Fednet Traffic Forecasting Eval

fednet-traffic-forecasting-eval

FedNET: Federated Learning for Proactive Traffic Management and Network Capacity Planning — Panda et al. (2025) (arXiv:2511.06797, 2025)

What this evaluates

Evaluates the ability of a federated learning model to perform multi-step time-series forecasting of node-level traffic in optical networks, measuring how prediction accuracy degrades with longer history windows and forecasting horizons.

Datasets

• Optical network traffic dataset — total ?; splits: train (-1), val (-1), test (-1)

Metrics

• R² score (primary) — range: (-∞, 1]
  • Coefficient of determination: $1 - \frac{\sum (y_i - \hat{y}_i)^2}{\sum (y_i - \bar{y})^2}$, where $y_i$ are true values, $\hat{y}_i$ are predictions, and $\bar{y}$ is the mean of true values.

Input / output format

Input: Sliding window time-series samples of historical node traffic with length $h$.

Output: Predicted traffic values for the next $p$ time intervals.

Scoring recipe

def compute_r2(y_true, y_pred):
    ss_res = sum((y - y_hat)**2 for y, y_hat in zip(y_true, y_pred))
    ss_tot = sum((y - mean(y_true))**2 for y in y_true)
    return 1 - (ss_res / ss_tot)

# Per client evaluation
client_r2 = [compute_r2(test_y, test_pred) for client in clients]
# Aggregate across clients
avg_r2 = sum(client_r2) / len(client_r2)

# Also report MSE
mse = mean((y_true - y_pred)**2)

Common pitfalls

• Data is split per client (non-IID), so the global test set is formed by aggregating individual client test sets rather than a single global split.
• Sliding window generation creates overlapping samples; evaluation must respect the temporal order and not shuffle the data.
• Performance drops significantly as prediction horizon $p$ increases, so comparing different $p$ values requires separate baselines and careful interpretation of error propagation.

Evidence (verbatim from paper)

Figures[3] and[4] illustrate the effect of the temporal window sizes on the model performance, using violin plots that demonstrate the distribution of clients’ $R^{2}$ scores. In Fig.[3], the history window length varies ($h=1,4,8,12$) while keeping the prediction window constant ($p=1$); conversely, Fig.[4] explores different prediction windows $(p=1,4,8,12)$ with a fixed history length ($h=1$). The red horizontal lines in each violin plot correspond to the mean $R^{2}$ scores shown numerically in Table[II], which provide comprehensive average scores for all $(h,p)$ pair combinations.

Citation

@misc{panda2025fednet,
  title={FedNET: Federated Learning for Proactive Traffic Management and Network Capacity Planning},
  author={Panda et al. (2025)},
  year={2025},
  note={arXiv:2511.06797}
}

• arXiv: 2511.06797