IFIP Open Access Digital Library
Not a member yet
22614 research outputs found
Sort by
Emerging Research Topics Identification Using Temporal Graph Neural Networks
No full textPart 2: Graphs/Neural Networks/Machine LearningInternational audienceThe dynamic landscape of research necessitates effective methods for the timely identification of emerging research topics, a critical pursuit for researchers and decision makers in both governmental and industrial spheres. Traditional approaches to this challenge have predominantly relied on retrospective analyses, limiting their applicability in real world scenarios where proactive foresight is paramount. This study addresses this constraint through the introduction of a novel methodology for the future prediction of emerging research topics, employing temporal graph neural networks. Our proposed framework revolves around the construction of co-word graphs, serving as input for our innovative machine learning model designed to forecast keyword frequencies in forthcoming time periods. To delineate emerging themes, keywords undergo clustering via a graph entropy algorithm that are subsequently sorted in terms of their “emergence score”. To validate the efficacy of our methodology, we apply it to forecast emerging research topics for the year 2022. The results showcase the potential of our approach, offering valuable insights into the trajectory of research themes poised to gain prominence in the near future
An Evaluation Framework for Synthetic Data Generation Models
No full textPart 3: Data Mining/ModelingInternational audienceNowadays, the use of synthetic data has gained popularity as a cost-efficient strategy for enhancing data augmentation for improving machine learning models performance as well as addressing concerns related to sensitive data privacy. Therefore, the necessity of ensuring quality of generated synthetic data, in terms of accurate representation of real data, consists of primary importance. In this work, we present a new framework for evaluating synthetic data generation models’ ability for developing high-quality synthetic data. The proposed approach is able to provide strong statistical and theoretical information about the evaluation framework and the compared models’ ranking. Two use case scenarios demonstrate the applicability of the proposed framework for evaluating the ability of synthetic data generation models to generated high quality data
Beyond Sentiment in Stock Price Prediction: Integrating News Sentiment and Investor Attention with Temporal Fusion Transformer
No full textPart 1: Deep LearningInternational audienceNews sentiment is attracting considerable interest in stock market prediction, given its crucial role in shaping stock prices. Previous research has mainly focused on improving prediction accuracy by exploiting news sentiment, without adequately considering the different levels of attention that individual news articles receive. Furthermore, despite the advanced predictive capabilities of deep learning models, there has been a lack of focus on the interpretability of these models, leading to predictions that are not transparent. This study presents an innovative prediction model that integrates a FinBERT-based analysis of news sentiment and investor attention metrics with an attention-based Temporal Fusion Transformer framework. This approach not only enables highly effective forecasting, but also provides insights into the temporal dynamics that influence the stock market. The effectiveness of the model is demonstrated by analyzing stock price data for 41 of the largest market capitalization companies over the period 2010 to 2021. The results confirm the superiority of the proposed model over existing deep learning approaches, and the attention analysis underscores the critical role of synthesizing news sentiment and attention metrics in predicting stock prices
A Prediction Analysis for the Case of a Korean Police Dataset
No full textPart 2: Machine LearningInternational audienceIn the evolving landscape of law enforcement, predictive policing, which leverages data analysis and machine learning to anticipate crimes and optimise police responses, has emerged as a critical tool. This paper explores the application of machine learning techniques in predictive policing through a detailed analysis of a Korean police dataset. Focusing on predicting the patterns and duration of police responses, the study employs various algorithms such as RandomForest, Gaussian Naive Bayes, Decision Tree and K-Nearest Neighbors. These models are evaluated based on accuracy, precision, recall, and F-score to determine their efficacy in different response scenarios. Our findings indicate that RandomForest has a much better performance in forecasting response duration, whilst Decision Tree and K-Nearest Neighbour models are particularly effective in predicting the type of response for incidents. The study underscores the significance of specific features like incident severity and police response type in influencing prediction outcomes. Through this research, we contribute to understanding the potential and challenges of machine learning in enhancing the efficiency of police operations in Korea, providing a framework applicable to broader contexts
SMT: Self-supervised Approach for Multiple Animal Detection and Tracking
No full textPart 2: Machine LearningInternational audienceIn the domain of animal farming and wildlife management, monitoring animal behavior and movement is crucial. This paper proposes an efficient online multi-object tracking framework named SMT (Self-supervised Multi-animal detection and Tracking) for a dynamic and complex environment. The framework is based on the tracking-by-detection approach and builds on the idea of employing self-supervised object detection and a bag of Bayesian trackers. We collected and annotated a custom dataset from an animal farm for training and validating the detection and tracking algorithms. Additionally, we utilized the public dataset Dancetrack to benchmark and compare the results against reference methods. The comparison with reference methods reveals substantial enhancements on standard tracking metrics, such as IDF1 and MOTA. The optimized combination of the self-supervised object detector and proposed tracker demonstrates robust performance by consistently preserving object identities and reducing identification errors throughout sequences. To reproduce the results, we made the code publically available at https://github.com/moosa1296/effdet_ocsort
Strategizing the Shallows: Leveraging Multi-Agent Reinforcement Learning for Enhanced Tactical Decision-Making in Littoral Naval Warfare
No full textPart 1: Reinforcement/Natural LanguageInternational audienceNaval engagements, though rare, present complex challenges for data-driven machine learning due to their intricate dynamics and the scarcity of empirical evidence. These conflicts are well-represented within the framework of Partially Observable Stochastic Games (POSG), which models the adversarial interactions between contending forces through decision-making agents, possible states, actions, observations, and probabilistic state transitions.This research delineates the implementation of Multi-Agent Reinforcement Learning (MARL) algorithms, particularly Double Deep Q-Networks (DDQN) and Proximal Policy Optimization (PPO), in navigating the complexities inherent in strategic naval operations. Despite the operational challenges encountered, the findings of this study underscore the effectiveness of MARL in formulating and assessing tactical strategies. This contributes substantially to the enhancement of tactical planning and the introduction of novel strategic paradigms. Significantly, this investigation illuminates the transformative potential of MARL in naval strategy and decision-making processes, asserting its pivotal role in contemporary warfare analysis. This study not only confirms the applicability of MARL in complex scenarios but also highlights its capacity to revolutionize traditional approaches to military strategy
Putting Authorization Servers on User-Owned Devices in User-Managed Access
No full textInternational audienceWhen a user uses a third-party application that relies on data from another platform, the platform must authorize access by that application based on the user’s consent for privacy. Although OAuth 2.0 is widely used for such authorization, it is a troublesome task for the users to manage authorization. User-Managed Access (UMA) is a framework for asynchronous user-centric authorization management. UMA aims to reduce the burden of users for handling authorization requests from third-party applications. In UMA, transparency of the authorization server, usually operated by a (believed to be) trusted third-party, is important because the authorization server may behave maliciously, such as authorizing access against the authorization policy set by users. Some proposals use blockchain for improving transparency, but blockchain may not be suitable because it makes information public although authorization policies and decisions will be sensitive to user privacy. In this paper, we propose an architecture that improves UMA’s transparency by putting the UMA authorization server on user-owned devices. By making authorization decisions on the devices, our architecture improves transparency while reducing the users’ burden. We also use STRIDE-per-Element to discuss discovered threats and how to mitigate risks
Enhancing Malware Detection Through Machine Learning Using XAI with SHAP Framework
No full textPart 2: Cyber Security/Anomaly DetectionInternational audienceMalware represents a significant cyber threat that can potentially disrupt any activities within an organization. There is a need to devise effective proactive methods for malware detection, thereby minimizing the associated risks. However, this task is challenging due to the ever-growing volume of malware data and the continuously evolving techniques employed by malicious actors. In this context, machine learning models offer a promising approach to identify key malware features and facilitate accurate detection. Machine learning has proven to be effective in detecting malware and has recently gained widespread attention from both the academic and research sectors. Despite their effectiveness, current research on machine learning (ML) models for malware detection often lacks necessary explanations for the selection of key features. This opacity of ML models can complicate the understanding of the outputs, errors, and decision-making processes. To address this challenge, this research uses Explainable AI (XAI), particularly the SHAP framework, to enhance transparency and interpretability. By providing extensive insights into how each feature contributes to the model’s conclusions, the approach further improves the model’s accountability. An experiment was conducted to demonstrate the applicability of the proposed method, beginning with the training of the chosen machine learning models, including Random Forest, Adaboost, Support Vector Machine and Artificial Neural Network, for detecting malware, and concluding with the explanation of the decision-making process using XAI techniques. The results showed high accuracy in malware detection, along with comprehensive explanations of the feature contributions, which justifies the outputs produced by the models
Hybrid Explanatory Interactive Machine Learning for Medical Diagnosis
No full textPart 1: Biomedical/ClassificationInternational audienceMachine learning (ML) models can be an effective assistance in medical diagnosis if they allow physicians to project their knowledge into model’s internal mechanism. Using model-agnostic explanatory interactive ML (XIML), physicians iteratively train a ML model and revise its decision-making mechanism depicted as local explanation. Counterexamples serve as additional training data and statistically outweigh the human feedback. Unfortunately, counterexamples alone do not guarantee that the feedback persists in subsequent optimization iterations – a form of catastrophic forgetting in XIML, which might cause serious consequences in sensitive domains such as medical diagnosis. To overcome this issue, we propose a hybrid approach: HyXiml collects the physicians’ feedback, learns a set of probabilistic logical rules, and substitutes the predictions for closely related instances by logical inferences. We show that the connection of XIML and probabilistic logic enhances the explanatory performance whilst retaining a stable predictive performance
The Role of Epigenetics in OCD: A Multi-order Adaptive Network Model for DNA-Methylation Pathways and the Development of OCD
No full textPart 1: Biomedical/ClassificationInternational audienceObsessive-Compulsive Disorder (OCD) can be classified as a psychiatric condition which is characterised by persistent intrusive thoughts, to be called obsessions, and repetitive behaviours and mental acts, to be called compulsions. Recent studies have linked OCD to increased methylation of the oxytocin receptor (OXTR) gene which potentially impacts OCD-linked social and emotional behaviours. Options to treat OCD include cognitive-behavioural therapy (CBT) and selective serotonin reuptake inhibitors (SSRI). Challenges to these treatment options exist due to some individuals not responding well to SSRIs. The paper introduces a hypothetical treatment option to restore OXTR functionality. The research focuses on epigenetics in OCD and presents simulations involving DNA methylation changes. In the first presented scenario OCD development is shown triggered by childhood trauma. Scenario two introduces an epigenetic therapy to counteract the OCD symptoms by demethylating the OXTR gene. The simulated hypothetical therapy shows the potential to relieve OCD symptoms. Epigenetic drugs used in diseases like cancer, do indeed suggest potential for usage in other disorders like OCD