We introduce COAT (Counterfactual Optimal Action Tree), a framework for learning interpretable prescriptive policies from observational data. COAT combines counterfactual outcome estimation with large-scale mixed-integer optimization, using column generation to translate causal predictions into feasible, transparent decisions under business and regulatory constraints. We apply COAT to airline ancillary pricing, a setting characterized by complex business rules and limited experimental flexibility. In a 17-week field pilot with a major global airline, COAT increased upsell revenue per booking by 6.9

We introduce a causal aware foundation-model framework for real time optimal decision making in discrete choice environments. We propose a constrained triple-head price optimization (C3PO) network to solve a bilevel decision problem in which a service provider selects an optimal assortment while heterogeneous users make personalized acceptance or rejection choices optimizing their own personalized preferences. C3PO integrates imitation learning of prices, multi-task learning of revenue responses, and in context learning of price elasticity to generate pricing recommendations while adhering to business constraints. During inference, frontier model prompting retrieves an enhanced elasticity prior for new products from behavioral economics literature, improving pricing effectiveness. We demonstrate strong in context learning performance using simulated, synthetic, and real-world datasets. C3PO is trained on simulated data generated from multiple classical discrete choice models in economics. The model is trained on data comprising simulated customer segments and counterfactual action and outcome pairs and evaluated on randomly generated choice environments with no access to the underlying preference structure. The trained model consistently improves the pricing KPIs, with gains increasing as customer price sensitivity increases. We also deploy the tuned foundation model for optimal pricing in real-world applications such as healthcare, tender pricing, airline ancillary pricing, and other domains, achieving substantial gains across multiple products, markets, and divisions.

This paper considers the pricing of multi-product request-for-quotes (RFQs) that are configured by a buyer based on a large number of products or services offered in a seller's product catalog. The buyer submits an RFQ for a desired bundle of line items in a bid configuration to a seller. The seller reviews the configuration and offers an approved price for each line item in the bundle. The buyer can selectively purchase any combination of products or services in the bid configuration at the seller's approved prices. In addition to the line item pricing approach, we propose a novel loss-leader model that uses machine learning to calibrate the buyer's preferences among correlated line items, and dynamically optimizes the prices of any configuration to maximize the seller's expected profit. The pricing strategies were implemented in a business-to-business (B2B) sales environment with a multinational technology company. Counterfactual analysis shows that loss-leader pricing can generate more than ten percent lift in gross profit over existing pricing practices.

Prescriptive artificial intelligence (AI) represents a transformative shift in decision-making by offering causal insights and actionable recommendations. Despite its huge potential, enterprise adoption of prescriptive AI faces several challenges. One such challenge is caused by the lack of experimental data for many enterprises, making it hard to attribute differences in outcomes to interventions alone. The second pertains to the explainability of AI recommendations, which is crucial for enterprise decision-making settings. The third challenge is contributed by the silos between technologists and business users, hindering effective collaboration. This paper outlines an initiative from IBM Research, PresAIse, aiming to address some of these challenges by offering a suite of prescriptive AI solutions. Leveraging insights from various research papers, the solutions include scalable causal inference methods, interpretable decision-making approaches, and the integration of large language models (LLMs) to bridge the communication gap via a conversation agent. A proof-of-concept demonstrates the solutions' potential by enabling non-ML experts to interact with prescriptive AI models via a natural language interface, democratizing advanced analytics for strategic decision-making.

Most airlines use dynamic pricing to optimize the price of their base economy product by maximizing the expected revenue. However, when it comes to pricing of premium products, airlines often uses a static price increments that are applied to the best available economy fare based on simple business rules for adjusting the price based on supply. In this paper, we present a suite of machine learning algorithms that take advantage of the rich booking session context available at the time of the booking to make its predictions. The challenge is to accurately predict bookings for new combinations of attributes by market and segment (departure time, length of stay, advance purchase, length of haul, …) while accounting for cross-product price effects in a scalable manner. To generate practical pricing policies, the approach accommodates a variety of real-world business requirements into the decision optimization problem. We present a scalable approach based on a novel path-based mixed-integer program (MIP) reformulation that can efficiently recover near-optimal pricing policies. We demonstrate the efficacy of our model with extensive experiments on synthetic and real-life data. Finally, we present an airline case study on deriving profitable prescriptive policies for premium cabin tickets based on easily interpretable pricing rules.

Attribute-based pricing---giving a price to potential product attributes individually and allowing customers to choose the attributes that form the final product---has been shown to improve customer satisfaction in the hospitality industry. In this paper, we consider the problem of finding optimal attribute prices to maximize the expected revenue from selling to a customer who chooses one product from a set of products that differ by only a few attributes. Because of complicated substitution effects among the final products that share common attributes, expected revenue is not concave in attribute prices. Nonetheless, we provide an algorithm to solve the attribute pricing problem and show that it converges to a stationary point that provides a high-quality solution to the problem. Through numerical experiments, we show our algorithm is, on average, ten times faster than gradient-based methods, both in terms of runtime and number of iterations. We also extend our algorithm to a setting where attribute prices are constrained through linear inequalities and prove convergence to a stationary point. We implement our algorithm on a real hotel data set and demonstrate its revenue benefits.

Customer satisfaction is crucial for the long term success of any travel service provider.Therefore, identifying situations that can lead to customer dissatisfaction is critical.The strongest evidence of customers dissatisfaction are their complaints.While complaints do not occur very often, they almost always lead to loss of customer goodwill which can cost travel providers millions of dollars in future revenues.In this paper, we describe an approach to proactively identify customers that have the highest propensity to complain as they encounter a travel disruption event.These are invaluable insights that can empower customer service teams with information to deliver a more timely, relevant and impactful service experience.We use three key aspects in this approach: (i) specialized feature engineering for the travel industry; (ii) handling extremely imbalanced data and (iii) adaptation of binary classification, anomaly detection and learning to rank models to our specific task.This research is an important step towards more individualized understanding of customer behavior, and potential service enhancements to further increase customer satisfaction.

With the abundance of available data, many enterprises seek to implement data-driven prescriptive analytics to help them make informed decisions. These prescriptive policies need to satisfy operational constraints, and proactively eliminate rule conflicts, both of which are ubiquitous in practice. It is also desirable for them to be simple and interpretable, so they can be easily verified and implemented. Existing approaches from the literature center around constructing variants of prescriptive decision trees to generate interpretable policies. However, none of the existing methods is able to handle constraints. In this paper, we propose a scalable method that solves the constrained prescriptive policy generation problem. We introduce a novel path-based mixed-integer program (MIP) formulation which identifies a (near) optimal policy efficiently via column generation. The policy generated can be represented as a multiway-split tree which is more interpretable and informative than binary-split trees due to its shorter rules. We demonstrate the efficacy of our method with extensive computational experiments on both synthetic and real datasets.

Data-driven pricing strategies are becoming increasingly common, where customers are offered a personalized price based on features that are predictive of their valuation of a product. It is desirable for this pricing policy to be simple and interpretable, so it can be verified, checked for fairness, and easily implemented. However, efforts to incorporate machine learning into a pricing framework often lead to complex pricing policies which are not interpretable, resulting in slow adoption in practice. We present a customized, prescriptive tree-based algorithm that distills knowledge from a complex black-box machine learning algorithm, segments customers with similar valuations and prescribes prices in such a way that maximizes revenue while maintaining interpretability. We quantify the regret of a resulting policy and demonstrate its efficacy in applications with both synthetic and real-world datasets.

Problem definition : The growing trend in online shopping has sparked the development of increasingly more sophisticated product recommendation systems. We construct a model that recommends a personalized discounted product bundle to an online shopper that considers the trade-off between profit maximization and inventory management, while selecting products that are relevant to the consumer’s preferences. Academic/practical relevance : We provide analytical performance guarantees that illustrate the complexity of the underlying problem, which combines assortment optimization with pricing. We implement our algorithms in two separate case studies on actual data from a large U.S. e-tailer and a premier global airline. Methodology : We focus on simultaneously balancing personalization through individualized functions of consumer propensity-to-buy, inventory management for long-run profitability, and tractability for practical business implementation. We develop two classes of approximation algorithms, multiplicative and additive, to produce a real-time output for use in an online setting. Results : Our computational results demonstrate significant lifts in expected revenues over current industry pricing strategies on the order of 2%–7% depending on the setting. We find that on average our best algorithm obtains 92% of the expected revenue of a full-knowledge clairvoyant strategy across all inventory settings, and in the best cases this improves to 98%. Managerial implications : We compare the algorithms and find that the multiplicative approach is relatively easier to implement and on average empirically obtains expected revenues within 1%–6% of the additive methods when both are compared with a full-knowledge strategy. Furthermore, we find that the greatest expected gains in revenue come from high-end consumers with lower price sensitivities, and that predicted improvements in sales volume depend on product category and are a result of providing relevant recommendations.

Consumers are increasingly navigating across sales channels to maximize the value of their purchase. The existing retail practices of pricing channels either independently or matching competitor prices are unable to achieve the desired profitable coordination between channels. We engaged with three major retailers over two years and developed omnichannel pricing (OCP) solutions in partnership with IBM Commerce to overcome these challenges. We implement an integrated data processing and machine learning framework that enables estimation of location-specific, cross-channel price elasticities and competitive effects. We develop an integrated OCP optimization formulation to profitably coordinate prices for nonperishable products offered across channels and store locations while satisfying practical constraints on volume and price. The resultant optimization formulations for discrete choice demand models are nonconvex and NP-hard, and we prescribe practically efficient mixed-integer programs that can be used to recover (near) optimal solutions. An OCP implementation in two categories for a major retail chain projected a 7% profit lift while preserving the sales volume. This benefit was achieved by lowering online prices and optimally raising and lowering location-specific store prices. This integrated pricing approach allows the retailer to be competitive and preserve market share without aggressively matching the low price of e-tail giants.

The commercialization of this work is a result of the combined efforts that includes the following team members from the IBM Commerce business unit: Saibal Bhattacharya, Irina Fedulova, Vitaly Grechko, Nikolay Murzin, Manjunath Pandit, Emily Port, Erich Schellhas, Maloney Whitfield, Oleg Sidorkin, Mahesh Virupakshaiah, Jin Jing Xie, Emrah Zarifoglu and Pavel Zelinsky. This list includes the data scientists, product managers and client executives. We sincerely acknowledge the Omni-channel Retailer who wishes to remain anonymous and more specifically, their pricing team, who provided the key business motivation and insights that helped in shaping the solution design.

This paper studies the pricing strategies for personalized product bundles. In such problems, a seller provides a variety of products for which customers can construct a personalized bundle and send a request for quote (RFQ) to the seller. The seller, after reviewing the RFQ, has to determine a price based on which the customer either purchases the whole bundle or nothing. Such problems are faced by many companies in practice, and they are very difficult because of the potential unlimited possible configurations of the bundle and the correlations among the individual products. In this paper, we propose a novel top-down and bottom-up approach to solve this problem. In the top-down step, we decompose the bundle into each component and calibrate a value score for each component. In the bottom-up step, we aggregate the components back to the bundle, define important features of the bundle, and segment different RFQs by those bundle features as well as customer attributes. Then we estimate a utility function for each segment based on historical sales data and derive an optimal price for each incoming RFQ. We show that such a model overcomes the aforementioned difficulties and can be implemented efficiently. We test our approach using empirical data from a major information technology service provider and the test result shows that the proposed approach can improve the effectiveness of pricing significantly.

An accredited biennial 2014 study by the Association of Certified Fraud Examiners claims that on average 5% of a company's revenue is lost because of unchecked fraud every year. The reason for such heavy losses are that it takes around 18 months for a fraud to be caught and audits catch only 3% of the actual fraud. This begs the need for better tools and processes to be able to quickly and cheaply identify potential malefactors. In this paper, we describe a robust tool to identify procurement related fraud/risk, though the general design and the analytical components could be adapted to detecting fraud in other domains. Besides analyzing standard transactional data, our solution analyzes multiple public and private data sources leading to wider coverage of fraud types than what generally exists in the marketplace. Moreover, our approach is more principled in the sense that the learning component, which is based on investigation feedback has formal guarantees. Though such a tool is ever evolving, a deployment of this tool over the past 12 months has found many interesting cases from compliance risk and fraud point of view across more than 150 countries and 65000+ vendors, increasing the number of true positives found by over 80\% compared with other state-of-the-art tools that the domain experts were previously using.

An accredited biennial 2012 study by the Association of Certified Fraud Examiners claims that on average 5% of a company’s revenue is lost because of unchecked fraud every year. The reason for such heavy losses are that it takes around 18 months for a fraud to be caught and audits catch only 3% of the actual fraud. This begs the need for better tools and processes to be able to quickly and cheaply identify potential malefactors. In this paper, we describe a robust tool to identify procurement related fraud/risk, though the general design and the analytical components could be adapted to detecting fraud in other domains. Besides analyzing standard transactional data, our solution analyzes multiple public and private data sources leading to wider coverage of fraud types than what generally exists in the marketplace. Moreover, our approach is more principled in the sense that the learning component, which is based on investigation feedback has formal guarantees. Though such a tool is ever evolving, an initial deployment of this tool over the past 6 months has found many interesting cases from compliance risk and fraud point of view, increasing the number of true positives found by over 80% compared with other state-of-the-art tools that the domain experts were previously using.