In recent years, the research community, but also the general public, has raised serious questions about the reproducibility and replicability of scientific work. Since many studies include some kind of computational work, these issues are also a technological challenge, not only in computer science, but also in most research domains. Computational replicability and reproducibility are not easy to achieve due to the variety of computational environments that can be used. Indeed, it is challenging to recreate the same environment via the same frameworks, code, programming languages, dependencies, and so on. We propose a framework, known as SciRep, that supports the configuration, execution, and packaging of computational experiments by defining their code, data, programming languages, dependencies, databases, and commands to be executed. After the initial configuration , the experiments can be executed any number of times, always producing exactly the same results. Our approach allows the creation of a reproducibility package for experiments from multiple scientific fields, from medicine to computer science, which can be re-executed on any computer. The produced package acts as a capsule, holding absolutely everything necessary to re-execute the experiment. To evaluate our framework, we compare it with three state-of-the-art tools and use it to reproduce 18 experiments extracted from published scientific articles. With our approach, we were able to execute 16 (89%) of those experiments, while the others reached only 61%, thus showing that our approach is effective. Moreover, all the experiments that were executed produced the results presented in the original publication. Thus, SciRep was able to reproduce 100% of the experiments it could run.

Computational reproducibility-the ability to reexecute a scientific experiment using the same code, data, and configuration-should be straightforward. However, researchers often struggle with inconsistencies in documentation, missing dependencies, and environment setup, which undermines the credibility of scientific results. To address this, we propose a conversational, text-based tool that aids researchers in reproducing and packaging computational experiments into a single file. This file can be re-executed with a double-click on any machine, requiring only a single tool. SciConv is designed to support two key scenarios: (i) enabling researchers to prepare their own experiments in a reproducible, shareable format, and (ii) helping other researchers reproduce existing experiments from shared code repositories. In both cases, the tool reduces technical overhead and simplifies environment configuration through conversational interaction. We evaluated the tool through two studies. In the first, we reproduced 15 of 18 published experiments, with most requiring little or no user interaction. In the second, we conducted a user study comparing our tool with a professional platform, using the System Usability Scale (SUS) and NASA Task Load Index (TLX). The results show a statistically significant advantage for our tool in both usability and workload, demonstrating its effectiveness in supporting reproducibility.

Computational reproducibility remains a critical yet unresolved issue across scientific disciplines, often hindered by complex configuration requirements and technical barriers. We present SciConv, a novel conversational tool designed to assist researchers in creating and executing reproducible computational experiments using natural language. By leveraging large language models (llMs), SciConv automates the detection of dependencies and programming languages, and packages experiments into portable artifacts with minimal manual input. Unlike traditional platforms based on graphical user interfaces (e.g., web-based), SciConv features a chat-based interface that guides researchers interactively through the reproducibility workflow. This paper introduces the architecture, design principles, and interaction model of SciConv, and discusses its potential to lower the technical barriers to reproducibility.

User studies are paramount for advancing science. In particular, the empirical evaluation of programmer-oriented tools is important to validate research ideas and prototypes, as well as production-ready tools. Previous research has collected several tools used by the software engineering and behavioral science communities to design and run studies. In this work, we study tools used in software engineering studies and identify their features. Furthermore, we analyze three behavioral science experiment tools to identify design ideas that might be adapted to programmer user studies. With this work, we present the set of features currently offered by software engineering tools to support researchers in the design and execution of programmer user studies. We also present the characteristics of some tools used in behavioral science experiments to identify design ideas that can be adapted to programmer user studies.

ReLU neural networks suffer from a problem of explainability because they partition the input space into a lot of polyhedrons. This paper proposes a constrained neural network model that replaces polyhedrons by orthotopes: each hidden neuron processes only a single component of the input signal. When the number of hidden neurons is large, we show that our neural network is equivalent to a logistic regression whose input is a non-linear transformation of the processed signal. Hence, the training of our neural network always converges to a unique solution. Numerical simulations show that the loss of performance with respect to state-of-the-art methods is negligible even though our neural network is strongly constrained on robustness and explainability.

This paper proposes an understandable neural network whose score function is modeled as an additive sum of univariate spline functions. It extends usual understandable models like generative additive models, spline-based models, and neural additive models. It is shown that this neural network can be approximated by a logistic regression whose inputs are obtained with a non-linear preprocessing of input data. This preprocessing depends on the neural network initialization but this paper establishes that it can be replaced by a non random kernel-based preprocessing that no longer depends on the initialization. Hence, the convergence of the training process is guaranteed and the solution is unique for a given training dataset.

In recent years, the research community has raised serious questions about the replicability and reproducibility of scientific work. In particular, since many studies include some kind of computing work, these are also technological challenges, not only in computer science but in most research domains. Replicability and reproducibility are not easy to achieve, not only because researchers have diverse proficiency in computing technologies, but also because of the variety of computational environments that can be used. Indeed, it is challenging to recreate the same environment using the same frameworks, code, programming languages, dependencies, and so on. In this work, we propose a vision for an Integrated Development Environment allowing the creation, configuration, execution, packaging, and sharing of scientific computational experiments. Such a framework should allow researchers to easily set the code and data used and define the programming languages, code, dependencies, databases, or commands to execute to achieve consistent results for each experiment. With this work, we intend to aid researchers by integrating into the same platform all the stages of the design, execution, and analysis of a computational experiment.

This paper aims to build a supervised classifier for dealing with imbalanced datasets, uncertain class proportions, dependencies between features, the presence of both numeric and categorical features, and arbitrary loss functions. The Bayes classifier suffers when prior probability shifts occur between the training and testing sets. A solution is to look for an equalizer decision rule whose class-conditional risks are equal. Such a classifier corresponds to a minimax classifier when it maximizes the Bayes risk. We develop a novel box-constrained minimax classifier which takes into account some constraints on the priors to control the risk maximization. We analyze the empirical Bayes risk with respect to the box-constrained priors for discrete inputs. We show that this risk is a concave non-differentiable multivariate piecewise affine function. A projected subgradient algorithm is derived to maximize this empirical Bayes risk over the box-constrained simplex. Its convergence is established and its speed is bounded. The optimization algorithm is scalable when the number of classes is large. The robustness of our classifier is studied on diverse databases. Our classifier, jointly applied with a clustering algorithm to process mixed attributes, tends to equalize the class-conditional risks while being not too pessimistic.