An Approach for Model Assissment for Activity Recognition1,2

摘要

This paper discusses quality metrics as well as procedure for parameters optimization and model assessment for human activity recognition based on sensors signals. We compare micro and macro performance measures in multiclass classification as well as various cross-validation techniques. The paper introduces general concept of Dual Leave-Group-of-Sources-Out cross-validation procedure. This technique provides reliable way for model parameters optimization in practical applications and prevents overestimation of recognition quality from point of view generalization capability.