🤖 machine learning

Stay or Stray - A Dynamical Systems Viewpoint of Popularity Bias

This paper employs a dynamical systems framework, specifically a two-time-scale stochastic approximation model, to theoretically characterize the emergence of popularity bias in recommendation systems and derive conditions for its provable occurrence versus symmetric user retention, validated through experiments on synthetic and real-world music platform data.

Original authors: Sarvesh Shashidhar, Lankireddy Prabhat, Arpit Agarwal, D. Manjunath, Karan Bhukar, Tanmay Khandelwal

Published 2026-08-12

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Sarvesh Shashidhar, Lankireddy Prabhat, Arpit Agarwal, D. Manjunath, Karan Bhukar, Tanmay Khandelwal

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 a digital town square where a giant, invisible librarian is constantly trying to guess what books you want to read. This librarian is a "recommendation system," a piece of software that learns your taste by watching what you click on. But here's the catch: the librarian is also watching the crowd. If a huge group of people (the "majority") all love the same pop songs, the librarian starts thinking, "Oh, everyone likes this!" and pushes those songs to everyone. Meanwhile, a smaller group of people who love obscure jazz might get ignored because the librarian is too busy listening to the loud crowd. This is called "popularity bias," and it's a big problem because it makes the system great for the many but terrible for the few.

To understand why this happens, scientists use a branch of math called "dynamical systems." Think of this as a way to study how things change over time when two things are pushing and pulling on each other. In our story, the two things are the librarian (the algorithm) and the crowd (the users). The librarian changes its mind very fast, learning from every single click. The crowd, however, is slower; people don't quit the town square instantly just because they got one bad book recommendation. They stick around for a while, but if the librarian keeps getting it wrong, they eventually leave. This paper asks a simple but deep question: If the librarian and the crowd keep reacting to each other, will the system eventually learn to serve everyone fairly, or will it inevitably get stuck favoring the loud majority and driving the quiet minority away?

The Great Digital Dance: Stay or Stray?

In this paper, the authors treat the relationship between a recommendation system and its users like a complex dance. They want to know: Will the dance partners stay together, or will one partner eventually walk away?

The researchers built a mathematical model to simulate this dance. They imagined two types of dancers: the Majority (popular users who love the hits) and the Minority (niche users who love the obscure stuff). The "music" they dance to is the recommendation algorithm. The algorithm is a fast learner; it updates its moves after every single step. The users are slower dancers; they only decide to leave the dance floor (churn) if the music has been bad for a long time.

The team used a clever trick from math called "two-timescale stochastic approximation." In plain English, this means they treated the algorithm as a hyperactive squirrel that changes its mind constantly, while the users are like slow-moving turtles. Because the squirrel changes so fast, the researchers could figure out exactly what the squirrel would be thinking at any moment based on where the turtles were standing. This allowed them to write down a set of rules (equations) that predict the long-term future of the dance floor.

The Four Corners of the Dance Floor

The researchers discovered that the system can only settle down in four specific "corners" of the dance floor. They mapped these out like a map of possible futures:

The Happy Ending (1, 1): Both the Majority and the Minority stay. Everyone is happy, and the system serves both groups well.
The Popularity Trap (1, 0): The Majority stays, but the Minority leaves. The system becomes obsessed with the popular stuff, and the niche users drift away. This is the dreaded "popularity bias."
The Reverse Trap (0, 1): The Minority stays, but the Majority leaves. (Theoretically possible, but less likely in real life where the majority is, well, the majority).
The Empty Room (0, 0): Everyone leaves. The system fails so badly that no one wants to use it anymore.

What the Math Says: The Rules of the Game

The paper proves some very specific things about how this dance plays out, using rigorous math to back up their claims.

First, the "Empty Room" is impossible.
The authors proved that if the system starts with any users (even just a few), it will never end up in the "Empty Room" where everyone quits at the same time. Even if the system is doing a terrible job, the math shows that at least one group of users will always find something they like enough to stay. The system might get biased, but it won't completely collapse.

Second, the "Popularity Trap" is a real danger.
The researchers found a specific "tipping point" (a number they call $p^*$ ). If the number of popular users in the crowd is higher than this tipping point, the system is mathematically guaranteed to drift toward the "Popularity Trap." The algorithm will get so good at pleasing the majority that it completely ignores the minority, causing the niche users to slowly walk away. It's like a radio station that plays only the top 10 hits because the ratings are high, eventually driving away everyone who likes jazz, rock, or classical music.

Third, there is a way to save the dance.
The paper also found the conditions needed to keep everyone happy (the "1, 1" corner). It turns out that if the musical tastes of the two groups are "different enough" (mathematically, if their average preferences point in opposite directions), the system can learn to serve both. However, if the groups are too similar in a specific way, or if the majority is just too huge, the system might get stuck favoring the majority no matter what.

Testing the Theory in the Real World

To make sure their math wasn't just a pretty theory, the authors tested it in two ways.

First, they ran thousands of computer simulations with fake data. They watched the "turtles" and "squirrels" dance for 100,000 steps. The results matched their predictions perfectly: when the majority was big enough, the niche users left. When the tastes were different enough, everyone stayed.

Second, and perhaps most excitingly, they tested their model on real data from a massive commercial music platform. They looked at about 410 million interactions between users and songs. They found that the real-world data behaved exactly like their model predicted. Users who liked niche music were indeed leaving the platform at a much higher rate than users who liked popular music. The system was, in fact, suffering from the popularity bias their equations had described.

The Solution: Balancing the Books

So, what's the fix? The authors suggest a strategy that sounds simple but is powerful: balance the accuracy. Instead of just trying to be right for the most people, the system should aim to be equally accurate for both the popular and the niche groups. They showed in their simulations that if you force the system to care about the minority just as much as the majority, you can stop the "Popularity Trap" and keep the dance floor full.

In the end, this paper gives us a clear, mathematical map of why recommendation systems sometimes go wrong. It shows that popularity bias isn't just a glitch; it's a natural outcome of how these systems learn when one group is much louder than the other. But it also gives us hope: by understanding the rules of the dance, we can change the steps to make sure everyone gets to dance.

Technical Summary: STAY OR STRAY – A DYNAMICAL SYSTEMS VIEWPOINT OF POPULARITY BIAS

1. Problem Statement

Popularity bias in recommendation systems (RS) occurs when a dominant user class generates a disproportionate volume of interaction data, causing the system to increasingly favor majority preferences while degrading recommendation quality for niche (minority) users. While extensive empirical evidence documents this phenomenon, the underlying mechanisms driving its emergence and amplification over time remain poorly understood. Existing literature largely examines popularity bias through a static lens, capturing snapshots of bias without modeling the coupled evolution of user behavior and model parameters. This work addresses the gap by studying the long-term dynamics of popularity bias and its downstream effects on niche user churn using a dynamical systems perspective.

2. Methodology

The authors formulate a stochastic process modeling the coupled evolution of a recommender model and user engagement, analyzed through the lens of dynamical systems.

2.1 System Formulation

User Classes: The system consists of two user classes: a majority (popular) class ( $+1$ ) and a minority (niche) class ($-1$). Items are similarly partitioned.
User Utility: Under the assumption that users prefer items matching their class identity, user utility is defined as a bilinear form: $Utility(u, i) = c_u \cdot c_i$ .
Feature Distribution: User feature vectors $X_t$ are modeled as samples from class-conditional Gaussian distributions ( $N(\mu_1, \Sigma_1)$ and $N(\mu_{-1}, \Sigma_{-1})$ ).
Dynamics:
- Model Updates: The RS acts as an online binary classifier parameterized by $\theta_t$ , updated via online gradient descent to minimize regularized Mean Squared Error (MSE).
- User Arrival (Walk-up Rates): Users update their arrival rates ( $\alpha_t$ ) based on recommendation quality. Correct classifications increase arrival rates; misclassifications decrease them.
Two-Time-Scale Stochastic Approximation: The framework assumes the model parameters ( $\theta_t$ ) evolve on a faster time scale than user arrival rates ( $\alpha_t$ ). This allows the system to be approximated by a continuous-time Ordinary Differential Equation (ODE) framework where $\theta_t$ is treated as quasi-static relative to $\alpha_t$ .

2.2 Mathematical Analysis

By substituting the optimal model parameters $\theta^*(\alpha)$ into the user arrival dynamics, the authors derive a reduced two-ODE system governing the evolution of walk-up rates $\alpha = (\alpha_1, \alpha_{-1})$ . The analysis focuses on the equilibrium points of this system within the domain $[0, 1]^2$ :

$(1, 1)$ : Retention of both user classes.
$(1, 0)$ or $(0, 1)$ : Selective disengagement of one class (Popularity Bias).
$(0, 0)$ : Simultaneous disengagement of both classes.

3. Key Contributions and Theoretical Results

The paper provides a rigorous characterization of the equilibrium structure and convergence properties of the coupled system, yielding four principal theoretical results:

Non-Attainability of Simultaneous Drop-off (Theorem 4.1): The system cannot asymptotically converge to the $(0, 0)$ equilibrium (total system collapse) from any interior initial condition. This is proven by showing that the conditions required for simultaneous decrease in arrival rates violate the Cauchy-Schwarz inequality given the model's optimization objective.
Emergence of Popularity Bias (Theorem 4.2): The authors derive a critical threshold $p^*$ on the fraction of majority users. If the majority fraction $p > p^*$ , the system asymptotically converges to $(1, 0)$ , where the niche class disengages entirely while the majority is retained. This formally characterizes the onset of popularity bias.
Conditions for Symmetric Retention (Theorem 4.3): A sufficient condition is identified for the system to converge to $(1, 1)$ (retention of both classes). This condition relies on the sign of the inner product of class-conditional means in a covariance-transformed space ( $\langle \mu_1, \mu_{-1} \rangle_{N^{-1}_\alpha} < 0$ ), implying that "well-separated" means allow the RS to accurately classify both groups.
Existence of Feasible Retention Intervals (Theorem 4.4): Even when the sufficient condition of Theorem 4.3 does not hold (e.g., when means are positively aligned), the authors identify a feasible interval of $p$ values that guarantees convergence to $(1, 1)$ . This is established by analyzing the monotonicity of update functions and applying the Intermediate Value Theorem.

4. Experimental Validation

The theoretical predictions are validated through two primary experimental avenues:

Synthetic Data Simulations: Extensive simulations on synthetic data confirm that system trajectories conform to the predicted equilibrium behaviors across a broad range of initializations and parameter configurations.
- Phase plots demonstrate convergence to $(1, 0)$ when $p > p^*$ and to $(1, 1)$ when conditions for symmetric retention are met.
- Scatter plots verify the impossibility of simultaneous drop-off (Theorem 4.1) by showing an empty third quadrant in shift plots.
- Experiments demonstrate that system trends can reverse if parameters evolve such that $p^*$ changes relative to the current state (e.g., initially converging to $(1, 1)$ before reversing to $(1, 0)$ ).
Real-World Production Logs: The framework is applied to a large-scale commercial music recommendation platform comprising approximately 410 million user-item interactions.
- Data Processing: Users were classified into "popular-heavy" and "niche-heavy" groups based on interaction with globally popular items. Gaussian distributions were fitted to user embeddings derived from matrix factorization.
- Findings: The empirical data exhibits disproportionate churn among niche-preferring users, consistent with the conditions identified by the theoretical framework. The system dynamics observed in the logs align with the predictions of Theorem 4.2, confirming the presence of popularity bias in the real-world setting.

5. Significance and Claims

The paper claims to provide a foundational understanding of how popularity bias emerges dynamically rather than just describing its static presence. By modeling the co-evolution of user behavior and model parameters, the work:

Identifies Mechanisms: It pinpoints the specific conditions (class imbalance thresholds and feature separation) under which bias becomes inevitable.
Offers Mitigation Strategy: The theoretical framework suggests a natural mitigation strategy focused on balancing the accuracy (error rates) across user classes. The authors empirically demonstrate that this strategy is effective in reducing popularity bias.
Validates Theory with Practice: The alignment between the ODE-derived predictions and real-world production logs provides strong evidence that the proposed dynamical system captures the essential mechanics of bias in commercial recommendation systems.

The authors maintain a modest scope, noting that the current work focuses on MSE loss and Gaussian assumptions, with future work planned to generalize to other convex loss functions and non-Gaussian distributions.