One-Shot Induction of Generalized Logical Concepts via Human Guidance, STARAI, 2020

Authors:

M. Das,

N. Ramanan,

J. Doppa and

S. Natarajan

Abstract: We consider the problem of learning generalized first-order representations of concepts from a single example. To address this challenging problem, we augment an inductive logic programming learner with two novel algorithmic contributions. First, we define a distance measure between candidate concept representations that improves the efficiency of search for target concept and generalization. Second, we leverage richer human inputs in the form of advice to improve the sample-efficiency of learning. We prove that the proposed distance measure is semantically valid and use that to derive a PAC bound. Our experimental analysis on diverse concept learning tasks demonstrates both the effectiveness and efficiency of the proposed approach over a first-order concept learner using only examples.

Discriminative Non-Parametric Learning of Arithmetic Circuits, ICML Workshop on Tractable Probabilistic Models, 2019

Authors:

N. Ramanan,

M. Das,

K. Kersting and

S. Natarajan

Abstract: Arithmetic Circuits (AC) and Sum-Product Networks (SPN) have recently gained significant interest by virtue of being tractable deep probabilistic models. Most previous work on learning AC structures, however, hinges on inducing a tree-structured AC and, hence, may potentially break loops that may exist in the true generative model. To repair such broken loops, we propose a gradient-boosted method for structure learning of discriminative ACs (DACs), called DACBOOST. Since, in discrete domains, ACs are essentially equivalent to mixtures of trees, DACBOOST decomposes a large AC into smaller tree-structured ACs and learns them in a sequential, additive manner. The resulting non-parametric manner of learning the DACs results in a model with very few tuning parameters making our learned model significantly more efficient. We demonstrate on standard data sets and some real-world data sets, the efficiency of DACBOOST compared to the state-of-the-art DAC learners without sacrificing the effectiveness. This makes it possible to employ DACs for large scale real-world tasks.

Ensemble Causal Learning, in AAAI Symposium,2019

Authors:

N. Ramanan and

S. Natarajan

Abstract: In this work-in-progress paper, we speculate a method for learning causal models directly from data without any interventions or inductive bias. Our ensemble approach uncovers some interesting relations for understanding post-partum depression
based on family and socio-economic factors.

Structure Learning for Relational Logistic Regression: An Ensemble Approach, International Conference on Principles of Knowledge Representation and Reasoning, 2018

Authors:

N. Ramanan,

G. Kunapuli,

T. Khot,

B. Fatemi,

S.M. Kazemi,

D. Poole,

K. Kersting and

S. Natarajan

Abstract: We consider the problem of learning Relational Logistic Regression (RLR). Unlike standard logistic regression, the features of RLRs are first-order formulae with associated weight vectors instead of scalar weights. We turn the problem of learning RLR to learning these vector-weighted formulae and develop a learning algorithm based on functional-gradient boosting methods for probabilistic logic models. Our empirical evaluation on standard data sets demonstrates the superiority of our approach over other methods for learning RLR.

On Whom Should I Perform the Lab Test on Next? An Active Feature Elicitation Approach, IJCAI, 2018

Authors:

S. Natarajan,

S. Das,

N. Ramanan,

G. Kunapuli and

P. Radivojac

Abstract: We consider the problem of active feature elicitation in which, given some examples with all the features (say, the full Electronic Health Record), and many examples with some of the features (say, demographics), the goal is to identify the set of examples
on which more information (say, lab tests) need to be collected. The observation is that some set of features may be more expensive, personal or cumbersome to collect. We propose a classifier independent, similarity metric-independent, general active learning approach which identifies examples that are dissimilar to the ones with the full set of data and acquire the complete set of features for these examples. Motivated by four real clinical tasks, our extensive evaluation demonstrates the effectiveness of this approach.

Discriminative Boosted Bayes Networks for Learning Multiple
Cardiovascular Procedures, International Conference on Bioinformatics and Biomedicine, 2017

Authors:

N. Ramanan,

S. Yang,

S. Grannis and

S. Natarajan

Abstract: We consider the problem of predicting three procedures, viz, EKG, Angioplasty and Valve Replacement procedures jointly from Electronic Health Records (EHR) and develop a discriminative boosted Bayesian network algorithm. Differences between our proposed approach and standard Bayes Net structure learners are (1) we do not assume that the number of features (observations) are uniform across training examples and (2) our method explicitly handles the precision-recall tradeoff. Our empirical evaluations on a real EHR data demonstrates the superiority of this proposed approach to learning these procedures individually.

Boosting for Postpartum Depression Prediction, ACM IEEE CHASE, 2017

Authors:

S. Natarajan,

A. Annu Prabhakar, N. Ramanan,

A. Bagilone,

K. Siek and

K. Connelly

Abstract: Pregnancy and childbirth are important transitional life events for women. Like many other transitional life events, the effects of pregnancy and childbirth can have a significant impact on a mother’s physical and mental well-being. Sometimes they can even lead to Postpartum Depression (PPD). If left untreated, PPD can be debilitating for the mother and can adversely affect her ability to take care of herself and her infant. Since PPD is not clinically diagnosable, we consider the problem of predicting PPD from survey data about demographics, depression, and pregnancy etc. We adapt the successful functional-gradient boosting algorithm that can handle class imbalance in a principled manner. Our results demonstrate that the proposed machine learning approach can outperform the baseline classifiers and, consequently, demonstrate the potential of machine learning in predicting PPD.