Stay or Stray - A Dynamical Systems Viewpoint of Popularity Bias

The paper Stay or Stray – A Dynamical Systems Viewpoint of Popularity Bias studies the emergence of popularity bias in recommendation systems through a dynamical systems lens. The authors formulate a model with two user classes—a majority (popular) class and a minority (niche) class—where the recommender system is an online learner that updates its parameters to match user preferences, while users adjust their arrival rates based on preference match. The system evolves on two timescales: the recommendation algorithm adapts rapidly with each user interaction, while the user population adjusts its arrival rate slowly. This two-timescale analysis permits treating arrival rates as quasi-static relative to model parameter changes, yielding a simplified ODE over arrival rates alone.

The authors characterize the equilibrium structure of this reduced system. Under a non-collinearity condition on class-conditional feature means, the four corners of [0,1]2 are valid equilibrium points: (1,1) corresponds to retention of both user classes, (1,0) and (0,1) correspond to selective disengagement of one class (popularity bias), and (0,0) corresponds to simultaneous disengagement. The analysis yields three principal theoretical results:

First, Theorem 4.1 proves that the equilibrium (0,0) is not asymptotically reachable from any interior initial condition, ruling out simultaneous disengagement as a feasible long-run outcome. The proof shows that convergence to (0,0) would require both TPR and TNR below 0.5, which violates the Cauchy-Schwarz inequality.

Second, Theorem 4.2 derives a threshold p* on the fraction of majority users such that when p > p*, the system asymptotically converges to (1,0)—the niche class disengages entirely while the majority class is retained—formally characterizing the onset of popularity bias. The threshold is given by:

p* = (α−1∥µ−1∥2 A−1 α) / (α−1∥µ−1∥2 A−1 α + α1⟨µ1, µ−1⟩ A−1 α)

Third, the authors identify two complementary regimes that guarantee symmetric retention (convergence to (1,1)): Theorem 4.3 provides a sufficient condition based on the sign of the inner product of the class-conditional means in a covariance-transformed space (specifically, ⟨µ1, µ−1⟩ N−1 α < 0 where N α = pα1Σ1 + (1-p)α−1Σ−1 + λI), and Theorem 4.4 provides an existence result identifying a feasible interval of p values ensuring both classes are retained, even when the sufficient condition of Theorem 4.3 does not hold.

The paper also demonstrates a reversal of trend phenomenon where the system's initial trajectory toward (1,1) can reverse and converge to (1,0) as p* evolves with the state.

The authors validate their theoretical predictions through simulations on synthetic data, demonstrating that system trajectories conform to predicted equilibrium behavior across a broad range of initializations and parameter configurations. They also construct a semi-synthetic evaluation using production engagement logs from a large-scale commercial music recommendation platform comprising approximately 410 million user–item interactions. By fitting their model to user embeddings derived from matrix factorization on this interaction data, they show that observed platform dynamics—including disproportionate churn among niche-preferring users—are consistent with the conditions identified by their framework. The real-world data analysis reveals that close to half of niche-heavy users churn out while only a quarter of popular-heavy users churn out.

The paper concludes by suggesting a natural mitigation strategy based on equalizing error rates across the two classes, similar in spirit to prior work, and empirically demonstrates its effectiveness in reducing popularity bias. Future work may extend the framework to general convex loss functions, relax assumptions on user feature distributions, and extend to multi-user class dynamics.

Improvements to AI Systems:

1. Adaptive Fairness-Aware Recommender with Dynamic Churn Prediction: The AI system can now explicitly model and predict user churn risk in real time as a function of the recommender’s own learning dynamics, not just static user features. By incorporating the two-timescale ODE framework, the system can detect early warning signs of niche-class disengagement (e.g., when the fraction of majority users p exceeds the threshold p*) and proactively adjust its loss function to rebalance error rates (TPR/TNR) between user classes, preventing the irreversible convergence to (1,0) popularity bias.

2. Online Calibration of Recommendation Algorithms via Error-Rate Equalization: The AI system can implement a closed-loop control mechanism that continuously monitors the true positive rate (TPR) and true negative rate (TNR) for each user class during online learning. When the system detects that the Cauchy-Schwarz-violating regime (TPR<0.5 and TNR<0.5) is approached, it can inject a corrective regularization term—e.g., adding a penalty proportional to the difference in class-wise error rates—to steer the trajectory away from the unstable (0,0) equilibrium and toward symmetric retention (1,1), even when the sufficient condition of Theorem 4.3 is not met.

3. Population-Aware Recommendation with Slow-Fast Timescale Decomposition: The AI system can separate its internal parameter updates (fast timescale) from user population arrival-rate shifts (slow timescale) to predict long-term equilibrium outcomes without expensive simulations. This allows the system to precompute the stability region of (1,1) for a given user feature distribution and then, at deployment, adjust its feature representation (e.g., via a learnable linear transform) to ensure that the inner product ⟨µ1, µ−1⟩ N−1 α becomes negative, thereby guaranteeing symmetric retention by construction.

4. Proactive Niche-User Retention via Threshold Monitoring: The AI system can compute the threshold p* in real time from the current model parameters and user embeddings. When the observed fraction of majority users approaches p*, the system can trigger a targeted intervention—such as reweighting the loss for niche-class samples or adjusting the learning rate for niche-relevant features—to shift the effective p below the threshold, preventing the system from sliding into the popularity-bias equilibrium.

5. Simulation-Driven Policy Selection for Multi-Class Recommenders: The AI system can use the reduced ODE (over arrival rates only) as a fast surrogate model to evaluate candidate mitigation policies (e.g., different regularization strengths, feature transforms, or loss reweightings) before deploying them. This enables the system to run thousands of policy simulations in milliseconds, selecting the one that maximizes the basin of attraction for (1,1) across a range of initial conditions and user heterogeneity, without needing full user-level simulations.

6. Churn-Aware Content Curation for Music/Media Platforms: The AI system can now forecast disproportionate churn among niche-preferring users (as observed in the 410M-interaction dataset) and automatically adjust recommendation diversity—e.g., by mixing niche items into popular-user feeds—to equalize TPR/TNR across classes. The system can also generate personalized “retention nudges” (e.g., recommending a niche item to a majority user) based on the predicted trajectory toward (1,0), thereby flattening the churn asymmetry.

What the improved AI system can do:

• Self-correcting recommender: It can detect and reverse its own drift toward popularity bias before it becomes irreversible, using the theoretical thresholds as early-warning signals.

• Guaranteed fairness under mild conditions: It can enforce symmetric retention (both user classes stay engaged) by actively shaping its feature space or loss landscape to satisfy the non-collinearity and inner-product conditions from Theorems 4.3 and 4.4.

• Explainable churn prediction: It can attribute user churn to the recommender’s own dynamics (e.g., “niche users disengage because p crossed p*”), rather than blaming user preferences, enabling more effective interventions.

• Fast policy iteration: It can evaluate and deploy fairness-aware mitigation strategies in near-real-time using the reduced ODE, making it suitable for large-scale production systems with millions of interactions.

• Robust to distribution shift: It can anticipate how changes in user arrival rates (e.g., due to external events) will affect the equilibrium, and preemptively adjust its learning rate or regularization to maintain a stable, inclusive equilibrium.

Popularity bias in recommendation systems arises when a majority user class generates disproportionate interaction data, causing the system to increasingly favour it while degrading recommendation quality for niche users. While extensive empirical evidence of popularity bias exists, the dynamics leading to its emergence are not well understood. In this work, we study the coupled evolution of recommender model updates and user engagement through the lens of dynamical systems. We formulate a stochastic process and analyse its asymptotic behaviour through an ordinary differential equation (ODE) framework grounded in two-time-scale stochastic approximation. We characterise the equilibrium points of this dynamical system, and derive conditions under which popularity bias is provably emergent, as well as conditions under which symmetric retention of all user classes is possible. We conduct experiments on synthetic data and real-world production logs derived from a large-scale commercial music recommendation platform to validate our theoretical results.