Churn Prediction
Churn prediction identifies customers, subscribers, or users who are likely to leave. This guide explains why graph neural networks (GNNs) can capture relational signals that traditional flat-table models miss, with examples from gaming and a practical implementation path.
Source title: Churn Prediction with Graph Neural Networks: A Complete Guide
Originally published: July 11, 2023
Authors: Kumo team and Ivaylo Bahtchevanov
What is churn and why it matters
Churn is the loss of a customer, subscriber, or user. The definition varies by industry: a subscription cancellation, an app uninstall, a lapsed purchase pattern, or a deactivated account. Acquiring a new customer costs 5 to 25 times more than retaining an existing one, which makes even small improvements in retention disproportionately valuable.
Churn takes many forms
- Hard churn: Explicit actions such as cancellations, unsubscribes, account closures, or app uninstalls.
- Soft churn: Gradual disengagement measured by declining sessions, fewer purchases, reduced time spent, or dropping click-through rates.
Soft churn is harder to detect because there is no single trigger event. It requires identifying patterns of declining activity across multiple signals over time.
Re-engaging an at-risk customer is less expensive than re-acquiring them after they leave. Even marginal retention improvements can materially improve profitability by converting low-engagement users into loyal, high-value customers.
The gaming industry: a case study in extreme churn
Gaming is an especially demanding environment for churn prediction. Game developers spend approximately $15 billion annually on player acquisition, yet 75% of players churn within 24 hours and 90% churn within 30 days. Games with millions of daily active users generate hundreds of millions of contextual data points from in-game actions, social interactions, purchase behaviors, session patterns, and progression metrics. The intervention window is often the first few play sessions.
Analogy: Churn prediction is like a doctor monitoring vital signs. One metric, such as blood pressure or heart rate, is informative, but diagnostic power comes from correlating multiple signals over time. Similarly, purchase frequency alone is informative, but combining it with session behavior, support interactions, and social connections reveals the full picture.
Traditional approaches to churn prediction
Most churn prediction relies on tabular ML models trained on hand-engineered features. A data scientist aggregates customer behavior into a flat feature table - days since last purchase, average order value, session count over 30 days, and support-ticket frequency - then trains a gradient-boosted model such as XGBoost or LightGBM.
Regression analysis
The simplest approach uses logistic regression or survival analysis on aggregate features. In gaming, that can mean sessions per week, in-app purchases, and level progression. These models can break when game mechanics change, and their features must be manually recalibrated when the product evolves.
Deep learning on flat features
LSTMs and attention networks can capture temporal patterns better than static features, but still rely on handcrafted feature pipelines. Creating these features is time-intensive, can introduce feature-designer bias, and is difficult to scale across millions of daily interactions. Each new use case requires a new feature set and training round.
The common limitation
Traditional approaches operate on a single flat feature table. To get there, relational data - customers, transactions, products, sessions, and support tickets - is flattened into one row per customer. That discards relational structure, multi-hop patterns, and temporal dynamics across tables.
The source reports that a data scientist with a Stanford CS Master’s degree and five years of experience needs approximately 12.3 hours and 878 lines of code for one prediction task. The process restarts for every new question.
The limitation is not the algorithm alone. Flattening relational tables into a feature row discards multi-hop, temporal, and structural signals that can drive churn behavior.
Why relational and graph approaches work better
Customer churn signals are scattered across tables: declining transaction frequency in orders, reduced session duration in activity, unresolved support issues, and changing product preferences in the catalog. Foreign keys connect those signals, and their combination tells a fuller story.
From tables to graphs
GNNs represent data as entities - customers, products, sessions, and transactions - connected by relationships such as purchased, viewed, contacted support, or played with. Rather than flattening these connections, a GNN learns directly from them. Each entity aggregates information from its neighbors to capture relationship context.
Churn risk can therefore depend on a customer’s own behavior, similar customers, products they buy, and support-interaction patterns. These multi-hop signals are often lost during feature engineering.
Analogy: Traditional ML predicts whether someone will leave a party by looking only at their facial expression. Graph-based ML also observes who they arrived with, who they are talking to, how engaged those people are, and whether their friends have already left.
Key advantages of graph-based churn prediction
- No manual feature engineering: The model learns which relationships and patterns matter directly from the data.
- Multi-hop pattern discovery: GNNs traverse paths such as customer -> orders -> products -> categories to find combinations that are difficult to manually encode.
- Cold-start capability: A new user’s limited history can be interpreted through their connections to similar entities, acquisition channel, first actions, or shared attributes.
- Temporal dynamics: Graph-based models can distinguish purchases that are accelerating from purchases that are decelerating, rather than only counting totals.
What the benchmarks show
On the H&M retail dataset, which has three tables and 16.6 million rows, the source describes a graph-based approach that finds churn signals a flat model misses. It identifies the interaction between low order count, no fashion-news subscription, and no club membership as a churn driver. The source reports that, on the RelBench user-churn task for this dataset, graph-based models achieve 69.88 AUROC compared with 55.21 for LightGBM with manual features.
The source reports graph-based models outperforming manually engineered feature pipelines by 15 to 18 AUROC points on real-world churn tasks. It attributes the improvement to retaining relational structure instead of compressing it into a flat table.
Churn prediction in gaming
Gaming combines massive data volumes, extreme churn rates, and intervention windows measured in hours rather than weeks. Players gradually stop logging in, reduce session length, stop purchasing, or abandon progression. Signals are distributed across gameplay data, social features, economic systems, and engagement metrics.
In a multiplayer game, a player can be connected to teammates, opponents, guild members, and friends. If a player’s guild becomes inactive, their churn probability can rise even if their individual metrics look healthy.
Real-world examples
The source describes Rovio, maker of Angry Birds with more than 10 million daily active users, using ML to predict churn and dynamically adjust game difficulty. When a player is likely to churn, a game can reduce difficulty, offer rewards, or trigger social features to re-engage them. The source also notes GNN use in state-of-the-art agents for games such as StarCraft.
The cold-start problem in gaming
New players have little or no behavioral history, which limits traditional models. Graph-based models can use connections such as acquisition channel, first actions, and patterns among players who entered the game in comparable ways. This is crucial when the first few sessions determine whether a player stays or leaves.
In gaming, a small improvement in early retention can materially increase average lifetime value. Graph-based models can leverage social connections, gameplay context, and similar-player patterns to predict churn with minimal user history.
How Kumo approaches churn prediction
Kumo’s approach moves from raw relational data to production churn predictions through a process that operates directly on connected tables.
Predictive Query Language for churn
Kumo’s Predictive Query Language (PQL) defines churn predictions without feature-engineering code.
Session-based churn
Predict active users who will have zero sessions in the next 90 days:
Transaction-based churn
Predict active customers who will stop transacting:
Scenario testing with interventions
Test the impact of a coupon on churn probability:
The ASSUMING clause can simulate an intervention, such as asking what churn would look like if a customer received a coupon in the next seven days.
PQL separates what to predict from how to predict it. Teams define the business question while the platform handles feature discovery, model selection, and training.
Understanding what drives churn
Churn predictions are most useful when teams can act on them. The source describes explainability at multiple levels to reveal not only who will churn but why they may be at risk.
Feature contribution analysis
Evaluation dashboards can show which features contribute most to churn predictions, including transaction frequency, postal code, membership status, session recency, and product-category preferences. For H&M, the source highlights the combination of order count, fashion-news subscription status, and club membership as a primary churn signal.
Temporal and segment-level analysis
Rather than only tracking days since last order, graph models can represent whether purchases are accelerating, decelerating, or clustering. Teams can compare performance with baseline models, identify patterns across user segments, and diagnose churn drivers such as feature gaps, expectation misalignment, or competitive pressure.
From prediction to action
- Outreach optimization: Determine which customers respond best to each retention offer.
- Segment response prediction: Forecast how user segments will react to product changes.
- High-value customer identification: Prioritize retention efforts for customers with the highest potential lifetime value.
- Gaming applications: Personalization, ARPU maximization, LTV forecasting, difficulty adjustment, and NPS improvement from the same platform and data graph.
Analogy: Traditional churn models give a list of at-risk customers. Graph-based models can add a diagnosis, such as a shifted purchasing pattern after a preferred product category was discontinued, combined with similar customers leaving recently. That diagnosis turns a prediction into an action plan.
Practical implementation
Deploying graph-based churn prediction in production requires connecting data, defining the prediction task, and configuring an AutoML pipeline.
Step 1: Graph construction
Connect customer information, transactions, subscriptions, product catalogs, and session logs through primary and foreign-key relationships. The source describes Kumo maintaining and updating the graph as data arrives and scaling to dozens of tables with terabytes of data and tens of billions of rows.
Step 2: Define predictions with PQL
Write predictive queries that specify the business question. The platform handles feature discovery without SQL aggregation pipelines, manual feature stores, or manual feature engineering.
Step 3: Configure AutoML
Step 4: Evaluate and iterate
Review the evaluation dashboard to compare performance with baselines, analyze feature contributions, and validate predictions across user segments. The source describes testing different churn definitions, such as session-based versus transaction-based churn, and 30-day or 90-day windows in parallel.
Step 5: Deploy and monitor
The source describes native Kumo deployment on Snowflake through Snowpark Container Services and Databricks through Lakehouse App. The Python SDK provides customization for integrating predictions into existing workflows.
Beyond churn: a reusable platform
Once data is connected, the same graph can support churn prediction, recommendation, lifetime-value forecasting, fraud detection, demand prediction, and personalization without rebuilding pipelines. For gaming, the source describes one platform supporting player churn, in-game recommendations, dynamic pricing, matchmaking optimization, and engagement scoring.
The operational benefit described by the source is reusability. A connected data graph can support multiple PQL prediction tasks, avoiding separate feature-engineering pipelines for every new question.