International audience

With the increasing adoption of Machine Learning (ML), the need to manage, in a tailored manner, the end-to-end engineering of ML-powered software arises. Machine Learning Operations (MLOps) extends DevOps to address the specific challenges of the development and lifelong management of this new type of software. Despite the growing attention to this domain, an accepted, accessible, and modular reference architecture for MLOps is still missing. This paper presents SkeltyMLOps, a reference architecture aimed at promoting collaboration among the diverse actors involved in MLOps. SkeltyMLOps is built as rigorously as possible from knowledge gathered through a systematic literature review. First, sets of MLOps actors and activities are manually extracted from the literature by two independent authors. The extracted terms are then clustered and normalized through a process involving human and Large Language Model (LLM) experts. To carefully build the artificial expertise, several experiments are conducted using various LLMs, with or without additional contextual knowledge and under different parameter settings. The resulting clusters and labels serve as the foundation for the design of the SkeltyMLOps architecture. There, special attention is paid to designing the orchestration of MLOps activities and to the coordination of contributions of various actors. Therefore, the originality of the contribution is twofold: (1) the careful knowledge gathering process that grounds the proposed architecture, and (2) the central role given to orchestration and coordination at the heart of the architecture.

Defect prediction aims at identifying software components that are likely to cause faults before a software is made available to the end-user. To date, this task has been modeled as a two-class classification problem, however its nature also allows it to be formulated as a one-class classification task. Previous studies show that One-Class Support Vector Machine (OCSVM) can outperform two-class classifiers for within-project defect prediction, however it is not effective when employed at a finer granularity (i.e., commit-level defect prediction). In this paper, we further investigate whether learning from one class only is sufficient to produce effective defect prediction model in two other different scenarios (i.e., granularity), namely cross-version and cross-project defect prediction models, as well as replicate the previous work at within-project granularity for completeness. Our empirical results confirm that OCSVM performance remain low at different granularity levels, that is, it is outperformed by the two-class Random Forest (RF) classifier for both cross-version and cross-project defect prediction. While, we cannot conclude that OCSVM is the best classifier, our results still show interesting findings. While OCSVM does not outperform RF, it still achieves performance superior to its two-class counterpart (i.e., SVM) as well as other two-class classifiers studied herein. We also observe that OCSVM is more suitable for both cross-version and cross-project defect prediction, rather than for within-project defect prediction, thus suggesting it performs better with heterogeneous data. We encourage further research on one-class classifiers for defect prediction as these techniques may serve as an alternative when data about defective modules is scarce or not available.

BACKGROUND: High-throughput experimental technologies can provide deeper insights into pathway perturbations in biomedical studies. Accordingly, their usage is central to the identification of molecular targets and the subsequent development of suitable treatments for various diseases. Classical interpretations of generated data, such as differential gene expression and pathway analyses, disregard interconnections between studied genes when looking for gene-disease associations. Given that these interconnections are central to cellular processes, there has been a recent interest in incorporating them in such studies. The latter allows the detection of gene modules that underlie complex phenotypes in gene interaction networks. Existing methods either impose radius-based restrictions or freely grow modules at the expense of a statistical bias towards large modules. We propose a heuristic method, inspired by Ant Colony Optimization, to apply gene-level scoring and module identification with distance-based search constraints and penalties, rather than radius-based constraints. RESULTS: We test and compare our results to other approaches using three datasets of different neurodegenerative diseases, namely Alzheimer's, Parkinson's, and Huntington's, over three independent experiments. We report the outcomes of enrichment analyses and concordance of gene-level scores for each disease. Results indicate that the proposed approach generally shows superior stability in comparison to existing methods. It produces stable and meaningful enrichment results in all three datasets which have different case to control proportions and sample sizes. CONCLUSION: The presented network-based gene expression analysis approach successfully identifies dysregulated gene modules associated with a certain disease. Using a heuristic based on Ant Colony Optimization, we perform a distance-based search with no radius constraints. Experimental results support the effectiveness and stability of our method in prioritizing modules of high relevance. Our tool is publicly available at github.com/GhadiElHasbani/ACOxGS.git.

In this work, we propose a distributed Memetic Algorithm for the traveling salesman problem focusing on small and medium-sized instances. The algorithm employs a new crossover operator that favors the fitter of the two parents and develops varying progeny from the same parents to evade premature convergence. In the implementation, we use the High-Level Architecture (HLA) to distribute laborious tasks and lower the lead time of the heuristic. Results show that our proposed approach significantly outperforms other algorithms when tested on instances from the TSPLIB benchmark.

In order to monitor and assess the spread of the Omicron variant of COVID-19, we propose a Distributed Digital Twin that virtually mirrors a hemodialysis unit in a hospital in Toronto, Canada. Since the solution involves heterogeneous components, we rely on the IEEE HLA distributed simulation standard. Based on the standard, we use an agent-based/discrete event simulator together with a virtual reality environment in order to provide to the medical staff an immersive experience that incorporates a platform showing predictive analytics during a simulation run. This can help professionals monitor the number of exposed, symptomatic, asymptomatic, recovered, and deceased agents. Agents are modeled using a redesigned version of the susceptible-exposed-infected-recovered (SEIR) model. A contact matrix is generated to help identify those agents that increase the risk of the virus transmission within the unit.

The hyper-parameters of a neural network are traditionally designed through a time-consuming process of trial and error that requires substantial expert knowledge. Neural Architecture Search algorithms aim to take the human out of the loop by automatically finding a good set of hyper-parameters for the problem at hand. These algorithms have mostly focused on hyper-parameters such as the architectural configurations of the hidden layers and the connectivity of the hidden neurons, but there has been relatively little work on automating the search for completely new activation functions, which are one of the most crucial hyperparameters to choose. There are some widely used activation functions nowadays that are simple and work well, but nonetheless, there has been some interest in finding better activation functions. The work in the literature has mostly focused on designing new activation functions by hand or choosing from a set of predefined functions while this work presents an evolutionary algorithm to automate the search for completely new activation functions. We compare these new evolved activation functions to other existing and commonly used activation functions. The results are favorable and are obtained from averaging the performance of the activation functions found over 30 runs, with experiments being conducted on 10 different datasets and architectures to ensure the statistical robustness of the study.

Intelligent Transportation System (ITS) help drivers by showing the shortest routes and some driving information such as congestion, accident and roadwork. Twitter traffic detection systems depend on real time collection of traffic and road-related data where users share real time events that can help extract traffic status on different roads. However, the current twitter-based approaches are not applied on Arabic traffic related tweets, and do not take into consideration tweets about roads that do not have congestion. In this paper, we address the aforementioned limitations by proposing a new proactive social learning approach for (1) detecting traffic related tweets, (2) extracting location using local dictionary and Google Maps API, (3) determining traffic status using Support Vector Machine (SVM), and (4) extracting the cause by classifying them into three different categories using incremental learning. Our experimental results show that our approach can identify tweets referring to traffic with an accuracy of 98%, determine jam status from those tweets by an accuracy of 91.1%, and identify the cause of traffic-related events with an accuracy of 84.7% per class label.

In this work, we present a genetic algorithm to optimize predictive models used to estimate software quality characteristics. Software quality assessment is crucial in the software development field since it helps reduce cost, time and effort. However, software quality characteristics cannot be directly measured but they can be estimated based on other measurable software attributes (such as coupling, size and complexity). Software quality estimation models establish a relationship between the unmeasurable characteristics and the measurable attributes. However, these models are hard to generalize and reuse on new, unseen software as their accuracy deteriorates significantly. In this paper, we present a genetic algorithm that adapts such models to new data. We give empirical evidence illustrating that our approach out-beats the machine learning algorithm C4.5 and random guess.

The goal of quality models is to predict a quality factor starting from a set of direct measures. Selecting an appropriate quality model for a particular software is a difficult, non-trivial decision. In this paper, we propose an approach to combine and/or adapt existing models (experts) in such way that the combined/adapted model works well on the particular system. Test results indicate that the models perform significantly better than individual experts in the pool.