Counterfactual Effect Subtraction for Debiased Recommendation

143 concepts78 papers1 proposal

When a user clicks on a recommended item, standard recommendation algorithms draw a straight line of inference: the user clicked the item, so the user must like the item. This observational correlation forms the bedrock of collaborative filtering.

However, this assumption conflates a user's genuine preference with systemic exposure biases. A user might click a video because it is universally viral (popularity bias), because it was placed at the absolute top of their feed (position or layout bias), or simply because its thumbnail is highly alluring (visual bias). When models train on these unadjusted historical logs, they inadvertently optimize for these shortcuts, trapping users in filter bubbles and over-promoting generic content.

To break this cycle, a powerful research direction has emerged that reformulates recommendation through the lens of causal inference. By mapping user interactions as Structural Causal Models (SCMs), researchers can isolate true user intent from superficial triggers using two primary mechanisms: counterfactual effect subtraction during inference and causal deconfounding during training.

The Causal Path Decomposition Framework

To understand why this approach works, we can visualize the recommendation process as a causal graph containing three primary nodes: the User ($U$), the Item ($I$), and the Interaction ($Y$, such as a click or purchase).

In a traditional, unbiased setting, we only care about the path $U \rightarrow Y$ and $I \rightarrow Y$, or their intersection: how well the user's taste aligns with the item's intrinsic properties. But in reality, biases introduce alternative pathways that distort the outcome:

    [ User (U) ] -----------> [ Matching (M) ] -----------> [ Interaction (Y) ]
         |                                                       ^
         |                                                       |
         v                                                       |
    [ Conformity ]                                               |
         |                                                       |
         v                                                       |
    [ Popularity / Visuals / Scenario Bias (B) ] ----------------+

In this graph, we can trace two distinct routes to an interaction:

  1. The Indirect Path ($U \rightarrow M \rightarrow Y$): This represents genuine preference. The user interacts with the item because its specific features match their personal taste.
  2. The Direct Path ($B \rightarrow Y$): This represents the shortcut or bias. The user interacts with the item purely because of an external factor—such as its high popularity, its visual flashiness, or its prominent layout placement—regardless of whether it fits their actual needs.

The total score generated by a standard machine learning model represents the Total Effect (TE) of the user and item on the click probability. Because the direct and indirect paths are entangled, a high TE score could be driven entirely by the direct bias path.

The core intuition of counterfactual debiasing is that we can isolate the indirect path by simulating a hypothetical, counterfactual world: What would the interaction score be if the item retained its bias, but the user-item matching signal was completely nullified?

This simulated score captures the Natural Direct Effect (NDE) of the bias alone. By subtracting this baseline from the Total Effect, we isolate the Total Indirect Effect (TIE), which reflects the user's true, unbiased preference:

$$\text{TIE} = \text{TE} - \text{NDE}$$

This simple yet elegant subtraction allows recommendation engines to train on highly biased historical data but surgically remove those distortions when ranking items in real time.

Subtracting Direct Biases in Practice

This framework has proven remarkably versatile, as different forms of recommendation bias can be mapped directly to specific causal paths and subtracted during inference.

Popularity and Conformity Bias

The authors of Model-Agnostic Counterfactual Reasoning for Eliminating Popularity Bias in Recommender System pioneered this approach by decomposing the prediction score into three distinct components: user-item matching, item-specific popularity, and user-specific conformity. By calculating the NDE—the score an item receives simply for being trending—and subtracting it from the total prediction, they successfully leveled the playing field for niche items.

This concept was expanded to news recommendation by Xiangfu He et al. in HPDM: A Hierarchical Popularity-aware Debiased Modeling Approach for Personalized News Recommender. Recognizing that popularity operates hierarchically (e.g., a catchy title versus a trending broad category), the authors used counterfactual reasoning to estimate and subtract a user's conformity effect, simulating user behavior in an environment free from herd mentality.

Visual and Presentation Biases

In visually-driven platforms, users frequently click items because the thumbnail looks appealing, a phenomenon known as visual bias. In CausalRec: Causal Inference for Visual Debiasing in Visually-Aware Recommendation, Ruihong Qiu et al. modeled the direct effect of an item's image on the click probability. By calculating the TIE, they subtracted the superficial visual attraction from the total interaction signal, ensuring final rankings were based on actual product compatibility rather than mere aesthetic appeal.

Environmental and Scenario Biases

When a platform serves recommendations across hundreds of different app environments or layouts, the scenario itself acts as a source of visibility bias. Yichao Wang et al. addressed this in M-scan: A Multi-Scenario Causal-driven Adaptive Network for Recommendation. They applied counterfactual inference during the prediction phase to isolate and subtract the direct influence of environmental factors (like layout or placement size) from the final output, ensuring that cross-scenario recommendations reflect intrinsic user interest rather than layout-driven visibility.

Data Completeness and Model Propensity

This framework even extends to data artifacts and model-induced noise. For instance, multimodal recommenders naturally favor items with complete data profiles over those with missing modalities. Jin Li 0002 et al. introduced Generating with Fairness: A Modality-Diffused Counterfactual Framework for Incomplete Multimodal Recommendations, which uses counterfactual inference to subtract the bias score attributed to the mere existence of complete multimodal content.

Similarly, when using Large Language Models (LLMs) to generate recommendation features, the models introduce their own stereotypical biases. Guixian Zhang demonstrates in Mitigating Propensity Bias of Large Language Models for Recommender Systems that calculating the pure natural direct effect of the LLM's internal tendencies allows for the surgical removal of unrelated propensity noise, leaving behind pure user preferences.

Deconfounding via Causal Interventions

While counterfactual subtraction operates primarily as a filter during the inference stage, a complementary strategy focuses on the training stage: using causal interventions to block backdoor paths.

In causal graph theory, a confounder is a variable that simultaneously influences both the treatment (which items are exposed to the user) and the outcome (whether the user clicks). Confounders create "backdoor paths" that inject spurious correlations into the training data. To eliminate this, researchers use Judea Pearl’s do-calculus—specifically the backdoor adjustment formula—to simulate a randomized controlled trial from purely observational data.

    [ Confounder (e.g., Video Age / Historical Popularity) ]
         /                                            \
        /                                              \
       v                                                v
  [ Exposure ] -----------------------------------> [ Click (Y) ]

Consider a video's age or release interval. Older videos naturally receive fewer impressions and fewer clicks simply because they are old, creating a release interval bias. Lulu Dong et al. tackled this in Not All Videos Become Outdated: Short-Video Recommendation by Learning to Deconfound Release Interval Bias. By treating the release interval as a statistical confounder and applying a backdoor adjustment, they mathematically isolated a video's intrinsic quality from its publication recency, preventing the system from blindly penalizing high-quality "evergreen" content.

Similarly, item popularity frequently acts as a confounder that distorts user history. In Taming Recommendation Bias with Causal Intervention on Evolving Personal Popularity, Shiyin Tan et al. utilized backdoor adjustments to isolate the direct effect of item quality on user interest, stripping away the artificial boost in clicks caused by high historical exposure.

This intervention strategy is equally vital in conversational systems, where the context of dialogue can introduce heavy confounding noise. In Do Mentioned Items Truly Matter? Enhancing Conversational Recommender Systems with Causal Intervention and Large Language Models, Lingzhi Wang et al. applied backdoor adjustments to decouple a user's true underlying preferences from items they merely mentioned in passing or in a negative context, preventing the system from making recommendations based on irrelevant historical dialogue noise.

Why Causal Debiasing Works

Traditional statistical debiasing methods, such as Inverse Propensity Weighting (IPW), typically rely on auxiliary models to estimate the probability that a user will be exposed to an item. These methods are notoriously sensitive; if the propensity model is slightly inaccurate, the resulting recommendation scores suffer from high variance and instability.

Causal path decomposition and interventions bypass this limitation by shifting the focus toward learning invariant representations. Instead of trying to guess exposure probabilities, these frameworks force the recommendation model to learn user and item embeddings that are structurally invariant to the presence of bias.

For example, Jun Wang et al. developed CounterCLR: Counterfactual Contrastive Learning with Non-random Missing Data in Recommendation, which enforces representation consistency between observed and counterfactual scenarios via contrastive learning. By ensuring that user representations remain invariant regardless of whether data is missing or present, the model extracts true preferences without requiring expensive, unbiased gold-standard datasets.

Furthermore, framing recommendation as a structural causal problem makes systems significantly more robust to out-of-distribution (OOD) shifts—situations where the real world changes between training and deployment (such as a sudden shift in global trends). As demonstrated by Chu Zhao et al. in Graph Representation Learning via Causal Diffusion for Out-of-Distribution Recommendation, incorporating causal backdoor adjustments strips away unstable environmental correlations. What remains are the invariant user preferences that hold true across different environments, ensuring long-term stability and genuine personalization.

Key Foundational Work

Estimating Attention Flow in Online Video Networks (2019) — contributed 2 concept elements to this direction.
History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 89 papers
  2. Version 1 · 2026-10-04 · 107 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_counterfactual_bias_mitigation_ca81c8a1,
  title = {counterfactual bias mitigation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/counterfactual-bias-mitigation/rdr_ca81c8a1a2ef424b9561a030fab62f3c},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 2, frozen text):

@misc{lacuna_counterfactual_bias_mitigation_ca81c8a1_v2,
  title = {counterfactual bias mitigation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/counterfactual-bias-mitigation/rdr_ca81c8a1a2ef424b9561a030fab62f3c/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake