Dismiss
InnovationQ will be updated on Sunday, Oct. 22, from 10am ET - noon. You may experience brief service interruptions during that time.
Browse Prior Art Database

A NOVEL ALGORITHM FOR MULTI-INSTANCE LEARNING UTILIZING A STACKED ENSEMBLE OF LAZY LEARNING CLASSIFIERS

IP.com Disclosure Number: IPCOM000238677D
Publication Date: 2014-Sep-11

Publishing Venue

The IP.com Prior Art Database

Abstract

The invention proposes a technique to classify groups and bags of instances on basis of a set of labeled bags. The technique includes an algorithm to classify such groups and bags. The proposed invention utilizes an ensemble of classifiers that work on a modification of nearest neighbor principle. For the modification of nearest neighbor principle each individual classifier is tuned to utilize a different combination of features, distance metric, neighborhood definition among others. The set of classifiers is selected to optimize trade-off between correctly identified all positive bags and reduced number of false positives. The classifiers are combined utilizing a second level classifier that benefits from diversity and achieves improved performance.

This text was extracted from a Microsoft Word document.
At least one non-text object (such as an image or picture) has been suppressed.
This is the abbreviated version, containing approximately 18% of the total text.

A NOVEL ALGORITHM FOR MULTI-INSTANCE LEARNING UTILIZING A STACKED ENSEMBLE OF LAZY LEARNING CLASSIFIERS

BRIEF ABSTRACT

The invention proposes a technique to classify groups and bags of instances on basis of a set of labeled bags. The technique includes an algorithm to classify such groups and bags. The proposed invention utilizes an ensemble of classifiers that work on a modification of nearest neighbor principle. For the modification of nearest neighbor principle each individual classifier is tuned to utilize a different combination of features, distance metric, neighborhood definition among others. The set of classifiers is selected to optimize trade-off between correctly identified all positive bags and reduced number of false positives. The classifiers are combined utilizing a second level classifier that benefits from diversity and achieves improved performance.

KEYWORDS

Bags, algorithm, ensemble classifier, instance label, leave-one-out


DETAILED DESCRIPTION

A group of instances is labeled as a bag. The bag is labeled positive when the bag encloses at least one positive instance and negative when the bag encloses at least one negative instance. Instance-level in some cases is available but such instances are not always available. This is applicable in number of complications where some difficulties exist. Acquiring precisely labeled instances is difficult or time-consuming whereas a precise labeling at lower level of granularity is easily obtained for a larger sample of instances. For instance, computer aided detection includes difficulty inmedical imaging where an expert labels an image with one or more abnormalities indicating certain diseases. The specific regions with abnormalities have diffused character or not specifically marked out. Acquiring more precise labeling is time consuming and suffers from operator variability. In such case, an image is considered as a bag of candidate region of interest (ROI) or pixels where the image labeling applies to the bag.But a specific label for each pixel/ROI is either imprecise or unavailable. Another difficultyis one of classifying a bag or sequence of instances rather than one of classifying a single instance. For instance, a gas turbine startup advisor predicts whether or not a sequence of the gas turbine startups results to a trip start. In such a case, historical data are precisely labeled as to which startup has a trip and which one does not has a trip. The complication is one of classifying sequence or bag of startups. Data on which the model is deployed does not contain sequences of consecutive startups. As a result, a particular unit data for startups result in the missing trips. Modeling the complication as one on instance bag classification is an efficient alternative.

A conventional technique includes single or multi instance learning algorithm to classify a group of instances as opposed to classification at instance level. However, the conventional technique does not optimize para...