Predicting a future quantitative value (congestion level) from historical, labeled data is a classic supervised regression task. The other options are less fit: clustering (unsupervised) finds patterns but doesn’t predict future values directly; reinforcement learning is for sequential decision-making, not one-shot prediction from static history; rule-based systems lack the ability to learn complex patterns from data.

Trust in AI often breaks down due to its ‘black box’ nature. Providing explanations for predictions (XAI) helps human operators understand the ‘why’ behind an alert, allowing them to contextualize false positives and build intuition about the model’s behavior. This is more effective and practical than seeking perfect accuracy, which is often unattainable, or abandoning the tool altogether.

Many ML algorithms (e.g., those using gradient descent, or distance-based like k-NN) are sensitive to the scale of input features. A feature with a broad range (0-1000 ms) can disproportionately influence the model compared to one with a small range (0-1 loss ratio). Standardization (mean=0, std=1) or normalization puts all features on a comparable scale, leading to more stable and faster training and better model performance.

The DMG is the foundational data layer. Its core responsibility is to manage the data pipeline: collecting raw network data from various sources, cleaning and storing it, maintaining a catalog so users know what data is available, and governing its usage. This reliable, managed data supply is critical for any subsequent AI/ML modeling or analytics.

A gradual degradation in model performance over time in a live environment is a classic sign of data drift (or concept drift). The real-world data distribution shifts (e.g., new user behavior, new services, network upgrades), so the model’s assumptions based on old training data become less valid. Continuous monitoring for drift and periodic retraining are necessary to maintain model relevance.

HITL is a risk mitigation and trust-building strategy. For high-stakes decisions (e.g., modifying slice resources for emergency services), allowing human oversight ensures accountability, provides a final sanity check, and helps operators build confidence in the AI’s reasoning. The goal is often to move towards full automation as trust is earned, not to permanently hinder the process.

This is a class imbalance problem. A model trained on such data can achieve misleadingly high overall accuracy by always predicting the majority class, while its precision and recall for the critical minority classes (like ‘fiber cut’) will be terrible. Techniques like resampling (oversampling minority/undersampling majority), using different evaluation metrics (F1-score, AUC-PR), or algorithmic cost-sensitive learning are required to address this.

ML systems are dynamic. A new model trained on new data can be deployed at any time. Versioning is critical for reproducibility, auditability, and reliable operations. If a new model version causes issues, teams must be able to quickly identify what changed and revert to a known-good version, just as they would with application code.

Real-time anomaly detection requires low-latency analysis of continuous data streams. Stream processing frameworks (e.g., Apache Flink, Kafka Streams) are designed for this, allowing the application to compute statistics, run models, and flag anomalies on-the-fly, enabling proactive mitigation. Batch processing introduces latency unsuitable for immediate response.

Predicting a future quantitative value (congestion level) from historical, labeled data is a classic supervised regression task. The other options are less fit: clustering (unsupervised) finds patterns but doesn’t predict future values directly; reinforcement learning is for sequential decision-making, not one-shot prediction from static history; rule-based systems lack the ability to learn complex patterns from data.0