Imagine walking into a coffee shop and ordering a highly advertised, trendy seasonal latte. Did you buy it because you genuinely love the flavor profile, or because it was splashed across every billboard and social media feed?
Traditional recommendation algorithms struggle with this exact question. When a user clicks on an item, standard collaborative filtering treats that interaction as a pure signal of personal preference. However, real-world behavior is heavily confounded by social conformity—the tendency to follow the crowd, react to external visibility, or succumb to peer influence. This creates an interest-conformity entanglement. If an algorithm mistakes a trend-driven click for genuine interest, it creates a feedback loop: it continues to recommend popular items, drowning out niche preferences and causing the system to fail when global trends inevitably shift.
To build more robust recommendation systems, researchers have developed frameworks to separate why a user interacts with an item into two distinct latent factors: intrinsic interest and external conformity.
The Causal Graph of Choice
To understand how we can separate interest from conformity, it helps to view user behavior through the lens of causal inference. In a standard recommendation system, we observe an interaction (like a click or a purchase), but we do not observe the underlying motivation.
We can model the interaction as the outcome of two competing causal pathways:
- The Intrinsic Path: User Taste $\rightarrow$ Item Attributes $\rightarrow$ Interaction.
- The Conformity Path: Item Popularity $\rightarrow$ Social Pressure/Visibility $\rightarrow$ Interaction.
Tianxin Wei et al. formalized this by isolating item popularity as a direct effect in a causal graph. By leveraging counterfactual reasoning, their framework (MACR) asks a fundamental question during inference: Would the user still have clicked this item if it were not popular? By mathematically subtracting the direct effect of popularity, the system uncovers the user's true, unbiased preference for the item's features.
This causal separation is crucial because clicks on trending content are frequently driven by herd mentality rather than long-term alignment with a user's taste. For example, Xiangfu He et al. applied counterfactual reasoning to news recommendations, demonstrating that separating viral headline conformity from genuine topic interest prevents recommendation feeds from becoming monopolized by transient, clickbait trends.
Architecting Dual Latent Spaces
How do we represent this separation inside a machine learning model? Instead of compressing a user's identity into a single embedding vector, modern architectures split user and item representations into separate, orthogonal latent spaces.
Standard Embedding: [ User Preference / Popularity Bias Confounded ]
│
▼
Disentangled Spaces: [ Intrinsic Interest ] X [ Social Conformity ]
In this dual-embedding paradigm, a user is represented by two vectors: one capturing their stable, long-term interest in item categories, and another capturing their susceptibility to social trends. Qing Li et al. implemented this by splitting user interaction embeddings into distinct latent factors. They utilized clustering-based prototypes to anchor the embeddings, effectively separating actions motivated by an item's position in the popularity distribution from actions driven by personal taste. Similarly, Yu Zheng et al. introduced the DICE framework, which uses specific user-item-item triplets to force the interest and conformity embeddings to learn independent, non-interfering signals.
This concept of splitting internal traits from external circumstances extends beyond popularity bias:
- Contextual and Environmental Factors: Cheng Yang et al. generalized this architecture into intrinsic traits (persistent internal preferences) and extrinsic circumstances (transient external contexts), allowing models to predict how a user's behavior changes when their situation shifts but their core tastes remain the same.
- Geographical Constraints: For point-of-interest recommendations, Huayu Li separated intrinsic personal tastes from extrinsic geographical constraints, distinguishing whether a user avoided a restaurant due to a lack of interest or simply because it was too far away.
- Educational Needs vs. Peer Influence: In online learning environments, Qingqing Liang et al. disentangled "cognitive homophily" (a student's actual knowledge gaps) from "social homophily" (following what classmates are studying), preventing peer noise from distorting educational resource recommendations.
Training Disentangled Models via Self-Supervision
The core challenge of training these dual-embedding models is that our training data consists entirely of raw, unlabeled interactions. We see that a user clicked an item, but the data never explicitly states why.
To train these models without explicit labels, researchers leverage contrastive learning and targeted data masking. The intuition is to create alternative "views" of the user's history that isolate one factor while holding the other constant.
For instance, Xiangfu He et al. selectively masked historical user interactions based on item popularity during training. By isolating highly popular engagements from niche, long-tail engagements during the representation learning phase, the model forces the conformity embedding to capture trend-following behavior and the interest embedding to capture specific personal preferences.
Because popularity is dynamic, Cheng Yang et al. introduced time-sensitive negative sampling in a contrastive learning framework (DPA). By treating shifting global trends as a dynamic background variable, they decoupled stable, long-term user interests from transient viral spikes. To handle the inherent data sparsity of the long-tail items that receive few clicks, Weiqi Zhao et al. combined causal graph structures with self-supervised contrastive loss, ensuring that both the interest and conformity pathways receive robust gradients even when interaction data is scarce.
The Payoff: Robustness, Discovery, and Clarity
Disentangling interest from conformity is not just an exercise in architectural elegance; it fundamentally changes how recommendation systems perform in production environments.
First, it grants robustness to trend shifts. When an item's popularity abruptly drops, standard models suffer from performance degradation because they relied on that transient trend to make predictions. Disentangled models retain an unadulterated representation of the user's core interests, allowing them to make accurate recommendations in a popularity-neutral environment, as demonstrated by Yu Zheng et al..
Second, it enables long-tail discovery. When popularity bias is stripped away from the user-item matching process, the recommendation engine can surface highly relevant, niche items that would otherwise be overshadowed by viral content (Qing Li et al.).
Finally, it provides behavioral clarity across different interaction types. In social networks, public actions and private actions stem from entirely different motivations. As Fanjin Zhang et al. discovered when analyzing WeChat user behaviors, private content consumption (clicking to read) and public endorsement (sharing with friends) are driven by distinct social mechanisms. Disentanglement provides the mathematical framework necessary to respect these boundaries, leading to systems that truly understand the difference between what we consume for ourselves and what we do to conform with the crowd.
Key Foundational Work
LA-CTR: A Limited Attention Collaborative Topic Regression for Social Media (2013) — contributed 1 concept elements to this direction.
Modeling Contextual Factors of Click Rates (2007) — contributed 1 concept elements to this direction.