This work investigates uncertainty-aware early prediction of TCP congestion events using sender-side socket telemetry collected under controlled Mininet emulation. Two supervised datasets are constructed for TCP Reno and TCP Cubic from a total of 600 single-bottleneck scenarios (300 per congestion-control algorithm), covering diverse conditions through variations in bandwidth, round-trip time, queue size, and background traffic mixes. Socket snapshots are sampled every 30 ms and include key TCP state variables such as congestion window, RTT, RTO, retransmission counter, pacing rate, slow-start threshold, and derived temporal features. Each sample is labeled as an imminent congestion event when a local maximum of the congestion window is followed by a decrease, targeting the pre-reduction phase preceding congestion manifestation.
We compare a standard point-logit LSTM with an Interval LSTM that produces per-class logit intervals. The proposed model generates 12 latent criteria logits per class and aggregates them via min/max operators to form interval bounds. Classification relies on midpoint logits, while interval widths are trained using an auxiliary calibration objective derived from midpoint probabilities, combined with a width regularization term to prevent degenerate solutions. This design enables the model to provide both predictions and a sample-wise reliability indicator.
Experiments are conducted using stratified 70/15/15 train/validation/test splits, median imputation, feature standardization, class-weighted cross-entropy, and early stopping. On the Reno dataset, the standard LSTM achieves 0.877 accuracy and 0.863 macro-F1, while the Interval LSTM achieves 0.879 accuracy and 0.864 macro-F1. On the Cubic dataset, the standard LSTM obtains 0.908 accuracy and 0.888 macro-F1, compared to 0.907 accuracy and 0.888 macro-F1 for the Interval model. Reliability analysis shows that interval width correlates positively with prediction error, with predicted-class width–error correlations of approximately 0.660 for Reno and 0.712 for Cubic. Probability RMSE values are 0.281 (Reno) and 0.246 (Cubic), confirming meaningful alignment between interval width and predictive uncertainty.
These results demonstrate that interval-valued logits provide a practical reliability signal without degrading classification performance. In a congestion-control context, predicted probabilities can trigger proactive responses, while interval width can inform the confidence of such actions, enabling more conservative behavior under uncertainty and improving robustness to variable network conditions.
