Abstract Enabling diabetic patients to predict their Blood Glucose Levels (BGL) is a crucial aspect of managing their metabolic condition, as it allows them to take appropriate measures to avoid hypo or hyperglycemia. Machine Learning (ML) and Deep Learning (DL) techniques have made this possible, and this paper evaluates and compares the performance of five distinct ML/DL models including: Convolutional Neural Network (CNN), Long Short Term Memory (LSTM), Support Vector Regression (SVR), Gated Reccurent Unit (GRU) and Deep Belief Network (DBN) for forecasting BGL, by applying two different forecasting methods, namely One Step Ahead (OSF) and Multi‐Step Ahead (MSF) comprising five different variants. The performance is evaluated based on four metrics: Mean Absolute Error (MAE), Mean Magnitude Relative Error (MMRE), Root Mean Square Error (RMSE) and Predictive Level (PRED). Additionally, the statistical significance of the regressors was evaluated using the Scott‐Knott (SK) test, while the Borda Count (BC) voting system was employed to rank them. The results indicate that the best performance was achieved with OSF using GRU. Furthermore, the effectiveness of an MSF strategy depends on the ML/DL technique used, and the best combinations were DBN with DirRec, DBN with Recursive, SVR with Recursive and SVR with DirRec. Additionally, DirRec was found to be the best strategy, as it consistently ranked first regardless of the ML/DL technique used.

Abstract Context Software development effort estimation (SDEE) is one of the most challenging aspects in project management. The presence of missing data (MD) in software attributes makes SDEE even more complex. K‐nearest neighbors imputation (KNNI) has been widely used in SDEE to deal with the MD issue. However, KNNI, in its classical process, has low tolerance to imprecision and uncertainty especially when dealing with categorical features. When dealing with categorical attributes, KNNI uses a classical approach, employing mainly numbers or classical intervals to represent software attributes and similarity measures originally designed for numerical attributes. Objectives This paper evaluates the use of an optimized fuzzy clustering‐based KNNI (FC‐KNNI) and compares it with classical KNN when dealing with mixed data in the context of SDEE. Methods We investigate the effect of two imputation techniques (FC‐KNNI and KNNI) on five SDEE techniques: case‐based reasoning, fuzzy case‐based reasoning, support vector regression, multilayer perceptron, and reduced‐error pruning tree. The evaluation is carried out using six publicly available datasets for SDEE using two performance measures, standardized accuracy (SA), and Pred (0.25). The Wilcoxon statistical test is also performed to assess the significance of results. Results The results are promising in the sense that using an imputation technique designed for mixed data is better than reusing methods originally designed for numerical data. We found that FC‐KNNI significantly outperforms KNNI regardless of the SDEE technique and dataset used. Another important finding is that F‐CBR improved the analogy process compared to CBR. Conclusion The introduction of fuzzy sets and fuzzy clustering in the analogy process improves its performances in terms of SA and Pred (0.25).

The best-performing machine learning (ML) models suffer from a lack of interpretability. This study conducted an empirical evaluation of two interpretability techniques, global surrogate and local interpretable model-agnostic explanations (LIME). Experiments on two black boxes, a multilayer perceptron, and support vector machines were carried out on two breast cancer tabular datasets. The results show that local interpretability can work along the global one to provide insights into the model. Quantitative evaluations show that the global surrogate slightly outperforms LIME. Interpretability techniques have the potential to fix the interpretability trade-off for opaque models.

Abstract Missing data is a serious issue in software engineering because it can lead to information loss and bias in data analysis. Several imputation techniques have been proposed to deal with both numerical and categorical missing data. However, most of those techniques used is simple reuse techniques originally designed for numerical data, which is a problem when the missing data are related to categorical attributes. This paper aims (a) to propose a new fuzzy case‐based reasoning (CBR) imputation technique designed for both numerical and categorical data and (b) to evaluate and compare the performance of the proposed technique with the k‐nearest neighbor (KNN) imputation technique in terms of error and accuracy under different missing data percentages and missingness mechanisms in four software engineering data sets. The results suggest that the proposed fuzzy CBR technique outperformed KNN in terms of imputation error and accuracy regardless of the missing data percentage, missingness mechanism, and data set used. Moreover, we found that the missingness mechanism has an important impact on the performance of both techniques. The results are encouraging in the sense that using an imputation technique designed for both categorical and numerical data is better than reusing methods originally designed for numerical data.

Missing data are a serious issue that influences the prediction accuracy of software development effort estimation (SDEE) techniques and especially analogy-based software effort estimation (ASEE). Hence, appropriate handling of missing data is necessary in order to ensure best performance. To deal with this issue K-nearest neighbors (KNN) imputation has been widely used. However, none of the studies investigating KNN imputation in SDEE have addressed the impact of parameter settings on the imputation process given that parameter optimization techniques are often used at the prediction level, as they highly impact the performance of SDEE techniques including ASEE. This paper proposes and evaluates an ensemble KNN imputation technique for ASEE. Thereafter, we compare ASEE performance using ensemble KNN imputation with those using either a grid search based single KNN imputation or KNN imputation without parameter optimization. For the six datasets used for comparison, the ensemble KNN imputation significantly improved ASEE performance compared with KNN imputation without optimization. Moreover, ensemble KNN imputation and grid search-based imputation behaved similarly. Given that grid search is time consuming, the ensemble KNN imputation may be an alternative to deal with missing data in the ASEE process.

Diabetes self-management relies on the blood glucose prediction as it allows taking suitable actions to prevent low or high blood glucose level.In this paper, we propose a deep learning neural network (NN) model for blood glucose prediction.It is a sequential one using a Long-Short-Term Memory (LSTM) layer with two fully connected layers.Several experiments were carried out over data of 10 diabetic patients to decide on the model's parameters in order to identify the best variant of it.The performance of the proposed LSTM NN measured in terms of root mean square error (RMSE) was compared with the ones of an existing LSTM and an autoregressive (AR) models.The results show that our LSTM NN is significantly more accurate; in fact, it outperforms the existing LSTM model for all patients and outperforms the AR model in 9 over 10 patients, besides, the performance differences were assessed by the Wilcoxon statistical test.Furthermore, the mean of the RMSE of our model was 12.38 mg/dl while it was 28.84 mg/dl and 50.69 mg/dl for AR and the existing LSTM respectively.

Estimation by analogy (EBA) is one of the most attractive software effort development estimation techniques.However, one of the critical issues when using EBA is the occurrence of missing data (MD) in the historical data sets.The absence of values of several relevant software attributes is a frequent phenomenon that may cause inaccurate EBA estimations.The MD can be numerical and/or categorical.This paper evaluates four MD techniques (toleration, deletion, k-nearest neighbors (KNN) imputation and support vector regression (SVR) imputation) over four mixed data sets.A total of 432 experiments were conducted involving four MD techniques, nine MD percentages (from 10% to 90%), three missingness mechanisms (MCAR: Missing Completely at Random, MAR: Missing at Random and NIM: Non-Ignorable Missing) and four data sets.The evaluation process consists of four steps and uses several accuracy measures such as standardized accuracy (SA) and prediction level (Pred).The results suggest that EBA with imputation techniques achieved significantly better SA values over EBA with toleration or deletion regardless of the mechanism of missingness.Moreover, no particular MD imputation technique outperformed the other techniques overall.However, according to Pred and other accuracy criteria, EBA with SVR was the best, followed by KNN imputation; we also found that toleration instead of deletion improves the accuracy of EBA.

Abstract Software development effort estimation (SDEE) plays a primary role in software project management. But choosing the appropriate SDEE technique remains elusive for many project managers and researchers. Moreover, the choice of a reliable estimation accuracy measure is crucial because SDEE techniques behave differently given different accuracy measures. The most widely used accuracy measures in SDEE are those based on magnitude of relative error (MRE) such as mean/median MRE (MMRE/MedMRE) and prediction at level p (Pred(p)), which counts the number of observations where an SDEE technique gave MREs lower than p. However, MRE has proven to be an unreliable accuracy measure, favoring SDEE techniques that underestimate. Consequently, an unbiased measure called standardized accuracy (SA) has been proposed. This paper deals with the Pred(p) and SA measures. We investigate (1) the consistency of Pred(p) and SA as accuracy measures and SDEE technique selectors, and (2) the relationship between Pred(p) and SA. The results suggest that Pred(p) is less biased towards underestimates and generally selects the same best technique as SA. Moreover, SA and Pred(p) measure different aspects of technique performance, and SA may be used as a predictor of Pred(p) by means of the 3 association rules.

Abstract Missing data (MD) is a widespread problem that can affect the ability to use data to construct effective software development effort estimation (SDEE) techniques. To deal with this challenge, several imputation techniques have been investigated in SDEE and k‐nearest neighbors (KNN)‐based imputation is still the most frequently used. To the best of our knowledge, no study has used support vector regression (SVR)‐based imputation to construct accurate estimation techniques, in particular those based on analogy. This paper introduces a new imputation technique based on SVR for handling MD in two analogy‐based SDEE techniques: classical analogy and fuzzy analogy. More specifically, we investigate whether the use of SVR instead of KNN in imputing MD improves the predictive performance of these two analogy‐based techniques. A total of 1134 experiments were conducted involving seven datasets, SVR/KNN MD imputation techniques (KNN with Euclidean and Manhattan distances), three missingness mechanisms (missing completely at random, missing at random, non‐ignorable missing), and MD percentages from 10% to 90%. The results suggest that the use of SVR imputation, rather than KNN imputation, may improve the prediction performance of both analogy‐based techniques. Furthermore, we found that the impact of MD percentage upon effort prediction performance is reduced when using SVR rather than KNN. Moreover, fuzzy analogy generates better estimates in terms of the standardized accuracy measure than classical analogy regardless of the MD technique, the dataset used, the missingness mechanism, or the MD percentage.

Missing Values (MV) present a serious problem facing research in software engineering (SE) which is mainly based on statistical and/or data mining analysis of SE data. The simple method of dealing with MV is to ignore data with missing observations. This leads to losing valuable information and then obtaining biased results. Therefore, various techniques have been developed to deal adequately with MV, especially those based on imputation methods. In this paper, a systematic mapping study was carried out to summarize the existing techniques dealing with MV in SE datasets and to classify the selected studies according to six classification criteria: research type, research approach, MV technique, MV type, data types and MV objective. Publication channels and trends were also identified. As results, 35 papers concerning MV treatments of SE data were selected. This study shows an increasing interest in machine learning (ML) techniques especially the K-nearest neighbor algorithm (KNN) to deal with MV in SE datasets and found that most of the MV techniques are used to serve software development effort estimation techniques.