Upskilling is a fast-growing segment of the education economy [31]. Yet, there is little algorithmic work that focuses on crafting dedicated strategies to reach high-skill mastery. In this paper, we formalize AdUp, an iterative upskilling problem that combines mastery learning [49] and Zone of Proximal Development [7]. We extend our previous work [9] and design two solutions for AdUp: MOO and MAB. MOO is a multiobjective optimization approach that relies on Hill Climbing to adapt the difficulty of recommended tests to three objectives: learner's predicted performance, aptitude, and skill gap. MAB is a meta approach based on Multi-Armed Bandits to learn the best combination of objectives to optimize at each iteration. We show how these solutions are combined with two common learner simulation models: BKT (KT-IDEM) [47] and Item Response Theory (IRT) [53]. Our simulation experiments demonstrate the necessity of leveraging all three objectives and the need to adapt the optimization objectives to the learner's progression ability as MAB offers a higher mastery rate and a better final skill gain than MOO.
The development and adoption of personal data management systems (PDMS) has been fueled by legal and technical means such as smart disclosure, data portability and data altruism. By using a PDMS, individuals can effortlessly gather and share data, generated directly by their devices or as a result of their interactions with companies or institutions. In this context, federated learning appears to be a very promising technology, but it requires secure, reliable, and scalable aggregation protocols to preserve user privacy and account for potential PDMS dropouts. Despite recent significant progress in secure aggregation for federated learning, we still lack a solution suitable for the fully decentralized PDMS context. This paper proposes a family of fully decentralized protocols that are scalable and reliable with respect to dropouts. We focus in particular on the reliability property which is key in a peer-to-peer system wherein aggregators are system nodes and are subject to dropouts in the same way as contributor nodes. We show that in a decentralized setting, reliability raises a tension between the potential completeness of the result and the aggregation cost. We then propose a set of strategies that deal with dropouts and offer different trade-offs between completeness and cost. We extensively evaluate the proposed protocols and show that they cover the design space allowing to favor completeness or cost in all settings.
JSON Schema is an expressive schema language for describing JSON documents which combines structural assertions with Boolean operators, including negation, and is able to express recursion. In typical situations, JSON Schema is used to validate JSON documents to ensure safety of operations processing these JSONdata. In many other situations, it is important to generate JSONdocuments starting from a JSON Schema schema as it allows for testing applications and for assessing interesting properties like scalability of processing engines. While existing work [6,8] addressed the witness generation problem, by discarding the uniqueItems operator, our goal lifts this restriction and addresses the problem of generating multiple instances from a given schema. Since the witness generation problem, which is equivalent to checking the satisfiability in JSON Schema, is known to be EXPTIME-complete even when uniqueItems is absent [ 10], our solution adopts an optimistic approach which trades completeness for soundness and efficiency. From a practical point of view, our approach is very satisfactory since, on most real-life scenarios, it is very efficient and highly precise, as testified by our experiments.
This paper presents an algorithmic family of dynamic topic models called Aligned Neural Topic Models (ANTM), which combine novel data mining algorithms to provide a modular framework for discovering evolving topics. ANTM maintains the temporal continuity of evolving topics by extracting time-aware features from documents using advanced pre-trained Large Language Models (LLMs) and employing an overlapping sliding window algorithm for sequential document clustering. This overlapping sliding window algorithm identifies a different number of topics within each time frame and aligns semantically similar document clusters across time periods. This process captures emerging and fading trends across different periods and allows for a more interpretable representation of evolving topics. Experiments on four distinct datasets show that ANTM outperforms probabilistic dynamic topic models in terms of topic coherence and diversity metrics. Moreover, it improves the scalability and flexibility of dynamic topic models by being accessible and adaptable to different types of algorithms. Additionally, a Python package is developed for researchers and scientists who wish to study the trends and evolving patterns of topics in large-scale textual data.
Spatio-temporal Predictive Queries encompass a spatio-temporal constraint, defining a region, a target variable, and an evaluation metric. The output of such queries presents the future values for the target variable computed by predictive models at each point of the spatio-temporal region. Unfortunately, especially for large spatio-temporal domains with millions of points, training temporal models at each spatial domain point is prohibitive. In this work, we propose a data-driven approach for selecting pre-trained temporal models to be applied at each query point. The chosen approach applies a model to a point according to the training and input time series similarity. The approach avoids training a different model for each domain point, saving model training time. Moreover, it provides a technique to decide on the best-trained model to be applied to a point for prediction. In order to assess the applicability of the proposed strategy, we evaluate a case study for temperature forecasting using historical data and auto-regressive models. Computational experiments show that the proposed approach, compared to the baseline, achieves equivalent predictive performance using a composition of pre-trained models at a fraction of the total computational cost.
Interplanetary File System (IPFS) is a file sharing network relying on a Distributed Hash Table (DHT) to locate data and a BitTorrent-like protocol to exchange blocks between the peers. However, in such a network, all nodes can access information about the files stored or accessed by others. In this article, we use these public pieces of information to try to characterize the IPFS network both in terms of nodes composing it and on the files stored in it. To that end, we set up an IPFS node connected to the public IPFS network and saved all the DHT requests forwarded through it. We show that nodes are mostly located in datacenters and not on the end-users’ computers and therefore that files are often accessed through public gateways. We also show that most files are not replicated and are not accessed frequently (cold data) which can question us about the relevance of using IPFS in such use case scenarios.
The aim of concept weighting in ontologies or in other data graphs is to characterise the importance of each concept in a specific domain, and to determine its selective power. This is particularly useful for data analysis tasks. Existing works on concept weighting mainly exploit either the graph’s structure or the frequency of the concept in the data instances. These works provide concept weights independently form the considered analysis task. We argue that these weights should vary according to the targeted task, and we introduce a neural network based approach which computes concept weights using regression on a customised multi-layered structure. The loss function used in the approach is specified according to a given labelling of the elements in the considered dataset. In this paper, we present the principles of our weighting approach and we report on some experiments showing its effectiveness on real data extracted from the national library of France describing the documents’ conservation histories.
Online Social Networks are built on relationships between individuals. Moreover, they are a space where particular events emerge more quickly than through traditional media such as newspapers or radio. These networks represent a good source for detecting indicators of changes, called weak signals, announcing future threats or opportunities for an organization. The identification of weak signals in a complex environment exposed to ongoing evolution is still a challenging task in research. This paper proposes an approach based on the topology of networks to identify weak signals. Our approach contrasts existing works that analyze significant themes and trends, i.e., strong signals prevalent in a social network at a particular time. We model social network data in the form of temporal interaction graphs. Then we choose graphlets (particular network motifs) as an operational tool to estimate the characteristics of weak signals. We demonstrate our approach accordingly by a study on retweets published in the last season of the TV series Game of Thrones, with which we evaluated and validated some properties of our approach.
The main objective of this work is to provide a unified, easy to configure and extensible end-to-end system that performs root cause analysis (RCA) methods on top of anomaly detection (AD) methods in an online setting. AD-focused RCA for online settings has not been investigated so far; therefore our work can be seen as an initial approach to this end. Inspired by the solutions developed in the ThirdEye project, which is coupled with the Apache Pinot data warehousing system, we re-engineer ThirdEye’s RCA components/techniques so that they operate in a manner that they can directly ingest input records from Apache Kafka and continuously compute aggregates at different level of granularity in a principled manner for both OLAP queries and provision of baselines to support RCA. To attain scalability, we build our solution in the Apache Flink stream processing engine. This work presents the main design choices when applying ThirdEye’s concepts on data streams and presents indicative examples and scalability experiments. Our solution is provided in open-source.
Tensors are powerful multi-dimensional mathematical objects, that easily embed various data models such as relational, graph or time series. Furthermore, tensor decomposition operators are of great utility to reveal hidden patterns and complex relationships in data. Among these decompositions, the Tucker decomposition allows to factorize a tensor into a smaller core tensor and a set of factor matrices. In this article, we propose to study the capabilities of the Tucker decomposition when it is used in data mining techniques such as exploratory analysis, clustering and classification of data. We apply these different techniques on practical examples using several datasets having a ground truth. It is a preliminary work to add the Tucker decomposition to the Tensor Data Model, a model aiming at making tensors data-centric, and at optimizing operators in order to enable the manipulation of large tensors.
Data driven science requires manipulating large datasets coming from various data sources through complex workflows based on a variety of models and languages. With the increasing number of big data sources and models developed by different groups, it is hard to relate models and data and use them in unanticipated ways for specific data analysis. Current solutions are typically ad-hoc, specialized for particular data, models and workflow systems. In this paper, we focus on data driven life science and propose an open service-based architecture, Life Science Workflow Services (LifeSWS), which provides data analysis workflow services for life sciences. We illustrate our motivations and rationale for the architecture with real use cases from life science.
Over the past decade, the volume of data has experienced a significant increase, and this growth is projected to accelerate in the coming years. Within the healthcare sector, various methods (such as liquid biopsies, medical images, and genome sequencing) generate substantial amounts of data, which can lead to the discovery of new biomarkers. Analyzing big data in healthcare holds the potential to advance precise diagnostics and effective treatments. However, healthcare data faces several complexity challenges, including volume, variety, and veracity, which necessitate innovative techniques for data management and knowledge discovery to ensure accurate insights and informed decision-making. This paper summarizes the results presented in the invited talk at BDA 2022 and addresses these challenges by proposing a knowledge-driven framework able to handle complexity issues associated with big data and their impact on analytics. In particular, we propose the use of Knowledge Graphs (KGs) as data structures that enable the integration of diverse healthcare data and facilitate the merging of data with ontologies that describe their meaning. We show the benefits of leveraging KGs to uncover patterns and associations among entities. Specifically, we illustrate the application of rule mining tasks that enhance the understanding of the role of biomarkers and previous cancers in lung cancer.
Graphs are a ubiquitous data model for capturing entities and their relationships. Since most graphs that model real-world networks evolve over time, efficiently managing temporal graphs is an important problem from both a theoretical and practical perspective. Querying the history of temporal graphs can lead to new applications such as object tracking, anomaly detection, and predicting future behavior. However, existing commercial graph databases lack native temporal support, hindering their usefulness in these use cases. This paper introduces Clock-G, a temporal graph management system designed to handle the history temporal graphs. What differentiates Clock-G from other temporal graph management systems is its comprehensive approach, covering query language, query processing, and physical storage. We define T-Cypher, a temporal extension of Cypher query language, enabling user-friendly and concise querying of the graph’s history. Additionally, we propose a query processor that utilizes temporal statistics collected from underlying temporal graphs to offer a good evaluation plan for T-Cypher queries. We also propose a novel storage technique that balances space usage and query evaluation time.
Prediction is one of the most important activities while working with time series. There are many alternative ways to model the time series. Finding the right one is challenging to model them. Most data-centric models (either statistical or machine learning) have hyperparameters to tune. Setting them right is mandatory for good predictions. It is even more complex since time series prediction also demands choosing a data preprocessing that complies with the chosen model. Many time series frameworks, such as Scikit Learning, have features to build models and tune their hyperparameters. However, only some works address tuning data preprocessing hyperparameters and model building. TSPredIT addresses this issue in this scope by providing a framework that seamlessly integrates data preprocessing activities with models’ hyperparameters. TSPredIT is made available as an R-package, which provides functions for defining and conducting time series prediction, including data pre(post)processing, decomposition, hyperparameter optimization, modeling, prediction, and accuracy assessment. Besides, TSPredIT is also extensible, which significantly expands the framework’s applicability, especially with other languages such as Python.
We describe some recent approaches to score-based explanations for query answers in databases. The focus is on work done by the author and collaborators. Special emphasis is placed on the use of counterfactual reasoning for score specification and computation. Several examples that illustrate the flexibility of these methods are shown.
The use of automated decision-making based on machine learning algorithms has raised concerns about potential discrimination against minority group defined by protected features such as gender, race, etc. Particularly, in some areas with high social impact such as justice, job search or healthcare, it has been observed that using protected feature in machine learning algorithms can lead to unfair decisions that favor one group (privileged) over another group (unprivileged). In order to improve fairness in decision-making with regard to protected features, many machine learning approaches focus either on discarding the protected features or maintaining an overall accuracy performance for both unprivileged and privileged groups. However, we notice that these approaches have limited efficiency in the case where the protected features are useful for the learning model or when dealing with imbalanced data. To overcome this limitation when dealing with such issues, we propose in this work FAPFID, a fairness-aware strategy based on the use of balanced and stable clusters. To do this, we divide our input data into stable clusters (subgroups) while ensuring that privileged and unprivileged groups are fairly represented in each cluster. Experiments on three real-world and biased datasets demonstrated that our proposed method outperforms state-of-the-art fairness-aware methods under comparison in terms of performance and fairness scores.
This paper addresses the information-theoretic definition of semantic similarity based on the notion of information content, and presents an evolution of a novel approach for evaluating semantic similarity in a taxonomy. Such an approach takes into account not only the generic sense of a concept but also its intended sense in a given context. In particular, a method for computing the semantic relatedness of concepts in RDF knowledge graphs is used for evaluating the relevance of the intended sense of a concept with respect to its generic sense. The experiment of this work shows that the relatedness method based on triple patterns adopted in this paper leads to higher correlation values with human judgment with respect to the ones obtained according to the original proposal that is based on a triple weights relatedness measure.
Thespis is a middleware that innovatively leverages the Actor model to implement causal consistency over an industry-standard database, whilst abstracting complexities for application developers behind a REST open-protocol interface. Our evaluation considers correctness, performance and scalability aspects. We also run empirical experiments using YCSB to show the efficacy of the approach for a variety of workloads.
AbstractTo ensure critical infrastructure is operating as expected, high-quality sensors are increasingly installed. However, due to the enormous amounts of high-frequency time series they produce, it is impossible or infeasible to transfer or even store these time series in the cloud when using state-of-the-practice compression methods. Thus, simple aggregates, e.g., 1–10-minutes averages, are stored instead of the raw time series. However, by only storing these simple aggregates, informative outliers and fluctuations are lost. Many Time Series Management System (TSMS) have been proposed to efficiently manage time series, but they are generally designed for either the edge or the cloud. In this paper, we describe a new version of the open-source model-based TSMS ModelarDB. The system is designed to be modular and the same binary can be efficiently deployed on the edge and in the cloud. It also supports continuously transferring high-frequency time series compressed using models from the edge to the cloud. We first provide an overview of ModelarDB, analyze the requirements and limitations of the edge, and evaluate existing query engines and data stores for use on the edge. Then, we describe how ModelarDB has been extended to efficiently manage time series on the edge, a novel file-based data store, how ModelarDB’s compression has been improved by not storing time series that can be derived from base time series, and how ModelarDB transfers high-frequency time series from the edge to the cloud. As the work that led to ModelarDB began in 2015, we also reflect on the lessons learned while developing it.
Large neural language models are achieving exceptional performance in question answering and other natural language processing tasks. However, these models can be costly to train and difficult to interpret. In this paper, we propose to investigate whether incorporating explicit linguistic information can boost model performance while improving model interpretability. We present a novel constituency-informed and constituency-constrained question answering model called SyHGT-CN. The linguistics-informed model integrates the symbolic information contained in constituency trees with the statistical knowledge of a neural language model. The integration of the linguistics graphic structures with the transformer-based neural language model is achieved by the adjunction to the latter of a heterogeneous graph neural network, in which the former is encoded. We comparatively and empirically show, with the SQuAD2.0 benchmark, that the proposed approach is more accurate than a constituency-oblivious BERT and the constituency-informed SyHGT-C model.
