Dans cette thèse, nous étudions des problèmes de prise de décisions séquentielles dans lesquels, pour chacune de ses décisions, l'apprenant reçoit une information qu'il utilise pour guider ses décisions futures. Pour aller au-delà du retour d’information conventionnel tel qu'il a été bien étudié pour des problèmes de prise de décision séquentielle tels que les bandits multi-bras, nous considérons des formes de retour d’information partielle motivées par des applications pratiques.En premier, nous considérons le problème des bandits duellistes, dans lequel l'apprenant sélectionne deux actions à chaque pas de temps et reçoit en retour une information relative (i.e. de préférence) entre les valeurs instantanées de ces deux actions.En particulier, nous proposons un algorithme optimal qui permet à l'apprenant d'obtenir un regret cumulatif quasi-optimal (le regret est la différence entre la récompense cumulative optimale et la récompense cumulative constatée de l’apprenant). Dans un second temps, nous considérons le problème des bandits corrompus, dans lequel un processus de corruption stochastique perturbe le retour d’information. Pour ce problème aussi, nous concevons des algorithmes pour obtenir un regret cumulatif asymptotiquement optimal. En outre, nous examinons la relation entre ces deux problèmes dans le cadre du monitoring partiel qui est un paradigme générique pour la prise de décision séquentielle avec retour d'information partielle. / The multi-armed bandit (MAB) problem is a mathematical formulation of the exploration-exploitation trade-off inherent to reinforcement learning, in which the learner chooses an action (symbolized by an arm) from a set of available actions in a sequence of trials in order to maximize their reward. In the classical MAB problem, the learner receives absolute bandit feedback i.e. it receives as feedback the reward of the arm it selects. In many practical situations however, different kind of feedback is more readily available. In this thesis, we study two of such kinds of feedbacks, namely, relative feedback and corrupt feedback.The main practical motivation behind relative feedback arises from the task of online ranker evaluation. This task involves choosing the optimal ranker from a finite set of rankers using only pairwise comparisons, while minimizing the comparisons between sub-optimal rankers. This is formalized by the MAB problem with relative feedback, in which the learner selects two arms instead of one and receives the preference feedback. We consider the adversarial formulation of this problem which circumvents the stationarity assumption over the mean rewards for the arms. We provide a lower bound on the performance measure for any algorithm for this problem. We also provide an algorithm called "Relative Exponential-weight algorithm for Exploration and Exploitation" with performance guarantees. We present a thorough empirical study on several information retrieval datasets that confirm the validity of these theoretical results.The motivating theme behind corrupt feedback is that the feedback the learner receives is a corrupted form of the corresponding reward of the selected arm. Practically such a feedback is available in the tasks of online advertising, recommender systems etc. We consider two goals for the MAB problem with corrupt feedback: best arm identification and exploration-exploitation. For both the goals, we provide lower bounds on the performance measures for any algorithm. We also provide various algorithms for these settings. The main contribution of this module is the algorithms "KLUCB-CF" and "Thompson Sampling-CF" which asymptotically attain the best possible performance. We present experimental results to demonstrate the performance of these algorithms. We also show how this problem setting can be used for the practical application of enforcing differential privacy.

Health monitoring devices, such as Fitbit, are gaining popularity both as wellness tools and as a source of information for healthcare decisions. Predicting such wellness goals accurately is critical for the users to make informed lifestyle choices. The core objective of this thesis is to design and implement such a system that takes energy consumption and privacy into account. This research is modelled as a time-series forecasting problem that makes use of Spiking Neural Networks (SNNs) due to their proven energy-saving capabilities. Thanks to their design that closely mimics natural neural networks (such as the brain), SNNs have the potential to significantly outperform classic Artificial Neural Networks in terms of energy consumption and robustness. In order to prove our hypotheses, a previous research by Sonia et al. [1] in the same domain and with the same dataset is used as our starting point, where a private forecasting system using Long short-term memory (LSTM) is designed and implemented. Their study also implements and evaluates a clustering federated learning approach, which fits well the highly distributed data. The results obtained in their research act as a baseline to compare our results in terms of accuracy, training time, model size and estimated energy consumed. Our experiments show that Spiking Neural Networks trades off accuracy (2.19x, 1.19x, 4.13x, 1.16x greater Root Mean Square Error (RMSE) for macronutrients, calories burned, resting heart rate, and active minutes respectively), to grant a smaller model (19% less parameters an 77% lighter in memory) and a 43% faster training. Our model is estimated to consume 3.36μJ per inference, which is much lighter than traditional Artificial Neural Networks (ANNs) [2]. The data recorded by health monitoring devices is vastly distributed in the real-world. Moreover, with such sensitive recorded information, there are many possible implications to consider. For these reasons, we apply the clustering federated learning implementation [1] to our use-case. However, it can be challenging to adopt such techniques since it can be difficult to learn from data sequences that are non-regular. We use a two-step streaming clustering approach to classify customers based on their eating and exercise habits. It has been shown that training different models for each group of users is useful, particularly in terms of training time; however this is strongly dependent on the cluster size. Our experiments conclude that there is a decrease in error and training time if the clusters contain enough data to train the models. Finally, this study addresses the issue of data privacy by using state of-the-art differential privacy. We apply e-differential privacy to both our baseline model (trained on the whole dataset) and our federated learning based approach. With a differential privacy of ∈= 0.1 our experiments report an increase in the measured average error (RMSE) of only 25%. Specifically, +23.13%, 25.71%, +29.87%, 21.57% for macronutrients (grams), calories burned (kCal), resting heart rate (beats per minute (bpm), and minutes (minutes) respectively. / Hälsoövervakningsenheter, som Fitbit, blir allt populärare både som friskvårdsverktyg och som informationskälla för vårdbeslut. Att förutsäga sådana välbefinnandemål korrekt är avgörande för att användarna ska kunna göra välgrundade livsstilsval. Kärnmålet med denna avhandling är att designa och implementera ett sådant system som tar hänsyn till energiförbrukning och integritet. Denna forskning är modellerad som ett tidsserieprognosproblem som använder sig av SNNs på grund av deras bevisade energibesparingsförmåga. Tack vare deras design som nära efterliknar naturliga neurala nätverk (som hjärnan) har SNNs potentialen att avsevärt överträffa klassiska artificiella neurala nätverk när det gäller energiförbrukning och robusthet. För att bevisa våra hypoteser har en tidigare forskning av Sonia et al. [1] i samma domän och med samma dataset används som utgångspunkt, där ett privat prognossystem som använder LSTM designas och implementeras. Deras studie implementerar och utvärderar också en klustringsstrategi för federerad inlärning, som passar väl in på den mycket distribuerade data. Resultaten som erhållits i deras forskning fungerar som en baslinje för att jämföra våra resultat vad gäller noggrannhet, träningstid, modellstorlek och uppskattad energiförbrukning. Våra experiment visar att Spiking Neural Networks byter ut precision (2,19x, 1,19x, 4,13x, 1,16x större RMSE för makronäringsämnen, förbrända kalorier, vilopuls respektive aktiva minuter), för att ge en mindre modell ( 19% mindre parametrar, 77% lättare i minnet) och 43% snabbare träning. Vår modell beräknas förbruka 3, 36μJ, vilket är mycket lättare än traditionella ANNs [2]. Data som registreras av hälsoövervakningsenheter är enormt spridda i den verkliga världen. Dessutom, med sådan känslig registrerad information finns det många möjliga konsekvenser att överväga. Av dessa skäl tillämpar vi klustringsimplementeringen för federerad inlärning [1] på vårt användningsfall. Det kan dock vara utmanande att använda sådana tekniker eftersom det kan vara svårt att lära sig av datasekvenser som är oregelbundna. Vi använder en tvåstegs streaming-klustringsmetod för att klassificera kunder baserat på deras mat- och träningsvanor. Det har visat sig att det är användbart att träna olika modeller för varje grupp av användare, särskilt när det gäller utbildningstid; detta är dock starkt beroende av klustrets storlek. Våra experiment drar slutsatsen att det finns en minskning av fel och träningstid om klustren innehåller tillräckligt med data för att träna modellerna. Slutligen tar denna studie upp frågan om datasekretess genom att använda den senaste differentiell integritet. Vi tillämpar e-differentiell integritet på både vår baslinjemodell (utbildad på hela datasetet) och vår federerade inlärningsbaserade metod. Med en differentiell integritet på ∈= 0.1 rapporterar våra experiment en ökning av det uppmätta medelfelet (RMSE) på endast 25%. Specifikt +23,13%, 25,71%, +29,87%, 21,57% för makronäringsämnen (gram), förbrända kalorier (kCal), vilopuls (bpm och minuter (minuter).

La publication de données soucieuse du respect de la vie privée est au coeur des préoccupations des organisations qui souhaitent publier leurs données. Un nombre croissant d'entreprises et d'organismes collectent et publient des données à caractère personnel pour diverses raisons (études démographiques, recherche médicale,...). Selon ces cas, celui qui publie les données fait face au dilemme suivant : comment permettre à un tiers l'analyse de ces données tout en évitant de divulguer des informations trop sensibles, relatives aux individus concernés? L'enjeu est donc la capacité à publier des jeux de données en maîtrisant ce risque de divulgation, c.a.d. de traiter l'opposition entre deux critères : d'un côté, on souhaite garantir la préservation de la confidentialité sur des données personnelles et, d'autre part, on souhaite préserver au maximum l'utilité du jeu de données pour ceux qui l'exploiteraient (notamment, des chercheurs). Dans ce travail, nous cherchons d'abord à élaborer plusieurs notions d'anonymisation des données selon plusieurs contextes. Nous montrons que les index spatiaux sont extrêmement efficaces dans le cadre de la publication de données, en raison de leur capacité à passer à l'échelle. Une évaluation empirique approfondie révèle qu'il est possible de diffuser des données de grande qualité et préservant un certain niveau de confidentialité dans les données. Il est de plus possible de traiter efficacement de très grands jeux de données en grandes dimensions et cette méthode peut être étendue à un niveau de confidentialité plus fort (differential privacy). Par ailleurs, la publication séquentielle de données (mise à jour du jeu de données) est cruciale dans un grand nombre d'applications. Nous proposons une technique menant à bien cette tâche, garantissant à la fois une forte confidentialité des données et une très bonne préservation de leur utilité.

Today we are surrounded by IOT devices that constantly generate different kinds of data about its environment and its users. Much of this data could be useful for different research purposes and development, but a lot of this collected data is privacy-sensitive for the individual person. To protect the individual's privacy, we have data protection laws. But these restrictions by laws also dramatically reduce the amount of data available for research and development. Therefore it would be beneficial if we could find a work around that respects people's privacy without breaking the laws while still maintaining the usefulness of data. The purpose of this thesis is to show how we can generate privacy-aware data from a dataset by using Generative Adversarial Networks (GANS) and Differential Privacy (DP), that maintains data utility. This is useful because it allows for the sharing of privacy-preserving data, so that the data can be used in research and development with concern for privacy. GANS is used for generating synthetic data. DP is an anonymization technique of data. With the combination of these two techniques, we generate synthetic-privacy-aware data from an existing open-source Fitbit dataset. The specific type of GANS model that is used is called CTGAN and differential privacy is achieved with the help of gaussian noise. The results from the experiments performed show many similarities between the original dataset and the experimental datasets. The experiments performed very well at the Kolmogorov Smirnov test, with the lowest P-value of all experiments sitting at 0.92. The conclusion that is drawn is that this is another promising methodology for creating privacy-aware-synthetic data, that maintains reasonable data utility while still utilizing DP techniques to achieve data privacy.

Dans les applications de traitement du langage naturel (NLP), la formation d’un modèle efficace nécessite souvent une quantité massive de données. Cependant, les données textuelles dans le monde réel sont dispersées dans différentes institutions ou appareils d’utilisateurs. Leur partage direct avec le fournisseur de services NLP entraîne d’énormes risques pour la confidentialité, car les données textuelles contiennent souvent des informations sensibles, entraînant une fuite potentielle de la confidentialité. Un moyen typique de protéger la confidentialité consiste à privatiser directement le texte brut et à tirer parti de la confidentialité différentielle (DP) pour protéger le texte à un niveau de protection de la confidentialité quantifiable. Par ailleurs, la protection des résultats de calcul intermédiaires via un mécanisme de privatisation de texte aléatoire est une autre solution disponible. Cependant, les mécanismes existants de privatisation des textes ne permettent pas d’obtenir un bon compromis entre confidentialité et utilité en raison de la difficulté intrinsèque de la protection de la confidentialité des textes. Leurs limitations incluent principalement les aspects suivants: (1) ces mécanismes qui privatisent le texte en appliquant la notion de dχ-privacy ne sont pas applicables à toutes les métriques de similarité en raison des exigences strictes; (2) ils privatisent chaque jeton (mot) dans le texte de manière égale en fournissant le même ensemble de sorties excessivement grand, ce qui entraîne une surprotection; (3) les méthodes actuelles ne peuvent garantir la confidentialité que pour une seule étape d’entraînement/ d’inférence en raison du manque de composition DP et de techniques d’amplification DP. Le manque du compromis utilité-confidentialité empêche l’adoption des mécanismes actuels de privatisation du texte dans les applications du monde réel. Dans ce mémoire, nous proposons deux méthodes à partir de perspectives différentes pour les étapes d’apprentissage et d’inférence tout en ne requérant aucune confiance de sécurité au serveur. La première approche est un mécanisme de privatisation de texte privé différentiel personnalisé (CusText) qui attribue à chaque jeton d’entrée un ensemble de sortie personnalisé pour fournir une protection de confidentialité adaptative plus avancée au niveau du jeton. Il surmonte également la limitation des métriques de similarité causée par la notion de dχ-privacy, en adaptant le mécanisme pour satisfaire ϵ-DP. En outre, nous proposons deux nouvelles stratégies de 5 privatisation de texte pour renforcer l’utilité du texte privatisé sans compromettre la confidentialité. La deuxième approche est un modèle Gaussien privé différentiel local (GauDP) qui réduit considérablement le volume de bruit calibrée sur la base d’un cadre avancé de comptabilité de confidentialité et améliore ainsi la précision du modèle en incorporant plusieurs composants. Le modèle se compose d’une couche LDP, d’algorithmes d’amplification DP de sous-échantillonnage et de sur-échantillonnage pour l’apprentissage et l’inférence, et d’algorithmes de composition DP pour l’étalonnage du bruit. Cette nouvelle solution garantit pour la première fois la confidentialité de l’ensemble des données d’entraînement/d’inférence. Pour évaluer nos mécanismes de privatisation de texte proposés, nous menons des expériences étendues sur plusieurs ensembles de données de différents types. Les résultats expérimentaux démontrent que nos mécanismes proposés peuvent atteindre un meilleur compromis confidentialité-utilité et une meilleure valeur d’application pratique que les méthodes existantes. En outre, nous menons également une série d’études d’analyse pour explorer les facteurs cruciaux de chaque composant qui pourront fournir plus d’informations sur la protection des textes et généraliser d’autres explorations pour la NLP préservant la confidentialité. / In Natural Language Processing (NLP) applications, training an effective model often requires a massive amount of data. However, text data in the real world are scattered in different institutions or user devices. Directly sharing them with the NLP service provider brings huge privacy risks, as text data often contains sensitive information, leading to potential privacy leakage. A typical way to protect privacy is to directly privatize raw text and leverage Differential Privacy (DP) to protect the text at a quantifiable privacy protection level. Besides, protecting the intermediate computation results via a randomized text privatization mechanism is another available solution. However, existing text privatization mechanisms fail to achieve a good privacy-utility trade-off due to the intrinsic difficulty of text privacy protection. The limitations of them mainly include the following aspects: (1) those mechanisms that privatize text by applying dχ-privacy notion are not applicable for all similarity metrics because of the strict requirements; (2) they privatize each token in the text equally by providing the same and excessively large output set which results in over-protection; (3) current methods can only guarantee privacy for either the training/inference step, but not both, because of the lack of DP composition and DP amplification techniques. Bad utility-privacy trade-off performance impedes the adoption of current text privatization mechanisms in real-world applications. In this thesis, we propose two methods from different perspectives for both training and inference stages while requiring no server security trust. The first approach is a Customized differentially private Text privatization mechanism (CusText) that assigns each input token a customized output set to provide more advanced adaptive privacy protection at the token-level. It also overcomes the limitation for the similarity metrics caused by dχ-privacy notion, by turning the mechanism to satisfy ϵ-DP. Furthermore, we provide two new text privatization strategies to boost the utility of privatized text without compromising privacy. The second approach is a Gaussian-based local Differentially Private (GauDP) model that significantly reduces calibrated noise power adding to the intermediate text representations based on an advanced privacy accounting framework and thus improves model accuracy by incorporating several components. The model consists of an LDP-layer, sub-sampling and up-sampling DP amplification algorithms 7 for training and inference, and DP composition algorithms for noise calibration. This novel solution guarantees privacy for both training and inference data. To evaluate our proposed text privatization mechanisms, we conduct extensive experiments on several datasets of different types. The experimental results demonstrate that our proposed mechanisms can achieve a better privacy-utility trade-off and better practical application value than the existing methods. In addition, we also carry out a series of analyses to explore the crucial factors for each component which will be able to provide more insights in text protection and generalize further explorations for privacy-preserving NLP.

This thesis investigates the effectiveness of Tabular Variational Auto Encoders (TVAEs) in generating high-quality synthetic tabular data and assesses their compliance with differential privacy principles. The study shows that while TVAEs are better than VAEs at generating synthetic data that faithfully reproduces the distribution of real data as measured by the Synthetic Data Vault (SDV) metrics, the latter does not guarantee that the synthetic data is up to the task in practical industrial applications. In particular, models trained on TVAE-generated data from the Creditcards dataset are ineffective. The author also explores various optimisation methods on TVAE, such as Gumbel Max Trick, Drop Out (DO) and Batch Normalization, while pointing out that techniques frequently used to improve two-dimensional TVAE, such as Kullback–Leibler Warm-Up and B Disentanglement, are not directly transferable to the one-dimensional context. However, differential privacy to TVAE was not implemented due to time constraints and inconclusive results. The study nevertheless highlights the benefits of stabilising training with the Differential Privacy - Stochastic Gradient Descent (DP-SGD), as with a dropout, and the existence of an optimal equilibrium point between the constraints of differential privacy and the number of training epochs in the model. / Denna avhandling undersöker hur effektiva Tabular Variational AutoEncoders (TVAE) är när det gäller att generera högkvalitativa syntetiska tabelldata och utvärderar deras överensstämmelse med differentierade integritetsprinciper. Studien visar att även om TVAE är bättre än VAE på att generera syntetiska data som troget återger fördelningen av verkliga data mätt med Synthetic Data Vault (SDV), garanterar det senare inte att de syntetiska data är upp till uppgiften i praktiska industriella tillämpningar. I synnerhet är modeller som tränats på TVAE-genererade data från Creditcards-datasetet ineffektiva. Författaren undersöker också olika optimeringsmetoder för TVAE, såsom Gumbel Max Trick, DO och Batch Normalization, samtidigt som han påpekar att tekniker som ofta används för att förbättra tvådimensionell TVAE, såsom Kullback-Leibler Warm-Up och B Disentanglement, inte är direkt överförbara till det endimensionella sammanhanget. På grund av tidsbegränsningar och redan ofullständiga resultat implementerades dock inte differentierad integritet för TVAE. Studien belyser ändå fördelarna med att stabilisera träningen med Differential Privacy - Stochastic Gradient Descent (DP-SGD), som med en drop-out, och förekomsten av en optimal jämviktspunkt mellan begränsningarna för differential privacy och antalet träningsepoker i modellen.

Nous présentons une nouvelle approche de préservation de la vie privée pour améliorer l’équité des éléments dans les systèmes de classement. Nous utilisons des techniques de post-traitement dans un environnement de recommandation multipartite afin d’équilibrer l’équité et la protection de la vie privée pour les producteurs et les consommateurs. Notre méthode utilise des serveurs de calcul multipartite sécurisés (MPC) et une confidentialité différentielle (DP) pour maintenir la confidentialité des utilisateurs tout en atténuant l’injustice des éléments sans compromettre l’utilité. Les utilisateurs soumettent leurs données sous forme de partages secrets aux serveurs MPC, et tous les calculs sur ces données restent cryptés. Nous évaluons notre approche à l’aide d’ensembles de données du monde réel, tels qu’Amazon Digital Music, Book Crossing et MovieLens-1M, et analysons les compromis entre confidentialité, équité et utilité. Notre travail encourage une exploration plus approfondie de l’intersection de la confidentialité et de l’équité dans les systèmes de recommandation, jetant les bases de l’intégration d’autres techniques d’amélioration de la confidentialité afin d’optimiser l’exécution et l’évolutivité pour les applications du monde réel. Nous envisageons notre approche comme un tremplin vers des solutions de bout en bout préservant la confidentialité et promouvant l’équité dans des environnements de recommandation multipartites. / We present a novel privacy-preserving approach to enhance item fairness in ranking systems. We employ post-processing techniques in a multi-stakeholder recommendation environment in order to balance fairness and privacy protection for both producers and consumers. Our method utilizes secure multi-party computation (MPC) servers and differential privacy (DP) to maintain user privacy while mitigating item unfairness without compromising utility. Users submit their data as secret shares to MPC servers, and all calculations on this data remain encrypted. We evaluate our approach using real-world datasets, such as Amazon Digital Music, Book Crossing, and MovieLens-1M, and analyze the trade-offs between privacy, fairness, and utility. Our work encourages further exploration of the intersection of privacy and fairness in recommender systems, laying the groundwork for integrating other privacy-enhancing techniques to optimize runtime and scalability for real-world applications. We envision our approach as a stepping stone towards end-to-end privacy-preserving and fairness-promoting solutions in multi-stakeholder recommendation environments.

This paper aims to determine the value of the product based on the mathematical concept of Differential Privacy, by assessing the value of the business opportunities it enables and the value of the possible GDPR-fines it prevents. To delimit the scope of the research the analysis will focus on what the value of personal data is for companies within the mobility sector. Mobility is a cross-industrial sector consisting of companies within connectivity-technology, transportation, and automotive. The method used to assess the final value of anonymizing personal data, such as consumer data, using a DP-solution (meaning, an implementation of the theory) has consisted of both quantitative and qualitative analysis. The quantitative analysis aims to assess the ‘Cost of Risk’ for mobility companies that are exposed to personal integrity regulation due to data processing. To further conclude the true cost of the financial impact caused by getting fined for infringing on privacy regulation because of unlawful data processing is done through a complementary qualitative assessment. Lastly, the 'Opportunity Cost', or rather the cost of missed financial opportunities, is determined qualitatively for a case study company within Sweden’s mobility ecosystem to conclude the overall value of a DP-solution for a specific company. The final product of this research paper is to provide a framework assessing the total value, for specifically companies in the mobility sector, of implementing differential privacy solutions. / Syftet med denna uppsats är att fastställa värdet av anonymisering baserat på det matematiska konceptet Differential Privacy, genom att bedöma värdet av de affärsmöjligheter det skapar, samt värdet av de möjliga GDPR- böter det förhindrar. För att avgränsa studiens omfattning består analysen endast av att uppskatta dessa värden för företag inom mobilitetssektorn. Mobilitetssektorn är en tvärindustriell sektor som består av företag inom uppkoppling-, transport- och bilindustrin. Metoden som använts för att ta fram det slutliga värdet av att anonymisera persondata genom en differential privacy lösning, består både av en kvantitativ och en kvalitativ analys. Målet med den kvantitativa analysen är att estimera kostnadsrisken för företag inom mobilitetssektorn som exponeras mot GDPR-böter med avseende på dess datahantering. För att vidare ta reda på den totala finansiella inverkan av sådana böter, kompletteras analysen av en kvalitativ studie, som delvis omfattas av de finansiella möjligheterna ett företag går miste om i en sådan situation. Den kvalitativa analysen består också av en fallstudie av ett svenskt företag inom mobilitetssektorn, med målet att estimera värdet av de affärsmöjligheter som uppstår med hjälp av anonymisering av data. Slutligen är målet med denna uppsats att förse läsaren med att ramverk för att estimera det totala värdet av att implementera differential privacy lösningar i företag inom mobilitetssektorn.

<p dir="ltr">This dissertation advances the field of causal inference by developing methodologies in the face of assumption violations. Traditional causal inference methodologies hinge on a core set of assumptions, which are often violated in the complex landscape of modern experiments and observational studies. This dissertation proposes novel methodologies designed to address the challenges posed by single or multiple assumption violations. By applying these innovative approaches to real-world datasets, this research uncovers valuable insights that were previously inaccessible with existing methods. </p><p><br></p><p dir="ltr">First, three significant sources of complications in causal inference that are increasingly of interest are interference among individuals, nonadherence of individuals to their assigned treatments, and unintended missing outcomes. Interference exists if the outcome of an individual depends not only on its assigned treatment, but also on the assigned treatments for other units. It commonly arises when limited controls are placed on the interactions of individuals with one another during the course of an experiment. Treatment nonadherence frequently occurs in human subject experiments, as it can be unethical to force an individual to take their assigned treatment. Clinical trials, in particular, typically have subjects that do not adhere to their assigned treatments due to adverse side effects or intercurrent events. Missing values also commonly occur in clinical studies. For example, some patients may drop out of the study due to the side effects of the treatment. Failing to account for these considerations will generally yield unstable and biased inferences on treatment effects even in randomized experiments, but existing methodologies lack the ability to address all these challenges simultaneously. We propose a novel Bayesian methodology to fill this gap. </p><p><br></p><p dir="ltr">My subsequent research further addresses one of the limitations of the first project: a set of assumptions about interference structures that may be too restrictive in some practical settings. We introduce a concept of the ``degree of interference" (DoI), a latent variable capturing the interference structure. This concept allows for handling arbitrary, unknown interference structures to facilitate inference on causal estimands. </p><p><br></p><p dir="ltr">While randomized experiments offer a solid foundation for valid causal analysis, people are also interested in conducting causal inference using observational data due to the cost and difficulty of randomized experiments and the wide availability of observational data. Nonetheless, using observational data to infer causality requires us to rely on additional assumptions. A central assumption is that of \emph{ignorability}, which posits that the treatment is randomly assigned based on the variables (covariates) included in the dataset. While crucial, this assumption is often debatable, especially when treatments are assigned sequentially to optimize future outcomes. For instance, marketers typically adjust subsequent promotions based on responses to earlier ones and speculate on how customers might have reacted to alternative past promotions. This speculative behavior introduces latent confounders, which must be carefully addressed to prevent biased conclusions. </p><p dir="ltr">In the third project, we investigate these issues by studying sequences of promotional emails sent by a US retailer. We develop a novel Bayesian approach for causal inference from longitudinal observational data that accommodates noncompliance and latent sequential confounding. </p><p><br></p><p dir="ltr">Finally, we formulate the causal inference problem for the privatized data. In the era of digital expansion, the secure handling of sensitive data poses an intricate challenge that significantly influences research, policy-making, and technological innovation. As the collection of sensitive data becomes more widespread across academic, governmental, and corporate sectors, addressing the complex balance between making data accessible and safeguarding private information requires the development of sophisticated methods for analysis and reporting, which must include stringent privacy protections. Currently, the gold standard for maintaining this balance is Differential privacy. </p><p dir="ltr">Local differential privacy is a differential privacy paradigm in which individuals first apply a privacy mechanism to their data (often by adding noise) before transmitting the result to a curator. The noise for privacy results in additional bias and variance in their analyses. Thus, it is of great importance for analysts to incorporate the privacy noise into valid inference.</p><p dir="ltr">In this final project, we develop methodologies to infer causal effects from locally privatized data under randomized experiments. We present frequentist and Bayesian approaches and discuss the statistical properties of the estimators, such as consistency and optimality under various privacy scenarios.</p>

Modern health self-monitoring devices and applications, such as Fitbit and MyFitnessPal, empower users to take concrete actions and set fitness and lifestyle goals based on their recorded trends and statistics. Predicting such trends is beneficial in the road of achieving long-time targets, as the individuals can adjust their diets and habits at any point to guarantee success. The design and implementation of such a system, which also respects user privacy, is the main objective of our work.This application is modelled as a time-series forecasting problem. Given the historical data of users, we aim to predict their eating and lifestyle habits in real-time. We apply the federated learning paradigm to our use-case be- cause of the highly-distributed nature of our data and the privacy concerns of such sensitive recorded information. However, federated learning from het- erogeneous sequences of data can be challenging, as even state-of-the-art ma- chine learning techniques for time-series forecasting can encounter difficulties when learning from very irregular data sequences. Specifically, in the pro- posed healthcare scenario, the machine learning algorithms might fail to cater to users with unique dietary patterns.In this work, we implement a two-step streaming clustering mechanism and group clients that exhibit similar eating and fitness behaviours. The con- ducted experiments prove that learning federatively in this context can achieve very high prediction accuracy, as our predictions are no more than 0.025% far from the ground truth value with respect to the range of each feature. Training separate models for each group of users is shown to be beneficial, especially in terms of the training time, but it is highly dependent on the parameters used for the models and the training process. Our experiments conclude that the configuration used for the general federated model cannot be applied to the clusters of data. However, a decrease in prediction error of more than 45% can be achieved, given the parameters are optimized for each case.Lastly, this work tackles the problem of data privacy by applying state-of- the-art differential privacy techniques. Our empirical study shows that noising the gradients sent to the server is unsuitable for small datasets and cancels out the benefits obtained by prior users’ clustering. On the other hand, noising the training data achieves remarkable results, obtaining a differential privacy level corresponding to an epsilon value of 0.1 with an increase in the observed mean absolute error by a factor of only 0.21. / Moderna apparater och applikationer för självövervakning av hälsa, som Fitbit och MyFitnessPal, ger användarna möjlighet att vidta konkreta åtgärder och sätta fitness- och livsstilsmål baserat på deras dokumenterade trender och statistik. Att förutsäga sådana trender är fördelaktigt för att uppnå långtidsmål, eftersom individerna kan anpassa sina dieter och vanor när som helst för att garantera framgång.Utformningen och implementeringen av ett sådant system, som dessutom respekterar användarnas integritet, är huvudmålet för vårt arbete. Denna appli- kation är modellerad som ett tidsserieprognosproblem. Med avseende på an- vändarnas historiska data är målet att förutsäga deras matvanor och livsstilsva- nor i realtid. Vi tillämpar det federerade inlärningsparadigmet på vårt använd- ningsfall på grund av den mycket distribuerade karaktären av vår data och in- tegritetsproblemen för sådan känslig bokförd information. Federerade lärande från heterogena datasekvenser kan emellertid vara utmanande, eftersom även de modernaste maskininlärningstekniker för tidsserieprognoser kan stöta på svårigheter när de lär sig från mycket oregelbundna datasekvenser. Specifikt i det föreslagna sjukvårdsscenariot kan maskininlärningsalgoritmerna misslyc- kas med att förse användare med unika dietmönster.I detta arbete implementerar vi en tvåstegsströmmande klustermekanism och grupperar användare som uppvisar liknande ät- och fitnessbeteenden. De genomförda experimenten visar att federerade lärande i detta sammanhang kan uppnå mycket hög nogrannhet i förutsägelse, eftersom våra förutsägelser in- te är mer än 0,025% ifrån det sanna värdet med avseende på intervallet för varje funktion. Träning av separata modeller för varje grupp användare visar sig vara fördelaktigt, särskilt gällande träningstiden, men det är mycket be- roende av parametrarna som används för modellerna och träningsprocessen. Våra experiment drar slutsatsen att konfigurationen som används för den all- männa federerade modellen inte kan tillämpas på dataklusterna. Dock kan en minskning av förutsägelsefel på mer än 45% uppnås, givet att parametrarna är optimerade för varje fall.Slutligen hanteras problemet med datasekretess genom att tillämpa bästa tillgängliga differentiell integritetsteknik. Vår empiriska studie visar att adde- ra brus till gradienter som skickas till servern är olämpliga för liten data och avbryter fördelarna med tidigare användares kluster. Däremot, genom att ad- dera brus till träningsdata uppnås anmärkningsvärda resultat. En differentierad integritetsnivå motsvarande ett epsilonvärde på 0,1 med en ökning av det ob- serverade genomsnittliga absoluta felet med en faktor på endast 0,21 erhölls.