<p>This study is aimed at developing a physics-informed and data-augmented machine learning (ML) approach which enables computational screening of wear-resistant Co-containing high-entropy alloys (HEAs) under an extremely small-sample-size constraint (<i>N</i> = 10). By integrating conventional HEA thermodynamic descriptors with an explicit precipitate volume fraction, the models capture both the high-entropy effect (entropy-assisted solid-solution strengthening) and precipitation hardening, which govern the hardness and tribological performance of CoCrFeNiW_x, CoCrFeNiMo_x HEAs and the Tribaloy HEAs created by mixing Tribaloy alloys. Initial canonical correlation analysis (CCA) confirms strong multivariate coupling between these descriptors and properties (hardness and wear resistance), while preliminary random forest (RF) models severely overfit the sparse dataset. Property-specific data augmentation strategy is then implemented, expanding the dataset of hardness and wear rate (experimental data on 10 HEAs) to <i>N</i> = 1000, characterized by asymmetric mixup with residual bootstrapping for approximately linear hardness response and ln-space Gaussian process regression (GPR) for the highly nonlinear wear-rate response. Using the augmented datasets, the improved CCA and RF models demonstrate excellent generalization for hardness and high predictive accuracy for wear rate. In predictive validation, the augmented RF model showed better agreement with the experimental measurements than CCA, particularly for the highly nonlinear wear response. The 3D landscapes (property—hardness and wear rate) mapped in (entropy, precipitation) space reveal an entropy–precipitation synergy, with optimal performance concentrated in high-entropy and high-precipitation domain. The proposed approach enables interpretable, data-efficient screening of wear-resistant Co-containing HEAs under severe data scarcity.</p>