Hybrid AI Model Hits 99.5% Accuracy in Classifying Sports Event Security Risks

A new hybrid deep learning model can sort sports events into low, medium and high security risk categories with 99.53% accuracy, according to a study published in Discover Artificial Intelligence. The model, called TabFT-Net, combines two transformer-based architectures designed for tabular data and was tested on a dataset of 40,000 event-level security scenarios drawn from the Kaggle platform. Its reported F1-score of 99.53% and ROC-AUC of 0.998 place it well above every baseline model the authors trained on the same data, including gradient-boosting methods that are usually formidable competitors on structured datasets.

The work addresses a persistent gap in stadium safety management. Large gatherings create complex, dynamic environments where crowd movement, congestion, queue formation and emergency accessibility must be monitored continuously, and traditional approaches based on manual surveillance and predefined safety protocols often struggle to respond promptly to rapidly changing crowd conditions. The authors argue that machine learning can recognize risk patterns across multiple variables that may not be obvious to human observers, and that correct classification into risk levels helps deploy security personnel and emergency resources appropriately.

The underlying dataset describes each record as a sports event security scenario, with 31 predictor variables covering attendance, crowd density, congestion, security resources, emergency accessibility, infrastructure and environmental conditions. The target variable, Security Risk Level, has three categories: Low, Medium and High. The distribution is reasonably balanced, with approximately 19,640 low-risk events (49.1%), 11,460 medium-risk events (28.7%) and 8,900 high-risk events (22.2%), and stratified sampling preserved these proportions across the training, validation and test splits of 70%, 15% and 15%.

Importantly, the researchers are candid about the provenance of this data. The dataset documentation describes the records as ‘realistic crowd-management scenarios’, but the authors note there is limited evidence to support this and it is unclear whether all records were gathered from real stadium events. They explicitly state that the study does not claim to use an actual operational dataset in the real world; rather, it serves as a structured benchmark for assessing machine learning techniques for security risk classification. The documentation also does not provide a full formal rule or algorithm for how each target label was generated, a restriction the authors state explicitly when interpreting their predictive results.

On top of the raw variables, the team engineered eight security indicators intended to capture crowd pressure, congestion severity and operational stress. These include a Crowd Pressure Index computed as actual attendance times crowd density divided by stadium capacity; a Congestion Severity Score given by current spectator flow divided by pathway capacity; a Security Stress Index equal to attendance divided by the number of security staff; an Accessibility Risk Score based on crowd density relative to emergency exit access; a Security Adequacy Score combining security staff and police deployment per attendee; an Emergency Preparedness Index based on medical and security staff per spectator; an Event Volatility Score multiplying incident count by crowd density; and a Composite Risk Feature summing several of these indices.

The model itself fuses two complementary components. TabNet, a deep learning architecture for structured data, uses a sequential attention mechanism with a sparsemax activation to select the most informative features at each decision step, progressively filtering out less relevant information. FT-Transformer then processes the selected features through self-attention, converting them into tokens and modeling the interactions among crowd, congestion and infrastructure variables. The two representations are concatenated in a feature fusion layer, passed through a residual dense block to preserve critical information and support gradient propagation, and finally classified by a softmax layer that outputs probabilities for the three risk categories.

Against seven baselines trained on identical data partitions and preprocessing, TabFT-Net dominated. TabNet alone reached 81.17% accuracy with a ROC-AUC of 0.871; SAINT Transformer achieved 83.10% accuracy (ROC-AUC 0.889); TabTransformer 89.07% (ROC-AUC 0.931); and FT-Transformer 92.62% (ROC-AUC 0.963). Among conventional algorithms, CatBoost was the strongest at 97.42% accuracy, with LightGBM at 97.16% and XGBoost at 96.84%. TabFT-Net’s 99.53% accuracy, 99.54% precision, 99.53% recall and 0.998 ROC-AUC exceeded all of them, and ten-fold stratified cross-validation using the same folds for all models yielded a mean accuracy of 99.50 ± 0.10%, compared with 97.40 ± 0.19% for CatBoost. Paired t-tests across the ten folds supported the differences, with estimated accuracy gains ranging from 2.10 percentage points over CatBoost to 18.30 over TabNet.

Because the target-generation mechanism is undocumented, the authors ran a series of sensitivity experiments to probe whether the engineered indicators might indirectly reconstruct the target label. A raw-features-only configuration achieved 97.0% accuracy, engineered features alone 98.5%, and the combination 99.6%. Removing the Composite Risk Feature lowered accuracy to 99.2%, and removing it together with other high-importance engineered indicators lowered it further to 98.0%. In a group-wise removal test, four variables a reviewer flagged as potentially related to the target construction — Crowd Density, Incident Count, Security Staff and Emergency Access — were discarded together, yet TabFT-Net retrained on the remaining raw predictors still reached 97.0% accuracy, 97.1% precision, 97.0% recall, 97.1% F1-score and 0.985 ROC-AUC, essentially matching the raw-features-only result. The authors stress that this is a sensitivity analysis only: good performance after feature removal does not prove the absence of target-generation dependence, and the undocumented label generation process means target leakage cannot be definitively ruled out.

Interpretability was addressed with two explainable AI techniques. SHAP analysis identified the engineered indicators CRF, CPI, CSS and SSI as the most influential variables, alongside attendance, crowd density and congestion measures, and dependence plots showed that higher crowd density and greater congestion severity generally push predicted risk upward. LIME provided local explanations for individual predictions, showing that attendance level, congestion severity, crowd density and the composite risk feature increased risk predictions, while higher security adequacy and emergency preparedness scores reduced them. The authors report that feature importance rankings remained stable across validation subsets, and that removing individual dominant features caused only gradual performance declines rather than collapse, suggesting the model learns complementary relationships among many crowd, infrastructure and operational factors rather than relying on a single dominant variable.

The study also quantifies the price of that accuracy. In a computational comparison, the gradient-boosting models trained and predicted far faster than the transformer-based architectures, and TabFT-Net consumed the most resources of all neural models, reflecting the combined cost of TabNet feature selection, FT-Transformer interaction learning and residual fusion. The authors argue the overhead is rewarded by the highest predictive performance, though the comparison with prior studies — which typically report accuracies of 78–91% on related but different tasks and datasets — is explicitly not direct, since no other study used the same data modality and task. The authors note the gap could mean the Kaggle dataset is easier, partly synthetic or label-leaky, or that their model benefits from richer features or more complex modeling.

Several limitations frame the contribution. The data contained only structured tabular event information; no real-time video streams, sensor network data or social media information were included, and because no other publicly available sports event security dataset with comparable features exists, external dataset validation was not possible. The authors identify multimodal data integration, graph neural networks, temporal transformers and real-time streaming frameworks for dynamic risk assessment in smart stadium environments as future directions, along with a more comprehensive evaluation against gradient-boosted decision tree models to test generalization. For now, the result stands as a strong benchmark performance on a structured crowd-management dataset — one whose real-world operational value will depend on validation beyond it.

Subject of Research: Machine learning classification of security risk levels at sports events using crowd flow, congestion and infrastructure data

Article Title: Hybrid transformer-based security risk classification for sports events using crowd flow and congestion analytics

Article References: Liu, F. (2026). Hybrid transformer-based security risk classification for sports events using crowd flow and congestion analytics. Discover Artificial Intelligence, 6(1), Article 1658. https://doi.org/10.1007/s44163-026-02618-w

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02618-w

Keywords: sports event security, machine learning, transformer models, TabNet, FT-Transformer, crowd management, risk classification, SHAP, LIME, tabular data, cross-validation, Kaggle dataset