With the huge volume of biological data generated by new technologies and the booming of new machine learning based analytical tools, we expect to advance life science and human health at an unprecedented pace. Unfortunately, there is a significant gap between the complex raw biological data from real life and the data required by mathematical and statistical tools. This gap is contributed by two fundamental and universal problems in biological data that are both related to confounding effects. The first is the intrinsic complexities of the data. An observed sample could be the mixture of multiple underlying sources and we may be only interested in one or part of the sources. The second type of complexities come from the acquisition process of the data. Different samples may be gathered at different time and/or from different locations. Therefore, each sample is associated with specific distortion that must be carefully addressed. These confounding effects obscure the signals of interest in the acquired data. Specifically, this dissertation will address the two major challenges in confounding effects removal: alignment and deconvolution. Liquid chromatography–mass spectrometry (LC-MS) is a standard method for proteomics and metabolomics analysis of biological samples. Unfortunately, it suffers from various changes in the retention time (RT) of the same compound in different samples, and these must be subsequently corrected (aligned) during data processing. Classic alignment methods such as in the popular XCMS package often assume a single time-warping function for each sample. Thus, the potentially varying RT drift for compounds with different masses in a sample is neglected in these methods. Moreover, the systematic change in RT drift across run order is often not considered by alignment algorithms. Therefore, these methods cannot effectively correct all misalignments. To utilize this information, we develop an integrated reference-free profile alignment method, neighbor-wise compound-specific Graphical Time Warping (ncGTW), that can detect misaligned features and align profiles by leveraging expected RT drift structures and compound-specific warping functions. Specifically, ncGTW uses individualized warping functions for different compounds and assigns constraint edges on warping functions of neighboring samples. We applied ncGTW to two large-scale metabolomics LC-MS datasets, which identifies many misaligned features and successfully realigns them. These features would otherwise be discarded or uncorrected using existing methods. When the desired signal is buried in a mixture, deconvolution is needed to recover the pure sources. Many biological questions can be better addressed when the data is in the form of individual sources, instead of mixtures. Though there are some promising supervised deconvolution methods, when there is no a priori information, unsupervised deconvolution is still needed. Among current unsupervised methods, Convex Analysis of Mixtures (CAM) is the most theoretically solid and strongest performing one. However, there are some major limitations of this method. Most importantly, the overall time complexity can be very high, especially when analyzing a large dataset or a dataset with many sources. Also, since there are some stochastic and heuristic steps, the deconvolution result is not accurate enough. To address these problems, we redesigned the modules of CAM. In the feature clustering step, we propose a clustering method, radius-fixed clustering, which could not only control the space size of the cluster, but also find out the outliers simultaneously. Therefore, the disadvantages of K-means clustering, such as instability and the need of cluster number are avoided. Moreover, when identifying the convex hull, we replace Quickhull with linear programming, which decreases the computation time significantly. To avoid the not only heuristic but also approximated step in optimal simplex identification, we propose a greedy search strategy instead. The experimental results demonstrate the vast improvement of computation time. The accuracy of the deconvolution is also shown to be higher than the original CAM. / Doctor of Philosophy / Due to the complexity of biological data, there are two major pre-processing steps: alignment and deconvolution. The alignment step corrects the time and location related data acquisition distortion by aligning the detected signals to a reference signal. Though many alignment methods are proposed for biological data, most of them fail to consider the relationships among samples carefully. This piece of structure information can help alignment when the data is noisy and/or irregular. To utilize this information, we develop a new method, Neighbor-wise Compound-specific Graphical Time Warping (ncGTW), inspired by graph theory. This new alignment method not only utilizes the structural information but also provides a reference-free solution. We show that the performance of our new method is better than other methods in both simulations and real datasets. When the signal is from a mixture, deconvolution is needed to recover the pure sources. Many biological questions can be better addressed when the data is in the form of single sources, instead of mixtures. There is a classic unsupervised deconvolution method: Convex Analysis of Mixtures (CAM). However, there are some limitations of this method. For example, the time complexity of some steps is very high. Thus, when facing a large dataset or a dataset with many sources, the computation time would be extremely long. Also, since there are some stochastic and heuristic steps, the deconvolution result may be not accurate enough. We improved CAM and the experimental results show that the speed and accuracy of the deconvolution is significantly improved.

Due to the high dimensionality of most biological data, it is a difficult task to directly analyze, model and visualize the data to gain biological insight. Thus, dimensionality reduction becomes an imperative pre-processing step in analyzing and visualizing high-dimensional biological data. Two major approaches to dimensionality reduction in genomic analysis and biomarker identification studies are: Feature extraction, creating new features by combining existing ones based on a mapping technique; and feature selection, choosing an optimal subset of all features based on an objective function. In this dissertation, we show how our innovative reduction schemes effectively reduce the dimensionality of DNA gene expression data to extract biologically interpretable and relevant features which result in enhancing the biomarker identification process. To construct biologically interpretable features and facilitate Muscular Dystrophy (MD) subtypes classification, we extract molecular features from MD microarray data by constructing sub-networks using a novel integrative scheme which utilizes protein-protein interaction (PPI) network, functional gene sets information and mRNA profiling data. The workflow includes three major steps: First, by combining PPI network structure and gene-gene co-expression relationship into a new distance metric, we apply affinity propagation clustering (APC) to build gene sub-networks; secondly, we further incorporate functional gene sets knowledge to complement the physical interaction information; finally, based on the constructed sub-network and gene set features, we apply multi-class support vector machine (MSVM) for MD sub-type classification and highlight the biomarkers contributing to the sub-type prediction. The experimental results show that our scheme could construct sub-networks that are more relevant to MD than those constructed by the conventional approach. Furthermore, our integrative strategy substantially improved the prediction accuracy, especially for those ‘hard-to-classify' sub-types. Conventionally, pathway-based analysis assumes that genes in a pathway equally contribute to a biological function, thus assigning uniform weight to genes. However, this assumption has been proven incorrect and applying uniform weight in the pathway analysis may not be an adequate approach for tasks like molecular classification of diseases, as genes in a functional group may have different differential power. Hence, we propose to use different weights for the pathway analysis which resulted in the development of four weighting schemes. We applied them in two existing pathway analysis methods using both real and simulated gene expression data for pathways. Weighting changes pathway scoring and brings up some new significant pathways, leading to the detection of disease-related genes that are missed under uniform weight. To help us understand our MD expression data better and derive scientific insight from it, we have explored a suite of visualization tools. Particularly, for selected top performing MD sub-networks, we displayed the network view using Cytoscape; functional annotations using IPA and DAVID functional analysis tools; expression pattern using heat-map and parallel coordinates plot; and MD associated pathways using KEGG pathway diagrams. We also performed weighted MD pathway analysis, and identified overlapping sub-networks across different weight schemes and different MD subtypes using Venn Diagrams, which resulted in the identification of a new sub-network significantly associated with MD. All those graphically displayed data and information helped us understand our MD data and the MD subtypes better, resulting in the identification of several potentially MD associated biomarker pathways and genes. / Ph. D.

Tissue heterogeneity, arising from intermingled cellular or tissue subtypes, significantly obscures the analyses of molecular expression data derived from complex tissues. Existing computational methods performing data deconvolution from mixed subtype signals almost exclusively rely on supervising information, requiring subtype-specific markers, the number of subtypes, or subtype compositions in individual samples. We develop a fully unsupervised deconvolution method to dissect complex tissues into molecularly distinctive tissue or cell subtypes directly from mixture expression profiles. We implement an R package, deconvolution by Convex Analysis of Mixtures (debCAM) that can automatically detect tissue or cell-specific markers, determine the number of constituent sub-types, calculate subtype proportions in individual samples, and estimate tissue/cell-specific expression profiles. We demonstrate the performance and biomedical utility of debCAM on gene expression, methylation, and proteomics data. With enhanced data preprocessing and prior knowledge incorporation, debCAM software tool will allow biologists to perform a deep and unbiased characterization of tissue remodeling in many biomedical contexts. Purified expression profiles from physical experiments provide both ground truth and a priori information that can be used to validate unsupervised deconvolution results or improve supervision for various deconvolution methods. Detecting tissue or cell-specific expressed markers from purified expression profiles plays a critical role in molecularly characterizing and determining tissue or cell subtypes. Unfortunately, classic differential analysis assumes a convenient test statistic and associated null distribution that is inconsistent with the definition of markers and thus results in a high false positive rate or lower detection power. We describe a statistically-principled marker detection method, One Versus Everyone Subtype Exclusively-expressed Genes (OVESEG) test, that estimates a mixture null distribution model by applying novel permutation schemes. Validated with realistic synthetic data sets on both type 1 error and detection power, OVESEG-test applied to benchmark gene expression data sets detects many known and de novo subtype-specific expressed markers. Subsequent supervised deconvolution results, obtained using markers detected by the OVESEG-test, showed superior performance when compared with popular peer methods. While the current debCAM approach can dissect mixed signals from multiple samples into the 'averaged' expression profiles of subtypes, many subsequent molecular analyses of complex tissues require sample-specific deconvolution where each sample is a mixture of 'individualized' subtype expression profiles. The between-sample variation embedded in sample-specific subtype signals provides critical information for detecting subtype-specific molecular networks and uncovering hidden crosstalk. However, sample-specific deconvolution is an underdetermined and challenging problem because there are more variables than observations. We propose and develop debCAM2.0 to estimate sample-specific subtype signals by nuclear norm regularization, where the hyperparameter value is determined by random entry exclusion based cross-validation scheme. We also derive an efficient optimization approach based on ADMM to enable debCAM2.0 application in large-scale biological data analyses. Experimental results on realistic simulation data sets show that debCAM2.0 can successfully recover subtype-specific correlation networks that is unobtainable otherwise using existing deconvolution methods. / Doctor of Philosophy / Tissue samples are essentially mixtures of tissue or cellular subtypes where the proportions of individual subtypes vary across different tissue samples. Data deconvolution aims to dissect tissue heterogeneity into biologically important subtypes, their proportions, and their marker genes. The physical solution to mitigate tissue heterogeneity is to isolate pure tissue components prior to molecular profiling. However, these experimental methods are time-consuming, expensive and may alter the expression values during isolation. Existing literature primarily focuses on supervised deconvolution methods which require a priori information. This approach has an inherent problem as it relies on the quality and accuracy of the a priori information. In this dissertation, we propose and develop a fully unsupervised deconvolution method - deconvolution by Convex Analysis of Mixtures (debCAM) that can estimate the mixing proportions and 'averaged' expression profiles of individual subtypes present in heterogeneous tissue samples. Furthermore, we also propose and develop debCAM2.0 that can estimate 'individualized' expression profiles of participating subtypes in complex tissue samples. Subtype-specific expressed markers, or marker genes (MGs), serves as critical a priori information for supervised deconvolution. MGs are exclusively and consistently expressed in a particular tissue or cell subtype while detecting such unique MGs involving many subtypes constitutes a challenging task. We propose and develop a statistically-principled method - One Versus Everyone Subtype Exclusively-expressed Genes (OVESEG-test) for robust detection of MGs from purified profiles of many subtypes.

<p dir="ltr">Maternal mortality (MM) is a pressing global health issue that results in thousands of mothers dying annually from pregnancy-related complications. Despite spending trillions of dollars on the healthcare industry, the U.S. continues to experience one of the highest rates of maternal death (MD) compared to other developed countries. This ongoing public health crisis highlights the urgent need for innovative strategies to detect and mitigate adverse maternal outcomes. This study introduces a novel approach, utilizing admission to the ICU as a proxy for MM. By analyzing 14 years of natality birth data, this study aims to explore the complex web of factors that elevate the chances of MD. The primary goal of this study is to identify features that are most influential in predicting ICU admission cases. These factors hold the potential to be applied to MM, as they can serve as early warning signs that complications may arise, allowing healthcare professionals to step in and intervene before adverse maternal outcomes occur. Two supervised machine learning models were employed in this study, specifically Logistic Regression (LR) and eXtreme Gradient Boosting (XGBoost). The models were executed twice for each dataset: once incorporating all available features and again utilizing only the most significant features. Following model training, XGBoost’s feature selection technique was employed to identify the top 10 influential features that are most important to the classification process. Our analysis revealed a diverse range of factors that are important for the prediction of ICU admission cases. In this study, we identified maternal transfusion, labor and delivery characteristics, delivery methods, gestational age, maternal attributes, and newborn conditions as the most influential factors to categorize maternal ICU admission cases. In terms of model performance, the XGBoost consistently outperformed LR across various datasets, demonstrating higher accuracy, precision, and F1 scores. For recall, however, LR maintained higher scores, surpassing those of XGBoost. Moreover, the models consistently achieved higher scores when trained with all available features compared to those trained solely with the top features. Although the models demonstrated satisfactory performance in some evaluation metrics, there were notable deficiencies in recall and precision, which suggests further model refinement is needed to effectively predict these cases.</p>

Die Etablierung von Hochdurchsatz-Technologien zur Durchführung von Genexpressionsmessungen führte in den letzten 20 Jahren zu einer stetig wachsende Menge an verfügbaren Daten. Sie ermöglichen durch Kombination einzelner Experimente neue Vergleichsstudien zu kombinieren oder Experimente aus verschiedenen Studien zu großen Datensätzen zu vereinen. Dieses Vorgehen wird als Meta-Analyse bezeichnet und in dieser Arbeit verwendet, um einen großen Genexpressionsdatensatz aus öffentlich zugänglichen T-Zell Experimenten zu erstellen. T-Zellen sind Immunzellen, die eine Vielzahl von unterschiedlichen Funktionen des Immunsystems inititiieren und steuern. Sie können in verschiedene Subtypen mit unterschiedlichen Funktionen differenzieren. Der mittels Meta-Analyse erstellte Datensatz beinhaltet nur Experimente zu einem T-Zell-Subtyp, den regulatorischen T-Zellen (Treg) bzw. der beiden Untergruppen, natürliche Treg (nTreg) und induzierte Treg (iTreg) Zellen. Eine bisher unbeantwortete Frage lautet, welche subtyp-spezifischen gen-regulatorische Mechanismen die T-Zell Differenzierung steuern. Dazu werden in dieser Arbeit zwei spezifische Herausforderungen der Treg Forschung behandelt: (i) die Identifikation von Zelloberflächenmarkern zur Unterscheidung und Charakterisierung der Subtypen, sowie (ii) die Rekonstruktion von Treg-Zell-spezifischen gen-regulatorischen Netzwerken (GRN), die die Differenzierungsmechanismen beschreiben. Die implementierte Meta-Analyse kombiniert mehr als 150 Microarray-Experimente aus über 30 Studien in einem Datensatz. Dieser wird benutzt, um mittels Machine Learning Zell-spezifische Oberflächenmarker an Hand ihres Expressionsprofils zu identifizieren. Mit der in dieser Arbeit entwickelten Methode wurden 41 Genen extrahiert, von denen sechs Oberflächenmarker sind. Zusätzliche Validierungsexperimente zeigten, dass diese sechs Gene die Experimenten beider T-Zell Subtypen sicher unterscheiden können. Zur Rekonstruktion von GRNs vergleichen wir unter Verwendung des erstellten Datensatzes 11 verschiedene Algorithmen und evaluieren die Ergebnisse mit Informationen aus Interaktionsdatenbanken. Die Evaluierung zeigt, dass die derzeit verfügbaren Methoden nicht in der Lage sind den Wissensstand Treg-spezifischer, regulatorsicher Mechanismen zu erweitern. Abschließend präsentieren wir eine Datenintegrationstrategie zur Rekonstruktion von GRN am Beispiel von Th2 Zellen. Aus Hochdurchsatzexperimenten wird ein Th2-spezifisches GRN bestehend aus 100 Genen rekonstruiert. Während 89 dieser Gene im Kontext der Th2-Zelldifferenzierung bekannt sind, wurden 11 neue Kandidatengene ohne bisherige Assoziation zur Th2-Differenzierung ermittelt. Die Ergebnisse zeigen, dass Datenintegration prinzipiell die GRN Rekonstruktion ermöglicht. Mit der Verfügbarkeit von mehr Daten mit besserer Qualität ist zu erwarten, dass Methoden zur Rekonstruktion maßgeblich zum besseren Verstehen der zellulären Differenzierung im Immunsystem und darüber hinaus beitragen können und so letztlich die Ursachenforschung von Dysfunktionen und Krankheiten des Immunsystems ermöglichen werden. / Within the last two decades high-throughput gene expression screening technologies have led to a rapid accumulation of experimental data. The amounts of information available have enabled researchers to contrast and combine multiple experiments by synthesis, one of such approaches is called meta-analysis. In this thesis, we build a large gene expression data set based on publicly available studies for further research on T cell subtype discrimination and the reconstruction of T cell specific gene regulatory events. T cells are immune cells which have the ability to differentiate into subtypes with distinct functions, initiating and contributing to a variety of immune processes. To date, an unsolved problem in understanding the immune system is how T cells obtain a specific subtype differentiation program, which relates to subtype-specific gene regulatory mechanisms. We present an assembled expression data set which describes a specific T cell subset, regulatory T (Treg) cells, which can be further categorized into natural Treg (nTreg) and induced Treg (iTreg) cells. In our analysis we have addressed specific challenges in regulatory T cell research: (i) discriminating between different Treg cell subtypes for characterization and functional analysis, and (ii) reconstructing T cell subtype specific gene regulatory mechanisms which determine the differences in subtype-specific roles for the immune system. Our meta-analysis strategy combines more than one hundred microarray experiments. This data set is applied to a machine learning based strategy of extracting surface protein markers to enable Treg cell subtype discrimination. We identified a set of 41 genes which distinguish between nTregs and iTregs based on gene expression profile only. Evaluation of six of these genes confirmed their discriminative power which indicates that our approach is suitable to extract candidates for robust discrimination between experiment classes. Next, we identify gene regulatory interactions using existing reconstruction algorithms aiming to extend the number of known gene-gene interactions for Treg cells. We applied eleven GRN reconstruction tools based on expression data only and compared their performance. Taken together, our results suggest that the available methods are not yet sufficient to extend the current knowledge by inferring so far unreported Treg specific interactions. Finally, we present an approach of integrating multiple data sets based on different high-throughput technologies to reconstruct a subtype-specific GRN. We constructed a Th2 cell specific gene regulatory network of 100 genes. While 89 of these are known to be related to Th2 cell differentiation, we were able to attribute 11 new candidate genes with a function in Th2 cell differentiation. We show that our approach to data integration does, in principle, allow for the reconstruction of a complex network. Future availability of more and more consistent data may enable the use of the concept of GRN reconstruction to improve understanding causes and mechanisms of cellular differentiation in the immune system and beyond and, ultimately, their dysfunctions and diseases.

The problem of feature selection is critical in several areas of machine learning and data analysis such as, for example, cancer classification using gene expression data, text categorization, etc. In this work, we consider feature selection for supervised learning problems, where one wishes to select a small set of features that facilitate learning a good prediction model in the reduced feature space. Our interest is primarily in filter methods that select features independently of the learning algorithm to be used and are generally faster to implement compared to other types of feature selection algorithms. Many common filter methods for feature selection make use of information-theoretic criteria such as those based on mutual information to guide their search process. However, even in simple binary classification problems, mutual information based methods do not always select the best set of features in terms of the Bayes error. In this thesis, we develop a general approach for selecting a set of features that directly aims to minimize the Bayes error in the reduced feature space with respect to the loss or performance measure of interest. We show that the mutual information based criterion is a special case of our setting when the loss function of interest is the logarithmic loss for class probability estimation. We give a greedy forward algorithm for approximately optimizing this criterion and demonstrate its application to several supervised learning problems including binary classification (with 0-1 error, cost-sensitive error, and F-measure), binary class probability estimation (with logarithmic loss), bipartite ranking (with pairwise disagreement loss), and multiclass classification (with multiclass 0-1 error). Our experiments suggest that the proposed approach is competitive with several state-of-the art methods.

At the time of writing, stress and stress related disease have become the most common reasons for absence in the workplace in Sweden. The purpose of the work presented here is to identify and notify people managing unhealthy levels of stress. Since symptoms of mental stress manifest through functions of the Sympathetic Nervous System (SNS), they are best measured through monitoring of SNS changes and phenomena. In this study, changes in the sympathetic control of heart rate were recorded and analyzed using heart rate variability analysis and a simple runner’s heart rate sensor connected to a smartphone. Mental stress data was collected through stressful video gaming. This was compared to data from non-stressful activities, physical activity and extremely stressful activities such as public speaking events. By using the period between heartbeats and selecting features from the frequency domain, a simple machine learning algorithm could differentiate between the types of data and thus could effectively recognize mental stress. The study resulted in a collection of 100 data points, an algorithm to extract features and an application to continuously collect and classify sequences of heart periods. It also revealed an interesting relationship in the data between different subjects. The fact that continuous stress monitoring can be achieved using minimally intrusive sensors is the greatest benefit of these results, especially when connsidering its potential value in the identification and prevention of stress related disease.

The metric of entropy provides a measure about the randomness of data and a measure of information gained by comparing different attributes. Intrusion detection systems can collect very large amounts of data, which are not necessarily manageable by manual means. Collected intrusion detection data often contains redundant, duplicate, and irrelevant entries, which makes analysis computationally intensive likely leading to unreliable results. Reducing the data to what is relevant and pertinent to the analysis requires the use of data mining techniques and statistics. Identifying patterns in the data is part of analysis for intrusion detections in which the patterns are categorized as normal or anomalous. Anomalous data needs to be further characterized to determine if representative attacks to the network are in progress. Often time subtleties in the data may be too muted to identify certain types of attacks. Many statistics including entropy are used in a number of analysis techniques for identifying attacks, but these analyzes can be improved upon. This research expands the use of Approximate entropy and Sample entropy for feature selection and attack analysis to identify specific types of subtle attacks to network systems. Through enhanced analysis techniques using entropy, the granularity of feature selection and attack identification is improved.

Technological advancement such as the increase in processing power, hard disk capacity and network bandwidth has opened up many exciting new techniques to synthesise sounds, one of which is Concatenative Sound Synthesis (CSS). CSS uses data-driven method to synthesise new sounds from a large corpus of small sound snippets. This technique closely resembles the art of mosaicing, where small tiles are arranged together to create a larger image. A ‘target’ sound is often specified by users so that segments in the database that match those of the target sound can be identified and then concatenated together to generate the output sound. Whilst the practicality of CSS in synthesising sounds currently looks promising, there are still areas to be explored and improved, in particular the algorithm that is used to find the matching segments in the database. One of the main issues in CSS is the basis of similarity, as there are many perceptual attributes which sound similarity can be based on, for example it can be based on timbre, loudness, rhythm, and tempo and so on. An ideal CSS system needs to be able to decipher which of these perceptual attributes are anticipated by the users and then accommodate them by synthesising sounds that are similar with respect to the particular attribute. Failure to communicate the basis of sound similarity between the user and the CSS system generally results in output that mismatches the sound which has been envisioned by the user. In order to understand how humans perceive sound similarity, several elements that affected sound similarity judgment were first investigated. Of the four elements tested (timbre, melody, loudness, tempo), it was found that the basis of similarity is dependent on humans’ musical training where musicians based similarity on the timbral information, whilst non-musicians rely on melodic information. Thus, for the rest of the study, only features that represent the timbral information were included, as musicians are the target user for the findings of this study. Another issue with the current state of CSS systems is the user control flexibility, in particular during segment matching, where features can be assigned with different weights depending on their importance to the search. Typically, the weights (in some existing CSS systems that support the weight assigning mechanism) can only be assigned manually, resulting in a process that is both labour intensive and time consuming. Additionally, another problem was identified in this study, which is the lack of mechanism to handle homosonic and equidistant segments. These conditions arise when too few features are compared causing otherwise aurally different sounds to be represented by the same sonic values, or can also be a result of rounding off the values of the features extracted. This study addresses both of these problems through an extended use of Artificial Intelligence (AI). The Analysis Hierarchy Process (AHP) is employed to enable order dependent features selection, allowing weights to be assigned for each audio feature according to their relative importance. Concatenation distance is used to overcome the issues with homosonic and equidistant sound segments. The inclusion of AI results in a more intelligent system that can better handle tedious tasks and minimize human error, allowing users (composers) to worry less of the mundane tasks, and focusing more on the creative aspects of music making. In addition to the above, this study also aims to enhance user control flexibility in a CSS system and improve similarity result. The key factors that affect the synthesis results of CSS were first identified and then included as parametric options which users can control in order to communicate their intended creations to the system to synthesise. Comprehensive evaluations were carried out to validate the feasibility and effectiveness of the proposed solutions (timbral-based features set, AHP, and concatenation distance). The final part of the study investigates the relationship between perceived sound similarity and perceived sound interestingness. A new framework that integrates all these solutions, the query-based CSS framework, was then proposed. The proof-of-concept of this study, ConQuer, was developed based on this framework. This study has critically analysed the problems in existing CSS systems. Novel solutions have been proposed to overcome them and their effectiveness has been tested and discussed, and these are also the main contributions of this study.

特徵選取為機器學習常見的資料前處理的方法，現今已有許多不同的特徵選取演算法，然而並不存在一個在所有資料上都優於其他方法的演算法，且由於現今的資料種類繁多，所以研發新的方法能夠帶來更多有關資料的資訊並且根據資料的特性採用不同的變數選取演算法是較好的做法。 本研究使用資訊理論entropy的概念依照變數之間資料雲幾何樹的分群結果定義變數之間的相關性，且依此選取資料的特徵，並與同樣使用entropy概念的FCBF方法、Lasso、F-score、隨機森林、基因演算法互相比較，本研究使用階層式分群法與多數決投票法套用在真實的資料上判斷預測率。結果顯示，本研究使用的entropy方法在各個不同的資料集上有較穩定的預測率提升表現，同時資料縮減的維度也相對穩定。 / Feature selection is a common preprocessing technique in machine learning. Although a large pool of feature selection techniques has existed, there is no such a dominant method in all datasets. Because of the complexity of various data formats, establishing a new method can bring more insight into data, and applying proper techniques to analyzing data would be the best choice. In this study, we used the concept of entropy from information theory to build a similarity matrix between features. Additionally, we constructed a DCG-tree to separate variables into clusters. Each core cluster consists of rather uniform variables, which share similar covariate information. With the core clusters, we reduced the dimension of a high-dimensional dataset. We assessed our method by comparing it with FCBF, Lasso, F-score, random forest and genetic algorithm. The performances of prediction were demonstrated through real-world datasets using hierarchical clustering with voting algorithm as the classifier. The results showed that our entropy method has more stable prediction performances and reduces sufficient dimensions of the datasets simultaneously.