Scoring Rounded Data: An Score Smoothing Method for Improving Outlier Detectors

Existing detectors calculate the outlier scores of objects independently, neglecting the consistency between score similarity and object similarity. Consequently, these detectors do not guarantee that the variance in scores among similar normal objects remains minimal. This can lead to normal objects receiving scores that surpass those of outliers, thereby increasing the likelihood of misclassification. Therefore, smoothing outlier scores becomes crucial for enhancing detectors' performance by generating more consistent scores for similar objects. Existing score smoothing techniques exhibit a computational complexity of at least $\mathcal {O}({n}\log {n})$ , where $n$ is the sample size, limiting their utility in scenarios involving large datasets. In response to this challenge, a pragmatic and effective score smoothing method called scoring rounded data (SRD) is proposed. Given a dataset with values between 0 and 1, SRD leverages established outlier detectors to compute outlier scores based on rounded data with reduced precision rather than original data with full precision. By assigning identical scores to objects that yield the same rounding results, which are defined as similar objects, SRD effectively reduces the variance of these scores, thereby facilitating score smoothing. Utilizing detectors on rounded data, which contains fewer bits, can lead to decreased memory usage and computational time during detector computing scores. SRD has a time complexity of $\mathcal {O}({n})$ , making it more suitable for large-scale data. Experiments involving 15 distinct outlier detectors, encompassing both deep-learning (DL)-based and non-DL-based methods, and 20 real-world datasets demonstrated that SRD can improve the average AUROC by 6.91% while using only 26.25% of the original data volume.