2021/07/15 by Saptarshi Bej, Bej, Saptarshi, Kristian Schultz +7
Computer Science · Health Professions · #Artificial Intelligence (cs.AI) #Artificial Intelligence in Healthcare #FOS: Computer and information sciences #Face and Expression Recognition #Imbalanced Data Classification Techniques #Machine Learning (cs.LG)
paper · pdf · doi:10.48550/arxiv.2107.07349
openalex publication_date 2021/07/15 · openalex created_date 2022/07/25 · openalex updated_date 2026/07/28
Over 85 oversampling algorithms, mostly extensions of the SMOTE algorithm,\nhave been built over the past two decades, to solve the problem of imbalanced\ndatasets. However, it has been evident from previous studies that different\noversampling algorithms have different degrees of efficiency with different\nclassifiers. With numerous algorithms available, it is difficult to decide on\nan oversampling algorithm for a chosen classifier. Here, we overcome this\nproblem with a multi-schematic and classifier-independent oversampling\napproach: ProWRAS(Proximity Weighted Random Affine Shadowsampling). ProWRAS\nintegrates the Localized Random Affine Shadowsampling (LoRAS)algorithm and the\nProximity Weighted Synthetic oversampling (ProWSyn) algorithm. By controlling\nthe variance of the synthetic samples, as well as a proximity-weighted\nclustering system of the minority classdata, the ProWRAS algorithm improves\nperformance, compared to algorithms that generate synthetic samples through\nmodelling high dimensional convex spaces of the minority class. ProWRAS has\nfour oversampling schemes, each of which has its unique way to model the\nvariance of the generated data. Most importantly, the performance of ProWRAS\nwith proper choice of oversampling schemes, is independent of the classifier\nused. We have benchmarked our newly developed ProWRAS algorithm against five\nsate-of-the-art oversampling models and four different classifiers on 20\npublicly available datasets. ProWRAS outperforms other oversampling algorithms\nin a statistically significant way, in terms of both F1-score and Kappa-score.\nMoreover, we have introduced a novel measure for classifier independence\nI-score, and showed quantitatively that ProWRAS performs better, independent of\nthe classifier used. In practice, ProWRAS customizes synthetic sample\ngeneration according to a classifier of choice and thereby reduces benchmarking\nefforts.\n