Daniel López-Sánchez

Deep neural networks and transfer learning applied to multimedia web mining

The growth in the amount of multimedia content available online supposes a challenge for search and recommender systems. This information in the form of visual elements is of great value to a variety of web mining tasks; however, the mining of these resources is a difficult task due to the complexity and variability of the images. In this paper, we propose applying a deep learning model to the problem of web categorization. In addition, we make use of a technique known as transfer or inductive learning to drastically reduce the computational cost of the training phase. Finally, we report experimental results on the effectiveness of the proposed method using different classification methods and features from various depths of the deep model.

Hybridizing metric learning and case-based reasoning for adaptable clickbait detection

The term clickbait is usually used to name web contents which are specifically designed to maximize advertisement monetization, often at the expense of quality and exactitude. The rapid proliferation of this type of content has motivated researchers to develop automatic detection methods, to effectively block clickbaits in different application domains. In this paper, we introduce a novel clickbait detection method. Our approach leverages state-of-the-art techniques from the fields of deep learning and metric learning, integrating them into the Case-Based Reasoning methodology. This provides the model with the ability to learn-over-time, adapting to different users’ criteria. Our experimental results also evidence that the proposed approach outperforms previous clickbait detection methods by a large margin.

A CBR framework with gradient boosting based feature selection for lung cancer subtype classification

Molecular subtype classification represents a challenging field in lung cancer diagnosis. Although different methods have been proposed for biomarker selection, efficient discrimination between adenocarcinoma and squamous cell carcinoma in clinical practice presents several difficulties, especially when the latter is poorly differentiated. This is an area of growing importance, since certain treatments and other medical decisions are based on molecular and histological features. An urgent need exists for a system and a set of biomarkers that provide an accurate diagnosis. In this paper, a novel Case Based Reasoning framework with gradient boosting based feature selection is proposed and applied to the task of squamous cell carcinoma and adenocarcinoma discrimination, aiming to provide accurate diagnosis with a reduced set of genes. The proposed method was trained and evaluated on two independent datasets to validate its generalization capability. Furthermore, it achieved accuracy rates greater than those of traditional microarray analysis techniques, incorporating the advantages inherent to the Case Based Reasoning methodology (e.g. learning over time, adaptability, interpretability of solutions, etc.).

Data-independent Random Projections from the feature-map of the Homogeneous Polynomial Kernel of degree two

Abstract This paper presents a novel non-linear extension of the Random Projection method based on the degree-2 homogeneous polynomial kernel. Our algorithm is able to implicitly map data points to the high-dimensional feature space of that kernel and from there perform a Random Projection to an Euclidean space of the desired dimensionality. Pairwise distances between data points in the kernel feature space are approximately preserved in the resulting representation. As opposed to previous kernelized Random Projection versions, our method is data-independent and preserves much of the computational simplicity of the original algorithm. This is achieved by focusing on a specific kernel function, what allowed us to analyze the effect of its associated feature mapping in the distribution of the Random Projection hyperplanes. Finally, we present empirical evidence that the proposed method outperforms alternative approaches in terms of pairwise distance preservation, while being significantly more efficient. Also, we show how our method can be used to approximate the accuracy of non-linear classifiers with efficient linear classifiers in some datasets.

A CBR System for Efficient Face Recognition Under Partial Occlusion

This work focuses on the design and validation of a CBR system for efficient face recognition under partial occlusion conditions. The proposed CBR system is based on a classical distance-based classification method, modified to increase its robustness to partial occlusion. This is achieved by using a novel dissimilarity function which discards features coming from occluded facial regions. In addition, we explore the integration of an efficient dimensionality reduction method into the proposed framework to reduce computational cost. We present experimental results showing that the proposed CBR system outperforms classical methods of similar computational requirements in the task of face recognition under partial occlusion.

Preliminary Study of Mobile Device-Based Speech Enhancement System Using Lip-Reading

This paper describes our preliminary study towards a new type of speech enhancement system. To avoid using odd-looking electrolarynx, we used lip-reading function. Our final image is to use a smart phone with camera and audio output to be able to convert the lip motion to speech output. We tested MLP, CNN, and MobileNets image recognition methods. 3k image datasets for training and testing were recorded from five persons. The preliminary experiment indicated that the MobileNets is the most adequate algorithm for smart phone apps. in terms of the recognition accuracy and the calculation cost.

Data-independent Random Projections from the feature-space of the homogeneous polynomial kernel

Abstract Performing a Random Projection from the feature space associated to a kernel function may be important for two main reasons. (1) As a consequence of the Johnson–Lindestrauss lemma, the resulting low-dimensional representation will preserve most of the structure of data in the kernel feature space and (2) an efficient linear classifier trained on transformed data might approximate the accuracy of its nonlinear counterparts. In this paper, we present a novel method to perform Random Projections from the feature space of homogeneous polynomial kernels. As opposed to other kernelized Random Projection proposals, our method focuses on a specific kernel family to preserve some of the beneficial properties of the original Random Projection algorithm (e.g. data independence and efficiency). Our extensive experimental results evidence that the proposed method efficiently approximates a Random Projection from the kernel feature space, preserving pairwise distances and enabling a boost on linear classification accuracies.

Dynamic Detection of Radical Profiles in Social Networks Using Image Feature Descriptors and a Case-Based Reasoning Methodology

Nowadays, security forces are challenged by a new type of terrorist propaganda which occurs in public social networks and targets vulnerable individuals. The current volume of online radicalization messages has rendered manual monitoring approaches unfeasible, and effective countermeasures can only be adopted through early detection by automatized tools. Some approaches focus on mining the information provided by social users in the form of interactions and textual content. However, radical users also tend to exhibit distinctive iconography in their profile images. In this work, we propose the use of local image descriptors over profile images to aid the detection and monitoring of online radicalization processes. In addition, we complement this approach with an interaction-based formula for risk assessment, so candidate profiles can be selected for image-analysis based on their interaction with confirmed radical profiles. These techniques are combined in the context of a Case-Based Reasoning framework which, together with the feedback provided by the end-user, enables a continuous monitoring of the activity of radical users and eases the discovery of new profiles with a radicalization agenda.

An Ensemble Framework Coping with Instability in the Gene Selection Process

This paper proposes an ensemble framework for gene selection, which is aimed at addressing instability problems presented in the gene filtering task. The complex process of gene selection from gene expression data faces different instability problems from the informative gene subsets found by different filter methods. This makes the identification of significant genes by the experts difficult. The instability of results can come from filter methods, gene classifier methods, different datasets of the same disease and multiple valid groups of biomarkers. Even though there is a wide number of proposals, the complexity imposed by this problem remains a challenge today. This work proposes a framework involving five stages of gene filtering to discover biomarkers for diagnosis and classification tasks. This framework performs a process of stable feature selection, facing the problems above and, thus, providing a more suitable and reliable solution for clinical and research purposes. Our proposal involves a process of multistage gene filtering, in which several ensemble strategies for gene selection were added in such a way that different classifiers simultaneously assess gene subsets to face instability. Firstly, we apply an ensemble of recent gene selection methods to obtain diversity in the genes found (stability according to filter methods). Next, we apply an ensemble of known classifiers to filter genes relevant to all classifiers at a time (stability according to classification methods). The achieved results were evaluated in two different datasets of the same disease (pancreatic ductal adenocarcinoma), in search of stability according to the disease, for which promising results were achieved.

Improving random projection with genetic algorithms: student research abstract

1. RANDOM PROJECTION: THE PROBLEM OF NON-DETERMINISM Within the field of machine learning, Random Projection (RP) is one of the simplest methods available to perform dimensionality reduction. This method relies on a classical result known as the Johnson–Lindenstrauss lemma, which states that a small set of points in a high dimensional feature space can be mapped into a space of much lower dimension in such a way that pairwise distances between the points are nearly preserved. Later on, it was proved that this mapping could be performed with a projection matrix whose elements are drawn from an extremely simple distribution [1], simplifying the projection computation to aggregate evaluation. However, due to the randomness introduced in the method during the construction of the projection matrix, the algorithm behaves non-deterministically. To illustrate this, Figure 1 shows the distribution of the stress measure over 200 runs of Random Projection on 500 samples from two typical machine learning datasets.