The 35th annual conference on Astronomical Data Analysis Software & Systems 2025 in Görlitz, Germany

We propose the ​Self-paced Ensemble (SPE)​​ algorithm to address extreme class imbalance in astronomical source classification. By leveraging dynamic sample selection and adaptive weighting, SPE enhances rare-class recognition capabilities. Applied to 868,371 ZTF DR4 sources (≥ 30 epochs in g/r bands), SPE identifies ​8,210 high-confidence YSO candidates​ (P ≥ 0.7). Cross-validation with LAMOST DR9 spectra confirms ​238 new YSOs​ showing distinct H𝛼 emission and [He I]/[N II] lines. Further analysis of variability morphology categorizes 657 sources into ​7 classes​ via quasi-periodicity (Q) and flux asymmetry (M) indices, revealing Quasi-Periodic Dippers (QPD, 42.5%) and Bursters (B, 16.9%) as dominant types. We integrate these findings with physical parameters to release the ​ZTF-LAMOST catalog​ (Teff, log L, mass, age, Macc, variability features) and develop an ANN-based evolutionary stage classifier achieving ​96.4% accuracy​ in Class II/III separation (10-band photometry). This work validates SPE’s efficacy in rare-class identification and establishes benchmarks for studying accretion dynamics, dust evolution, and disk physics in star-forming regions.