Kernel-Based Ranking. Methods for Learning and Performance Estimation

Machine learning provides tools for automated construction of predictive models in data intensive areas of engineering and science. The family of regularized kernel methods have in the recent years become one of the mainstream approaches to machine learning, due to a number of advantages the methods share. The approach provides theoretically well-founded solutions to the problems of under- and overfitting, allows learning from structured data, and has been empirically demonstrated to yield high predictive performance on a wide range of application domains. Historically, the problems of classification and regression have gained the majority of attention in the field. In this thesis we focus on another type of learning problem, that of learning to rank. In learning to rank, the aim is from a set of past observations to learn a ranking function that can order new objects according to how well they match some underlying criterion of goodness. As an important special case of the setting, we can recover the bipartite ranking problem, corresponding to maximizing the area under the ROC curve (AUC) in binary classification. Ranking applications appear in a large variety of settings, examples encountered in this thesis include document retrieval in web search, recommender systems, information extraction and automated parsing of natural language. We consider the pairwise approach to learning to rank, where ranking models are learned by minimizing the expected probability of ranking any two randomly drawn test examples incorrectly. The development of computationally efficient kernel methods, based on this approach, has in the past proven to be challenging. Moreover, it is not clear what techniques for estimating the predictive performance of learned models are the most reliable in the ranking setting, and how the techniques can be implemented efficiently. The contributions of this thesis are as follows. First, we develop RankRLS, a computationally efficient kernel method for learning to rank, that is based on minimizing a regularized pairwise least-squares loss. In addition to training methods, we introduce a variety of algorithms for tasks such as model selection, multi-output learning, and cross-validation, based on computational shortcuts from matrix algebra. Second, we improve the fastest known training method for the linear version of the RankSVM algorithm, which is one of the most well established methods for learning to rank. Third, we study the combination of the empirical kernel map and reduced set approximation, which allows the large-scale training of kernel machines using linear solvers, and propose computationally efficient solutions to cross-validation when using the approach. Next, we explore the problem of reliable cross-validation when using AUC as a performance criterion, through an extensive simulation study. We demonstrate that the proposed leave-pair-out cross-validation approach leads to more reliable performance estimation than commonly used alternative approaches. Finally, we present a case study on applying machine learning to information extraction from biomedical literature, which combines several of the approaches considered in the thesis. The thesis is divided into two parts. Part I provides the background for the research work and summarizes the most central results, Part II consists of the five original research articles that are the main contribution of this thesis.

Découverte et application de nouveaux motifs d'association propres à l'hexaphénylbenzène et à ses dérivés

Les propriétés des matériaux moléculaires proviennent à la fois de la structure des composantes individuelles et de la façon dont elles s’associent. Ce dernier aspect reste difficile à contrôler, malgré de grandes avancées en science des matériaux. Pour mieux comprendre la relation structure-propriétés, nous avons entrepris une étude systématique de l'hexaphénylbenzène et de ses dérivés, qui offrent une charpente symétrique et rigide. En premier lieu, nous avons attaché six groupements diaminotriazinyles sur l’hexaphénylbenzène afin de produire des réseaux tridimensionnels hautement poreux maintenus par des ponts hydrogène. En modifiant systématiquement le coeur moléculaire, nous avons excisé près du tiers de la molécule-mère, générant des réseaux supramoléculaires dont la porosité s’est élevée graduellement jusqu’à 75%, équivalant ainsi le record pour ce type de matériaux. Ensuite, nous avons étudié le comportement de l’hexakis(4-nitrophényl)benzène. Dans les structures cristallines obtenues, des interactions non-covalentes entre groupements nitro démontrent leur potentiel en chimie supramoléculaire. Le coeur moléculaire ne joue qu’un rôle secondaire dans l’empilement des molécules : seules quelques interactions C-H•••π impliquant le cycle aromatique central de l’hexaphénylbenzène sont évidentes. Cette dernière observation nous a poussés à étudier le comportement à l’état cristallin de l’hexaphénylbenzène et ses dérivés. En scrutant attentivement neuf structures cristallines de ces composés, nous avons décerné la présence récurrente d’interactions C-H•••π impliquant le cycle aromatique central. Cette association caractéristique a été exploitée pour créer des réseaux supramoléculaires maintenus par des interactions C-H•••π sélectives entre un groupement éthynyle et le cycle aromatique central de l’hexaphénylbenzène. Finalement, nous avons joint le côté sombre de l’ingénierie cristalline en utilisant nos connaissances dans le but d’empêcher la formation d’interactions directionnelles. En protégeant le cycle aromatique central de l’hexaphénylbenzène à l’aide de groupements alkyles, les interactions C-H•••π ont été pratiquement éliminées. Ces résultats offrent la possibilité de créer de nouveaux matériaux amorphes. Dans ces études, focalisées sur le système hexaphénylbenzène, nous avons mis en relief des phénomènes qui sont obscurcis dans d'autres familles de molécules. De plus, ce système a grandement facilité l’utilisation d’une approche méthodique pour explorer la relation structure-propriétés. Nos travaux nous ont amenés à des conclusions de valeur universelle en science des matériaux moléculaires.

Surface acidity/basicity of yttria and its mixed oxides with alumina catalysts

Acid/base strength distribution of Y2O3 and its mixed oxides with alumina catalysts are measured on Hammett acidity function scale and expressed in terms of H0max value. Basicity of Y2O3 increases with increase in activation temperature and for mixed oxides the basicity increases with increase in concentration of Y2O3, in the catalyst.

Acid base characteristics of binary oxides of Zr with Ce and La

The surface acidity and basicity of binary oxides of Zr with Ce and La are determined using a series of Hammet indicators and Ho,,max values are reported. The generation of new acid sites habe been ascribed to the charge imbalance of M1-O-M2 bonds, where M1 and M2 are metal atoms. Both Bronsted and Lewis acid sites contribute to the acidity of the oxides

Surface acidity/basicity and catalytic activity of ternary oxides of Al, Ce and Dy

The surface acidity and basicity of ternary oxides of AI, Ce and Dy have been determined using a set of Hammett indicators. The data have been correlated with the catalytic activity of these oxides towards the liquid phase reduction of cyclohexanone in 2-propanol. The reaction is favoured by the higher basicity of the ternary oxide.

Synthesis of nickel–rubber nanocomposites and evaluation of their dielectric properties

Nanocomposites based on natural rubber and nano-sized nickelwere synthesized by incorporating nickel nanoparticles in a natural rubber matrix for various loadings of the filler. Structural, morphological, magnetic and mechanical properties of the compositeswere evaluated along with a detailed study of dielectric properties. Itwas found that nickel particleswere uniformly distributed in the matrix without agglomeration resulting in a magnetic nanocomposite. The elastic properties showed an improvement with increase in filler content but breaking stress and breaking strain were found to decrease. Dielectric permittivity was found to decrease with increase in frequency, and found to increase with increase in nickel loading. The decrease in permittivity with temperature is attributed to the high volume expansivity of rubber at elevated temperatures. Dielectric loss of blank rubber as well as the composites was found to increase with temperature.

Structural and electrical studies on tetrameric cobalt phthalocyanine and polyaniline composites

Polyaniline and oligomeric cobalt phthalocyanine are blended in different proportions by chemical methods. These blends are characterised by spectroscopic methods and dielectric measurements. Dielectric studies on the conducting polymer blends are carried out in the frequency range of 100 kHz to 5MHz from room temperature (300 K) to 373 K. Dielectric permittivity and dielectric loss of these blends are explained on the basis of interfacial polarisation. From the dielectric permittivity studies, ac conductivity of the samples were calculated and the results are correlated. In order to understand the exact conduction mechanism of the samples, dc electrical conductivity of the blends is carried out in the temperature range of 70–300 K. By applying Mott’s theory, it is found that the conducting polymer composites obey a 3D variable range hopping mechanism. The values of Mott’s temperature (T0), density of states at the Fermi energy (N(EF)), range of hopping (R) and hopping energy (W) for the composites are calculated and presented

Effect of preparation on magnetic properties of Mn-Zn ferrite

Mixed ferrites belonging to the type Mn0.9Zn0.1Fe2O4 have been prepared by the double sintering method and by the chemical co-precipitation for comparing their magnetic properties. Sintered and precipitated ferrites exhibit different characteristics, especially in their magnetic properties like magnetization (Ms), coercive field (Hc) and Curie temperature (Tc). The sintered particles were size reduced in order to compare with the nanosized co-precipitated particles. The effect of grinding has also been studied. Particles have been collected at regular intervals of grinding and their properties have been studied. The increase in the coercive field has been recorded by a hysteresis curve tracer confirming size reduction. X-ray diffraction studies confirmed the structure and consequent size reduction

Modification of Polypropylene/High Density Polyethylene Blend Using Nanokaolinite Clay and E-Glass Fibre: Preparation, Characterization and Micromechanical Modelling

Upgrading two widely used standard plastics, polypropylene (PP) and high density polyethylene (HDPE), and generating a variety of useful engineering materials based on these blends have been the main objective of this study. Upgradation was effected by using nanomodifiers and/or fibrous modifiers. PP and HDPE were selected for modification due to their attractive inherent properties and wide spectrum of use. Blending is the engineered method of producing new materials with tailor made properties. It has the advantages of both the materials. PP has high tensile and flexural strength and the HDPE acts as an impact modifier in the resultant blend. Hence an optimized blend of PP and HDPE was selected as the matrix material for upgradation. Nanokaolinite clay and E-glass fibre were chosen for modifying PP/HDPE blend. As the first stage of the work, the mechanical, thermal, morphological, rheological, dynamic mechanical and crystallization characteristics of the polymer nanocomposites prepared with PP/HDPE blend and different surface modified nanokaolinite clay were analyzed. As the second stage of the work, the effect of simultaneous inclusion of nanokaolinite clay (both N100A and N100) and short glass fibres are investigated. The presence of nanofiller has increased the properties of hybrid composites to a greater extent than micro composites. As the last stage, micromechanical modeling of both nano and hybrid A composite is carried out to analyze the behavior of the composite under load bearing conditions. These theoretical analyses indicate that the polymer-nanoclay interfacial characteristics partially converge to a state of perfect interfacial bonding (Takayanagi model) with an iso-stress (Reuss IROM) response. In the case of hybrid composites the experimental data follows the trend of Halpin-Tsai model. This implies that matrix and filler experience varying amount of strain and interfacial adhesion between filler and matrix and also between the two fillers which play a vital role in determining the modulus of the hybrid composites.A significant observation from this study is that the requirement of higher fibre loading for efficient reinforcement of polymers can be substantially reduced by the presence of nanofiller together with much lower fibre content in the composite. Hybrid composites with both nanokaolinite clay and micron sized E-glass fibre as reinforcements in PP/HDPE matrix will generate a novel class of high performance, cost effective engineering material.

Study on Granular Dynamics in Vertically Vibrated Beds using Tracking Technique

The study of granular material is of great interest to many researchers in both engineering and science communities. The importance of such a study derives from its complex rheological character and also its significant role in a wide range of industrial applications, such as coal, food, plastics, pharmaceutical, powder metallurgy and mineral processing. A number of recent reports have been focused on the physics of non-cohesive granular material submitted to vertical vibration in either experimental or theoretical approaches. Such a kind of system can be used to separate, mix and dry granular materials in industries. It exhibits different instability behaviour on its surface when under vertical vibration, for example, avalanching, surface fluidization and surface wave, and these phenomena have attracted particular interest of many researchers. However, its fundamental understanding of the instability mechanism is not yet well-understood. This paper is therefore to study the dynamics of granular motion in such a kind of system using Positron Emission Particle Tracking (PEPT), which allows the motion of a single tracer particle to be followed in a non-invasive way. Features of the solids motion such as cycle frequency and dispersion index were investigated via means of authors’ specially-written programmes. Regardless of the surface behaviour, particles are found to travel in rotational movement in horizontal plane. Particle cycle frequency is found to increase strongly with increasing vibration amplitude. Particle dispersion also increased strongly with vibration amplitude. Horizontal dispersion is observed to always exceed vertical dispersion.

Resiliencia organizacional y capital psicológico

Desde la noción universal sobre la empresa como un sistema de interacción con un entorno determinado para alcanzar un objetivo, de manera planificada y en función de satisfacer las demandas de un mercado mediante la actividad económica, su viabilidad, sostenibilidad y crecimiento dependerán, por supuesto, de una serie de estrategias adecuadas no solo para tales fines, sino también para enfrentar diversidad de agentes endógenos y exógenos que puedan afectar el normal desempeño de su gestión. Estamos hablando de la importancia de la resiliencia organizacional y del Capital Psicológico. En un escenario tan impredecible como el de la economía mundial, donde la constante son los cambios en su comportamiento —unos propios de su dinámica e interdependencia, naturales de fenómenos como la globalización, y otros derivados de eventos disruptivos— hoy más que nunca es necesario implementar el modelo de la empresa resiliente, que es aquella entidad capaz de adaptarse y recuperarse frente a una perturbación. Al mismo tiempo, más allá de su tamaño, naturaleza u objeto social, es indispensable reconocer básicamente que toda organización está constituida por personas, lo cual implica la trascendencia que para su funcionamiento tiene el factor humano-dependiente, y por lo tanto se crea la necesidad de promover el Capital Psicológico y la resiliencia a nivel de las organizaciones a través de una cultura empresarial.