샤플리값을 이용한 소규모 파킨슨 음성 데이터에 대한 데이터 평가

초록

In the complex domain of machine learning, this study navigates through the important aspect of data valuation, particularly emphasizing its role in discriminating the relative impact of training samples on model performance within the context of multi- stage Parkinson's disease (PD) classification. The study meticulously evaluates two Data Shapley value estimation methodologies, namely truncated Monte Carlo Shapley (tmc-shapley) and Class-wise Shapley (cs-shapley), aiming to proficiently discriminate between training instances based on their respective in-class and out-of-class contributions. By integrating dysphonia acoustic features and considering the unique speech impairments across various PD stages, the evaluation revealed a superior performance by the cs-shapley method in maintaining model accuracy. This method exhibited a weighted accuracy drop of 0.239, in contrast to the tmc-shapley method which encountered a more pronounced drop of 0.465.