Assessing user bias in affect detection within context-based spoken dialog systems

This paper presents an empirical evidence of user bias within a laboratory-oriented evaluation of a Spoken Dialog System. Specifically, we addressed user bias in their satisfaction judgements. We question the reliability of this data for modeling user emotion, focusing on contentment and frustration in a spoken dialog system. This bias is detected through machine learning experiments that were conducted on two datasets, users and annotators, which were then compared in order to assess the reliability of these datasets. The target used was the satisfaction rating and the predictors were conversational/dialog features. Our results indicated that standard classifiers were significantly more successful in discriminating frustration and contentment and the intensities of these emotions (reflected by user satisfaction ratings) from annotator data than from user data. Indirectly, the results showed that conversational features are reliable predictors of the two abovementioned emotions.

Analysis of pattern recognition and dimensionality reduction techniques for odor biometrics

In this paper, we analyze the performance of several well-known pattern recognition and dimensionality reduction techniques when applied to mass-spectrometry data for odor biometric identification. Motivated by the successful results of previous works capturing the odor from other parts of the body, this work attempts to evaluate the feasibility of identifying people by the odor emanated from the hands. By formulating this task according to a machine learning scheme, the problem is identified with a small-sample-size supervised classification problem in which the input data is formed by mass spectrograms from the hand odor of 13 subjects captured in different sessions. The high dimensionality of the data makes it necessary to apply feature selection and extraction techniques together with a simple classifier in order to improve the generalization capabilities of the model. Our experimental results achieve recognition rates over 85% which reveals that there exists discriminatory information in the hand odor and points at body odor as a promising biometric identifier.

Plataforma de evaluación de algoritmos para la identificación de usuarios

La minería de datos es un campo de las ciencias de la computación referido al proceso que intenta descubrir patrones en grandes volúmenes de datos. La minería de datos busca generar información similar a la que podría producir un experto humano. Además es el proceso de descubrir conocimientos interesantes, como patrones, asociaciones, cambios, anomalías y estructuras significativas a partir de grandes cantidades de datos almacenadas en bases de datos, data warehouses o cualquier otro medio de almacenamiento de información. El aprendizaje automático o aprendizaje de máquinas es una rama de la Inteligencia artificial cuyo objetivo es desarrollar técnicas que permitan a las computadoras aprender. De forma más concreta, se trata de crear programas capaces de generalizar comportamientos a partir de una información no estructurada suministrada en forma de ejemplos. La minería de datos utiliza métodos de aprendizaje automático para descubrir y enumerar patrones presentes en los datos. En los últimos años se han aplicado las técnicas de clasificación y aprendizaje automático en un número elevado de ámbitos como el sanitario, comercial o de seguridad. Un ejemplo muy actual es la detección de comportamientos y transacciones fraudulentas en bancos. Una aplicación de interés es el uso de las técnicas desarrolladas para la detección de comportamientos fraudulentos en la identificación de usuarios existentes en el interior de entornos inteligentes sin necesidad de realizar un proceso de autenticación. Para comprobar que estas técnicas son efectivas durante la fase de análisis de una determinada solución, es necesario crear una plataforma que de soporte al desarrollo, validación y evaluación de algoritmos de aprendizaje y clasificación en los entornos de aplicación bajo estudio. El proyecto planteado está definido para la creación de una plataforma que permita evaluar algoritmos de aprendizaje automático como mecanismos de identificación en espacios inteligentes. Se estudiarán tanto los algoritmos propios de este tipo de técnicas como las plataformas actuales existentes para definir un conjunto de requisitos específicos de la plataforma a desarrollar. Tras el análisis se desarrollará parcialmente la plataforma. Tras el desarrollo se validará con pruebas de concepto y finalmente se verificará en un entorno de investigación a definir. ABSTRACT. The data mining is a field of the sciences of the computation referred to the process that it tries to discover patterns in big volumes of information. The data mining seeks to generate information similar to the one that a human expert might produce. In addition it is the process of discovering interesting knowledge, as patterns, associations, changes, abnormalities and significant structures from big quantities of information stored in databases, data warehouses or any other way of storage of information. The machine learning is a branch of the artificial Intelligence which aim is to develop technologies that they allow the computers to learn. More specifically, it is a question of creating programs capable of generalizing behaviors from not structured information supplied in the form of examples. The data mining uses methods of machine learning to discover and to enumerate present patterns in the information. In the last years there have been applied classification and machine learning techniques in a high number of areas such as healthcare, commercial or security. A very current example is the detection of behaviors and fraudulent transactions in banks. An application of interest is the use of the techniques developed for the detection of fraudulent behaviors in the identification of existing Users inside intelligent environments without need to realize a process of authentication. To verify these techniques are effective during the phase of analysis of a certain solution, it is necessary to create a platform that support the development, validation and evaluation of algorithms of learning and classification in the environments of application under study. The project proposed is defined for the creation of a platform that allows evaluating algorithms of machine learning as mechanisms of identification in intelligent spaces. There will be studied both the own algorithms of this type of technologies and the current existing platforms to define a set of specific requirements of the platform to develop. After the analysis the platform will develop partially. After the development it will be validated by prove of concept and finally verified in an environment of investigation that would be define.

DAEDALUS at SemEval-2014 Task 9: Comparing approaches for sentiment analysis in twitter

This paper describes our participation at SemEval- 2014 sentiment analysis task, in both contextual and message polarity classification. Our idea was to com- pare two different techniques for sentiment analysis. First, a machine learning classifier specifically built for the task using the provided training corpus. On the other hand, a lexicon-based approach using natural language processing techniques, developed for a ge- neric sentiment analysis task with no adaptation to the provided training corpus. Results, though far from the best runs, prove that the generic model is more robust as it achieves a more balanced evaluation for message polarity along the different test sets.

Autonomous Acquisition of Natural Situated Communication

An important part of human intelligence, both historically and operationally, is our ability to communicate. We learn how to communicate, and maintain our communicative skills, in a society of communicators – a highly effective way to reach and maintain proficiency in this complex skill. Principles that might allow artificial agents to learn language this way are in completely known at present – the multi-dimensional nature of socio-communicative skills are beyond every machine learning framework so far proposed. Our work begins to address the challenge of proposing a way for observation-based machine learning of natural language and communication. Our framework can learn complex communicative skills with minimal up-front knowledge. The system learns by incrementally producing predictive models of causal relationships in observed data, guided by goal-inference and reasoning using forward-inverse models. We present results from two experiments where our S1 agent learns human communication by observing two humans interacting in a realtime TV-style interview, using multimodal communicative gesture and situated language to talk about recycling of various materials and objects. S1 can learn multimodal complex language and multimodal communicative acts, a vocabulary of 100 words forming natural sentences with relatively complex sentence structure, including manual deictic reference and anaphora. S1 is seeded only with high-level information about goals of the interviewer and interviewee, and a small ontology; no grammar or other information is provided to S1 a priori. The agent learns the pragmatics, semantics, and syntax of complex utterances spoken and gestures from scratch, by observing the humans compare and contrast the cost and pollution related to recycling aluminum cans, glass bottles, newspaper, plastic, and wood. After 20 hours of observation S1 can perform an unscripted TV interview with a human, in the same style, without making mistakes.

Multisensor data fusion for accurate modelling of mobile objects

In the last decade, multi-sensor data fusion has become a broadly demanded discipline to achieve advanced solutions that can be applied in many real world situations, either civil or military. In Defence,accurate detection of all target objects is fundamental to maintaining situational awareness, to locating threats in the battlefield and to identifying and protecting strategically own forces. Civil applications, such as traffic monitoring, have similar requirements in terms of object detection and reliable identification of incidents in order to ensure safety of road users. Thanks to the appropriate data fusion technique, we can give these systems the power to exploit automatically all relevant information from multiple sources to face for instance mission needs or assess daily supervision operations. This paper focuses on its application to active vehicle monitoring in a particular area of high density traffic, and how it is redirecting the research activities being carried out in the computer vision, signal processing and machine learning fields for improving the effectiveness of detection and tracking in ground surveillance scenarios in general. Specifically, our system proposes fusion of data at a feature level which is extracted from a video camera and a laser scanner. In addition, a stochastic-based tracking which introduces some particle filters into the model to deal with uncertainty due to occlusions and improve the previous detection output is presented in this paper. It has been shown that this computer vision tracker contributes to detect objects even under poor visual information. Finally, in the same way that humans are able to analyze both temporal and spatial relations among items in the scene to associate them a meaning, once the targets objects have been correctly detected and tracked, it is desired that machines can provide a trustworthy description of what is happening in the scene under surveillance. Accomplishing so ambitious task requires a machine learning-based hierarchic architecture able to extract and analyse behaviours at different abstraction levels. A real experimental testbed has been implemented for the evaluation of the proposed modular system. Such scenario is a closed circuit where real traffic situations can be simulated. First results have shown the strength of the proposed system.

Automatic physical activity pattern recognition and meal intake modelling in type 1 diabetes for their inclusion in artificial pancreas systems

La diabetes comprende un conjunto de enfermedades metabólicas que se caracterizan por concentraciones de glucosa en sangre anormalmente altas. En el caso de la diabetes tipo 1 (T1D, por sus siglas en inglés), esta situación es debida a una ausencia total de secreción endógena de insulina, lo que impide a la mayoría de tejidos usar la glucosa. En tales circunstancias, se hace necesario el suministro exógeno de insulina para preservar la vida del paciente; no obstante, siempre con la precaución de evitar caídas agudas de la glucemia por debajo de los niveles recomendados de seguridad. Además de la administración de insulina, las ingestas y la actividad física son factores fundamentales que influyen en la homeostasis de la glucosa. En consecuencia, una gestión apropiada de la T1D debería incorporar estos dos fenómenos fisiológicos, en base a una identificación y un modelado apropiado de los mismos y de sus sorrespondientes efectos en el balance glucosa-insulina. En particular, los sistemas de páncreas artificial –ideados para llevar a cabo un control automático de los niveles de glucemia del paciente– podrían beneficiarse de la integración de esta clase de información. La primera parte de esta tesis doctoral cubre la caracterización del efecto agudo de la actividad física en los perfiles de glucosa. Con este objetivo se ha llevado a cabo una revisión sistemática de la literatura y meta-análisis que determinen las respuestas ante varias modalidades de ejercicio para pacientes con T1D, abordando esta caracterización mediante unas magnitudes que cuantifican las tasas de cambio en la glucemia a lo largo del tiempo. Por otro lado, una identificación fiable de los periodos con actividad física es un requisito imprescindible para poder proveer de esa información a los sistemas de páncreas artificial en condiciones libres y ambulatorias. Por esta razón, la segunda parte de esta tesis está enfocada a la propuesta y evaluación de un sistema automático diseñado para reconocer periodos de actividad física, clasificando su nivel de intensidad (ligera, moderada o vigorosa); así como, en el caso de periodos vigorosos, identificando también la modalidad de ejercicio (aeróbica, mixta o de fuerza). En este sentido, ambos aspectos tienen una influencia específica en el mecanismo metabólico que suministra la energía para llevar a cabo el ejercicio y, por tanto, en las respuestas glucémicas en T1D. En este trabajo se aplican varias combinaciones de técnicas de aprendizaje máquina y reconocimiento de patrones sobre la fusión multimodal de señales de acelerometría y ritmo cardíaco, las cuales describen tanto aspectos mecánicos del movimiento como la respuesta fisiológica del sistema cardiovascular ante el ejercicio. Después del reconocimiento de patrones se incorpora también un módulo de filtrado temporal para sacar partido a la considerable coherencia temporal presente en los datos, una redundancia que se origina en el hecho de que en la práctica, las tendencias en cuanto a actividad física suelen mantenerse estables a lo largo de cierto tiempo, sin fluctuaciones rápidas y repetitivas. El tercer bloque de esta tesis doctoral aborda el tema de las ingestas en el ámbito de la T1D. En concreto, se propone una serie de modelos compartimentales y se evalúan éstos en función de su capacidad para describir matemáticamente el efecto remoto de las concetraciones plasmáticas de insulina exógena sobre las tasas de eleiminación de la glucosa atribuible a la ingesta; un aspecto hasta ahora no incorporado en los principales modelos de paciente para T1D existentes en la literatura. Los datos aquí utilizados se obtuvieron gracias a un experimento realizado por el Institute of Metabolic Science (Universidad de Cambridge, Reino Unido) con 16 pacientes jóvenes. En el experimento, de tipo ‘clamp’ con objetivo variable, se replicaron los perfiles individuales de glucosa, según lo observado durante una visita preliminar tras la ingesta de una cena con o bien alta carga glucémica, o bien baja. Los seis modelos mecanísticos evaluados constaban de: a) submodelos de doble compartimento para las masas de trazadores de glucosa, b) un submodelo de único compartimento para reflejar el efecto remoto de la insulina, c) dos tipos de activación de este mismo efecto remoto (bien lineal, bien con un punto de corte), y d) diversas condiciones iniciales. ABSTRACT Diabetes encompasses a series of metabolic diseases characterized by abnormally high blood glucose concentrations. In the case of type 1 diabetes (T1D), this situation is caused by a total absence of endogenous insulin secretion, which impedes the use of glucose by most tissues. In these circumstances, exogenous insulin supplies are necessary to maintain patient’s life; although caution is always needed to avoid acute decays in glycaemia below safe levels. In addition to insulin administrations, meal intakes and physical activity are fundamental factors influencing glucose homoeostasis. Consequently, a successful management of T1D should incorporate these two physiological phenomena, based on an appropriate identification and modelling of these events and their corresponding effect on the glucose-insulin balance. In particular, artificial pancreas systems –designed to perform an automated control of patient’s glycaemia levels– may benefit from the integration of this type of information. The first part of this PhD thesis covers the characterization of the acute effect of physical activity on glucose profiles. With this aim, a systematic review of literature and metaanalyses are conduced to determine responses to various exercise modalities in patients with T1D, assessed via rates-of-change magnitudes to quantify temporal variations in glycaemia. On the other hand, a reliable identification of physical activity periods is an essential prerequisite to feed artificial pancreas systems with information concerning exercise in ambulatory, free-living conditions. For this reason, the second part of this thesis focuses on the proposal and evaluation of an automatic system devised to recognize physical activity, classifying its intensity level (light, moderate or vigorous) and for vigorous periods, identifying also its exercise modality (aerobic, mixed or resistance); since both aspects have a distinctive influence on the predominant metabolic pathway involved in fuelling exercise, and therefore, in the glycaemic responses in T1D. Various combinations of machine learning and pattern recognition techniques are applied on the fusion of multi-modal signal sources, namely: accelerometry and heart rate measurements, which describe both mechanical aspects of movement and the physiological response of the cardiovascular system to exercise. An additional temporal filtering module is incorporated after recognition in order to exploit the considerable temporal coherence (i.e. redundancy) present in data, which stems from the fact that in practice, physical activity trends are often maintained stable along time, instead of fluctuating rapid and repeatedly. The third block of this PhD thesis addresses meal intakes in the context of T1D. In particular, a number of compartmental models are proposed and compared in terms of their ability to describe mathematically the remote effect of exogenous plasma insulin concentrations on the disposal rates of meal-attributable glucose, an aspect which had not yet been incorporated to the prevailing T1D patient models in literature. Data were acquired in an experiment conduced at the Institute of Metabolic Science (University of Cambridge, UK) on 16 young patients. A variable-target glucose clamp replicated their individual glucose profiles, observed during a preliminary visit after ingesting either a high glycaemic-load or a low glycaemic-load evening meal. The six mechanistic models under evaluation here comprised: a) two-compartmental submodels for glucose tracer masses, b) a single-compartmental submodel for insulin’s remote effect, c) two types of activations for this remote effect (either linear or with a ‘cut-off’ point), and d) diverse forms of initial conditions.

Clasificación de tráfico de Internet mediante análisis de datos

Poder clasificar de manera precisa la aplicación o programa del que provienen los flujos que conforman el tráfico de uso de Internet dentro de una red permite tanto a empresas como a organismos una útil herramienta de gestión de los recursos de sus redes, así como la posibilidad de establecer políticas de prohibición o priorización de tráfico específico. La proliferación de nuevas aplicaciones y de nuevas técnicas han dificultado el uso de valores conocidos (well-known) en puertos de aplicaciones proporcionados por la IANA (Internet Assigned Numbers Authority) para la detección de dichas aplicaciones. Las redes P2P (Peer to Peer), el uso de puertos no conocidos o aleatorios, y el enmascaramiento de tráfico de muchas aplicaciones en tráfico HTTP y HTTPS con el fin de atravesar firewalls y NATs (Network Address Translation), entre otros, crea la necesidad de nuevos métodos de detección de tráfico. El objetivo de este estudio es desarrollar una serie de prácticas que permitan realizar dicha tarea a través de técnicas que están más allá de la observación de puertos y otros valores conocidos. Existen una serie de metodologías como Deep Packet Inspection (DPI) que se basa en la búsqueda de firmas, signatures, en base a patrones creados por el contenido de los paquetes, incluido el payload, que caracterizan cada aplicación. Otras basadas en el aprendizaje automático de parámetros de los flujos, Machine Learning, que permite determinar mediante análisis estadísticos a qué aplicación pueden pertenecer dichos flujos y, por último, técnicas de carácter más heurístico basadas en la intuición o el conocimiento propio sobre tráfico de red. En concreto, se propone el uso de alguna de las técnicas anteriormente comentadas en conjunto con técnicas de minería de datos como son el Análisis de Componentes Principales (PCA por sus siglas en inglés) y Clustering de estadísticos extraídos de los flujos procedentes de ficheros de tráfico de red. Esto implicará la configuración de diversos parámetros que precisarán de un proceso iterativo de prueba y error que permita dar con una clasificación del tráfico fiable. El resultado ideal sería aquel en el que se pudiera identificar cada aplicación presente en el tráfico en un clúster distinto, o en clusters que agrupen grupos de aplicaciones de similar naturaleza. Para ello, se crearán capturas de tráfico dentro de un entorno controlado e identificando cada tráfico con su aplicación correspondiente, a continuación se extraerán los flujos de dichas capturas. Tras esto, parámetros determinados de los paquetes pertenecientes a dichos flujos serán obtenidos, como por ejemplo la fecha y hora de llagada o la longitud en octetos del paquete IP. Estos parámetros serán cargados en una base de datos MySQL y serán usados para obtener estadísticos que ayuden, en un siguiente paso, a realizar una clasificación de los flujos mediante minería de datos. Concretamente, se usarán las técnicas de PCA y clustering haciendo uso del software RapidMiner. Por último, los resultados obtenidos serán plasmados en una matriz de confusión que nos permitirá que sean valorados correctamente. ABSTRACT. Being able to classify the applications that generate the traffic flows in an Internet network allows companies and organisms to implement efficient resource management policies such as prohibition of specific applications or prioritization of certain application traffic, looking for an optimization of the available bandwidth. The proliferation of new applications and new technics in the last years has made it more difficult to use well-known values assigned by the IANA (Internet Assigned Numbers Authority), like UDP and TCP ports, to identify the traffic. Also, P2P networks and data encapsulation over HTTP and HTTPS traffic has increased the necessity to improve these traffic analysis technics. The aim of this project is to develop a number of techniques that make us able to classify the traffic with more than the simple observation of the well-known ports. There are some proposals that have been created to cover this necessity; Deep Packet Inspection (DPI) tries to find signatures in the packets reading the information contained in them, the payload, looking for patterns that can be used to characterize the applications to which that traffic belongs; Machine Learning procedures work with statistical analysis of the flows, trying to generate an automatic process that learns from those statistical parameters and calculate the likelihood of a flow pertaining to a certain application; Heuristic Techniques, finally, are based in the intuition or the knowledge of the researcher himself about the traffic being analyzed that can help him to characterize the traffic. Specifically, the use of some of the techniques previously mentioned in combination with data mining technics such as Principal Component Analysis (PCA) and Clustering (grouping) of the flows extracted from network traffic captures are proposed. An iterative process based in success and failure will be needed to configure these data mining techniques looking for a reliable traffic classification. The perfect result would be the one in which the traffic flows of each application is grouped correctly in each cluster or in clusters that contain group of applications of similar nature. To do this, network traffic captures will be created in a controlled environment in which every capture is classified and known to pertain to a specific application. Then, for each capture, all the flows will be extracted. These flows will be used to extract from them information such as date and arrival time or the IP length of the packets inside them. This information will be then loaded to a MySQL database where all the packets defining a flow will be classified and also, each flow will be assigned to its specific application. All the information obtained from the packets will be used to generate statistical parameters in order to describe each flow in the best possible way. After that, data mining techniques previously mentioned (PCA and Clustering) will be used on these parameters making use of the software RapidMiner. Finally, the results obtained from the data mining will be compared with the real classification of the flows that can be obtained from the database. A Confusion Matrix will be used for the comparison, letting us measure the veracity of the developed classification process.