💻 computer science

CHIMERA-Tab: A Scalable AI-Driven Framework for Automated Feature Selection and Heterogeneous Stacking in Imbalanced Banking Data Classification

This paper introduces CHIMERA-Tab, a scalable AI framework that integrates chaotic metaheuristic preprocessing with a heterogeneous stacking ensemble of TabNet and gradient boosting models to achieve state-of-the-art performance in imbalanced banking classification, demonstrating that the stacking architecture is the primary driver of accuracy while chaotic preprocessing enhances feature selection and interpretability.

Original authors: Anilkumar Vishwanath Brahmane, P. William, Shashikant Raghunathrao Deshmukh, Elviz Ismayilov, Smita Nirkhi

Published 2026-08-28

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Anilkumar Vishwanath Brahmane, P. William, Shashikant Raghunathrao Deshmukh, Elviz Ismayilov, Smita Nirkhi

Original paper licensed under CC BY 4.0 (https://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

In the world of banking, predicting whether a customer will say "yes" to a new financial product is a high-stakes game of probability. Banks rely on telemarketing campaigns to reach potential clients, but calling everyone is expensive and inefficient. The goal is to identify the few people likely to subscribe to a term deposit before the phone even rings. This is a classic problem of classification: sorting a massive pile of data into two groups, "will subscribe" and "will not." The challenge is that the data is messy. It is heavily skewed, with far more people saying no than yes, and it contains a mix of numbers and categories that are difficult for standard computer programs to process together. Furthermore, the data often includes information that is only available after the call has finished, which creates a trap for models that learn from the past but must predict the future. To solve this, researchers need tools that can sift through noise, find the few signals that truly matter, and combine different types of mathematical reasoning to make a reliable guess.

A team of researchers has introduced a new framework called CHIMERA-Tab, designed specifically to tackle these messy, unbalanced banking datasets. Their approach is not about inventing a single, perfect algorithm, but rather about building a smart system that orchestrates several different methods to work together. The system begins by cleaning the data, stripping away unnecessary details to focus on the most critical factors. It then uses a sophisticated search process to tune the settings of a deep learning model, ensuring it is perfectly calibrated for the task. Finally, it brings together four distinct types of predictive models—a deep learning network and three powerful tree-based models—and teaches a simple "meta-learner" how to weigh their individual opinions to form a single, superior prediction. This combination allows the system to see patterns that any single model would miss, particularly in the difficult task of spotting the rare "yes" responses among thousands of "no" responses.

The researchers tested their system on a well-known dataset containing over 41,000 records of past banking interactions. The data was heavily imbalanced, with only about 11 percent of customers actually subscribing to a term deposit. In this environment, the CHIMERA-Tab framework achieved a remarkable level of accuracy, correctly ranking potential subscribers higher than non-subscribers in 95 percent of cases. This performance was not a fluke; the system was tested five times with different random starting points, and it consistently outperformed nine other standard methods, including popular machine learning tools and deep learning models used in isolation. The study confirmed that the system's success came primarily from its ability to combine different model families. When the researchers removed the final step of combining the models, the performance dropped significantly, proving that the synergy between the deep learning network and the tree-based models was the true engine of success.

One of the most practical contributions of this work is how it handles the sheer volume of information. The original dataset had 21 different features, ranging from the customer's age and job to economic indicators and call history. The system's first stage automatically identified that only eight of these features were truly necessary to make an accurate prediction. By reducing the data from 21 variables down to 8, the system became much faster and easier to interpret, without losing any predictive power. This reduction was achieved using a specialized search method inspired by the hunting behavior of eagles, which was guided by a chaotic mathematical engine to explore the best combinations of features efficiently. The researchers found that while this feature selection made the model more efficient and transparent, it did not significantly change the final accuracy, suggesting that the powerful combination of models at the end of the pipeline was robust enough to handle extra information if needed.

The study also took a rigorous look at the reliability of its results, moving beyond simple accuracy scores to use statistical tests that confirm the findings are real and not just luck. The analysis showed that the new framework was statistically superior to almost all the other methods tested. However, the researchers were careful to point out a critical limitation regarding real-world deployment. The dataset included a feature called "call duration," which measures how long a customer stayed on the phone. This information is incredibly powerful for prediction, but it is only available after the call has already happened. In a real marketing campaign, a bank cannot know the duration before deciding whether to make the call. When the researchers removed this feature to simulate a realistic pre-call scenario, the system's accuracy dropped, though it still remained competitive with other advanced methods. This honest assessment highlights the difference between a model that performs well in a lab and one that can be deployed in the field.

Ultimately, the paper demonstrates that the best way to solve complex data problems is not to choose between deep learning and traditional tree-based models, but to use both. By letting a deep learning network and several gradient boosting models vote on the answer, and then using a simple learning algorithm to decide how much to trust each vote, the system captures the strengths of each approach. The research also introduces a novel way to understand how the model reacts to changes in the data, using chaotic mathematical patterns to test the system's sensitivity. While the framework requires significant computing power to train, the authors suggest that once trained, it can make predictions instantly. This work provides a clear path forward for financial institutions looking to improve their marketing strategies, offering a tool that is not only highly accurate but also transparent about what it knows and how it knows it.

Technical Summary of CHIMERA-Tab: A Scalable AI-Driven Framework for Automated Feature Selection and Heterogeneous Stacking in Imbalanced Banking Data Classification

Problem Statement
The paper addresses the persistent challenges in financial decision support, specifically within the domain of bank marketing (telemarketing). The core problem involves binary classification on tabular data characterized by two primary difficulties: (1) Heterogeneous feature spaces mixing numerical and categorical variables, and (2) Severe class imbalance, where the target variable (term deposit subscription) has a positive rate of approximately 11.3% (a 7.9:1 ratio). These conditions often cause standard classifiers to favor the majority class, reducing their effectiveness in identifying potential subscribers. Furthermore, existing literature often treats deep learning architectures and gradient boosting models in isolation, lacks integrated wrapper-based feature selection, and frequently relies on single-run evaluations without rigorous statistical validation.

Methodology: The CHIMERA-Tab Framework
The authors propose CHIMERA-Tab (Chaotic Hybrid Intelligent Metaheuristic Ensemble with Residual-Augmented Tabular Learning), a unified pipeline comprising three sequential, chaotically enhanced stages:

Chaotic Feature Selection (CBAOFS):
- Utilizes a Chaotic Binary Aquila Optimizer to perform wrapper-based feature selection.
- The fitness function balances classification error (AUC-ROC via 3-fold LightGBM) against model parsimony.
- An entropy-guided chaotic map engine dynamically selects among four generators (Logistic, Tent, Sine, Chebyshev) based on Shannon entropy monitoring to optimize the search dynamics.
- A V-shaped transfer function binarizes the continuous search agents.
- Outcome: Reduces the feature set from 21 to 8 features (a 61.9% reduction).
Chaotic Hyperparameter Optimization (CSCA):
- Employs a Chaotic Sine-Cosine Algorithm to optimize eight hyperparameters for the TabNet deep learning architecture.
- The objective function is a weighted composite of AUC-ROC (60%) and Macro F1-score (40%) to prevent the optimizer from ignoring the minority class.
- Chaotic sequences replace random parameters in the SCA update rule to improve space coverage.
- Outcome: Identifies an optimal configuration favoring a wide prediction layer ( $N_d=64$ ) and narrow attention layer ( $N_a=8$ ) with seven decision steps.
Heterogeneous Stacking Ensemble:
- Constructs a four-model base ensemble consisting of:
  - RA-TabNet: A Residual-Augmented TabNet where a three-layer residual network corrects TabNet's outputs, blending predictions at a 60:40 ratio.
  - LightGBM, XGBoost, and CatBoost: Gradient boosting models trained on the reduced 8-feature set.
- A Logistic Regression meta-learner combines the validation-set predictions of the base models, learning optimal combination weights while applying L2 regularization to prevent dominance by a single model.

Key Contributions
The paper outlines five specific contributions:

Heterogeneous Stacking Architecture: Demonstrates that combining deep tabular (RA-TabNet) and tree-based models (LightGBM, XGBoost, CatBoost) via a learned meta-learner yields a primary performance advantage (+0.71 percentage points AUC over standalone TabNet).
CBAOFS: Introduces a Chaotic Binary Aquila Optimizer with an entropy-guided four-map ensemble, achieving significant dimensionality reduction (21 to 8 features) for improved interpretability.
CSCA: Applies a Chaotic Sine-Cosine Algorithm for TabNet hyperparameter optimization, moving beyond standard grid or random search.
Entropy-Guided Chaotic Engine: Develops an adaptive mechanism that selects the most appropriate chaotic map (Logistic, Tent, Sine, or Chebyshev) at each iteration based on real-time entropy monitoring.
Comprehensive Evaluation Protocol: Establishes a rigorous validation methodology including five independent runs with distinct seeds, Friedman and Wilcoxon statistical tests, Cohen's d effect sizes, bootstrap confidence intervals, a seven-variant ablation study, and a "without-duration" assessment to simulate realistic deployment constraints.

Results
Evaluated on the UCI Bank Marketing dataset (41,188 samples), CHIMERA-Tab achieved:

Mean AUC-ROC: 0.9507 ± 0.0026
Mean Macro F1-score: 0.7949 ± 0.0056

Statistical Validation:

CHIMERA-Tab significantly outperformed eight of nine baselines (including CatBoost, LightGBM, XGBoost, and MLP) via Wilcoxon signed-rank tests ( $p = 0.0313$ ).
It tied with CatBoost on AUC-ROC but surpassed it on Macro F1-score.
The framework demonstrated robustness, with the lowest variance across five runs compared to other methods.

Ablation and Sensitivity Analysis:

Primary Driver: The heterogeneous stacking architecture was identified as the main performance driver. Removing the stacking (A5-NoRA) resulted in a 0.71 percentage point drop in AUC.
Preprocessing Impact: The chaotic preprocessing (feature selection and hyperparameter tuning) provided marginal gains in raw classification accuracy (e.g., removing feature selection actually slightly increased AUC by 0.24pp) but offered substantial benefits in dimensionality reduction, inference speed, and interpretability.
Deployment Realism: When the "duration" feature (a post-call variable causing data leakage) was removed, AUC dropped to 0.8201. This highlights the gap between benchmark performance and realistic pre-call targeting capabilities.

Scalability and Training Time:

Scalability: The framework was tested on subsampled datasets. At 10% of the data (2,883 samples), the AUC-ROC was 0.9458, increasing monotonically to 0.9542 at 100% of the data (41,188 samples).
Training Time: The training time for the scaling analysis increased approximately linearly from 90.5s (at 10% data) to 334.7s (at 100% data). The complete pipeline, including all five runs, ablation study, scalability analysis, and explainability computations, consumed approximately 310 minutes on a CPU runtime.

Significance and Claims
The paper claims that CHIMERA-Tab addresses critical gaps in the literature by integrating deep learning and tree-based ensembles rather than evaluating them in isolation. It emphasizes that while chaotic metaheuristics may not always maximize the primary accuracy metric in isolation, their value lies in automation, dimensionality reduction, and interpretability.

The authors position the work as a methodological contribution that avoids the "deep learning versus trees" dichotomy. By transparently identifying the contribution hierarchy through ablation, the paper argues that the stacking ensemble is the critical innovation, while chaotic preprocessing serves as a complementary tool for efficiency and explainability. The study concludes that this framework offers a scalable, statistically validated approach for imbalanced tabular classification in financial contexts, with the "without-duration" metric serving as a crucial benchmark for real-world applicability.

More like this