06Data · ML Track

Deep, Graph & Federated ML

When a deep, graph or federated method beats a classical one on network data — and when it does not.

Training  /  Course 06

Overview Lab Environment Topics & Cases

The honest case for deep learning in a telco: when a deep, graph or federated method genuinely beats the classical baseline from Course 05 — and, just as often, when it does not. CNNs/ROCKET for KPI time series, sequence models and self-attention for signalling ladders, graph neural nets that localize a fault on the service-based topology itself, deep anomaly detection, federated learning across sites that never share raw data, and a trust layer — conformal prediction, drift detection and a human in the loop. Offline, no GPU required.

You leave able to: You can pick the right model for a telco problem on evidence, and wrap it in the calibration and drift checks that make it safe to trust.

Aims — what you can do after

  • Decide when a deep, graph or federated method beats a classical one — and when it does not.
  • Apply a CNN to time series and use convolutional / ROCKET features to classify or score KPIs.
  • Model signalling as sequences — turn call ladders into features, and read a self-attention map.
  • Localize faults with a graph neural network over the service-based-architecture topology.
  • Detect anomalies with a deep autoencoder, and compare it fairly to PCA and Isolation Forest.
  • Train a model with federated learning across simulated sites, and explain what non-IID data does.
  • Wrap any model in a trust layer: conformal prediction, drift detection and a correct human-in-the-loop.
  • Translate every method to its production (PyTorch) form and to the reference analytics function.

The decision that opens the course

  Is the signal in STRUCTURE a classical model can't see?
   |
   +-- a shape over time (KPI curve) ........... CNN / ROCKET
   +-- an ordered sequence (call ladder) ....... attention / sequence model
   +-- a topology (which NF talks to which) .... graph neural network
   +-- 'normal' is complex & unlabelled ........ deep autoencoder
   |
   +-- small, flat, tabular data .............. a tree still wins — use it
   #  every deep result is judged against a classical baseline, honestly

The first question the course trains — reach for deep learning only when it earns its keep.

The deep toolkit, mapped to network data

CNN · time series → Attention · ladders → GNN · topology → Autoencoder · anomaly

Each family matches a shape in the data: a curve, a sequence, a graph, or a hard-to-model “normal”. The course builds each one and compares it fairly to a simpler method.

Federated learning — train without moving the data

Local trainEach site / slice trains on its own data — the data never leaves.
↓
Send weightsOnly the model updates go to the server — not the traffic, not the subscribers.
↓
AverageThe server combines them (FedAvg) into one shared model.
↓
BroadcastThe improved model goes back to every site — and the loop repeats.

Sites keep their traffic private and still build one shared model. You also see what non-IID data (every site different) does to that average.

The trust layer around any model

Prediction → Conformal set → Drift check → Human decides

A prediction is not enough: conformal prediction gives calibrated confidence, a drift check flags when the model has gone stale, and a human owns the consequential calls.

Module breakdown

0
Foundation & the synthetic telemetry engine25 min
The stack and the data generator — the course runs fully offline, no GPU.
1
Why deep? Representations vs hand-crafted features30 min
What deep learning buys you, and what it costs.
2
CNNs for time series45 min
The convolution mechanism and fast ROCKET features for KPI curves.
3
Sequence models & self-attention45 min
Signalling ladders as sequences; reading an attention map.
4
Graph neural networks45 min
Fault localization over the service-based-architecture topology.
5
Deep anomaly detection40 min
Autoencoders vs PCA vs Isolation Forest — a fair comparison.
6
Federated learning45 min
FedAvg across sites and slices, and the non-IID problem.
7
Trust35 min
Conformal prediction, drift detection and the human in the loop.

What this course adds to your ML

  • Reach for deep learning only when data and structure justify it — and prove it against a baseline.
  • Read a self-attention map and a graph model's node scores as explanations, not black boxes.
  • Ship a model with calibrated confidence and a drift alarm, not just an accuracy number.
← Course 05 — Data Engineering & ML Course 07 — IMS Roaming & Security →