Learning from N-Tuple Data with M Positive Instances: Unbiased Risk Estimation and Theoretical Guarantees