Mixed Popularity: Multi-Signal Dynamics

• Mixed popularity is defined as the study of integrating diverse social signals—such as quality, review counts, and content heterogeneity—to model online engagement beyond a single popularity count.
• It leverages various methodologies, including probabilistic simulation, logistic regression, and mixed-effects models, to assess the interplay between social influence and intrinsic quality.
• Applications span recommendation calibration, matching theory, and predictive modeling, leading to practical innovations such as bias mitigation and dynamic embedding in scalable systems.

Searching arXiv for recent and relevant papers on "mixed popularity" and related uses of the term. Mixed popularity denotes a family of research constructs in which popularity is neither modeled nor evaluated as a single standalone count. Across the literature, it appears as the joint use of multiple social signals in ranking, the interaction between quality and cumulative advantage in popularity growth, the dependence of post-level attention on content heterogeneity, the calibration of recommendation popularity to user-specific taste, the use of mixed-effects models for hierarchical popularity data, and the extension of majority-based popularity to randomized or feasibility-constrained outcomes in matching and coalition formation. This suggests an umbrella concept: popularity is “mixed” when it emerges from, or is normatively compared through, more than one mechanism, signal, or structural layer [(Analytis et al., 2017); (Krafft et al., 2014); (Bessi et al., 2015); (Schirmer et al., 1 Apr 2026); (Ge et al., 2022)].

1. Conceptual range of the term

The literature does not supply a single canonical definition of mixed popularity. Instead, distinct fields use the term, or closely related formulations, for different technical objects. In some studies, popularity is mixed with review quality; in others, with content heterogeneity, user hierarchy, or combinatorial voting constraints. The common structure is the rejection of one-dimensional popularity.

Domain	Mixed element	Representative formalization
Online choice and ranking	Average review score + number of reviews	Relative-logit choice rule
Social-media attention	Heavy-tailed user activity + content heterogeneity	Beta-distributed post attractiveness
Recommendation alignment	User history popularity + recommendation popularity	Popularity Quantile Calibration
Popularity dynamics	Social influence + past popularity + intrinsic quality	Coupled Friedkin–Johnsen update
Coalition formation and matching	Majority popularity + randomization or feasibility restriction	Mixed popular outcomes; popular max-matchings

This range matters methodologically. Some formulations are descriptive, asking how popularity distributions arise; some are predictive, asking how popularity can be forecast; some are normative, asking what a popularity-aware system should optimize; and some are axiomatic, defining popularity as a majority relation over feasible outcomes. The resulting literature is best understood as a cluster of formally related but non-equivalent theories rather than a unified doctrine (Bessi et al., 2015, Jeong et al., 2024, Kavitha, 2020).

2. Content heterogeneity and distributional popularity

A canonical empirical formulation of mixed popularity patterns appears in the study of Facebook pages that contrast heterogeneous content with a near-perfect homogeneous control. The dataset comprises 74 pages in total—34 science pages, 39 conspiracy pages, and 1 baseline page posting the same picture every day—over 49,354 posts, 2,095,677 likes, 192,967 comments, 3,782,480 shares, 344,367 likers, and 64,903 commenters. The baseline page serves as a control for minimal content variation: “La stessa foto di Toto Cutugno ogni giorno” posts the same image every day, allowing the effect of content heterogeneity on popularity to be isolated (Bessi et al., 2015).

The central empirical result is distributional. User activity, particularly likes per user, follows a heavy-tailed distribution for all page types, and user lifetime differs only slightly between the baseline and the science/conspiracy pages. Post popularity, however, behaves differently. On heterogeneous science and conspiracy pages, likes per post are heavy-tailed: a small number of posts attract very high engagement while most receive relatively little attention. On the homogeneous baseline page, likes per post are approximately Gaussian, centered around a typical level. The result is that heavy-tailed user behavior can map either to heavy-tailed post popularity or to approximately Gaussian post popularity, depending on content diversity rather than on user activity alone (Bessi et al., 2015).

The paper formalizes this with a probabilistic model in which post attractiveness is drawn as

$v \sim \mathcal{Be}(1,\beta),$

with $\beta=1$ corresponding to homogeneous content and large $\beta$ to highly heterogeneous content. User activity volume $a$ and preference threshold $b$ are drawn from a power-law distribution

$p(x)=x^{-\gamma}, \qquad \gamma=1.5,$

normalized to $[0,1]$, and a user likes a post iff $b<v$, subject to the user’s activity budget. Simulations with $P=10{,}000$ posts, $U=20{,}000$ users, $\beta=1$0, and 100 averaged iterations reproduce the empirical transition: heavy-tailed user activity persists throughout, but post consumption becomes approximately Gaussian at $\beta=1$1 and skewed for large $\beta=1$2 (Bessi et al., 2015).

3. Quality, review scores, and cumulative advantage

In online ranking and choice, mixed popularity often refers to the joint use of perceived quality and social endorsement. One formulation studies binary choice between Amazon books and IMDb films when the only visible attributes are average review score and number of reviews. The experiments use 83 fiction books from Amazon and 98 feature movies from IMDb; each participant makes 200 binary choices, of which 90% are dilemmas and 10% are dominated comparisons. The best parsimonious account is a logistic regression on score difference and log popularity ratio,

$\beta=1$3

Average score matters more than popularity in both domains, but most participants accept somewhat lower average scores for more popular items. The fitted trade-off is stronger for books than for movies: the average-score difference has about 2.72 times the weight of the popularity term for books and about 10.2 times the weight for movies. The same study also reports substantial heterogeneity across participants, grouped into five behavioral types ranging from “Rev. inclined” to “Dissenters,” motivating personalized ranking rather than a global popularity rule (Analytis et al., 2017).

A second formulation studies mixed popularity dynamically as the joint effect of quality and cumulative advantage. In eToro.com, where users can mirror traders and observe both performance-like signals and current popularity, the model assumes Bayesian updating and probability matching. With $\beta=1$4 denoting current popularity and $\beta=1$5 an observed quality signal, the expected popularity change is derived as

$\beta=1$6

The empirical test uses 24,587 users over 100 consecutive days from September 09, 2011 to December 29, 2011. The key hypothesis is that the interaction between past popularity and past performance is positive; the reported coefficients for both the marginal popularity effect and the interaction term are significant with $\beta=1$7. The interpretation is not merely that quality and popularity both matter, but that popularity amplifies the effect of quality on future popularity (Krafft et al., 2014).

A third formulation places the interaction inside an idealized online cultural market. Each item $\beta=1$8 has hidden intrinsic quality $\beta=1$9, and the platform pre-selects items with probability

$\beta$0

where $\beta$1 is current popularity rank and $\beta$2 controls popularity bias. Users inspect $\beta$3 items and choose the best-quality one. The resulting theory identifies a harmful regime for mild popularity bias, $\beta$4, in which the minimum discrimination effort $\beta$5 diverges with catalog size $\beta$6, and a benign regime for stronger bias, $\beta$7, in which $\beta$8 remains bounded. In the unbiased case $\beta$9, one has $a$0: comparing two uniformly sampled items is already sufficient for asymptotic quality-popularity alignment. The same model shows that when a fraction $a$1 of naive users always picks the most popular item, alignment cannot be restored if $a$2 (Gaeta et al., 2022).

4. Recommendation systems, calibration, and popularity-aware allocation

In recommender systems, mixed popularity is increasingly treated as a calibration problem rather than as a purely global bias. One recent formulation models a user’s historical popularity preference as a conditional distribution $a$3 over item popularity scores and the recommender’s induced distribution as $a$4. Popularity Quantile Calibration defines alignment through quantile matching, and the summary statistic

$a$5

measures user-level miscalibration. On top of this metric, SPREE uses activation steering in SASRec-based sequential recommenders to move internal representations along a learned popularity direction. The steering update is

$a$6

where $a$7 is a Lasso-based bias estimator that sets both sign and magnitude of the intervention per user. Experiments on Foursquare Tokyo, MovieLens-1M, MovieLens-20M, and RateBeer report that SPREE consistently improves user-level popularity alignment measured by PCE@100 while preserving recommendation quality measured by NDCG@100 more effectively than global debiasing baselines such as IPR and PopSteer (Schirmer et al., 1 Apr 2026).

A related line of work studies popularity calibration in point-of-interest recommendation. Using Brightkite, Foursquare, Gowalla, and Yelp, it compares a non-contextual baseline, BPR, against context-aware models LORE and USG and two calibrated re-ranking variants, $a$8 and $a$9. Users are partitioned into LowPop, MedPop, and HighPop groups according to average profile popularity, and items into Tail, Mid, and Head groups. The re-ranking objective is

$b$0

with $b$1 tuned separately for each user group. The empirical pattern is asymmetric: context-awareness is not uniformly bias-mitigating, because LORE tends to push toward less popular items while USG can either reduce or worsen popularity bias depending on the dataset. Calibration improves alignment with user popularity profiles but introduces an accuracy–bias trade-off, and the combination of LORE with calibration produces the closest match to users’ mixed-popularity profiles among the studied methods (Forster et al., 4 Jul 2025).

A more architectural use of the idea appears in mixed-dimension embeddings for large-scale recommendation. Here popularity is query frequency, and representational capacity is allocated accordingly rather than uniformly. The practical sizing rule is

$b$2

so that frequent IDs receive larger internal dimensions than rare IDs. The motivation is extreme skew: on MovieLens, the top 10% of users receive as many queries as the remaining 90%, while on the Criteo Kaggle dataset the top $b$3 of indices receive as many queries as the remaining $b$4 million. In CTR experiments with DLRM on the full Criteo dataset, a mixed-dimension model with $b$5 achieves a learning curve comparable to a uniform $b$6 model while using the parameter count of a uniform $b$7 model, i.e. a $b$8 reduction; with $b$9, it improves accuracy by about $p(x)=x^{-\gamma}, \qquad \gamma=1.5,$0 using half as many parameters; and it can train more than $p(x)=x^{-\gamma}, \qquad \gamma=1.5,$1 faster on GPU at a given test loss (Ginart et al., 2019).

5. Prediction, hierarchy, and multidimensional determinants

Mixed popularity also appears in predictive settings where hierarchical dependence or multi-source signal integration is explicit. For Instagram post popularity, one formulation defines the response as

$p(x)=x^{-\gamma}, \qquad \gamma=1.5,$2

with $p(x)=x^{-\gamma}, \qquad \gamma=1.5,$3, and models the nested user–post structure through a Linear Mixed Model,

$p(x)=x^{-\gamma}, \qquad \gamma=1.5,$4

where $p(x)=x^{-\gamma}, \qquad \gamma=1.5,$5 captures user-specific random effects. The study combines non-image covariates with Google Cloud Vision API image labels, image properties, Seeded-LDA topic variables, and Munsell color categories. Empirically, the image and color features achieve 6.8% higher accuracy than non-image covariates alone. LMM outperforms ordinary linear regression, confirming the importance of user hierarchy, but nonlinear models perform better still: with all covariates, XGBoost attains RMSE $p(x)=x^{-\gamma}, \qquad \gamma=1.5,$6 and MAE $p(x)=x^{-\gamma}, \qquad \gamma=1.5,$7, versus RMSE $p(x)=x^{-\gamma}, \qquad \gamma=1.5,$8 and MAE $p(x)=x^{-\gamma}, \qquad \gamma=1.5,$9 for LMM and RMSE $[0,1]$0 and MAE $[0,1]$1 for linear regression (Jeong et al., 2024).

A distinct predictive literature argues that much apparent success in popularity prediction is driven by temporal peeking rather than by robust understanding. In datasets spanning Last.fm, Flickr, Goodreads, and Twitter, popularity is the number of adopters, and the standard task observes the first $[0,1]$2 adoptions and predicts whether future popularity exceeds the median at horizon $[0,1]$3. Logistic regression usually performs best, but the dominant signal is simply early adoption speed. The single feature $[0,1]$4 yields more than 70% accuracy on all datasets and accounts for nearly 97% of the full model’s accuracy; temporal-feature models trained on one dataset transfer to other datasets with accuracy typically within 5 percentage points of same-domain performance. By contrast, non-temporal features are unstable: for 12 of 25 features, the logistic coefficient flips sign across datasets. When the task is reformulated to control for early speed through Temporally Matched Balanced Classification, overall accuracy falls below 65%, revealing a gap between prediction and explanatory understanding (Shulman et al., 2016).

At the level of dynamical systems, mixed popularity can be written directly as a convex combination of social influence, recommendation feedback, and intrinsic quality. In a coupled Friedkin–Johnsen model for influencer competition, user $[0,1]$5’s attention to influencer $[0,1]$6 evolves as

$[0,1]$7

with $[0,1]$8. Here $[0,1]$9 is normalized aggregate attention and $b<v$0 is intrinsic quality. The theory isolates three regimes. With $b<v$1, long-run popularity is proportional to quality. With $b<v$2, convergence is to a consensus profile determined by social influence, recommendation feedback, and initial conditions rather than quality. With all terms active, the limit is linear in $b<v$3 and modulated by a shared amplification profile induced by network and recommendation structure (Cocca et al., 19 Mar 2025).

Cultural-product studies add a further dimension by showing that popularity may mix several forms of novelty rather than a single novelty–familiarity scalar. In a dataset of 51,411 songs, negative binomial models of Spotify popularity find that lyrical uniqueness and audio uniqueness are both negatively associated with popularity, while chord uniqueness is not a meaningful predictor. In genre-specific models, audio uniqueness is numerically the strongest predictor, though not significantly stronger than lyrical uniqueness in the Wald test. Mediation analysis further shows that theme and repetitiveness explain about a quarter of the association between lyrical uniqueness and popularity, while sentiment does not materially mediate it. The reported pattern is linear and negative rather than inverted-U, contradicting optimal distinctiveness theory in this setting (Yu et al., 2022).

6. Majority-based popularity under randomization and feasibility constraints

In algorithmic social choice and matching theory, mixed popularity has a precise axiomatic meaning unrelated to social-media metrics. In the roommate diversity problem, a deterministic outcome $b<v$4 is popular if no other outcome defeats it by majority margin. A mixed outcome is a probability distribution over deterministic outcomes,

$b<v$5

and its expected popularity margin against another mixed outcome $b<v$6 is

$b<v$7

A mixed outcome is popular if $b<v$8 for every mixed $b<v$9. The existence result is unconditional: every roommate diversity game has a mixed popular outcome by a minimax argument on the antisymmetric payoff matrix of pairwise popularity margins. Computation is harder: finding a mixed popular outcome in polynomial time is impossible unless $P=10{,}000$0, even with dichotomous preferences. By contrast, when room size is fixed to $P=10{,}000$1, a deterministic popular partition always exists and can be computed in polynomial time via maximum-weight perfect matching (Ge et al., 2022).

A related constrained notion appears in bipartite matching. Given strict preferences, a maximum matching $P=10{,}000$2 is a popular max-matching if

$P=10{,}000$3

Popularity is therefore evaluated only within the feasible family of maximum-cardinality solutions. The paper characterizes such matchings through alternating cycles and alternating paths with a matching-dependent weight function, constructs an auxiliary graph $P=10{,}000$4 whose stable matchings project exactly to popular max-matchings, and derives a compact extended formulation for the popular max-matching polytope. The algorithmic consequence is strong: a minimum-cost popular max-matching can be found in polynomial time. This stands in contrast to the min-cost popular matching problem, which is NP-hard, and to minimum-cost Pareto-optimal matching and max-matching, which are also NP-hard (Kavitha, 2020).

Taken together, these axiomatic formulations show that mixed popularity need not refer to the mixing of social signals at all. It can instead denote either randomization over outcomes or popularity comparison under hard admissibility constraints. In that sense, the term spans both empirical popularity formation and formal majority-based optimality, with the unifying theme that popularity is defined relative to a richer structure than a raw popularity count.