No description available.

This dissertation presents the results of three research studies investigating the role trust plays in an auditor's decisions. The first study examines whether auditors develop trust in a client's management after working with the client during prior audit engagements. The results indicate that auditors have higher trust in the client's management after a positive, overall satisfying experience working with the client compared to a negative, overall unsatisfying experience. The first study also investigates whether auditors" trust in a client affects their audit decisions. The results show a negative relationship between auditors" trust and their fraud risk assessment. Specifically, lower levels of trust are associated with higher levels of risk, and vice versa. Together, the results suggest that auditors may indeed develop trust in a client's management and this trust may affect their audit decisions. The second study examines whether Certified Public Accountants’ (CPAs) level of moral reasoning affects their decision to trust a client's management and the extent to which to trust them. The results show that CPAs with relatively higher levels of moral reasoning have less trust in the client's management than CPAs with relatively lower levels of moral reasoning. The findings indicate that an auditor's decision to trust a client's management is, at least in part, an ethical judgment. Also, because the decision is an ethical one, the findings suggest that trust beyond some threshold would be considered unethical. The third study extends the results of the first study by simultaneously examining how an auditor's trust and the financial importance of the client affect the auditor's decision to accept the client's preferred method of recognizing revenue. The results indicate that auditors" trust in the client's management is positively related to their commitment to the goal of supporting the client's preferred reporting methods (goal commitment), which in turn is positively related to the auditors" assessments of the acceptability of the client's methods for reporting purposes. The importance of the client did not affect auditors" goal commitment or their acceptability assessments. The findings suggest that auditors with higher levels of trust may be more likely to accept the client's preferred method of financial reporting. Overall, these results add to our knowledge of audit judgment and decision-making by providing evidence that auditors do indeed develop trust in a client's management; that the decision and extent to trust the client is in part an ethical judgment; and that auditors" trust may affect their audit decisions. This dissertation highlights the important role that an auditor's trust plays in his or her audit decisions. / Ph. D.

Biological network data is plentiful with practically every experimental methodology giving 'network views' into cellular function and behavior. Bioinformatic screens that yield network data include, for example, genome-wide deletion screens, protein-protein interaction assays, RNA interference experiments, and methods to probe metabolic pathways. Efficient and comprehensive computational approaches are required to model these screens and gain insight into the nature of biological networks. This thesis presents three new algorithms to model and mine network datasets. First, we present an algorithm that models genome-wide perturbation screens by deriving relations between phenotypes and subsequently using these relations in a local manner to derive genephenotype relationships. We show how this algorithm outperforms all previously described algorithms for gene-phenotype modeling. We also present theoretical insight into the convergence and accuracy properties of this approach. Second, we define a new data mining problem–constrained minimal separator mining—and propose algorithms as well as applications to modeling gene perturbation screens by viewing the perturbed genes as a graph separator. Both of these data mining applications are evaluated on network datasets from S. cerevisiae and C. elegans. Finally, we present an approach to model the relationship between metabolic pathways and operon structure in prokaryotic genomes. In this approach, we present a new pattern class—biclusters over domains with supplied partial orders—and present algorithms for systematically detecting such biclusters. Together, our data mining algorithms provide a comprehensive arsenal of techniques for modeling gene perturbation screens and metabolic pathways. / Ph. D.

No / Model interpretation is one of the key aspects of the model evaluation process. The explanation of the relationship between model variables and outputs is relatively easy for statistical models, such as linear regressions, thanks to the availability of model parameters and their statistical significance . For “black box” models, such as random forest, this information is hidden inside the model structure. This work presents an approach for computing feature contributions for random forest classification models. It allows for the determination of the influence of each variable on the model prediction for an individual instance. By analysing feature contributions for a training dataset, the most significant variables can be determined and their typical contribution towards predictions made for individual classes, i.e., class-specific feature contribution “patterns”, are discovered. These patterns represent a standard behaviour of the model and allow for an additional assessment of the model reliability for new data. Interpretation of feature contributions for two UCI benchmark datasets shows the potential of the proposed methodology. The robustness of results is demonstrated through an extensive analysis of feature contributions calculated for a large number of generated random forest models.

Le risque attribuable est un outil épidémiologique apparu dans les années 1950 aujourd’hui encore assez peu utilisé. Il permet d’estimer la proportion de cas d’une maladie potentiellement évitable par suppression ou réduction de l’exposition d’une population à un facteur de risque. Son principal intérêt réside dans la prise en compte concomitante de l’ampleur d’effet du facteur de risque et de la distribution de ce facteur au sein de la population. Après une présentation des caractéristiques essentielles du risque attribuable et des principes de son estimation à partir d’une étude cas-témoins, nous proposons un cadre conceptuel qui permet d’estimer l’impact d’une intervention de santé publique dans une nouvelle population dont l’exposition à certains facteurs de risque diffère de celle observée dans la population d’étude. Une décomposition du risque attribuable permet alors de prendre en compte l’action combinée, ou synergie, des facteurs de risque dans la survenue de la maladie. Parce qu’il donne une dimension populationnelle à l’estimation de l’effet d’une variable, le risque attribuable est particulièrement intéressant pour quantifier l’importance relative des différentes variables explicatives d’un modèle de régression. La question de l’importance relative des biomarqueurs classiques et de ceux issus des technologies à haut débit dans les modèles diagnostiques est actuellement centrale pour établir les apports respectifs de ces deux niveaux d’information. À partir de simulations, nous montrons comment l’apport des nouvelles technologies, quantifié en termes de risque attribuable, peut être faussé par l’utilisation de méthodologies inadaptées / The attributable risk is an epidemiologic tool that dates back to the fifties but is still relatively seldom used. It estimates the proportion of cases of a given disease that could be avoided if the exposure to a specific risk factor was removed or reduced. Its major interest is that it combines the magnitude of the effect of the risk factor to the distribution of this factor within the population. After a review of the attributable risk main features and the principles of its estimation from case-control studies data, we propose a conceptual framework that allows estimating the impact of a public health intervention in a new population with different exposure to certain risk factors than those observed in the study population. To reach this goal, we used a splitting of the attributable risk that takes into account the combined action –or synergy– of the risk factors on the occurrence of the disease. Because the attributable risk allows estimating the effect of a variable at the population level, it is particularly interesting to quantify the relative importance of the covariates of a regression model. In diagnostic models, the estimation of the relative importance of classic biomarkers and biomarkers obtained from high-throughput technologies is currently crucial in establishing the contribution of each of these two levels of information. Using simulations we have demonstrated the way the role of high-throughput-technologies –quantified in terms of attributable risk– may be wrongly assessed through the use of unsuitable methodology

Segmentation of magnetic resonance images is an important part of planning radiotherapy treat-ments for patients with brain tumours but due to the number of images contained within a scan and the level of detail required, manual segmentation is a time consuming task. Convolutional neural networks have been proposed as tools for automated segmentation and shown promising results. However, the data sets used for training these deep learning models are often imbalanced and contain data that does not contribute to the performance of the model. By carefully selecting which data to train on, there is potential to both speed up the training and increase the network’s ability to detect tumours. This thesis implements the method of importance sampling for training a convolutional neural network for patch-based segmentation of three dimensional multimodal magnetic resonance images of the brain and compares it with the standard way of sampling in terms of network performance and training time. Training is done for two different patch sizes. Features of the most frequently sampled volumes are also analysed. Importance sampling is found to speed up training in terms of number of epochs and also yield models with improved performance. Analysis of the sampling trends indicate that when patches are large, small tumours are somewhat frequently trained on, however more investigation is needed to confirm what features may influence the sampling frequency of a patch. / Segmentering av magnetröntgenbilder är en viktig del i planeringen av strålbehandling av patienter med hjärntumörer. Det höga antalet bilder och den nödvändiga precisionsnivån gör dock manuellsegmentering till en tidskrävande uppgift. Faltningsnätverk har därför föreslagits som ett verktyg förautomatiserad segmentering och visat lovande resultat. Datamängderna som används för att träna dessa djupinlärningsmodeller är ofta obalanserade och innehåller data som inte bidrar till modellensprestanda. Det finns därför potential att både skynda på träningen och förbättra nätverkets förmåga att segmentera tumörer genom att noggrant välja vilken data som används för träning. Denna uppsats implementerar importance sampling för att träna ett faltningsnätverk för patch-baserad segmentering av tredimensionella multimodala magnetröntgenbilder av hjärnan. Modellensträningstid och prestanda jämförs mot ett nätverk tränat med standardmetoden. Detta görs förtvå olika storlekar på patches. Egenskaperna hos de mest valda volymerna analyseras också. Importance sampling uppvisar en snabbare träningsprocess med avseende på antal epoker och resulterar också i modeller med högre prestanda. Analys av de oftast valda volymerna indikerar att under träning med stora patches förekommer små tumörer i en något högre utsträckning. Vidareundersökningar är dock nödvändiga för att bekräfta vilka aspekter som påverkar hur ofta en volym används.

Payment service providers have unique credit portfolios with different characteristics than many other credit providers. It is therefore important to study if common credit risk estimation methods are applicable to their setting. By comparing simulation based methods for credit risk estimation it was found that combining Monte Carlo simulation with importance sampling and the asymptotic single risk factor model is the most suitable model amongst those analyzed. It allows for a combination of variance reduction, scenario analysis and correlation checks, which all are important for estimating credit risk in a payment service provider portfolio. / Betaltjänstleverantörer har unika kreditportföljer med andra egenskaper än många andra kreditgivare. Det är därför viktigt att undersöka om vanliga metoder för uppskattning av kreditrisk går att tillämpa på betaltjänstleverantörer. Genom att jämföra olika simulationsbaserade metoder för uppskattning av kreditrisk fann man att att kombinationen av Monte Carlo-simulering med Importance Sampling och en ASRF-modell är den mest lämpliga bland de analyserade metoderna. Det möjliggör en kombination av variansminskning, scenarioanalys och korrelationskontroller som alla är viktiga för att uppskatta kreditrisk i en betaltjänstleverantörsportfölj.

Få studier om ljudsymbolik, d.v.s. kopplingen mellan ords form och betydelse, har baserats på statistisk analys. I denna studie använder vi random forests med måttet permutation variable importance för att utforska vilka fonem (språkljud) som är prevalenta i engelska ord som beskriver hårdhet eller mjukhet. Denna icke-parametriska maskininlärningsmetod har funnits vara användbar för identifiering av ett fåtal inflytelserika förklaringsvariabler i situationer där n < p eller interkorrelationer förekommer. Vårt material och val av metod grundar sig på en tidigare studie, som fann att r-ljud hade starkt samband med betydelsen ‘strävhet’, men som inte kontrollerade för betydelsen ‘hårdhet’ trots att dessa korrelerar med varandra. Vi kontrollerar för dimensionen strävhet-lenhet genom att utföra random forest-analysen på två delmängder: ord som används för att beskriva hårdhet eller mjukhet (n = 81), samt den delmängd av dessa ord som inte beskriver strävhet eller lenhet (n  = 40). Samtliga regressorer är binära variabler, som anger förekomsten eller avsaknaden av varsitt fonem; vi utförde separata analyser på respektive datamängd för att se vilka fonem som hade störst effekt, då man betraktade specifika stavelsekomponenter. Vi fann att r-ljuden hade starkt samband med betydelsen ‘hårdhet’ både före och efter kontrollen för ‘strävhet’. Vi fann även att ljudet med symbolen i (t.ex. sista vokalen i fluffy) hade starkt samband med betydelsen ‘mjukhet’ före och efter kontroll, men vi misstänker att detta egentligen reflekterar sambandet mellan ‘mjukhet’ och exkluderade bakgrundsvariabler. / Few studies about sound symbolism, i.e. the association between the shape and meaning of words, have been based on statistical analysis. In this study, we use random forests and the permutation variable importance measure to explore which phonemes (language sounds) are prevalent in English descriptors of hardness or softness. This non-parametric machine learning method has been found useful for identification of a few influential predictors in situations where n < p or intercorrelations are present. Our materials and choice of method are based on an earlier study, in which a strong association was found between r-sounds and ‘roughness’, but which did not control for the meaning ‘hardness’ despite the correlation between them. We control for the dimension ‘roughness-smoothness’ by performing the random forest-analysis on two subsets of data: descriptors of hardness or softness (n = 81), and descriptors of hardness or softness which are not used to describe roughness or smoothness (n = 40). All regressors are binary variables indicating the presence or absence of a phoneme. Separate analyses were conducted on each subset to see which phonemes had the largest effect when specific syllable compontents were considered. We found that r-sounds had a strong association with ‘hardness’ both before and after controlling for ‘roughness’. We also found that the sound here symbolized by i (e.g. the last vowel of fluffy) had a strong association with ‘softness’ before and after control, but we suspect that this might instead reflect an association between ‘softness’ and excluded variables.

Mémoire numérisé par la Division de la gestion de documents et des archives de l'Université de Montréal.

Cette thèse porte sur l'indexation et l'interrogation de pages Web. Dans ce cadre, nous proposons un nouveau modèle : BlockWeb, qui s'appuie sur une décomposition de pages Web en une hiérarchie de blocs visuels. Ce modèle prend en compte, l'importance visuelle de chaque bloc et la perméabilité des blocs au contenu de leurs blocs voisins dans la page. Les avantages de cette décomposition sont multiples en terme d'indexation et d'interrogation. Elle permet notamment d'effectuer une interrogation à une granularité plus fine que la page : les blocs les plus similaires à une requête peuvent être renvoyés à la place de la page complète. Une page est représentée sous forme d'un graphe acyclique orienté dont chaque nœud est associé à un bloc et étiqueté par l'importance de ce bloc et chaque arc est étiqueté la perméabilité du bloc cible au bloc source. Afin de construire ce graphe à partir de la représentation en arbre de blocs d'une page, nous proposons un nouveau langage : XIML (acronyme de XML Indexing Management Language), qui est un langage de règles à la façon de XSLT. Nous avons expérimenté notre modèle sur deux applications distinctes : la recherche du meilleur point d'entrée sur un corpus d'articles de journaux électroniques et l'indexation et la recherche d'images sur un corpus de la campagne d'ImagEval 2006. Nous en présentons les résultats. / This thesis is about indexing and querying Web pages. We propose a new model called BlockWeb, based on the decomposition of Web pages into a hierarchy of visual blocks. This model takes in account the visual importance of each block as well as the permeability of block's content to their neighbor blocks on the page. Splitting up a page into blocks has several advantages in terms of indexing and querying. It allows to query the system with a finer granularity than the whole page: the most similar blocks to the query can be returned instead of the whole page. A page is modeled as a directed acyclic graph, the IP graph, where each node is associated with a block and is labeled by the coefficient of importance of this block and each arc is labeled by the coefficient of permeability of the target node content to the source node content. In order to build this graph from the bloc tree representation of a page, we propose a new language : XIML (acronym for XML Indexing Management Language), a rule based language like XSLT. The model has been assessed on two distinct dataset: finding the best entry point in a dataset of electronic newspaper articles, and images indexing and querying in a dataset drawn from web pages of the ImagEval 2006 campaign. We present the results of these experiments.