An Invitation to Greater Use of Matthews Correlation Coefficient in Robotics and Artificial Intelligence

A binary classification is a computational procedure that labels data elements as members of one or another category. In machine learning and computational statistics, input data elements which are part of two classes are usually encoded as 0's or -1's (negatives) and 1's (positives). During a binary classification, a method assigns each data element to one of the two categories, usually after a machine learning phase. The computational procedure then usually creates a 2 × 2 contingency table called confusion matrix, where the positive elements correctly predicted positive are called true positives (TP), the negative elements correctly predicted negative are called true negatives (TN), the positive elements wrongly labeled as negatives are called false negatives (FN), and the negative elements wrongly labeled as positives are called false positives (FP).Since it would be difficult to always analyze the four categories of the confusion matrix for each test, scientists defined statistical rates that summarize TP, FP, FN, and TN in one value. Accuracy (Equation 1), for example, is a rate that indicates the ratio of correct positives (Zliobaite, 2015), while F 1 score (Equation 2), is the harmonic mean of positive predictive value and true positive rate (Huang et al., 2015;Lipton et al., 2014). accuracy = T P + T N T N + T P + F P + F N(1) (worst value = 0; best value = 1)F 1 score = 2 • T P 2 • T P + F P + F N(2)(worst value = 0; best value = 1)Even if accuracy and F 1 score are very common in machine learning studies, they can be misleading (Chicco and Jurman, 2020) in several situations.The Matthews correlation coefficient (Equation 3) (Matthews, 1975), instead, is the only statistical rate that generates a high score only if the values of the four basic rates (sensitivity, specificity, precision, negative predictive value) are high (Zhu, 2020;Yao and Shepperd, 2020). For this reason, the MCC results being more informative and reliable than accuracy, F 1 score, and many other rates (Jurman et al., 2012;Chicco, 2017;Chicco and Jurman, 2020;Chicco et al., 2021;Chicco et al., 2021a,b). Robotics). The average number of articles including MCC results in these ten journals is 1.40 (Table 1).On the contrary, we found hundreds and thousands of articles mentioning the accuracy rate (Table 1), ranging from 135 articles of Science Robotics to 2,390 studies published in IEEE Robotics and AutomationLetters. The average number of articles including accuracy results in these ten journals is 1,077.2 (Table 1).The number of articles including the F 1 score was smaller than the accuracy ones, but definetely more than the MCC studies. The number of F 1 score articles ranged from none (Science Robotics) to 71 (IEEE Robotics and Automation Letters), with an overall average value of 19.30 (Table 1). Almost all the journals had at least ten published articles containing results measured by F 1 score, except IEEE Transactions on Robotics with eight articles and the already mentioned Science Robotics with zero.Our results clearly show that the Matthews correlation coefficient is almost unknown in robotics. F 1 score is clearly underused with respect to accuracy, but it is still known for all the journals except Science Robotics.The MCC, instead, is clearly out of radar for most of the robotics researchers that published articles in these ten robotics journals. The MCC is unknown probably also to the reviewers and the associate editors who handled the review of these manuscripts and did not invite the authors to include results measured by this statistical rate.All the authors of all the manuscripts published in five robotics journals (Frontiers in Robotics and AI, IEEE Robotics and Automation Letters, IEEE Transactions on Robotics, Robotics and Intelligent Systems, and Science Robotics) decided not to include any result measured by the MCC.Regarding Frontiers in Artificial Intelligence, we notice that the Matthews correlation coefficient was employed by the authors of three original research studies (Li et al., 2021;Wu et al., 2021;Bhatt et al., 2021), two methods articles (Fletcher et al., 2021;Weerawardhana et al., 2022), and one review (Tripathi et al., 2021). The study of Li et al. (2021) presents a deep learning application on chemoinformatics data for the prediction of carcinogenicity. Chemical data analysis is also the topic of the article by Wu et al. (2021), which employs natural language processing techniques for drug labeling and indexing. Fletcher et al. (2021), instead, present a study on fairness in artificial intelligence applied to public health, reporting a case study on machine learning applied to data of pulmonary disease. Weerawardhana et al. (2022) employed the MCC to measure the results in a human-aware intervention and behavior classification study,In their review article, Tripathi et al. (2021) reported some AI best practices in manufacturing, indicating the MCC as one of the confusion matrix rates employed in this field.Among the five articles published in the Robotics and Autonomous Systems journal, three are about robots' visual activities (Bosse and Zlot, 2009;Özbilge, 2016, 2019, one is about swarm robotics (Lau et al., 2011), and one is about human-robot verbal interaction (Grassi et al., 2022).The only article of Frontiers in Neurorobotics including results measured by the MCC is a study on visual perception of robots (Layher et al., 2017), while the only MCC study in International Journal of RoboticsResearch describes a dataset on urban point cloud obtained acquired by mobile laser scanning (Roynard et al., 2018). The article of the Journal of Field Robotics including MCC results is about the robotics visual obstacle detection (Santana et al., 2011). The presence of the MCC in these studies does not seem to follow a precise trend, but rather be occasionally employed by authors who are aware of MCC's assets, for reasons we do not know.Regarding dates, it is interesting to notice that, except one article published in 2009 and one in 2011, all the other studies were published after 2016, showing an increased interest towards the Matthews correlation coefficient. Eight articles out of fourteen have been published in 2021 and 2022, suggesting a greater use of the MCC in future studies.As we explained earlier, the amount of articles including MCC results is very low compared to the number of published studies involving accuracy and F 1 score (Table 1). And we think this is a serious drawback: as we explained in our study (Chicco and Jurman, 2020), the Matthews correlation coefficient is more informative and reliable than accuracy and F 1 score, because it takes into account the ratio of positive data instances, negative data instances, positive predictions, and negative predictions.Accuracy and F 1 score both range between 0 and 1, with 0 meaning worst result possible and 1 meaning perfect prediction. An accuracy value of 0.9 and a F 1 score of 0.95, for example, suggest a very good binary classification. If the original dataset consisted of 91 positive elements and 9 negative elements, these results could be generated by a cracked classifier that labels everything as positive. Please notice: if your classifier assigned the "positive" label to all the 100 data elements, you would get accuracy = 0.9 and F 1 score = 0.95, which are clearly misleading results and could let you think your binary classification was excellent. The MCC, instead, would have been -0.03, that in the [−1, +1] interval indicates a poor prediction similar to random guessing: the MCC would inform you that your binary classification was quite bad, while accuracy and F 1 score tried to make you believe it was great.We therefore invite the robotics and artificial intelligence communities to include results measured through the MCC for any binary classification analysis.Bosse, M. and Zlot, R. (2009). Keypoint design and evaluation for place recognition in 2D lidar maps.Robotics and Autonomous Systems 57, 1211-1224 Frontiers

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.