Koller & Friedman · 2009
Probabilistic Graphical Models: Principles and Techniques. A comprehensive reference on model representation, inference and learning.
MIT Press ↗NETWORK SCIENCE · INTERACTIVE DEMONSTRATION
When one node
moves, who feels it?
A synthetic network makes one idea visible: a change in one influential member can alter the risk signal around them.
Explore the message-passing intuition behind network-aware churn analysis. The graph and values below are invented for demonstration.
Explore the simulation ↓WEIGHTED TIES→PASSING SIGNALS→CHANGING EXPOSURE
01 / PLAY WITH THE NETWORK
Start with one signal.
Choose a starting node. Step through a few rounds and watch its neighbors—and their neighbors—receive weighted messages.
SYNTHETIC NETWORKROUND 0
12 MEMBERS / 18 TIESLower exposureRising exposureElevated exposureTie strength
02 / THE MECHANISM
A simple
message rule.
Each member sends a fraction of their current signal along each tie. A noisy-OR combines incoming messages while preserving an individual baseline.
pᵢ (t+1)=1 − (1 − pᵢ⁰) ∏j ∈ N(i) (1 − β · wⱼᵢ · pⱼ(t))
pᵢ⁰synthetic baseline risk for member i
βthe message-strength control
wⱼᵢtie strength from member j to i
pⱼ(t)member j’s signal in the current round
This compact rule is for explanation, not inference. Real churn models need validated data, leakage controls, calibration, time-aware evaluation, governance and a clear intervention policy.
03 / THE PROFESSIONAL STORY
Built from a real question.
Shown with a new example.
At Nextel, I developed a proprietary graph-based churn model called Favorites. It drew on network influence, message-passing ideas and probabilistic graphical models to examine how churn by an influential member could change risk across their close network. I explored graph tooling including NetworkX and graph databases.
≈0.80–0.85Historical ROC-AUC, as recalled by the author.
The original model, evaluation data and artifacts are unavailable, so this figure has not been independently verified or reproduced here.
Important boundaryThis public demo is a new, synthetic teaching example. It is not the proprietary Favorites algorithm, its source code, its parameters, or a reproduction of its performance. No customer, carrier or subscriber data is included.
04 / RESEARCH CONTEXT
Ideas with a research trail.
These sources ground the modeling concepts and the telecom use case. They are context for this demonstration, not validation of the historical Favorites model.
Kschischang, Frey & Loeliger · 2001
“Factor graphs and the sum-product algorithm.” A foundational treatment of local message passing for probabilistic inference. The demo’s noisy-OR update is a different, simplified rule.
IEEE Transactions on Information Theory ↗Verbeke et al. · 2014
“Social network analysis for customer churn prediction.” Studies relational learning and network effects in real telecom churn data.
Applied Soft Computing ↗Óskarsdóttir et al. · 2017
“Social network analytics for churn prediction in telco.” Across eight call-detail datasets, network construction affected results; collective inference did not improve performance in their benchmark.
Expert Systems with Applications ↗Predictive association does not by itself establish causal peer influence. These publications provide scientific context; they do not verify or reproduce Favorites or its recalled 0.80–0.85 ROC-AUC.
KEEP EXPLORING