Searching for high-temperature ambient-pressure superconductors is a challenge in materials science. Machine learning has a promising application in materials discovery. A data-driven approach that overcomes low-data limitations by computationally inexpensive descriptors based on the Bardeen-Cooper-Schrieffer (BCS) theory and semi-supervised learning was proposed. The accuracy of the classification mode is 72%. This approach can screen over 10 000 binary and ternary BCS compounds in the Material Project database, thus identifying some promising superconductors at ambient pressure. The compounds in B-C and B-C-N systems have a maximum superconducting critical temperature (TC) of 60 K, which is greater than that for MgB2 (i.e., TC=39 K)