An application to the evaluation of an active labour market programme shows the value of the new methods for applied research.

We investigate the finite sample performance of causal machine learning estimators for heterogeneous causal effects at different aggregation levels. We employ an Empirical Monte Carlo Study that relies on arguably realistic data generation processes (DGPs) based on actual data. We consider 24 different DGPs, eleven different causal machine learning estimators, and three aggregation levels of the estimated effects. In the main DGPs, we allow for selection into treatment based on a rich set of observable covariates. We provide evidence that the estimators can be categorized into three groups. The first group performs consistently well across all DGPs and aggregation levels. These estimators have multiple steps to account for the selection into the treatment and the outcome process. The second group shows competitive performance only for particular DGPs. The third group is clearly outperformed by the other estimators.

"I choose this restaurant because they have vegan sandwiches" could be a typical explanation we would expect from a human. However, current Reinforcement Learning (RL) techniques are not able to provide such explanations, when trained on raw pixels. RL algorithms for state-of-the-art benchmark environments are based on neural networks, which lack interpretability, because of the very factor that makes them so versatile – they have many parameters and intermediate representations. Enforcing safety guarantees is important when deploying RL agents in the real world, and guarantees require interpretability of the agent. Humans use short explanations that capture only the essential parts and often contain few causes to explain an effect. In our thesis, we address the problem of making RL agents understandable by humans. In addition to the safety concerns, the quest to mimic human-like reasoning is of general scientific interest, as it sheds light on the easy problem of consciousness. The problem of providing interpretable and simple causal explanations of agent's behavior is connected to the problem of learning good state representations. If we lack such a representation, any reasoning algorithm's outputs would be useless for interpretability, since even the "referents" of the "thoughts" of such a system would be obscure to us. One way to define simplicity of causal explanations via the sparsity of the Causal Model that describes the environment: the causal graph has the fewest edges connecting causes to their effects. For example, a model for choosing the restaurant that only depends on the cause "vegan" is simpler and more interpretable than a model that looks at each pixel of a photo of the menu of a restaurant, and possibly relies as well on spurious correlations, such the style of the menu. In this thesis, we propose a framework "CauseOccam" for model-based Reinforcement Learning where the model is regularized for simplicity in terms of sparsity of the causal graph it corresponds to. The framework contains a learned mapping from observations to latent features, and a model predicting latent features at the next time-steps given ones from the current time-step. The latent features are regularized with the sparsity of the model, compared to a more traditional regularization on the features themselves, or via a hand-crafted interpretability loss. To achieve sparsity, we use discrete Bernoulli variables with gradient estimation, and to find the best parameters, we use the primal-dual constrained formulation to achieve a target model quality. The novelty of this work is in learning jointly a sparse causal graph and the representation taking pixels as the input on RL environments. We test this framework on benchmark environments with non-trivial high-dimensional dynamics and show that it can uncover the causal graph with the fewest edges in the latent space. We describe the implications of our work for developing priors enforcing interpretability.