Home
Machine learning models for workload prediction typically require training data from the same time period as deployment. In CI/CD environments, obtaining representative real-world training data is often difficult due to privacy concerns, data availability constraints, or the need to predict for newly established infrastructure. This paper presents a transfer learning methodology that enables transferring seasonality patterns and workload characteristics from real billing data collected during one-time period to real workflow execution logs from a completely non-overlapping period. The article approach combines two techniques: (1) Quantile Mapping to align duration distributions between the target- and source-domain data, preserving structural characteristics (shape, variability, skewness), and (2) Cross-Temporal Seasonality Transfer to map weekly activity patterns across time windows. Validated on GitHub Actions data spanning two distinct quarters (January-April 2023 billing data transferred to June-October 2023 real execution logs), our method achieved a strong correlation (r=0.8739, p=0.0101) for weekly patterns, explaining 76.4% of the variance in seasonality (R2=0.764), and aligned the duration distribution, reducing the CV error from 0.26 to 0.01, skewness from 1.06 to 0.03, and kurtosis from 2.33 to 0.07 relative to the source domain – all without overlapping temporal coverage between the two data sources.