IBN-DriftV2: A Multi-Dimensional Benchmark Dataset for Intent Drift

Our dataset is an enhanced version of the publicly available IBN-Drift benchmark, extending its data generation pipeline to improve the representation of path-semantic drift scenarios. Specifically, compared with the original IBN-Drift dataset, we substantially increase the proportion of path drift samples while preserving the original data format, feature representation, and labeling strategy. This enhancement alleviates the severe class imbalance of path-semantic violations in the original benchmark and provides a more suitable dataset for evaluating path-related intent drift detection and prediction methods.<br/><br/>The enhanced dataset includes traces generated on two representative network topologies, GÉANT (22 nodes and 58 edges) and Abilene (12 nodes and 15 edges). Following the original benchmark, the dataset is divided into training, validation, and test sets. For GÉANT, the three splits contain 7,400, 1,586, and 1,586 sequences, respectively, while the corresponding numbers for Abilene are 10,807, 2,270, and 2,271. The overall drift ratios remain stable across different splits, ranging from 40.5%–40.7% for GÉANT and 34.7%–35.1% for Abilene, ensuring a consistent data distribution.<br/><br/>Each sequence is annotated with three independent intent-drift dimensions, namely performance drift, path-semantic drift, and energy drift. In the GÉANT topology, the dataset contains 29.9% performance drift samples, 9.9%–10.0% path-semantic drift samples, and 25.5% energy drift samples across all splits. For the Abilene topology, the corresponding proportions are 23.2%–24.0%, 11.2%–11.4%, and 21.2%–21.7%, respectively. Compared with the original IBN-Drift benchmark, the proportion of path-semantic drift has been significantly increased, resulting in a more balanced multi-dimensional label distribution while maintaining similar distributions for the other drift categories.