Reports de recerca
http://hdl.handle.net/2117/3688
2016-05-25T07:26:22ZA variational formulation for GTM through time: Theoretical foundations
http://hdl.handle.net/2117/86323
A variational formulation for GTM through time: Theoretical foundations
Olier Caparroso, Iván; Vellido Alcacena, Alfredo
Generative Topographic Mapping (GTM) is a latent variable model that, in its standard version, was conceived to provide clustering and visualization of multivariate, real-valued, i.i.d. data. It was also extended to deal with non-i.i.d. data such as multivariate time series in a variant called GTM Through Time (GTMTT), defined as a constrained Hidden Markov Model (HMM). In this technical report, we provide the theoretical foundations of the reformulation of GTM-TT within the Variational Bayesian framework. This approach, in its application, should naturally handle the presence of noise in the time series, helping to avert the problem of data overfitting.
2016-04-28T09:20:38ZOlier Caparroso, IvánVellido Alcacena, AlfredoGenerative Topographic Mapping (GTM) is a latent variable model that, in its standard version, was conceived to provide clustering and visualization of multivariate, real-valued, i.i.d. data. It was also extended to deal with non-i.i.d. data such as multivariate time series in a variant called GTM Through Time (GTMTT), defined as a constrained Hidden Markov Model (HMM). In this technical report, we provide the theoretical foundations of the reformulation of GTM-TT within the Variational Bayesian framework. This approach, in its application, should naturally handle the presence of noise in the time series, helping to avert the problem of data overfitting.A variational Bayesian formulation for GTM: Theoretical foundations
http://hdl.handle.net/2117/86314
A variational Bayesian formulation for GTM: Theoretical foundations
Olier Caparroso, Iván; Vellido Alcacena, Alfredo
Generative Topographic Mapping (GTM) is a non-linear latent variable model of the manifold learning family that provides simultaneous visualization and clustering of high-dimensional data. It was originally formulated as a constrained mixture of Gaussian distributions, for which the adaptive parameters were determined by Maximum Likelihood (ML), using the Expectation-Maximization (EM) algorithm. In this paper, we define an alternative variational formulation of GTM that provides a full Bayesian treatment to a Gaussian Process (GP) - based variation of the model.
2016-04-28T08:12:40ZOlier Caparroso, IvánVellido Alcacena, AlfredoGenerative Topographic Mapping (GTM) is a non-linear latent variable model of the manifold learning family that provides simultaneous visualization and clustering of high-dimensional data. It was originally formulated as a constrained mixture of Gaussian distributions, for which the adaptive parameters were determined by Maximum Likelihood (ML), using the Expectation-Maximization (EM) algorithm. In this paper, we define an alternative variational formulation of GTM that provides a full Bayesian treatment to a Gaussian Process (GP) - based variation of the model.Predictive models in churn data mining: a review
http://hdl.handle.net/2117/86182
Predictive models in churn data mining: a review
García, David L.; Vellido Alcacena, Alfredo; Nebot Castells, M. Àngela
The development of predictive models of customer abandonment plays a central role in any churn management strategy. These models can be developed using either qualitative approaches or can take a data-centred point of view. In the latter case, the use of Data Mining procedures and techniques can provide useful and actionable insights into the processes leading to abandonment. In this report, we provide a brief and structured review of some of the Data Mining approaches that have been put forward in recent academic literature for customer abandonment prediction.
2016-04-26T09:44:49ZGarcía, David L.Vellido Alcacena, AlfredoNebot Castells, M. ÀngelaThe development of predictive models of customer abandonment plays a central role in any churn management strategy. These models can be developed using either qualitative approaches or can take a data-centred point of view. In the latter case, the use of Data Mining procedures and techniques can provide useful and actionable insights into the processes leading to abandonment. In this report, we provide a brief and structured review of some of the Data Mining approaches that have been put forward in recent academic literature for customer abandonment prediction.Customer continuity management as a foundation for churn data mining
http://hdl.handle.net/2117/86180
Customer continuity management as a foundation for churn data mining
García, David L.; Vellido Alcacena, Alfredo; Nebot Castells, M. Àngela
This report lays the first theoretical foundations for a research program on analytical churn management. In the current hypercompetitive business scenario, firms have to bend over backwards in their strategies both to retain their customers and to lure those from the competition. For this reason, understanding how customer loyalty construction mechanisms work, anticipating the customer’s intention to abandon a service provider, becomes a critical business acumen to ensure a company’s continuity in the market. To this end, in this report we propose a complete customer loyalty management model referred to as Customer Continuity Management. This type of knowledge can always be qualitatively obtained, but quantitative analysis of actual customer data should increase the certainty on its reliability and business actionability. This is why Data Mining techniques, applied to market surveyed information, can provide valuable assistance to churn (customer attrition) management processes. This report focuses on the prevention side of customer loyalty management and, especially, on customer satisfaction and switching barriers. Some of principals existing explanatory models of customer loyalty building are review in detail.
2016-04-26T09:27:07ZGarcía, David L.Vellido Alcacena, AlfredoNebot Castells, M. ÀngelaThis report lays the first theoretical foundations for a research program on analytical churn management. In the current hypercompetitive business scenario, firms have to bend over backwards in their strategies both to retain their customers and to lure those from the competition. For this reason, understanding how customer loyalty construction mechanisms work, anticipating the customer’s intention to abandon a service provider, becomes a critical business acumen to ensure a company’s continuity in the market. To this end, in this report we propose a complete customer loyalty management model referred to as Customer Continuity Management. This type of knowledge can always be qualitatively obtained, but quantitative analysis of actual customer data should increase the certainty on its reliability and business actionability. This is why Data Mining techniques, applied to market surveyed information, can provide valuable assistance to churn (customer attrition) management processes. This report focuses on the prevention side of customer loyalty management and, especially, on customer satisfaction and switching barriers. Some of principals existing explanatory models of customer loyalty building are review in detail.Elements of generative manifold learning for semi-supervised tasks
http://hdl.handle.net/2117/86178
Elements of generative manifold learning for semi-supervised tasks
Cruz, Raúl; Vellido Alcacena, Alfredo
For many real-world application problems, the availability of data labels for supervised learning is rather limited. It is often the case that a limited number of labelled cases is accompanied by a larger number of unlabeled ones. This is the setting for semi-supervised learning, in which unsupervised approaches assist the supervised problem and viceversa. In this report, we outline some basic theoretical foundations of semi-supervised learning using models of the generative manifold-learning family.
2016-04-26T09:01:38ZCruz, RaúlVellido Alcacena, AlfredoFor many real-world application problems, the availability of data labels for supervised learning is rather limited. It is often the case that a limited number of labelled cases is accompanied by a larger number of unlabeled ones. This is the setting for semi-supervised learning, in which unsupervised approaches assist the supervised problem and viceversa. In this report, we outline some basic theoretical foundations of semi-supervised learning using models of the generative manifold-learning family.Método multiobjetivo de aprendizaje para razonamiento inductivo difuso
http://hdl.handle.net/2117/86170
Método multiobjetivo de aprendizaje para razonamiento inductivo difuso
Acosta, Jesús; Nebot Castells, M. Àngela; Fuertes Armengol, José Mª
It has been recognized in various studies that the variations in the granularity (number of classes per variable) and the membership functions have a significant effect in the behaviour of the fuzzy systems. The FIR methodology is not an exception. The efficiency of the qualitative model identification and fuzzy forecast processes of FIR is very influenced by the fuzzification parameters of the system variables (i.e. number of classes and shape of the membership functions). To resolve this problematic we have been presented in previous works hybrid methodologies called Genetic Fuzzy Systems (GFSs) that try to learn in a joint way or by separated those parameters. These methods have used monoobjetive functions for the evaluation of the chromosomes. In this investigation another method of automatic learning is presented. This new method permits to obtain at the same time the fuzzification parameters of the FIR methodology but using Multiobjective Genetic Algorithms. Its main components are described and the results obtained on an environmental application are presented.
2016-04-26T08:08:06ZAcosta, JesúsNebot Castells, M. ÀngelaFuertes Armengol, José MªIt has been recognized in various studies that the variations in the granularity (number of classes per variable) and the membership functions have a significant effect in the behaviour of the fuzzy systems. The FIR methodology is not an exception. The efficiency of the qualitative model identification and fuzzy forecast processes of FIR is very influenced by the fuzzification parameters of the system variables (i.e. number of classes and shape of the membership functions). To resolve this problematic we have been presented in previous works hybrid methodologies called Genetic Fuzzy Systems (GFSs) that try to learn in a joint way or by separated those parameters. These methods have used monoobjetive functions for the evaluation of the chromosomes. In this investigation another method of automatic learning is presented. This new method permits to obtain at the same time the fuzzification parameters of the FIR methodology but using Multiobjective Genetic Algorithms. Its main components are described and the results obtained on an environmental application are presented.Object recognition and tracking in video sequences: a new integrated methodology
http://hdl.handle.net/2117/86159
Object recognition and tracking in video sequences: a new integrated methodology
Amézquita Gómez, Nicolás; Alquézar Mancho, René; Serratosa, Francesc
This paper describes a methodology that integrates recognition and segmentation, simultaneously with image tracking in a cooperative manner, for recognition of objects (or parts of them) in image sequences. A probabilistic general approach at pixel level is depicted together with a practical heuristic simplification in which pixels’ class probabilities are approximated by a finite small set of class possibility values. These possibility values are updated iteratively along the image sequence for each class and each pixel taking into account both the prior tracking information and the spot-based object recognition results provided by a trained neural network. A further segmentation of the class possibility images allows the tracking of each object of interest in the sequence. The good experimental results obtained so far show the viability of the approach under certain conditions.
2016-04-25T14:21:12ZAmézquita Gómez, NicolásAlquézar Mancho, RenéSerratosa, FrancescThis paper describes a methodology that integrates recognition and segmentation, simultaneously with image tracking in a cooperative manner, for recognition of objects (or parts of them) in image sequences. A probabilistic general approach at pixel level is depicted together with a practical heuristic simplification in which pixels’ class probabilities are approximated by a finite small set of class possibility values. These possibility values are updated iteratively along the image sequence for each class and each pixel taking into account both the prior tracking information and the spot-based object recognition results provided by a trained neural network. A further segmentation of the class possibility images allows the tracking of each object of interest in the sequence. The good experimental results obtained so far show the viability of the approach under certain conditions.Combining neural networks and clustering techniques for object recognition in indoor video sequences
http://hdl.handle.net/2117/86156
Combining neural networks and clustering techniques for object recognition in indoor video sequences
Serratosa, Francesc; Amézquita Gómez, Nicolás; Alquézar Mancho, René
This paper presents the results obtained in a real experiment for object recognition in a sequence of images captured by a mobile robot in an indoor environment. Objects are simply represented as an unstructured set of spots (image regions) for each frame, which are obtained from the result of an image segmentation algorithm applied on the whole sequence. In a previous work, neural networks were used to classify the spots independently as belonging to one of the objects of interest or the background from different spot features (color, size and invariant moments). In this work, clustering techniques are applied afterwards taking into account both the neural net outputs (class probabilities) and geometrical data (spot mass centers). In this way, context information is exploited to improve the classification performance. The experimental results of this combined approach are quite promising and better than the ones obtained using only the neural nets.
2016-04-25T14:09:55ZSerratosa, FrancescAmézquita Gómez, NicolásAlquézar Mancho, RenéThis paper presents the results obtained in a real experiment for object recognition in a sequence of images captured by a mobile robot in an indoor environment. Objects are simply represented as an unstructured set of spots (image regions) for each frame, which are obtained from the result of an image segmentation algorithm applied on the whole sequence. In a previous work, neural networks were used to classify the spots independently as belonging to one of the objects of interest or the background from different spot features (color, size and invariant moments). In this work, clustering techniques are applied afterwards taking into account both the neural net outputs (class probabilities) and geometrical data (spot mass centers). In this way, context information is exploited to improve the classification performance. The experimental results of this combined approach are quite promising and better than the ones obtained using only the neural nets.Un algoritmo para la extracción automática de reglas lógicas a partir de modelos FIR
http://hdl.handle.net/2117/85858
Un algoritmo para la extracción automática de reglas lógicas a partir de modelos FIR
Castro Espinoza, Félix Agustín; Nebot Castells, M. Àngela
In this report the LR-FIR (Logical Rules with FIR) algorithm is described. The main goal of LR-FIR is to extract, in an automatic way, a set of logical rules that explain system’s behaviour. The algorithm starts from the model identified by the Fuzzy Inductive Reasoning (FIR) methodology and obtains a compacted set of logical rules. A FIR model is composed of the mask, that represents system’s structure, and the pattern rule base, that contains system’s behaviour. This report is organized in two sections. The first one presents FIR methodology in detail, whereas the second one describes the LR-FIR algorithm developed in an accurate way.
En este reporte se describe el algoritmo LR-FIR (Logical Rules with FIR), que tiene como objetivo extraer de manera automática un conjunto de reglas lógicas que expliquen el comportamiento del sistema. LR-FIR parte del modelo del sistema identificado mediante la metodología del Razonamiento Inductivo Difuso (FIR, por sus siglas en inglés). Un modelo FIR está compuesto de la máscara que describe la estructura del sistema y la base de reglas patrón que aglutina el comportamiento de éste. Este reporte está organizado en dos secciones. En la primera de ellas se presenta en la metodología FIR mientas que en la segunda se describe en detalle el algoritmo LR-FIR desarrollado.
2016-04-19T07:42:44ZCastro Espinoza, Félix AgustínNebot Castells, M. ÀngelaIn this report the LR-FIR (Logical Rules with FIR) algorithm is described. The main goal of LR-FIR is to extract, in an automatic way, a set of logical rules that explain system’s behaviour. The algorithm starts from the model identified by the Fuzzy Inductive Reasoning (FIR) methodology and obtains a compacted set of logical rules. A FIR model is composed of the mask, that represents system’s structure, and the pattern rule base, that contains system’s behaviour. This report is organized in two sections. The first one presents FIR methodology in detail, whereas the second one describes the LR-FIR algorithm developed in an accurate way.
En este reporte se describe el algoritmo LR-FIR (Logical Rules with FIR), que tiene como objetivo extraer de manera automática un conjunto de reglas lógicas que expliquen el comportamiento del sistema. LR-FIR parte del modelo del sistema identificado mediante la metodología del Razonamiento Inductivo Difuso (FIR, por sus siglas en inglés). Un modelo FIR está compuesto de la máscara que describe la estructura del sistema y la base de reglas patrón que aglutina el comportamiento de éste. Este reporte está organizado en dos secciones. En la primera de ellas se presenta en la metodología FIR mientas que en la segunda se describe en detalle el algoritmo LR-FIR desarrollado.A sequential algorithm for feed-forward neural networks with optimal coefficients and interacting frequencies
http://hdl.handle.net/2117/85824
A sequential algorithm for feed-forward neural networks with optimal coefficients and interacting frequencies
Romero Merino, Enrique; Alquézar Mancho, René
An algorithm for sequential approximation with optimal coefficients and interacting frequencies (SAOCIF) for feed-forward neural networks is presented. SAOCIF combines two key ideas. The first one is the optimization of the coefficients (the linear part of the approximation). The second one is the strategy to choose the frequencies (the non-linear weights), taking into account the interactions with the previously selected ones. The resulting method combines the locality of sequential approximations, where only one frequency is found at every step, with the globality of non-sequential methods, where every frequency interacts with the others. The idea behind SAOCIF can be theoretically extended to general Hilbert spaces. Experimental results show a very satisfactory performance.
2016-04-18T14:53:10ZRomero Merino, EnriqueAlquézar Mancho, RenéAn algorithm for sequential approximation with optimal coefficients and interacting frequencies (SAOCIF) for feed-forward neural networks is presented. SAOCIF combines two key ideas. The first one is the optimization of the coefficients (the linear part of the approximation). The second one is the strategy to choose the frequencies (the non-linear weights), taking into account the interactions with the previously selected ones. The resulting method combines the locality of sequential approximations, where only one frequency is found at every step, with the globality of non-sequential methods, where every frequency interacts with the others. The idea behind SAOCIF can be theoretically extended to general Hilbert spaces. Experimental results show a very satisfactory performance.