Martin Shepperd

Lech Madeyski, Barbara Kitchenham, Martin Shepperd

Context: Large language models (LLMs) are released faster than users' ability to evaluate them rigorously. When LLMs underpin research, such as identifying relevant literature for systematic reviews (SRs), robust empirical assessment is essential. Objective: We identify and discuss key challenges in assessing LLM performance for selecting relevant literature, identify good (evaluation) practices, and propose recommendations. Method: Using a recent large-scale study as an example, we identify problems with the use of traditional metrics for assessing the performance of Gen-AI tools for identifying relevant literature in SRs. We analyzed 27 additional papers investigating this issue, extracted the performance metrics, and found both good practices and widespread problems, especially with the use and reporting of performance metrics for SR screening. Results: Major weaknesses included: i) a failure to use metrics that are robust to imbalanced data and do not directly indicate whether results are better than chance, e.g., the use of Accuracy, ii) a failure to consider the impact of lost evidence when making claims concerning workload savings, and iii) pervasive failure to report the full confusion matrix (or performance metrics from which it can be reconstructed) which is essential for future meta-analyses. On the positive side, we extract good (evaluation) practices on which our recommendations for researchers and practitioners, as well as policymakers, are built. Conclusions: SR screening evaluations should prioritize lost evidence/recall alongside chance-anchored and cost-sensitive Weighted MCC (WMCC) metric, report complete confusion matrices, treat unclassifiable outputs as referred-back positives for assessment, adopt leakage-aware designs with non-LLM baselines and open artifacts, and ground conclusions in cost-benefit analysis where FNs carry higher penalties than FPs.

Jingxiu Yao, Martin Shepperd

Context: Software engineering researchers have undertaken many experiments investigating the potential of software defect prediction algorithms. Unfortunately, some widely used performance metrics are known to be problematic, most notably F1, but nevertheless F1 is widely used. Objective: To investigate the potential impact of using F1 on the validity of this large body of research. Method: We undertook a systematic review to locate relevant experiments and then extract all pairwise comparisons of defect prediction performance using F1 and the un-biased Matthews correlation coefficient (MCC). Results: We found a total of 38 primary studies. These contain 12,471 pairs of results. Of these, 21.95% changed direction when the MCC metric is used instead of the biased F1 metric. Unfortunately, we also found evidence suggesting that F1 remains widely used in software defect prediction research. Conclusions: We reiterate the concerns of statisticians that the F1 is a problematic metric outside of an information retrieval context, since we are concerned about both classes (defect-prone and not defect-prone units). This inappropriate usage has led to a substantial number (more than one fifth) of erroneous (in terms of direction) results. Therefore we urge researchers to (i) use an unbiased metric and (ii) publish detailed results including confusion matrices such that alternative analyses become possible.

Martin Shepperd, Stephen G. MacDonell

Context: Software engineering has a problem in that when we empirically evaluate competing prediction systems we obtain conflicting results. Objective: To reduce the inconsistency amongst validation study results and provide a more formal foundation to interpret results with a particular focus on continuous prediction systems. Method: A new framework is proposed for evaluating competing prediction systems based upon (1) an unbiased statistic, Standardised Accuracy, (2) testing the result likelihood relative to the baseline technique of random 'predictions', that is guessing, and (3) calculation of effect sizes. Results: Previously published empirical evaluations of prediction systems are re-examined and the original conclusions shown to be unsafe. Additionally, even the strongest results are shown to have no more than a medium effect size relative to random guessing. Conclusions: Biased accuracy statistics such as MMRE are deprecated. By contrast this new empirical validation framework leads to meaningful results. Such steps will assist in performing future meta-analyses and in providing more robust and usable recommendations to practitioners.

Jingxiu Yao, Martin Shepperd

Context: There is considerable diversity in the range and design of computational experiments to assess classifiers for software defect prediction. This is particularly so, regarding the choice of classifier performance metrics. Unfortunately some widely used metrics are known to be biased, in particular F1. Objective: We want to understand the extent to which the widespread use of the F1 renders empirical results in software defect prediction unreliable. Method: We searched for defect prediction studies that report both F1 and the Matthews correlation coefficient (MCC). This enabled us to determine the proportion of results that are consistent between both metrics and the proportion that change. Results: Our systematic review identifies 8 studies comprising 4017 pairwise results. Of these results, the direction of the comparison changes in 23% of the cases when the unbiased MCC metric is employed. Conclusion: We find compelling reasons why the choice of classification performance metric matters, specifically the biased and misleading F1 metric should be deprecated.

Martin Shepperd, Yuchen Guo, Ning Li et al.

Context: Conducting experiments is central to research machine learning research to benchmark, evaluate and compare learning algorithms. Consequently it is important we conduct reliable, trustworthy experiments. Objective: We investigate the incidence of errors in a sample of machine learning experiments in the domain of software defect prediction. Our focus is simple arithmetical and statistical errors. Method: We analyse 49 papers describing 2456 individual experimental results from a previously undertaken systematic review comparing supervised and unsupervised defect prediction classifiers. We extract the confusion matrices and test for relevant constraints, e.g., the marginal probabilities must sum to one. We also check for multiple statistical significance testing errors. Results: We find that a total of 22 out of 49 papers contain demonstrable errors. Of these 7 were statistical and 16 related to confusion matrix inconsistency (one paper contained both classes of error). Conclusions: Whilst some errors may be of a relatively trivial nature, e.g., transcription errors their presence does not engender confidence. We strongly urge researchers to follow open science principles so errors can be more easily be detected and corrected, thus as a community reduce this worryingly high error rate with our computational experiments.

Ning Li, Martin Shepperd, Yuchen Guo

Background: Unsupervised machine learners have been increasingly applied to software defect prediction. It is an approach that may be valuable for software practitioners because it reduces the need for labeled training data. Objective: Investigate the use and performance of unsupervised learning techniques in software defect prediction. Method: We conducted a systematic literature review that identified 49 studies containing 2456 individual experimental results, which satisfied our inclusion criteria published between January 2000 and March 2018. In order to compare prediction performance across these studies in a consistent way, we (re-)computed the confusion matrices and employed the Matthews Correlation Coefficient (MCC) as our main performance measure. Results: Our meta-analysis shows that unsupervised models are comparable with supervised models for both within-project and cross-project prediction. Among the 14 families of unsupervised model, Fuzzy CMeans (FCM) and Fuzzy SOMs (FSOMs) perform best. In addition, where we were able to check, we found that almost 11% (262/2456) of published results (contained in 16 papers) were internally inconsistent and a further 33% (823/2456) provided insufficient details for us to check. Conclusion: Although many factors impact the performance of a classifier, e.g., dataset characteristics, broadly speaking, unsupervised classifiers do not seem to perform worse than the supervised classifiers in our review. However, we note a worrying prevalence of (i) demonstrably erroneous experimental results, (ii) undemanding benchmarks and (iii) incomplete reporting. We therefore encourage researchers to be comprehensive in their reporting.

Martin Shepperd

I welcome the contribution from Falessi et al. [1] hereafter referred to as F++ , and the ensuing debate. Experimentation is an important tool within empirical software engineering, so how we select participants is clearly a relevant question. Moreover as F++ point out, the question is considerably more nuanced than the simple dichotomy it might appear to be at first sight. This commentary is structured as follows. In Section 2 I briefly summarise the arguments of F++ and comment on their approach. Next, in Section 3, I take a step back to consider the nature of representativeness in inferential arguments and the need for careful definition. Then I give three examples of using different types of participant to consider impact. I conclude by arguing, largely in agreement with F++, that the question of whether student participants are representative or not depends on the target population. However, we need to give careful consideration to defining that population and, in particular, not to overlook the representativeness of tasks and environment. This is facilitated by explicit description of the target populations.

Martin Shepperd, Carolyn Mair, Magne Jørgensen

CONTEXT: The role of expert judgement is essential in our quest to improve software project planning and execution. However, its accuracy is dependent on many factors, not least the avoidance of judgement biases, such as the anchoring bias, arising from being influenced by initial information, even when it's misleading or irrelevant. This strong effect is widely documented. OBJECTIVE: We aimed to replicate this anchoring bias using professionals and, novel in a software engineering context, explore de-biasing interventions through increasing knowledge and awareness of judgement biases. METHOD: We ran two series of experiments in company settings with a total of 410 software developers. Some developers took part in a workshop to heighten their awareness of a range of cognitive biases, including anchoring. Later, the anchoring bias was induced by presenting low or high productivity values, followed by the participants' estimates of their own project productivity. Our hypothesis was that the workshop would lead to reduced bias, i.e., work as a de-biasing intervention. RESULTS: The anchors had a large effect (robust Cohen's $d=1.19$) in influencing estimates. This was substantially reduced in those participants who attended the workshop (robust Cohen's $d=0.72$). The reduced bias related mainly to the high anchor. The de-biasing intervention also led to a threefold reduction in estimate variance. CONCLUSIONS: The impact of anchors upon judgement was substantial. Learning about judgement biases does appear capable of mitigating, although not removing, the anchoring bias. The positive effect of de-biasing through learning about biases suggests that it has value.

Tim Menzies, Martin Shepperd

CONTEXT: There has been a rapid growth in the use of data analytics to underpin evidence-based software engineering. However the combination of complex techniques, diverse reporting standards and poorly understood underlying phenomena are causing some concern as to the reliability of studies. OBJECTIVE: Our goal is to provide guidance for producers and consumers of software analytics studies (computational experiments and correlation studies). METHOD: We propose using "bad smells", i.e., surface indications of deeper problems and popular in the agile software community and consider how they may be manifest in software analytics studies. RESULTS: We list 12 "bad smells" in software analytics papers (and show their impact by examples). CONCLUSIONS: We believe the metaphor of bad smell is a useful device. Therefore we encourage more debate on what contributes to the validty of software analytics studies (so we expect our list will mature over time).

Martin Shepperd

CONTEXT: There is growing interest in establishing software engineering as an evidence-based discipline. To that end, replication is often used to gain confidence in empirical findings, as opposed to reproduction where the goal is showing the correctness, or validity of the published results. OBJECTIVE: To consider what is required for a replication study to confirm the original experiment and apply this understanding in software engineering. METHOD: Simulation is used to demonstrate why the prediction interval for confirmation can be surprisingly wide. This analysis is applied to three recent replications. RESULTS: It is shown that because the prediction intervals are wide, almost all replications are confirmatory, so in that sense there is no 'replication crisis', however, the contributions to knowledge are negligible. CONCLUSIONS: Replicating empirical software engineering experiments, particularly if they are under-powered or under-reported, is a waste of scientific resources. By contrast, meta-analysis is strongly advocated so that all relevant experiments are combined to estimate the population effect.