Source-linked AI summary
Finding an Accurate Early Forecasting Model from Small Dataset: A Case of 2019-nCoV Novel Coronavirus Outbreak
Simon James Fong, Gloria Li, Nilanjan Dey, Rubén González Crespo, Enrique Herrera-Viedma
TL;DR
Early epidemic forecasting is difficult because data are scarce, disease characteristics are uncertain, and containment decisions have substantial consequences. The paper proposes GROOMS, which combines optimized candidate models, panel selection, and multiple information sources. In the experiment, PNN-based models achieved the lowest reported error, with PNN obtaining RMSE 136.547.
Problem
Early epidemic forecasting must support decisions despite scarce training data, uncertain disease characteristics, and strong societal and economic consequences.
Method
GROOMS combines multiple forecasting methods, parameter optimization, panel selection, and relevant additional time-series inputs for small-dataset forecasting.
Results
PNN-based models performed best in panel selection, with PNN achieving the lowest observed RMSE of 136.547.
Takeaways & Limitations
The study presents GROOMS as a best-effort forecasting methodology for early-stage epidemics under low data availability.
Abstract
from arXiv · showhide
Epidemic is a rapid and wide spread of infectious disease threatening many lives and economy damages. It is important to fore-tell the epidemic lifetime so to decide on timely and remedic actions. These measures include closing borders, schools, suspending community services and commuters. Resuming such curfews depends on the momentum of the outbreak and its rate of decay. Being able to accurately forecast the fate of an epidemic is an extremely important but difficult task. Due to limited knowledge of the novel disease, the high uncertainty involved and the complex societal-political factors that influence the widespread of the new virus, any forecast is anything but reliable. Another factor is the insufficient amount of available data. Data samples are often scarce when an epidemic just started. With only few training samples on hand, finding a forecasting model which offers forecast at the best efforts is a big challenge in machine learning. In the past, three popular methods have been proposed, they include 1) augmenting the existing little data, 2) using a panel selection to pick the best forecasting model from several models, and 3) fine-tuning the parameters of an individual forecastingmodel for the highest possible accuracy. In this paper, a methodology that embraces these three virtues of data mining from a small dataset is proposed...
I. Introduction S
The paper addresses early epidemic forecasting when data are scarce, disease characteristics are uncertain, and decisions must balance transmission control against economic costs. It proposes combining model selection, parameter tuning, and additional data sources to improve forecasting from small datasets.
- Motivation: Containment measures limit physical contact but impose increasing economic losses, making their timing difficult to determine.Lifting restrictions too early may precede epidemic subsidence, while extending them increases societal costs.
- Motivation: Authorities need forecasts of whether the epidemic is worsening and when it may end, despite limited knowledge and scarce early-stage data.The uncertainty complicates decisions about containment measures and their eventual removal.
- Related approaches: Small-dataset forecasting commonly uses data augmentation, ensemble selection, or parameter tuning of an individual prediction algorithm.These approaches address limited training data through different forms of expansion, selection, or optimization.
- Contribution: GROOMS combines multiple candidate forecasting algorithms, tuned parameters, and relevant additional information in multiple-regression candidates.The selected model is the candidate with the lowest error.
- Paper organization: The paper is organized around methodology, experiments, results discussion, and conclusions with future work.
II. Our Proposed Methodology
GROOMS is a small-dataset forecasting methodology that optimizes candidate models, selects the lowest-error model, and can incorporate multiple data sources. It organizes five method groups by complexity and recommends starting with traditional forecasting and simple analytics when rapid results are needed.
- GROOMS process: GROOMS passes the available dataset through optimized candidate forecasting models and selects the model with the lowest curve-fitting error.The winning model provides the best-effort prediction technically available from the candidate panel.
- GROOMS process: The methodology targets highest possible prediction accuracy under low data availability and limited knowledge.
- Method groups: Group 01 uses complex machine-learning models with multiple regressions, requiring both parameter tuning and selection of relevant time-series inputs.Neural-network candidates may use multivariate data and feature selection to minimize prediction error.
- Method groups: Groups 02 and 03 use complex and lightweight machine-learning base learners, respectively, with lightweight models running faster but requiring tuning to avoid underfitting or overfitting.
- Method groups: Group 04 provides descriptive statistics, while Group 05 uses traditional univariate methods such as autoregression, moving average, and ARIMA variants.
- Recommended order: GROOMS recommends applying traditional forecasting first, alongside simple statistical analysis, then attempting more complex groups as resources and time permit.The sequence prioritizes rapid forecasting before higher-complexity methods.
B. Polynomial Neural Network with Corrective Feedback (PNN+cf)
The paper uses Polynomial Neural Network with corrective feedback as an evolving candidate within GROOMS. PNN expands its structure iteratively, while PNN+cf adds lagged data and prior training residuals to adapt the regression inputs.
- PNN: PNN is an evolutionary neural network based on GMDH that expands iteratively until adding neurons no longer improves performance.
- PNN: PNN increases polynomial complexity and self-tunes coefficients through iterative data sampling and controlled network expansion.This seeks marginal model complexity alongside high predictive ability.
- PNN formulation: The PNN polynomial maps sequential input variables to a predicted outcome while candidate beta models are evaluated using output error.
- Corrective feedback: PNN+cf adds lagged data and training errors from previous iterations as corrective-feedback inputs during polynomial formation.Residuals represent deviations between predicted and actual values and are intended to compensate for nonlinear bias.
- Input selection: Adding redundant variables can cause overfitting and longer computation, so extra inputs should be statistically significant or relevant.
- Corrective feedback: A rolling window feeds recent training data to the network for one-step-ahead prediction, while residuals from previous cycles are fed back at each prediction.
III. Experiment
The experiment evaluates three forecasting groups on a small, evolving epidemic dataset using RMSE, finding that PNN-based methods perform best overall while input selection affects performance.
- Data and setup: Daily epidemic data were extracted from Chinese health-authority archives and organized into confirmed, cured, died, suspected, and critical trends.The data were updated daily from 21 January 2020.
- Data and setup: GROOMS tested classical time-series, machine-learning, and polynomial-neural-network forecasting algorithms against the limited epidemic data.The comparison examined how well each group fit the early-stage epidemic curve.
- Evaluation: RMSE was recorded after running the same small dataset through the algorithms, with automatic tuning enabled when available and default parameters otherwise.Results were compared using bar charts in Fig. 5.
- Evaluation: Classical forecasting algorithms generally performed poorly because sharp curve changes created large gaps between fitted and actual values.The initial hump was a particularly difficult region for curve-fitting methods.
- Results: PNN achieved the lowest observed RMSE of 136.547, while Fast Decision Tree Learner was an exception among machine-learning models with RMSE 1744.526.The reported PNN result was the best in the panel selection.
- Results: PNN+cf won panel selection, and combining suspected, confirmed, and critical data produced superior performance for both Combi and MIA PNN architectures.The selected statuses follow the reported progression of patient cases and were used as related input sources.
IV. Conclusion
GROOMS addresses epidemic forecasting with very limited data by combining multiple forecasting methods under panel selection, with PNN reported as the relatively lowest-error model. The methodology remains limited because accuracy alone cannot capture the broader multifactor nature of epidemic decisions.
- GROOMS ensembles five forecasting-method types under a panel selector for small-data epidemic forecasting.The methods range from classical time-series forecasting to self-evolving polynomial neural networks.
- PNN produced a forecasting model with relatively lowest error in the experiment.The paper extends PNN to PNN+cf with corrective feedback during optimization.
- PNN incrementally grows its internal structure from simple to optimal, making over- or under-fitting less likely.
- GROOMS currently treats accuracy or error as its sole criterion, although epidemic decisions involve multiple factors.Historical data provide only one aspect of the decision, despite remaining important.
- Future work should investigate algorithm differences, test additional panel-selection algorithms, and connect technical forecasts with nontechnical decision processes.