Kelwin Fernandes

A proactive intelligent decision support system for predicting the popularity of online news

Due to the Web expansion, the prediction of online news popularity is becoming a trendy research topic. In this paper, we propose a novel and proactive Intelligent Decision Support System (IDSS) that analyzes articles prior to their publication. Using a broad set of extracted features (e.g., keywords, digital media content, earlier popularity of news referenced in the article) the IDSS first predicts if an article will become popular. Then, it optimizes a subset of the articles features that can more easily be changed by authors, searching for an enhancement of the predicted popularity probability. Using a large and recently collected dataset, with 39,000 articles from the Mashable website, we performed a robust rolling windows evaluation of five state of the art models. The best result was provided by a Random Forest with a discrimination power of 73%. Moreover, several stochastic hill climbing local searches were explored. When optimizing 1000 articles, the best optimization method obtained a mean gain improvement of 15 percentage points in terms of the estimated popularity probability. These results attest the proposed IDSS as a valuable tool for online news authors.

Transfer Learning with Partial Observability Applied to Cervical Cancer Screening

Cervical cancer remains a significant cause of mortality in low-income countries. As in many other diseases, the existence of several screening/diagnosis methods and subjective physician preferences creates a complex ecosystem for automated methods. In order to diminish the amount of labeled data from each modality/expert we propose a regularization-based transfer learning strategy that encourages source and target models to share the same coefficient signs. We instantiated the proposed framework to predict cross-modality individual risk and cross-expert subjective quality assessment of colposcopic images for different modalities. Thus, we are able to transfer knowledge gained from one expert/modality to another.

Tackling Class Imbalance with Ranking

In classification, when there is a disproportion in the number of observations in each class, the data is said to be class imbalance. Class imbalance is pervasive in real world applications of data classification and has been the focus of much research. The minority class contributes too little to the decision boundary because the learning process learns from each observation in isolation. In this paper, we discuss the application of learning pairwise rankers as a solution to class imbalance. We compare ranking models to alternatives from the literature.

Temporal Segmentation of Digital Colposcopies

Cervical cancer remains a significant cause of mortality in low-income countries. Digital colposcopy is a promising and inexpensive technology for the detection of cervical intraepithelial neoplasia. However, diagnostic sensitivity varies widely depending on the doctor expertise. Therefore, automation of this process is needed in both, detection and visualization. Colposcopies cover four steps: macroscopic view with magnifier white light, observation under green light, Hinselmann and Schiller. Also, there are transition intervals where the specialist manipulates the observed area. In this paper, we focus on the temporal segmentation of the video in these steps. Using our solution, physicians may focus on the step of interest and lesion detection tools can determine the interval to diagnose. We solved the temporal segmentation problem using Weighted Automata. Images were described by their chromacity histograms and labeled using a KNN classifier with a precision of 97 %. Transition frames were recognized with a precision of 91 %.

Pavement pathologies classification using graph-based features

Pavement cracks involve important information to measure road quality. Crack classification is a challenging problem given the diversity of possible cracks, therefore, it is needed to retrieve good features in order to facilitate the learning of predictive models with as few samples as possible. In this paper, we propose a graph-based set of features to efficiently describe cracks. These features proved to have high degree of expressiveness and robustness when used for crack classification. We show that the proposed features succeed in the assessment of 525 images with different kinds of cracks. We proved the robustness of the approach applying different levels of noise to the images and evaluating the classification accuracy.

Combining Ranking with Traditional Methods for Ordinal Class Imbalance

In classification problems, a dataset is said to be imbalanced when the distribution of the target variable is very unequal. Classes contribute unequally to the decision boundary, and special metrics are used to evaluate these datasets. In previous work, we presented pairwise ranking as a method for binary imbalanced classification, and extended to the ordinal case using weights. In this work, we extend ordinal classification using traditional balancing methods. A comparison is made against traditional and ordinal SVMs, in which the ranking adaption proposed is found to be competitive.

Ordinal Class Imbalance with Ranking

Classification datasets, which feature a skewed class distribution, are said to be class imbalance. Traditional methods favor the larger classes. We propose pairwise ranking as a method for imbalance classification so that learning compares pairs of observations from each class, and therefore both contribute equally to the decision boundary. In previous work, we suggested treating the binary classification as a ranking problem, followed by a threshold mapping to convert back the ranking score to the original classes. In this work, the method is extended to multi-class ordinal classification, and a new mapping threshold is proposed. Results are compared with traditional and ordinal SVMs, and ranking obtains competitive results.

Learning and Ensembling Lexicographic Preference Trees with Multiple Kernels

We study the problem of learning lexicographic preferences on multiattribute domains, and propose Rankdom Forests as a compact way to express preferences in learning to rank scenarios. We start generalizing Conditional Lexicographic Preference Trees by introducing multiple kernels in order to handle non-categorical attributes. Then, we define a learning strategy for inferring lexicographic rankers from partial pairwise comparisons between options. Finally, a Lexicographic Ensemble is introduced to handle multiple weak partial rankers, being Rankdom Forests one of these ensembles. We tested the performance of the proposed method using several datasets and obtained competitive results when compared with other lexicographic rankers.

Discriminative directional classifiers

In different areas of knowledge, phenomena are represented by directional-angular or periodic-data; from wind direction and geographical coordinates to time references like days of the week or months of the calendar. These values are usually represented in a linear scale, and restricted to a given range (e.g. 0 , 2 π ) ), hiding the real nature of this information. Therefore, dealing with directional data requires special methods. So far, the design of classifiers for periodic variables adopts a generative approach based on the usage of the von Mises distribution or variants. Since for non-periodic variables state of the art approaches are based on non-generative methods, it is pertinent to investigate the suitability of other approaches for periodic variables. We propose a discriminative Directional Logistic Regression model able to deal with angular data, which does not make any assumption on the data distribution. Also, we study the expressiveness of this model for any number of features. Finally, we validate our model against the previously proposed directional naive Bayes approach and against a Support Vector Machine with a directional Radial Basis Function kernel with synthetic and real data obtaining competitive results.

A deep learning approach for the forensic evaluation of sexual assault

Abstract Despite the existence of patterns able to discriminate between consensual and non-consensual intercourse, the relevance of genital lesions in the corroboration of a legal rape complaint is currently under debate in many countries. The testimony of the physicians when assessing these lesions has been questioned in court due to several factors (e.g., a lack of comprehensive knowledge of lesions, wide spectrum of background area, among others). Therefore, it is relevant to provide automated tools to support the decision process in an objective manner. In this work, we evaluate the performance of state-of-the-art deep learning architectures for the forensic assessment of sexual assault. We propose a deep architecture and learning strategy to tackle the class imbalance on deep learning using ranking. The proposed methodologies achieved the best results when compared with handcrafted feature engineering and with other deep architectures .