Iker Gondra

Applying machine learning to software fault-proneness prediction

The importance of software testing to quality assurance cannot be overemphasized. The estimation of a modules fault-proneness is important for minimizing cost and improving the effectiveness of the software testing process. Unfortunately, no general technique for estimating software fault-proneness is available. The observed correlation between some software metrics and fault-proneness has resulted in a variety of predictive models based on multiple metrics. Much work has concentrated on how to select the software metrics that are most likely to indicate fault-proneness. In this paper, we propose the use of machine learning for this purpose. Specifically, given historical data on software metric values and number of reported errors, an Artificial Neural Network (ANN) is trained. Then, in order to determine the importance of each software metric in predicting fault-proneness, a sensitivity analysis is performed on the trained ANN. The software metrics that are deemed to be the most critical are then used as the basis of an ANN-based predictive model of a continuous measure of fault-proneness. We also view fault-proneness prediction as a binary classification task (i.e., a module can either contain errors or be error-free) and use Support Vector Machines (SVM) as a state-of-the-art classification method. We perform a comparative experimental study of the effectiveness of ANNs and SVMs on a data set obtained from NASAs Metrics Data Program data repository.

Content-based image retrieval with the normalized information distance

The main idea of content-based image retrieval (CBIR) is to search on an images visual content directly. Typically, features (e.g., color, shape, texture) are extracted from each image and organized into a feature vector. Retrieval is performed by image example where a query image is given as input by the user and an appropriate metric is used to find the best matches in the corresponding feature space. We attempt to bypass the feature selection step (and the metric in the corresponding feature space) by following what we believe is the logical continuation of the CBIR idea of searching visual content directly. It is based on the observation that, since ultimately, the entire visual content of an image is encoded into its raw data (i.e., the raw pixel values), in theory, it should be possible to determine image similarity based on the raw data alone. The main advantage of this approach is its simplicity in that explicit selection, extraction, and weighting of features is not needed. This work is an investigation into an image dissimilarity measure following from the theoretical foundation of the recently proposed normalized information distance (NID) [M. Li, X. Chen, X. Li, B. Ma, P. Vitanyi, The similarity metric, in: Proceedings of the 14th ACM-SIAM Symposium on Discrete Algorithms, 2003, pp. 863-872]. Approximations of the Kolmogorov complexity of an image are created by using different compression methods. Using those approximations, the NID between images is calculated and used as a metric for CBIR. The compression-based approximations to Kolmogorov complexity are shown to be valid by proving that they create statistically significant dissimilarity measures by testing them against a null hypothesis of random retrieval. Furthermore, when compared against several feature-based methods, the NID approach performed surprisingly well.

Learning in Region-Based Image Retrieval with Generalized Support Vector Machines

Relevance feedback approaches based on support vector machine (SVM) learning have been applied to significantly improve retrieval performance in content-based image retrieval (CBIR). Those approaches require the use of fixed-length image representations because SVM kernels represent an inner product in a feature space that is a non-linear transformation of the input space. Many region-based CBIR approaches create a variable length image representation and define a similarity measure between two variable length representations. The standard SVM approach cannot be applied to this approach because it violates the requirements that SVM places on the kernel. Fortunately, a generalized SVM (GSVM) has been developed that allows the use of an arbitrary kernel. In this paper, we present an initial investigation into utilizing a GSVM-based relevance feedback learning algorithm. Since GSVM does not place restrictions on the kernel, any image similarity measure can be used. In particular, the proposed approach uses an image similarity measure developed for region-based, variable length representations. Experimental results over real world images demonstrate the efficacy of the proposed method.

A multiple instance learning based framework for semantic image segmentation

Most image segmentation algorithms extract regions satisfying visual uniformity criteria. Unfortunately, because of the semantic gap between low-level features and high-level semantics, such regions usually do not correspond to meaningful parts. This has motivated researchers to develop methods that, by introducing high-level knowledge into the segmentation process, can break through the performance ceiling imposed by the semantic gap. The main disadvantage of those methods is their lack of flexibility due to the assumption that such knowledge is provided in advance. In content-based image retrieval (CBIR), relevance feedback (RF) learning has been successfully applied as a technique aimed at reducing the semantic gap. Inspired by this, we present a RF-based CBIR framework that uses multiple instance learning to perform a semantically-guided context adaptation of segmentation parameters. A partial instantiation of this framework that uses mean shift-based segmentation is presented. Experiments show the effectiveness and flexibility of the proposed framework on real images.

Improving image retrieval performance by inter-query learning with one-class support vector machines

Relevance feedback (RF) is an iterative process which improves the performance of content-based image retrieval by modifying the query and similarity metric based on the user’s feedback on the retrieval results. This short-term learning within a single query session is called intra-query learning. However, the interaction history of previous users over all past queries may also be potentially exploited to help improve the retrieval performance for the current query. The long-term learning accumulated over the course of many query sessions is called inter-query learning. We present a novel RF framework that learns one-class support vector machines (1SVM) from retrieval experience to represent the set memberships of users’ high-level concepts and stores them in a “concept database”. The “concept database” provides a mechanism for accumulating inter-query learning obtained from previous queries. By doing a fuzzy classification of a query into the regions of support represented by the 1SVMs, past experience is merged with current intra-query learning. The geometric view of 1SVM allows a straightforward interpretation of the density of past interaction in a local area of the feature space and thus allows the decision of exploiting past information only if enough past exploration of the local area has occurred. The proposed approach is evaluated on real data sets and compared against both traditional intra-query-learning-only RF approaches and other methods that also exploit inter-query learning.

MRI segmentation fusion for brain tumor detection

Abstract The process of manually generating precise segmentations of brain tumors from magnetic resonance images (MRI) is time-consuming and error-prone. We present a new algorithm, Potential Field Segmentation (PFS), and propose the use of ensemble approaches that combine the results generated by PFS and other methods to achieve a fused segmentation. For the PFS method, we build on our recently proposed clustering algorithm, Potential Field Clustering, which is based on an analogy with the concept of potential field in Physics. We view the intensity of a pixel in an MRI as a “mass” that creates a potential field. Specifically, for each pixel in the MRI, the potential field is computed and, if smaller than an adaptive potential threshold, the pixel is associated with the tumor region. This “small potential” segmentation criterion is intuitively valid because tumor pixels have larger “mass” and thus the potential of surrounding regions is also much larger than in other regions of smaller or no “mass”. We evaluate the performance of the different methods, including the ensemble approaches, on the publicly available Brain Tumor Image Segmentation (BRATS) MRI benchmark database.

Multi-Dimensional Dynamic Time Warping for Image Texture Similarity

Modern content-based image retrieval systems use different features to represent properties (e.g., color, shape, texture) of the visual content of an image. Retrieval is performed by example where a query image is given as input and an appropriate metric is used to find the best matches in the corresponding feature space. Both selecting the features and the distance metric continue to be active areas of research. In this paper, we propose a new approach, based on the recently proposed Multidimensional Dynamic Time Warping (MD-DTW) distance [1], for assessing the texture similarity of images with structured textures. The MD-DTW allows the detection and comparison of arbitrarily shifted patterns between multi-dimensional series, such as those found in structured textures. Chaos theory tools are used as a preprocessing step to uncover and characterize regularities in structured textures. The main advantage of the proposed approach is that explicit selection and extraction of texture features is not required (i.e., similarity comparisons are performed directly on the raw pixel data alone). The method proposed in this preliminary investigation is shown to be valid by proving that it creates a statistically significant image texture similarity measure.

Adaptive kernel diverse density estimate for multiple instance learning

We present AKDDE, an adaptive kernel diverse density estimate scheme for multiple instance learning. AKDDE revises the definition of diverse density as the kernel density estimate of diverse positive bags. We show that the AKDDE is inversely proportional to the least bound that contains at least one instance from each positive bag. In order to incorporate the influence of negative bags an objective function is constructed as the difference between the AKDDE of positive bags and the kernel density estimate of negative ones. This scheme is simple in concept and has better properties than other MIL methods. We validate AKDDE on both synthetic and real-world benchmark MIL datasets.

A Bayesian network-based tunable image segmentation algorithm for object recognition

We present a Bayesian network-based tunable image segmentation algorithm that can be used to segment a particular object of interest (OOI). In tasks such as object recognition, semantically accurate segmentation of the OOI is a critical step. Due to the OOI consisting of different-looking fragments, traditional image segmentation algorithms that are based on the identification of homogeneous regions tend to oversegment. The algorithm presented in this paper uses Multiple Instance Learning to learn prototypical representations of each fragment of the OOI and a Bayesian network to learn the spatial relationships that exist among those fragments. The Bayesian network, as a probabilistic graphical model, in turn becomes evidence that is used for the process of semantically accurate segmentation of future instances of the OOI. The key contribution of this paper is the inclusion of domain-specific information in terms of spatial relationships as an input to a conventional Bayesian network structure learning algorithm. Preliminary results indicate that the proposed method improves segmentation performance.

Improving the Initial Image Retrieval Set by Inter-Query Learning with One-Class SVMs

Relevance Feedback attempts to reduce the semantic gap between a user’s perception of similarity and a feature-based representation of an image by asking the user to provide feedback regarding the relevance or non-relevance of the retrieved images. This is intra-query learning. However, in most current systems, all prior experience is lost whenever a user generates a new query thus inter-query information is not used. In this paper, we focus on the possibility of incorporating prior experience (obtained from the historical interaction of users with the system) to improve the retrieval performance on future queries. We propose learning one-class SVMs from retrieval experience to represent the set memberships of users’ query concepts. Using a fuzzy classification approach, this historical knowledge is then incorporated into future queries to improve the retrieval performance. In order to learn the set membership of a user’s query concept, a one-class SVM maps the relevant or training images into a nonlinearly transformed kernel-induced feature space and attempts to include most of those images into a hyper-sphere. The use of kernels allows the one-class SVM to deal with the non-linearity of the distribution of training images in an efficient manner, while at the same time, providing good generalization. The proposed approach is evaluated against real data sets and the results obtained confirm the effectiveness of using prior experience in improving retrieval performance.