Data-Driven, Personalized Usable Privacy

We live in the "inverse-privacy" world, where service providers derive insights from users' data that the users do not even know about. This has been fueled by the advancements in machine learning technologies, which allowed providers to go beyond the superficial analysis of users' transactions to the deep inspection of users' content. Users themselves have been facing several problems in coping with this widening information discrepancy. Although the interfaces of apps and websites are generally equipped with privacy indicators (e.g., permissions, policies, ...), this has not been enough to create the counter-effect. We particularly identify three of the gaps that hindered the effectiveness and usability of privacy indicators: - Scale Adaptation: The scale at which service providers are collecting data has been growing on multiple fronts. Users, on the other hand, have limited time, effort, and technological resources to cope with this scale. - Risk Communication: Although providers utilize privacy indicators to announce what and (less often) why they need particular pieces of information, they rarely relay what can be potentially inferred from this data. Without this knowledge, users are less equipped to make informed decisions when they sign in to a site or install an application. - Language Complexity: The information practices of service providers are buried in complex, long privacy policies. Generally, users do not have the time and sometimes the skills to decipher such policies, even when they are interested in knowing particular pieces of it. In this thesis, we approach usable privacy from a data perspective. Instead of static privacy interfaces that are obscure, recurring, or unreadable, we develop techniques that bridge the understanding gap between users and service providers. Towards that, we make the following contributions: - Crowdsourced, data-driven privacy decision-making: In an effort to combat the growing scale of data exposure, we consider the context of files uploaded to cloud services. We propose C3P, a framework for automatically assessing the sensitivity of files, thus enabling realtime, fine-grained policy enforcement on top of unstructured data. - Data-driven app privacy indicators: We introduce PrivySeal, which involves a new paradigm of dynamic, personalized app privacy indicators that bridge the risk under- standing gap between users and providers. Through PrivySeal's online platform, we also study the emerging problem of interdependent privacy in the context of cloud apps and provide a usable privacy indicator to mitigate it. - Automated question answering about privacy practices: We introduce PriBot, the first automated question-answering system for privacy policies, which allows users to pose their questions about the privacy practices of any company with their own language. Through a user study, we show its effectiveness at achieving high accuracy and relevance for users, thus narrowing the complexity gap in navigating privacy policies. A core aim of this thesis is paving the road for a future where privacy indicators are not bound by a specific medium or pre-scripted wording. We design and develop techniques that enable privacy to be communicated effectively in an interface that is approachable to the user. For that, we go beyond textual interfaces to enable dynamic, visual, and hands-free privacy interfaces that are fit for the variety of emerging technologies.

Time- and Space-Efficient Spatial Data Analytics

Mirjana Pavlovic

Advances in data acquisition technologies and supercomputing for large-scale simulations have led to an exponential growth in the volume of spatial data. This growth is accompanied by an increase in data complexity, such as spatial density, but also by more varied data distributions. As data evolves, so do the needs of applications. Recently, we notice a shift from predefined to ad-hoc workloads, as a result of the recent data exploration trend among data-driven applications. At the same time, given the massive volume of data, it has become imperative to use computational and storage resources efficiently, where efficiency requirements typically vary across applications. In this thesis, we show that traditional spatial data management techniques underperform as data size and complexity increase: they waste both computational and storage resources. They are also inefficient in supporting ad-hoc workloads. To achieve time- and space-efficiency, we design spatial data management algorithms and storage layouts that leverage and adapt to data characteristics and workload access patterns. In particular, we revisit the design of spatial join algorithms, indexing techniques and point cloud data management solutions. First, we propose data-aware spatial joins that leverage and adapt to dataset characteristics to avoid wasting computational resources and achieve time-efficiency on non-uniform data distributions. GIPSY is designed to efficiently join two datasets with contrasting densities. GIPSY uses the sparser dataset to guide the join process and therefore, by leveraging dataset characteristics, selectively retrieves and joins only the data needed. TRANSFORMERS achieves robust performance and time-efficiency on non-uniform data distributions, by adapting to dataset characteristics. It detects local variations in distributions and adapts the join strategy and data layout to local dataset characteristics at run-time. We next introduce incremental indexing approaches that take into account workload access patterns. This way, they minimize the data-to-insight time and avoid unnecessary preprocessing costs, substantially accelerating the exploratory analysis of spatial data. Incremental indexes are built as a side-effect of query execution and only for the parts of the data queried. Space Odyssey is designed for exploratory analyses of multiple spatial datasets that reside on disk. It takes advantage of workload access patterns to incrementally index the datasets and optimize accesses to parts frequently queried together. QUASII supports spatial data exploration in main memory. QUASII reduces the data-to-insight time and curbs the cost of incremental indexing, by gradually and partially sorting the data, while simultaneously producing a data-oriented hierarchical structure. Finally, we propose a time- and space-efficient solution to storing and managing point cloud data in main memory column-store database management systems. Our approach leverages point cloud data properties to employ dictionary-based compression in the spatial data management domain and enhances it with indexing capabilities by using space-filling curves. The proposed scheme also represents a partitioning strategy. It is a middle ground between data- and space-oriented partitioning, accounting for the data distribution, while preserving the simplicity of grid-like structures.

EPFL2019

Multi-Centroid Hyperdimensional Computing Approach for Epileptic Seizure Detection

David Atienza Alonso, Una Pale, Tomas Teijeiro Campo

Long-term monitoring of patients with epilepsy presents a challenging problem from the engineering perspective of real-time detection and wearable devices design. It requires new solutions that allow continuous unobstructed monitoring and reliable detection and prediction of seizures. A high variability in the electroencephalogram (EEG) patterns exists among people, brain states, and time instances during seizures, but also during non-seizure periods. This makes epileptic seizure detection very challenging, especially if data is grouped under only seizure (ictal) and non-seizure (inter-ictal) labels. Hyperdimensional (HD) computing, a novel machine learning approach, comes in as a promising tool. However, it has certain limitations when the data shows a high intra-class variability. Therefore, in this work, we propose a novel semi-supervised learning approach based on a multi-centroid HD computing. The multi-centroid approach allows to have several prototype vectors representing seizure and non-seizure states, which leads to significantly improved performance when compared to a simple single-centroid HD model. Further, real-life data imbalance poses an additional challenge and the performance reported on balanced subsets of data is likely to be overestimated. Thus, we test our multi-centroid approach with three different dataset balancing scenarios, showing that performance improvement is higher for the less balanced dataset. More specifically, up to 14% improvement is achieved on an unbalanced test set with 10 times more non-seizure than seizure data. At the same time, the total number of sub-classes is not significantly increased compared to the balanced dataset. Thus, the proposed multi-centroid approach can be an important element in achieving a high performance of epilepsy detection with real-life data balance or during online learning, where seizures are infrequent.

2022

Textual Explanations and Critiques in Recommendation Systems

Diego Matteo Antognini

Artificial intelligence and machine learning algorithms have become ubiquitous. Although they offer a wide range of benefits, their adoption in decision-critical fields is limited by their lack of interpretability, particularly with textual data. Moreover, with more data available than ever before, it has become increasingly important to explain automated predictions.Generally, users find it difficult to understand the underlying computational processes and interact with the models, especially when the models fail to generate the outcomes or explanations, or both, correctly. This problem highlights the growing need for users to better understand the models' inner workings and gain control over their actions. This dissertation focuses on two fundamental challenges of addressing this need. The first involves explanation generation: inferring high-quality explanations from text documents in a scalable and data-driven manner. The second challenge consists in making explanations actionable, and we refer to it as critiquing. This dissertation examines two important applications in natural language processing and recommendation tasks.First, we present two models for extracting high-quality and personalized justifications from text documents under the paradigm of selective rationalization - a binary selection of input features from which the model computes the outcome. We propose the concept of multi-dimensional rationales and emphasize that a single overall selection does not accurately capture the multi-faceted nature of useful rationales. The first method relies on multi-task learning and learns the rationales in an unsupervised manner, with no prior in the data. Inspired by the role of concept-based thinking in human reasoning, our second model is a generalization: it assumes that only one label is observed. We empirically demonstrate the efficiency of the models in terms of explanation and predictive performance as well as their scalability. It opens new doors for applications, such as recommendation or summarization.Second, we use the high-level idea of multi-dimensional rationalization to mine massive corpora and construct large personalized justification datasets. We show that human users significantly prefer our explanations over those produced by state-of-the-art methods.Last, we propose two conversational explainable recommendation systems. The first explains a user rating by inferring a set of keyphrases. Conditioned on them, it generates an abstractive personal justification. We allow users to interact iteratively with the explanations and refine the recommendation through an unsupervised critiquing method. Our second model is based on multimodal modeling and self-supervision. It enables fast and efficient multi-step critiquing. Using real-world datasets, we show that our models exhibit better performance than state-of-the-art models in terms of recommendation, explanation, and critiquing.Overall, we demonstrate that interpretability does not come at the cost of reduced performance in two consequential applications. Our framework is applicable to other fields as well. This dissertation presents an effective means of closing the gap between promise and practice in artificial intelligence.

EPFL2022