- Number
- 10614379
- Published
- 2020-04-07
- Filed
- 2016-09-27
- Assignee
- Disney Enterprises, Inc.
- Inventors
- Fu; Yanwei, Sigal; Leonid
- CPC
- G06N20/00; G06F16/285; G06F16/9024; G06F16/51; G06F16/55; G06F16/583
- Verdict
- Set aside generic ML classification research, no application hook
- Source
- Google Patents · FreePatentsOnline
Abstract
Techniques are disclosed for identifying and filtering outliers from a sample set of data prior to training a classifier on an object using the sample set. A data set including a plurality of samples used to train a classification model is retrieved. The samples in the data set have a feature dimensionality. A graph of the data set is built. Each node in the graph corresponds to a sample in the data set and edges connecting the nodes correspond to a measure of similarity between the nodes. The feature dimensionality of the sample data set is reduced based on a topology of the graph. One or more outliers in the data set are identified based on the reduced feature dimensionality.
Background
BACKGROUNDField(1) Embodiments presented herein generally relate to data classification (e.g., image classification, audio classification, etc.), and more specifically, to identifying and filtering outliers from a data set.Description of the Related Art(2) Many modern computer data recognition approaches use supervised learning techniques that rely on a large set of labeled data to train classification models. Such data sets have been collected from the web by searching for query terms relevant to a particular object label and verifying consistency through crowdsourced labeling, i.e., soliciting verification of the label to a data set from a large group of people. Such an approach is often costly and difficult to scale. For example, image data sets often include multiple labels, and each image in the set needs to be verified or labeled by a group of individual annotators.(3) To address such difficulties of image data sets, automated methods may collect, cluster and annotate the images based on characteristics or features found in the images. The annotated clusters can be used to train a classifier to recognize one or more objects in subsequently presented example images where examples of that object appeared in the annotated clusters on a given image-based object (i.e., an object that appears in the image). However, the resulting annotated clusters tend to include a number of outliers, i.e., images that do not correspond to an instance of the object associated with the cluste
Claims
1. A computer-implemented method of outlier detection and removal for classifier training, the computer-implemented method comprising: retrieving a data set including a plurality of samples used to train a classification model, wherein the samples in the data set have an original feature dimensionality; building a graph data structure of the data set, the graph data structure comprising a transductive graph that is a k-nearest-neighbor graph having a plurality of nodes, each node of which corresponds to a respective sample of the plurality of samples, wherein an edge connecting first and second nodes of the plurality of nodes corresponds to a measure of similarity between the first and second nodes, wherein building the graph data structure comprises defining a similarity weight between nodes as a function of an inner product of a plurality of features of the nodes with a free parameter; determining, by operation of one or more computer processors, a reduced feature dimensionality, relative to the original feature dimensionality, of the data set using a transductive diffusion component analysis over the plurality of nodes of the graph data structure, wherein determining the reduced feature dimensionality of the data set includes determining a softmax approximation over the transductive graph; identifying, based on the reduced feature dimensionality, one or more outliers in the plurality of samples using a pre-conditioned least absolute shrinkage and selection operator, wherein the one or more outliers are identified based on a determined regularization path of the least absolute shrinkage and selection operator; modifying the plurality of samples by removing the identified one or more outliers from the plurality of samples; training the classification model on an inlier object using the modified plurality of samples; and classifying the inlier object using the trained classification model.
9. A non-transitory computer-readable medium storing instructions executable to perform an operation of outlier detection and removal for classifier training, the operation comprising: retrieving a data set including a plurality of samples used to train a classification model, wherein the samples in the data set have an original feature dimensionality; building a graph data structure of the data set, the graph data structure comprising a transductive graph that is a k-nearest-neighbor graph having a plurality of nodes, each node of which corresponds to a sample of the plurality of samples, wherein an edge connecting first and second nodes of the plurality of nodes corresponds to a measure of similarity between the first and second nodes, wherein building the graph data structure comprises defining a similarity weight between nodes as a function of an inner product of a plurality of features of the nodes with a free parameter; determining, by operation of one or more computer processor when executing the instructions, a reduced feature dimensionality, relative to the original feature dimensionality, of the data set using a transductive diffusion component analysis over the plurality of nodes of the graph data structure, wherein determining the reduced feature dimensionality of the data set includes determining a softmax approximation over the transductive graph; and identifying, based on the reduced feature dimensionality, one or more outliers in the plurality of samples using a pre-conditioned least absolute shrinkage and selection operator, wherein the one or more outliers are identified based on a determined regularization path of the least absolute shrinkage and selection operator; modifying the plurality of samples by removing the identified one or more outliers from the plurality of samples; training the classification model on an inlier object using the modified plurality of samples; and classifying the inlier object using the trained classification model.
15. A system of outlier detection and removal for classifier training, the system comprising: one or more computer processors; and a memory storing program code which, when executed on the one or more computer processors, performs an operation comprising: retrieving a data set including a plurality of samples used to train a classification model, wherein the samples in the data set have an original feature dimensionality; building a graph data structure of the data set, the graph data structure comprising a transductive graph that is a k-nearest-neighbor graph having a plurality of nodes, each node of which corresponds to a respective sample of the plurality of samples, wherein an edge connecting first and second nodes of the plurality of nodes corresponds to a measure of similarity between the first and second nodes, wherein building the graph data structure comprises defining a similarity weight between nodes as a function of an inner product of a plurality of features of the nodes with a free parameter; determining a reduced feature dimensionality, relative to the original feature dimensionality, of the data set using a transductive diffusion component analysis over the plurality of nodes of the graph data structure, wherein determining the reduced feature dimensionality of the data set includes determining a softmax approximation over the transductive graph; and identifying, based on the reduced feature dimensionality, one or more outliers in the plurality of samples using a pre-conditioned least absolute shrinkage and selection operator, wherein the one or more outliers are identified based on a determined regularization path of the least absolute shrinkage and selection operator; modifying the plurality of samples by removing the identified one or more outliers from the plurality of samples; training the classification model on an inlier object using the modified plurality of samples; and classifying the inlier object using the trained classification model.