🤖 machine learning

Estimating Treatment Effects in Networks under Unknown Exposure Mappings

This paper introduces HINet, a neural network method that leverages expressive graph neural networks and domain-adversarial training to estimate heterogeneous treatment effects in network settings without relying on potentially misspecified, prespecified exposure mappings.

Original authors: Daan Caljon, Jente Van Belle, Wouter Verbeke

Published 2026-08-04

📖 8 min read🧠 Deep dive

CC BY 4.0

Original authors: Daan Caljon, Jente Van Belle, Wouter Verbeke

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to figure out why some plants in a garden grow taller than others. In a simple world, you might just look at the water and sunlight each plant gets. But in a real garden, plants are neighbors. If you water one plant, it might shade its neighbor, or its roots might steal nutrients, or maybe the extra water makes the soil soggy for everyone nearby. This is the tricky world of "interference": where what happens to one thing changes what happens to its friends. Scientists call this "causal inference," and they want to know the true effect of a treatment (like a new medicine or a marketing campaign) without getting confused by these neighborly side effects. Usually, to solve this, researchers have to make a big guess about how the neighbors influence each other. They might assume, for example, that "the more neighbors you have who got the treatment, the more you are affected." But what if that guess is wrong? What if the influence depends on which specific neighbors are treated, or their unique personalities? If the guess is wrong, the whole calculation falls apart.

This is exactly the problem tackled by a new method called HINet, proposed by researchers Daan Caljon, Jente Van Belle, and Wouter Verbeke. Instead of forcing a pre-made guess about how neighbors interact, HINet is like a super-smart detective that learns the rules of the neighborhood on its own. The researchers built a neural network (a type of computer brain) that looks at the entire web of connections and figures out how treatments ripple through the group without needing a manual. They tested this on fake data and real-world social networks, and found that HINet consistently gets the answer right, even when the "rules" of the neighborhood are complex and unknown. In contrast, older methods that rely on guessing the rules often fail miserably when their guess is slightly off. The paper suggests that by letting the computer learn the neighborhood dynamics directly, we can make much better decisions about who to treat, whether in medicine or business, without getting tripped up by the messy reality of social influence.

The Problem: The "Guessing Game" of Neighbors

Imagine you are a doctor trying to test a new vaccine. You give it to some people and not others. In a perfect, isolated world, you could just compare the sick rates of the two groups. But people aren't isolated; they are part of a social network. If your vaccinated friend gets the shot, they might not get sick, but they also won't pass the virus to you. You are protected even though you didn't get the shot. This is interference: the treatment of one person changes the outcome of another.

To measure this, scientists usually use something called an exposure mapping. Think of this as a recipe for how to mix up the "neighborhood influence." A common recipe is simple: "Count how many of your immediate neighbors got the treatment, and divide by the total number of neighbors." If 50% of your friends got the shot, your "exposure" is 0.5. This works great if the influence is truly just a simple average. But in the real world, influence is messy. Maybe you only care if your best friend got the shot, or maybe you are only affected if three specific neighbors got it. If the scientist uses the simple "average" recipe when the real rule is "best friend only," their calculation of the vaccine's effect will be wrong. This is called misspecification, and it's a huge headache because, in most real-world situations, nobody actually knows the true recipe.

The Solution: HINet, the Neighborhood Detective

The authors propose HINet (Heterogeneous Interference Network) to solve this. Instead of asking the scientist to write down the recipe, HINet brings a powerful tool called a Graph Neural Network (GNN).

Think of a GNN as a detective who doesn't just look at one person; they look at the whole party. When HINet tries to predict what will happen to a person, it doesn't just look at whether that person got the treatment. It looks at the person's own traits, and then it scans their entire circle of friends. It asks: "Who in this circle got the treatment? What are their traits? How does that specific mix affect our target?"

Crucially, HINet does this jointly. It learns the "neighborhood representation" (the summary of who is around you and what they are doing) at the exact same time it learns to predict the outcome. It's like a student who learns to solve a math problem by simultaneously figuring out the rules of the game. Because it learns the rules from the data itself, it doesn't need a pre-written exposure mapping. It can discover that the influence depends on specific neighbors, or complex combinations, without being told to look for them.

The Balancing Act: Avoiding the "Fake Friend" Trap

There is a second problem the paper addresses. In real life, people don't get treatments randomly. Maybe the doctor gives the vaccine to healthy people, or a marketing campaign targets rich neighborhoods. This creates a bias called treatment-configuration imbalance. If the treated group is very different from the untreated group, it's hard to tell if the outcome was caused by the treatment or just by the fact that they were different to begin with.

In a network, this gets even weirder. If your friends are all similar to you (a concept called homophily), and you are all likely to get the same treatment, then your "local treatment configuration" (you + your friends' treatments) is predictable based on your traits. HINet uses a clever trick called adversarial training to fix this.

Imagine a game of "Hide and Seek."

The Predictor (HINet's main brain) tries to predict the outcome (e.g., "Will this person buy the product?").
The Adversary (The "Seeker") tries to guess who got the treatment just by looking at the data HINet has processed.

HINet is trained to make the Predictor accurate while simultaneously tricking the Adversary. It wants to create a "representation" of the data that is so balanced that the Adversary cannot tell who got the treatment and who didn't. By doing this, HINet ensures that the differences in outcomes are actually due to the treatment, not because the treated group was just a different type of person to begin with. The paper shows that this network-aware balancing is crucial, especially when the treatment assignment depends heavily on the local environment.

What They Found: The "No-Guess" Method Wins

The researchers tested HINet on several datasets, including fake networks they built to simulate different types of interference (like the "Barabási–Albert" model and "Homophily" models) and real-world data from social networks like Flickr and co-authorship graphs. They compared HINet against other methods that rely on guessing the exposure mapping (like NetEst and TNet) and methods that ignore the network entirely.

The results were clear:

When the guess was right: Methods that guessed the exposure mapping correctly performed well, but so did HINet.
When the guess was wrong: This is where the magic happened. When the researchers changed the rules of the simulation (e.g., making the influence depend on a specific neighbor rather than the average), the guessing methods crashed. Their error rates skyrocketed because their pre-set recipe didn't match reality.
HINet's consistency: HINet performed consistently well across all scenarios. It didn't matter if the interference was simple, complex, or based on specific neighbors. Because it learned the pattern from the data, it didn't get tripped up by a wrong assumption.

The paper also introduced two new ways to measure success called CNEE (Counterfactual Network Estimation Error) and PEHNE (Precision in Estimation of Heterogeneous Network Effects). These metrics check how well a model predicts outcomes across many different possible "what-if" scenarios, not just the one that actually happened. HINet scored the lowest errors on these metrics, proving it is robust.

The Takeaway

The main finding is that you don't need to know the rules of interference to measure its effects. By using a flexible neural network that learns the neighborhood dynamics on the fly and balances the data to remove bias, HINet can estimate treatment effects accurately even when the underlying mechanism is a mystery.

The authors suggest that while older methods are useful if you are absolutely sure about how neighbors influence each other, they are risky in the real world where those rules are unknown. HINet offers a safer, more consistent alternative. It suggests that in the messy, interconnected world of social networks, the best approach is to let the data teach us the rules, rather than forcing our own guesses onto it. The paper doesn't claim to have solved every problem in causal inference, but it provides a strong, evidence-backed tool for navigating the uncertainty of network interference.

Technical Summary: Estimating Treatment Effects in Networks under Unknown Exposure Mappings

Problem Statement

Estimating heterogeneous treatment effects in network settings is complicated by interference, where the outcome of an instance is influenced by the treatment status of its neighbors. While recent causal machine learning methods address interference, they typically rely on a prespecified exposure mapping—a function that aggregates neighbors' treatments into a single variable (e.g., the proportion of treated neighbors) to summarize their influence on an instance's outcome.

This reliance on a prespecified mapping presents two critical issues:

Misspecification: In many real-world scenarios, the mechanism of spillover is unknown or complex (e.g., depending on specific neighbor characteristics rather than just the count of treated neighbors). A prespecified mapping that does not match the underlying mechanism leads to inaccurate estimates.
Treatment-Configuration Imbalance: In observational data, treatment assignment often depends on covariates. Under interference, the relevant "treatment" for a node is its local configuration (its own treatment plus neighbors' treatments). If this configuration distribution varies with the local covariate environment (covariate homophily), certain configurations become rare for specific node types, forcing models to extrapolate and introducing bias.

Existing methods either assume a known, often homogeneous, exposure mapping or attempt to learn mappings through concatenation or attention mechanisms that still require candidate specifications. There is a need for a method that learns the relevant neighborhood representation directly from data without prespecifying how neighbors' treatments combine.

Methodology: HINet

The authors propose HINet (Heterogeneous Interference Network), a neural method designed to estimate treatment effects without a prespecified exposure mapping. The architecture consists of two main components trained jointly:

1. Expressive GNN Outcome Model

HINet employs a Graph Neural Network (specifically a Graph Isomorphism Network, GIN) to learn an exposure-relevant neighborhood representation jointly with outcome prediction.

Input: For each node $i$ , the model takes the node's covariates $X_i$ and treatment $T_i$ , along with the covariates and treatments of its one-hop neighbors $N_i$ .
Mechanism: Instead of aggregating neighbor treatments via a fixed function (like a sum or mean), the GIN learns a representation $h^Y_i$ that captures how the specific configuration of neighbor treatments and covariates affects the outcome.
Flexibility: The GIN is chosen for its theoretical ability to represent count-based exposure mappings and retain information about neighborhood size, making it a robust default when the functional form of interference is unknown.

2. Network-Aware Domain-Adversarial Training

To address treatment-configuration imbalance, HINet incorporates an adversarial objective.

Architecture: A second branch (upper branch) uses a GIN to predict the treatment assignment $T_i$ of node $i$ based on the learned representations of the node and its neighbors.
Gradient Reversal: Gradient Reversal Layers (GRLs) are used to reverse the gradient from the treatment prediction branch during backpropagation. This forces the shared encoder to learn representations that are invariant to the treatment assignment.
Network Awareness: Unlike standard domain adaptation which balances individual node representations, HINet's adversarial branch utilizes neighborhood information (representations of $i$ and $N_i$ ) to predict $T_i$ . This targets the invariance of the local representation $\Phi_i = (\phi_i, \phi_{N_i})$ with respect to the node's own treatment, which, under the paper's causal assumptions, implies marginal invariance with respect to neighbors' treatments as well.

Loss Function

The model is trained by minimizing a combined loss:
$\mathcal{L}_{comb} = \mathcal{L}_y + \alpha \cdot \mathcal{L}_t$
Where $\mathcal{L}_y$ is the outcome prediction loss (MSE), $\mathcal{L}_t$ is the treatment prediction loss (Binary Cross-Entropy), and $\alpha$ controls the strength of the adversarial balancing.

Key Contributions

HINet Architecture: A novel method for estimating heterogeneous treatment effects in networks that does not require a prespecified exposure mapping, learning the exposure-relevant representation jointly with outcome prediction.
Theoretical Motivation & Bounds: The authors derive a population-level generalization bound relating counterfactual estimation error to factual outcome risk and the discrepancy between observed and target distributions of representations and local treatment configurations. This motivates the network-aware adversarial objective.
Evaluation Metrics: Introduction of CNEE (Counterfactual Network Estimation Error) and PEHNE (Precision in Estimation of Heterogeneous Network Effects). These metrics evaluate performance across a distribution of counterfactual networks rather than a single counterfactual assignment, addressing the variability in performance across different treatment rates.
Empirical Robustness: Demonstration that HINet performs consistently across diverse datasets and exposure mappings, whereas methods relying on prespecified mappings degrade significantly when the mapping is misspecified.

Experimental Results

The authors evaluated HINet on five datasets: two fully synthetic (Barabási–Albert and Homophily Sim) and three semi-synthetic (Flickr, BlogCatalog, and Coauthor-CS).

Performance under Misspecification: When the data-generating process (DGP) used an exposure mapping different from the one assumed by competing methods (e.g., using a sum of treatments while competitors assumed a proportion), methods like NetEst and TNet performed poorly. HINet maintained low error rates across all mappings.
Comparison to Baselines: HINet achieved the lowest PEHNE and CNEE on all five datasets under the default exposure mapping. It ranked first or second in 40 out of 50 comparisons across different dataset-mapping combinations.
Sensitivity to Imbalance: In experiments with increasing strength of covariate-dependent treatment assignment, HINet outperformed variants without balancing ( $\alpha=0$ ) and variants without neighborhood information in the adversarial branch. This confirms that network-aware balancing is crucial when treatment configurations are imbalanced due to homophily.
Architecture Choice: While HINet is modular, the GIN architecture was found to be the most consistent performer across different exposure mappings, supporting its selection as the default.

Significance and Claims

The paper claims that HINet offers a practical solution for causal inference in networks where the mechanism of interference is unknown or complex. By avoiding the need to prespecify an exposure mapping, HINet mitigates the risk of bias introduced by model misspecification. The authors emphasize that while the method relies on the assumption of strong ignorability conditional on the local covariate environment, its ability to learn the relevant exposure representation directly from data makes it robust to the "unknown" nature of real-world spillover effects.

The derived generalization bound and the new evaluation metrics (CNEE/PEHNE) provide a theoretical and practical framework for assessing treatment effect estimators in network settings more comprehensively than previous single-counterfactual approaches. The authors conclude that HINet is a strong choice for practice because the true interference mechanisms are typically difficult to specify a priori.