Ramasubramanian (Ramsu) Sundararajan

We consider the problem of product bundling (seats and ancillaries) in the context of offering the right products to airline customers at the right price and in the right manner, so as to best satisfy customer needs and maximize airline revenue. This problem falls on the cusp of airline revenue management (apropos controlling price and availability) and retail e-commerce (apropos bundle design and shopping session management); therefore, we synthesize ideas from both domains to devise a… Show more

We consider the problem of product bundling (seats and ancillaries) in the context of offering the right products to airline customers at the right price and in the right manner, so as to best satisfy customer needs and maximize airline revenue. This problem falls on the cusp of airline revenue management (apropos controlling price and availability) and retail e-commerce (apropos bundle design and shopping session management); therefore, we synthesize ideas from both domains to devise a solution framework. Our proposed solution is designed in a modular manner, so as to allow incremental and independent improvements to product design, pricing, and shopping session management. In this paper, we specifically focus on methodologies for offer construction: creating product bundles and estimating willingness to pay. We demonstrate the utility of these methodologies through illustrative results on real and simulated datasets. Show less

This document describes a novel learning algorithm that classifies "bags" of instances rather than individual instances. A bag is labeled positive if it contains at least one positive instance (which may or may not be specifically identified), and negative otherwise. This class of problems is known as multi-instance learning problems, and is useful in situations where the class label at an instance level may be unavailable or imprecise or difficult to obtain, or in situations where the problem… Show more

This document describes a novel learning algorithm that classifies "bags" of instances rather than individual instances. A bag is labeled positive if it contains at least one positive instance (which may or may not be specifically identified), and negative otherwise. This class of problems is known as multi-instance learning problems, and is useful in situations where the class label at an instance level may be unavailable or imprecise or difficult to obtain, or in situations where the problem is naturally posed as one of classifying instance groups. The algorithm described here is an ensemble-based method, wherein the members of the ensemble are lazy learning classifiers learnt using the Citation Nearest Neighbour method. Diversity among the ensemble members is achieved by optimizing their parameters using a multi-objective optimization method, with the objectives being to maximize Class 1 accuracy and minimize false positive rate. The method has been found to be effective on the Musk1 benchmark dataset. Show less

We consider the problem of finding common behavioral patterns among travelers in an airline network through the process of clustering. Travelers can be characterized at relational or transactional level. In this article, we focus on the transactional level characterization; our unit of analysis is a single trip, rather than a customer relationship comprising multiple trips. We begin by characterizing a trip in terms of a number of features that pertain to the booking and travel behavior. Trips… Show more

We consider the problem of finding common behavioral patterns among travelers in an airline network through the process of clustering. Travelers can be characterized at relational or transactional level. In this article, we focus on the transactional level characterization; our unit of analysis is a single trip, rather than a customer relationship comprising multiple trips. We begin by characterizing a trip in terms of a number of features that pertain to the booking and travel behavior. Trips thus characterized are then grouped using an ensemble clustering algorithm that aims to find stable clusters as well as discover subgroup structures within groups. A multidimensional analysis of trips based on these groupings leads us to discover non-trivial patterns in traveler behaviour that can then be exploited for better revenue management. Show less

This paper deals with grouping entities in a population which behave similarly. The behavior of each entity is represented by a regression function estimated using observational data of the entity comprising a dependent variable and one or more independent variables. No restriction is imposed on the form and nature of individual regression function, but it is assumed that the same variables within similar range are present in each dataset. These functions may represent stimulus-response models… Show more

This paper deals with grouping entities in a population which behave similarly. The behavior of each entity is represented by a regression function estimated using observational data of the entity comprising a dependent variable and one or more independent variables. No restriction is imposed on the form and nature of individual regression function, but it is assumed that the same variables within similar range are present in each dataset. These functions may represent stimulus-response models or growth curves in practical applications. The regression model of each entity is tested on
the dataset of every other entity. The corresponding validation error is used as a measure of difference in behavior between two entities and also as the edge weight in a graph having entities as nodes. This graph is clustered by finding the optimal split into communities using an iterative method and a proposed measure called the average meta-validation accuracy. The algorithm is tested on synthetic data such that the performance can be judged visually. Show less

We consider the problem of building classifiers with the option to reject i.e., not return a prediction on a given test example. Adding a reject option to a classifier is well-known in practice; traditionally, this has been accomplished in two different ways. One is the decoupled method where an optimal base classifier (without the reject option) is build first and then the rejection boundary is optimized, typically in terms of a band around the separating surface. The coupled method is based… Show more

We consider the problem of building classifiers with the option to reject i.e., not return a prediction on a given test example. Adding a reject option to a classifier is well-known in practice; traditionally, this has been accomplished in two different ways. One is the decoupled method where an optimal base classifier (without the reject option) is build first and then the rejection boundary is optimized, typically in terms of a band around the separating surface. The coupled method is based on finding both the classifier as well as the rejection band at the same time. Existing coupled approaches are based on minimizing risk under an extension of the classical 0−1 loss function wherein a loss d∈(0,.5) is assigned to a rejected example. In this paper, we propose a double ramp loss function which gives a continuous upper bound for (0−d−1) loss described above. Our coupled approach is based on minimizing regularized risk under the double ramp loss which is done using difference of convex (DC) programming. We show the effectiveness of our approach through experiments on synthetic and benchmark datasets. Show less

We consider the problem of automatically prescribing oblique planes (short axis, 4 chamber and 2 chamber views) in Cardiac Magnetic Resonance Imaging (MRI). A concern with technologist-driven acquisitions of these planes is the quality and time taken for the total examination. We propose an automated solution incorporating anatomical features external to the cardiac region. The solution uses support vector machine regression models wherein complexity and feature selection are optimized using… Show more

We consider the problem of automatically prescribing oblique planes (short axis, 4 chamber and 2 chamber views) in Cardiac Magnetic Resonance Imaging (MRI). A concern with technologist-driven acquisitions of these planes is the quality and time taken for the total examination. We propose an automated solution incorporating anatomical features external to the cardiac region. The solution uses support vector machine regression models wherein complexity and feature selection are optimized using multi-objective genetic algorithms. Additionally, we examine the robustness of our approach by training our models on images with additive Rician-Gaussian mixtures at varying Signal to Noise (SNR) levels. Our approach has shown promising results, with an angular deviation of less than 15 degrees on 90% cases across oblique planes, measured in terms of average 6-fold cross validation performance -- this is generally within acceptable bounds of variation as specified by clinicians. Show less

We consider the problem of propensity modeling in consumer finance. These modeling problems are characterized by the two aspects: the model needs to optimize a business objective which may be nonstandard, and the rate of occurence of the event to be modeled may be very low. Traditional methods such as logistic regression are ill-equipped to deal with nonstandard objectives and low event rates. Methods which deal with the low event rate problem by learning on biased samples face the problem of… Show more

We consider the problem of propensity modeling in consumer finance. These modeling problems are characterized by the two aspects: the model needs to optimize a business objective which may be nonstandard, and the rate of occurence of the event to be modeled may be very low. Traditional methods such as logistic regression are ill-equipped to deal with nonstandard objectives and low event rates. Methods which deal with the low event rate problem by learning on biased samples face the problem of overlearning. We propose a parallel genetic algorithm method that addresses these challenges. Each parallel process evolves propensity models based on a different biased sample, while a mechanism for validation and cross-pollination between the islands helps address the overlearning issue. We demonstrate the utility of the method on a real-life dataset. Show less

In this paper, we address the problem of making optimal product offers to customers of a retail bank by using techniques including Markov chains, genetic algorithms, mathematical programming, and design of experiments. Our challenges were large problem size, uncertainty about estimates of customer responses to product offers, and practical issues in training and implementation. The solution had an estimated financial impact of around $20 million; it also provided other intangible benefits… Show more

In this paper, we address the problem of making optimal product offers to customers of a retail bank by using techniques including Markov chains, genetic algorithms, mathematical programming, and design of experiments. Our challenges were large problem size, uncertainty about estimates of customer responses to product offers, and practical issues in training and implementation. The solution had an estimated financial impact of around $20 million; it also provided other intangible benefits, including structured decision making, the capability of performing what-if analysis, and portability to other markets and portfolios. Show less

We consider the problem of detecting the presence of pneumoconiosis in a patient on the basis of evidence found in chest radiographs. Abnormalities pertaining to pneumoconiosis appear in the form of opacities of various sizes; the profusion of these opacities determines the stage of the disease. We present a multiresolution approach whereby we segment regions of interest (ROIs) from the X-Ray image at two levels - lung field and lung zone. We characterize each of these regions using a set of… Show more

We consider the problem of detecting the presence of pneumoconiosis in a patient on the basis of evidence found in chest radiographs. Abnormalities pertaining to pneumoconiosis appear in the form of opacities of various sizes; the profusion of these opacities determines the stage of the disease. We present a multiresolution approach whereby we segment regions of interest (ROIs) from the X-Ray image at two levels - lung field and lung zone. We characterize each of these regions using a set of features and build support vector machine (SVM) classifiers that can predict whether or not the region contains any abnormalities. We combine these ROI-level predictions with a second stage SVM in order to get a prediction for the entire chest. Experimental validation shows that this approach provides good results. Show less

We consider the problem of selecting the optimal list of customers to target for a cross-sell campaign in a retail bank. Target selection involves taking estimates of several parameters (response propensity, expected volume, expected profit from a customer, etc) and deciding on the list of customers to whom the offer should be sent such that a certain set of business objectives are met/optimized. We discuss some of the issues related to the target selection process, namely those of unreliable… Show more

We consider the problem of selecting the optimal list of customers to target for a cross-sell campaign in a retail bank. Target selection involves taking estimates of several parameters (response propensity, expected volume, expected profit from a customer, etc) and deciding on the list of customers to whom the offer should be sent such that a certain set of business objectives are met/optimized. We discuss some of the issues related to the target selection process, namely those of unreliable estimates and computational complexity of the problem. We propose a fuzzy mathematical programming technique to address these issues. The imprecise parameters and constraints are represented as triangular fuzzy numbers, while the problem of computational complexity is addressed through a group-level formulation. We use an example of a real-life cross-sell problem for a bank to demonstrate the method. We also provide some sensitivity analyses on critical resources. Show less