Date of Submission


Date of Award


Institute Name (Publisher)

Indian Statistical Institute

Document Type

Doctoral Thesis

Degree Name

Doctor of Philosophy

Subject Name

Computer Science


Computer Vision and Pattern Recognition Unit (CVPR-Kolkata)


Palit, Sarbani (CVPR-Kolkata; ISI)

Abstract (Summary of the Work)

Data processing by the human sensory system comes naturally. This processing, commonly denoted as pattern recognition and analysis are carried out spontaneously by humans. In day to day life, in most cases, decision making by humans come without any conscious effort. From the middle of the past century, humans have shown interest to render their abstraction capabilities (pattern recognition and analysis) to the machine. The abstraction capability of the machine is ’machine intelligence’ or ’machine learning’ [87].The primary goal of machine learning methods is to extract some meaningful information from the ’data’. Data refers to the information or attributes that are fed to a machine learning algorithm or method. The two main types of learning are – i] Summarization and ii] Generalization. An algorithm makes a summarization of the given information to understand the key components of the data. The algorithm might aim to learn the key features, key data point which can provide the particulars of the data. The other aspect of machine learning is generalization of data – extracting the underlying structure of the data to make correct prediction of upcoming new data. Classification of data by an algorithm requires it to make a generalization first.Classification of objects into different categories is a fundamental element of decision making. The machine learning community has taken a keen interest in developing competent classification algorithms (classifiers) since it’s outset. Thereupon, diversified paradigms like naive bayes [77], knn [65], neural networks [35] have emerged to facilitate effective classification of data. A classifier is modeled through the generalization of the given data or the training data. The purpose of a good classifier is to make correct predictions of the future data (or the test data) on the basis of the generated model. Consequently, the prediction capability of a classifier is dependent on how good one hasmodeled the classifier. The modeling in turn depends on the training data. To integrate and facilitate the intertwining of efficient modeling and predictions, some standard assumptions are made by the machine learning community.Classification algorithms, being mathematical formulations are devised under several such assumptions on the datasets. The assumptions can be i] the different classes of a dataset have similar cardinalities ( if the standard deviations in class cardinalities is high, we have the class-imbalance issue ), ii] the training and test partitions of a dataset will have similar attribute profile (same number of classes and distributions), iii] an instance can belong to one class only, iv] information about all attributes of the instances are known and more. The continuous and intelligent efforts of the machine learning community has given ( and is still providing ) a class of robust classifiers in the past and present times. However, often, it is observed that a ’potent’ classifier which gives accurate predictions under a given assumption fails to deliver optimal performance whenever it is exposed to situations where these assumptions are not fulfilled. To get admissible performance from the conventional classifiers, it is essential to comply with the assumptions.Real-world data from diversified domains like medical, biology, security, banking, social networking, web-data, news articles show breaches of several of these assumptions.


ProQuest Collection ID:

Control Number


Creative Commons License

Creative Commons Attribution 4.0 International License
This work is licensed under a Creative Commons Attribution 4.0 International License.


Included in

Mathematics Commons