Causal Disentanglement of User Preference and Popularity Bias

This direction is split into other directions: interest-conformity entanglement. Its last version stays readable and citable here.

Version 1 of 1149 papers205 conceptsfrozen 2026-10-04

When a user clicks on a movie or buys a product, recommendation systems typically record this as a "positive signal." However, this signal is deeply ambiguous. Did the user click because they genuinely liked the item’s attributes, or simply because it was trending at the top of their feed?

Traditional collaborative filtering conflates these motivations into a single representation. Recent research has shifted toward causal disentanglement, a paradigm that uses causal inference to separate a user’s intrinsic interest from external confounding factors like popularity, social conformity, and exposure bias.

The Confounding Nature of Popularity

In the language of causal inference, item popularity acts as a confounder. As Yang Zhang and Xiangnan He argue in Causal Intervention for Leveraging Popularity Bias in Recommendation, popularity influences both the probability that an item is shown to a user (exposure) and the probability that the user interacts with it.

If a system doesn't account for this, it learns a "spurious correlation": it assumes users like popular items because of their content, when they might actually be clicking due to high visibility. This creates a feedback loop where popular items get even more exposure, while niche "long-tail" items—which might be a better match for a user's true taste—are buried. Domokos M. Kelen and András A. Benczúr demonstrate that decoupling these signals using Bayesian inference allows models to surface personalized content that global trends would otherwise obscure.

Disentangling Interest from Conformity

To solve this, researchers have developed frameworks to split a user’s behavior into two distinct latent factors: intrinsic interest and conformity.

The DICE framework, introduced by Yu Zheng and Chen Gao in Disentangling User Interest and Conformity for Recommendation with Causal Embedding, uses causal embeddings to isolate these drivers. By training separate embeddings for interest and popularity-driven conformity, the model becomes robust to shifts in trends. If a blockbuster movie stops trending, a disentangled model still knows which users liked it for its genre versus those who just followed the crowd.

Similarly, the DIRAC framework by Qidong Liu and Feng Tian applies this to session-based recommendations, treating popularity as a legitimate behavioral signal rather than just noise. This allows the system to satisfy both a user's niche preferences and their natural human drive to stay current with trends.

Beyond Clicks: Relevance and Context

The need for disentanglement extends beyond popularity. As early as 2005, Yi Zhang and Jamie Callan noted that topical relevance and user preference are distinct concepts. A document might be relevant to a search query but still not be "liked" by the user due to style or quality.

Modern e-commerce search models like DRP, proposed by Yejing Wang and Xiangyu Zhao, use causal modeling to reconstruct the behavior space, separating short-term query relevance from long-term user preference. This prevents a single "emergency purchase" (like buying a generic umbrella during a storm) from biasing a user's long-term profile toward generic goods.

Other environmental factors also act as confounders:

  • Positional Bias: Hila Becker showed that ad clicks are heavily influenced by page position, necessitating models that separate the probability of observing an item from the probability of liking it.
  • Geographical Bias: In location recommendation, UKGC uses counterfactual learning to "subtract" the effect of mere proximity, identifying places a user actually wants to visit rather than just the ones that are nearby.

Personalization in the Age of LLMs

This causal perspective is now being applied to Large Language Models (LLMs). In NextQuill, Xiaoyan Zhao and Juntao You use causal preference modeling to distinguish between words a user writes because of their unique personal style and words necessitated by general grammar rules. By isolating the causal influence of a user's identity, they achieve much higher alignment in personalized text generation.

Why It Works: The Power of Counterfactuals

The intuition behind these methods is often counterfactual reasoning: asking "Would the user have clicked this item if it weren't popular?"

By answering this question, models like DCRec by Da Luo can refine self-supervised signals in sequential recommendation. Instead of treating every item in a sequence as equally representative of taste, the model weights interactions with unpopular or niche items more heavily, as these are "high-fidelity" indicators of true preference. As summarized by Yaochen Zhu and Jundong Li in their survey of causal strategies, this shift from correlation to causation allows systems to be more explainable, robust to environment changes, and ultimately more helpful to the user.

Papers in this version

Showing 60 of 149 papers.

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_causal_disentanglement_of_user_preferenc_40939136,
  title = {Causal Disentanglement of User Preference and Popularity Bias},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/causal-disentanglement-of-user-preference-and-popularity-bias/rdr_40939136ab804f3b85e004a872f7ead8},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_causal_disentanglement_of_user_preferenc_40939136_v1,
  title = {Causal Disentanglement of User Preference and Popularity Bias},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/causal-disentanglement-of-user-preference-and-popularity-bias/rdr_40939136ab804f3b85e004a872f7ead8/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}