Source-linked AI summary
Modeling and Predicting Popularity Dynamics via Reinforced Poisson Processes
Hua-Wei Shen, Dashun Wang, Chaoming Song, Albert-László Barabási
TL;DR
Predicting the popularity of individual items is difficult because their attention dynamics are noisy and existing approaches provide limited probabilistic, item-specific modeling. The paper introduces a reinforced Poisson process with fitness, aging, and reinforcement, then adds Bayesian treatment through a conjugate prior. On a longitudinal citation dataset, the RPP model with prior consistently outperforms competing methods and is presented as adaptable to other competing-process domains.
Problem
Existing approaches provide limited probabilistic frameworks for modeling and predicting the popularity dynamics of individual items, despite the importance of such predictions.
Method
The paper uses a generative reinforced Poisson process combining item fitness, temporal relaxation, and rich-get-richer reinforcement, with a gamma conjugate prior for Bayesian prediction.
Results
The RPP model with prior consistently outperforms competing methods on a longitudinal citation dataset.
Takeaways & Limitations
The framework models individual attention arrivals directly and can be adapted, with suitable adjustments, to other domains driven by competing processes.
Abstract
from arXiv · showhide
An ability to predict the popularity dynamics of individual items within a complex evolving system has important implications in an array of areas. Here we propose a generative probabilistic framework using a reinforced Poisson process to model explicitly the process through which individual items gain their popularity. This model distinguishes itself from existing models via its capability of modeling the arrival process of popularity and its remarkable power at predicting the popularity of individual items. It possesses the flexibility of applying Bayesian treatment to further improve the predictive power using a conjugate prior. Extensive experiments on a longitudinal citation dataset demonstrate that this model consistently outperforms existing popularity prediction methods.
I. INTRODUCTION
The paper frames individual popularity as a noisy stochastic process that existing approaches do not adequately model or predict. It proposes a reinforced Poisson-process framework combining item fitness, aging, and rich-get-richer reinforcement, with direct modeling of attention arrivals.
- Motivation: Individual popularity dynamics are noisy, while existing models either aggregate items or treat popularity as deterministic time series.These paradigms provide limited item-specific predictive power and ignore the underlying arrival process of attention.
- Proposed framework: The reinforced Poisson process models each item’s attention-arrival process using fitness, temporal relaxation, and reinforcement.Fitness captures intrinsic competitiveness, relaxation captures aging, and reinforcement captures the rich-get-richer effect.
- Proposed framework: The framework is generative and probabilistic, enabling Bayesian treatment and flexible adaptation through different relaxation functions.Its design contrasts with approaches based only on aggregated popularity time series.
- Empirical scope: A longitudinal Physical Review citation dataset is used to evaluate the framework, which consistently outperforms competing methods.The dataset spans over 100 years and includes all papers published by the American Physical Society.
- Model formulation: The model represents popularity through ordered attention-arrival times and a time-varying rate that depends on prior attention.The total attention count includes real attentions and an effective prior count m.
B. Parameter Estimation and Prediction
The model estimates an item’s fitness parameter from observed popularity dynamics and uses it to predict future attention counts. Prediction after the training period is formulated through a differential equation with the observed count as its boundary condition.
- Parameter estimation: Maximum likelihood estimation yields the most likely fitness parameter λ*_d for each item in closed form.The specific estimation of relaxation-function parameters depends on the chosen form of f_d(t; θ_d).
- Prediction: For t ≥ T, predicting an item’s expected future attention count c_d(t) is equivalent to solving a differential equation.The prediction uses the model’s rate function beyond the training endpoint.
- Prediction: The prediction differential equation is solved subject to the boundary condition c_d(T) = n_d.Here n_d is the number of attentions observed by the end of the training period.
III. REINFORCED POISSON PROCESS WITH PRIOR
The unregularized fitness estimate can overfit, especially with sparse observations, producing unstable or null forecasts. Introducing a conjugate prior provides a Bayesian alternative intended to improve prediction accuracy.
- Limitation of maximum likelihood: Maximum likelihood estimation suffers from overfitting when the observed sample size is small.The issue arises in estimating the fitness parameter from limited attention data.
- Bayesian treatment: The exponential dependence of c_d(t) on λ*_d creates large uncertainty in predicted popularity.The paper introduces a conjugate prior for λ_d to address this parameter-estimation drawback.
- Bayesian treatment: The Bayesian treatment with a conjugate prior is presented as a way to further improve the RPP model’s prediction accuracy.It replaces reliance on the problematic single maximum-likelihood fitness estimate.
A. Conjugate Prior
The paper uses a gamma conjugate prior for the fitness parameter because its likelihood has matching power and exponential dependence. The resulting posterior retains the prior’s form while updating its parameters using observed arrival data.
- Prior construction: The conjugate prior for fitness λ_d is a gamma distribution because the likelihood is a power function times an exponential function of λ_d.This choice preserves conjugacy for Bayesian updating.
- Prior construction: The prior distribution applies to the fitness parameters of all N items rather than to one particular item.The paper treats it as a population-level prior across items.
- Posterior updating: After combining the likelihood with the prior, the posterior has the same form as the prior but with likelihood-mediated parameter updates.This is the stated benefit of the conjugate prior.
- Posterior updating: The Bayesian prediction function uses the posterior distribution of λ_d instead of a single λ*_d value.This connects posterior updating directly to future popularity prediction.
B. Bayesian Prediction
Bayesian prediction replaces a single maximum-likelihood fitness estimate with posterior averaging, producing both expected popularity and prediction confidence. The paper compares this Bayesian formulation with the no-prior predictor in experiments.
- Bayesian prediction: The Bayesian prediction function estimates expected attention using the posterior distribution of fitness rather than a single maximum-likelihood estimate.The posterior mean is used to predict the expected number of attentions cd(t).
- Bayesian prediction: Posterior-based prediction also provides the variance of predicted popularity as a confidence measure.The variance of cd(t) is calculated from the posterior distribution of λd.
- Experimental comparison: The Bayesian and no-prior prediction functions are compared through experiments on a longitudinal real-world dataset.The comparison evaluates Eq. (11) against Eq. (8).
C. Parameter estimation
The paper estimates prior and relaxation parameters by likelihood-based optimization while discussing practical difficulties in tuning prior parameters. It uses latent-variable maximum likelihood to avoid relying directly on future validation information.
- Prior-parameter estimation: Tuning prior parameters on a validation set requires future popularity information and repeated model training, making the process impractical and time-consuming.The validation procedure checks prediction accuracy against prior parameters and retrains the model many times.
- Prior-parameter estimation: A fully Bayesian alternative introduces hyperpriors but requires approximation or Monte Carlo methods because inference is usually intractable.Approximation introduces an approximation gap, while Monte Carlo methods have high computational cost.
- Prior-parameter estimation: The paper estimates α and β by maximizing a logarithmic likelihood with a latent variable.These prior parameters are selected through maximum likelihood estimation.
- Model-parameter estimation: The prior parameters adjust the fitness parameter λd, while each relaxation-function parameter θd is estimated by maximizing the likelihood.The calculation of θd depends on the specific relaxation function used in the experiments.
IV. EXPERIMENTS
The experiments evaluate the effectiveness of the proposed reinforced Poisson process model in versions with and without a prior.
- Experiments: The experimental section demonstrates the effectiveness of the proposed RPP model with and without a prior.
A. Experiment setup
The experiments use a longitudinal citation dataset from 11 American Physical Society journals and define model, baseline, and evaluation settings for popularity prediction. The dataset spans 117 years and treats citations as paper popularity.
- Dataset: The dataset contains all papers and citations from 11 American Physical Society journals between 1893 and 2009.
- Dataset: The dataset spans 117 years, making it suitable for modeling and predicting temporal dynamics.
- Dataset: Papers are treated as items whose popularity is characterized by citations within a relatively homogeneous physics community.
- Model specification: The model uses a general temporal relaxation function, with a log-normal form whose parameters are estimated by maximizing logarithmic likelihood.The parameters µd and σd are optimized for the selected relaxation function.
- Baselines: Baseline comparisons include autoregression and linear regression of logarithmic popularity.These methods are identified as competing approaches for modeling and predicting popularity dynamics.
- Evaluation metrics: MAPE measures average deviation between predicted and empirical popularity, while accuracy measures correctly predicted papers under an error tolerance of 0.1.
B. Experimental Results
Experiments show that the reinforced Poisson process with a conjugate prior consistently improves popularity prediction across citation collections, while its performance depends on training length and reinforcement strength. The generative model also reproduces citation distributions, although comparisons with AR and SH use future popularity information unavailable in many real settings.
- Popularity prediction: The RPP model with prior consistently outperforms AR, SH, and RPP without prior across three citation collections.The collections cover Physical Review papers from the 1960s, Physical Review Letters from the 1970s, and Physical Review B from the 1980s.
- Popularity prediction: The prior avoids overfitting caused by the exponential dependence on the fitness parameter, improving performance over the RPP model without prior.Without a prior, the model performs modestly in MAPE because a handful of papers produce large errors.
- Popularity prediction: The superiority of RPP with prior over AR and SH increases with the number of years after the training period.RPP explicitly models attention arrivals and reinforcement, whereas AR and SH capture linear and logarithmic correlations between early and future popularity.
- Generative modeling: The RPP model with prior matches real citation distributions across all studied collections, extending its use beyond individual prediction to global citation-system properties.This distributional result follows from the model's generative formulation.
- Factors affecting performance: Average MAPE decreases as the training period increases, but the rate of decrease quickly slows, indicating diminishing marginal gains.The experiments vary training periods on papers published in Physical Review from 1960 to 1969.
- Factors affecting performance: The learned mean fitness remains nearly constant from 5- to 15-year training periods, while longer training reduces the prior's role in prediction.The increasing variance of prior distributions with training length partly explains how the prior helps overcome overfitting.
- Factors affecting performance: Decreasing m reduces average MAPE, while the mean and variance of the prior distribution decay with m, confirming reinforcement's importance for citation dynamics.Here m is the effective number of attention and controls reinforcement strength; larger m implies weaker reinforcement.
V. RELATED WORK
Prior work studies popularity dynamics through aggregate statistical patterns, time-series prediction, and diffusion-oriented models, but few approaches explain how individual items accrue popularity or predict it directly.
- Popularity research spans influence spread, trust propagation, information access, group formation, online content, user behavior, and citation prediction.
- Most existing approaches do not model how an individual item accrues popularity, leaving individual-item popularity difficult to model and predict.
- Time-series approaches predict future popularity by exploiting temporal patterns and correlations in item-level popularity trajectories.
- Epidemic-spread and diffusion-of-innovation models are descriptive, while reaction-diffusion and branching stochastic processes characterize popularity dynamics.
VI. CONCLUSIONS
The paper presents a reinforced Poisson-process framework for stochastic popularity dynamics, combining item fitness, aging, and reinforcement. Bayesian treatment with a conjugate prior improves prediction robustness and accuracy, and experiments on a century-scale citation dataset show advantages over competing methods.
- The reinforced Poisson process models individual popularity growth as a stochastic process rather than using existing deterministic approaches.
- The model combines fitness, a temporal relaxation function for aging, and reinforcement for the “rich-get-richer” effect.
- Bayesian treatment with a prior produces more robust and accurate popularity predictions.
- Empirical validation on a longitudinal citation dataset spanning over one hundred years demonstrates clear advantages over competing methods.
- Its relaxation function is flexible, supporting a general framework for popularity dynamics beyond a single application domain.