Special Issue: Ensemble Learning for Operations Research and Business Analytics

The growing role of machine learning in operational research and business analytics is illustrated by the increasing attention in academic work as well as the widespread adoption by businesses, governments, and communities. McKinsey & Company recently estimated that roughly 25 percent of companies reinvest more than 5 percent of their earnings in data analytics. The popularity of analytics is accelerated through the digitization of markets and operations, global crises such as the COVID-19 pandemic, cheaper data storage, and more powerful and more affordable computing environments.

In business analytics and operations research (OR) the need for high-performance, scalable predictive methods is more pronounced than ever. In this context, ensemble learning quickly became an established modeling paradigm in the past two to three decades. Ample theoretical and empirical research has indeed demonstrated the beneficial effect of combining many classification models into aggregated ensemble classifiers on classification accuracy. Many contributions are inspired by classical ensemble strategies including bagging, boosting, error-correcting output codes, and stacking. These straightforward strategies all aim to create more capable and robust predictive models from weak or high-variance learners such as decision trees or stumps. However, these basic algorithms have spawned a plethora of now well-established methods such as Random Forest, Rotation Forest, and XGBoost. These methods have proven to yield superior performance in several benchmark studies in operations management and business analytics.

Ensemble algorithm innovation generally occurs on three levels. First, the base classifier can be varied. In an attempt to bring diversity to a pool of learners, one can rely on unstable learners or pursue variability by varying and optimizing base learner hyperparameters or even create heterogeneous ensembles where many different learners are combined. At the data level, diversification can be introduced by data transformation schemes, ranging from simple sampling approaches to feature extraction, etc. Finally, at the fusion level, one decides which ensemble members contribute to the model's predictions, and how. Recently, much attention was given to optimizing this choice through ensemble pruning, or even creating dynamic approaches. On each of these levels further innovation is possible to tailor solutions to specific applications in business analytics as well as OR. For example, several studies have explored the potential of changing the objective function of the underlying fusion mechanism or base classifiers to make them more attuned to the specific business problem.

Furthermore, recent developments have introduced a need for further contributions in a number of areas. First, in an era characterized by a quickly growing popularity of neural networks and deep learning to digest large volumes of often non-structured data, much potential lies in exploring the interface between deep and ensemble learning. Second, the growing importance of XAI, or explainable artificial intelligence, introduces a need for introducing more easily interpretable, justifiable, and actionable ensemble learning methods. Finally, more tailor-made applications of ensemble learning to business such as shifting from error-based to cost-sensitive or profit-driven learning and evaluation.

Specific topics of interest include, but are not limited to the following:

• Innovative applications of ensemble learning in OR or business analytics (e.g., healthcare operations, learning analytics, customer churn, credit scoring, fraud detection, social media analytics, sales forecasting, supply chain, production and planning, etc.)

• Combinatorial strategies for creating ensemble learners of deep learning models

• Exploring synergies between ensemble learning and conventional OR-methodologies, such as data envelopment analysis or methods for multi-criteria decision making

• Exploring the interface of ensemble and deep learning

• Ensemble learning for non-standard data and/or modeling paradigms: unstructured data, time series data, active learning, multilabel classification, semi-supervised learning, reinforcement learning

• Exploring the merit of using deep learners as base learners in ensemble models over conventional models (decision trees, shallow neural networks, etc.)

• Novel approaches to ensemble selection or pruning

• Cost-sensitive or profit-driven ensemble learning

• Integrating optimization (single or multi-criteria) in ensemble learning

• Explainable ensemble learning

• Investigating synergy in integrating feature engineering or model evaluation in ensemble learning methods

Call for Papers Flyer: Ensemble Learning for Operations Research and Business Analytics