Analysis of Metabolomic Data Using Support Vector Machines
Sankar Mahadevan,?Sirish L.Shah,?Thomas J.Marrie,?and Carolyn M.Slupsky*,?Department of Chemical and Materials Engineering and Department of Medicine,University of Alberta, Edmonton,Canada
Metabolomics is an emerging?eld providing insight into physiological processes.It is an effective tool to investigate disease diagnosis or conduct toxicological studies by observing changes in metabolite concentrations in various bio?uids.Multivariate statistical analysis is generally employed with nuclear magnetic resonance(NMR)or mass spectrometry(MS)data to determine differences between groups(for instance diseased vs healthy).Char-acteristic predictive models may be built based on a set of training data,and these models are subsequently used to predict whether new test data falls under a speci?c class.In this study,metabolomic data is obtained by doing a1H NMR spectroscopy on urine samples obtained from healthy subjects(male and female)and patients suffering from Streptococcus pneumoniae.We compare the per-formance of traditional PLS-DA multivariate analysis to support vector machines(SVMs),a technique widely used in genome studies on two case studies:(1)a case where nearly complete distinction may be seen(healthy versus pneumonia)and(2)a case where distinction is more ambiguous(male versus female).We show that SVMs are superior to PLS-DA in both cases in terms of predictive accuracy with the least number of features.With fewer number of features,SVMs are able to give better predictive model when compared to that of PLS-DA.
Metabolomics can be de?ned as the?eld of science that deals with the measurement of metabolites in an organism for the study of the physiological processes and their reactions to various stimuli such as infection,disease,or drug use.1The?eld of metabolomics has shown great promise for early diagnosis of diseases or preclinical screening of candidate drugs in the pharmaceutical industry.2In order to carry out such studies,analytical processes such as NMR spectroscopy and mass spectrometry are combined with statistical techniques,such as multivariate analysis and machine learning tools.1
In general,metabolomic NMR data consists of observations (study subjects)with associated variables(bins,metabolites,etc.). The main task in the analysis of this type of data is to extract meaningful information and hopefully facilitate an understanding of the complex biological processes.In general,for metabolomic and other such chemometric data,the distinguishing factors between any two classes(for instance diseased versus healthy) are a combination of variables or features rather than a single variable.Thus,multivariate analysis,using techniques such as principal components analysis(PCA),3,4clustering analysis,or PLS-DA(partial least-squares-discriminant analysis),5is necessary to discriminate between different classes of observations.
In general,there are two types of multivariate analysis techniques:unsupervised,such as PCA,and supervised,such as PLS-DA.6Supervised multivariate analysis uses class information to build predictive models.Machine learning algorithms are a more recent class of multivariate analysis technique.These algorithms can be trained to learn rules and form patterns from the input data and subsequently be applied to analyze new data. Machine learning tools often provide as good or better classi?ca-tion than PCA or PLS-DA.In particular,Bayesian classi?ers, hidden Markov models,and evolutionary algorithms such as genetic algorithm,simulated annealing,and particle swarm opti-mization7have also been used to ef?ciently and effectively interpret biological data such as microarray databases,protein structures,and networks.In this work,the applicability of one such machine learning technique,namely support vector machines (SVMs),is explored to analyze and classify metabolomic data.
SVMs are known to have excellent generalization abilities when compared to other statistical multivariate methods,such as PCA or PLS-DA.Unlike PCA or PLS-DA,SVMs can be extended to nonlinear cases with the help of kernels,whereas PCA and PLS-DA have an assumption of linearity.For nearly a decade,SVMs have been used in the?eld of bioinformatics for classifying and evaluating gene expression microarray data8-11and identifying
*To whom correspondence should be addressed.E-mail:cslupsky@ ualberta.ca.
?Department of Chemical and Materials Engineering.
?Department of Medicine.
(1)Nicholson,J.;Lindon,J.;Holmes,E.Xenobiotica1999,29,1181–1189.
(2)Lindon,J.;et al.Toxicol.Appl.Pharmacol.2003,187,137–146.
(3)Holmes,E.;Antti,H.Analyst2002,127,1549–1557.
(4)Holmes,E.;Nicholson,J.;Nicholls,A.;Lindon,J.;Connor,S.;Polley,S.;
Connelly,https://www.doczj.com/doc/0016115356.html,b.Syst.1998,44,245–255.
(5)Keun,H.;Ebbels,T.;Antti,H.;Bollard,M.;Beckonert,O.;Holmes,E.Anal.
Chim.Acta2003,490,265–276.
(6)Eriksson,L.;Johansson,E.;Wold,H.;Wold,S.Introduction to Multi and
Megavariate Analysis using Projection Methods(PCA and PLS);Umetrics Academy:Sweden,1999.
(7)Baldi,P.;Brunak,S.Bioinformatics:The Machine Learning Approach,Second
Edition(Adaptive Computation and Machine Learning);The MIT Press: Cambridge,MA,2001.
(8)Furey,T.;Cristianini,N.;Duffy,N.;Bednarski,D.;Schummer,M.;Haussler,
D.Bioinformatics2000,10,906–914.
(9)Mukherjee,S.;Tamayo,P.;Mesirov,J.;Slonim,D.;Verri,A.;Poggio,T.
Support vector machine classi?cation of microarray data,1999.
(10)Brown,M.;Grundy,W.;Lin,D.;Cristianini,N.;Sugnet,C.;Furey,T.;Ares,
J.;Haussler,D.Proc.Natl.Acad.Sci.U.S.A.2000,97,262–267. (11)Guyon,I.;Weston,S.B.;Vapnik,V.Machine Learning2002,46,389–422.
Anal.Chem.2008,80,7562–7570
10.1021/ac800954c CCC:$40.75 2008American Chemical Society 7562Analytical Chemistry,Vol.80,No.19,October1,2008
Published on Web09/04/2008
protein homologies.12SVMs are a relatively new machine learning supervised technique for classi?cation of data.The foundation for SVMs was laid in the year1982by Vapnik13and formally proposed in1992.14SVMs have seen great success in the task of classifying handwritten digit recognition of US postal codes,15as they are quite robust when it comes to handling noisy data,and are generally not susceptible to the presence of outliers.The basic principle of SVMs,which are essentially a binary support vector classi?er,is as follows:given a set of data with two classes,an optimal linear classi?er is constructed in the form of a hyperplane which has the maximum margin(the simultaneous minimization of the empirical classi?cation error and maximization of the geometric margin).The margin of a classi?er is de?ned as the width up to which the boundary can be extended on both sides before it hits one of the data points.The points onto which this margin hits are called the“support vectors”.In the case of data sets that are not linearly separable,the original data is mapped into higher dimensional feature space and a linear classi?er is constructed in this feature space(this is known as the“kernel trick”)which is equivalent to constructing a nonlinear classi?er in the original input space.This mapping is implicitly given by the kernel function.
Consider a training data set x i∈R n,i)1,...,m where each of the x i fall in one of the two categories y i∈{-1,1}.For this case, SVMs determine the hyperplane,the parameters of which are given by(w,b)as obtained from the solution of14the following convex optimization problem:
min w,b, 1
2
w t w+c∑
i)1
m
i
(1)
subject to y
i (w t x
i
+b)g1- i
i
g0
Here c is the regularization parameter,which is a trade off between the training accuracy and the prediction term,and is a measure of the number of misclassi?cations and known as the slack variable.The inclusion of the regularization term reduces the problem of over?tting which is the fundamental principle of soft margin classi?ers.16
The optimization problem in eq1can be solved.The expres-sion for the maximum margin classi?er involves the term?x i,x j?(inner product).Letφbe the term that maps the data to the feature space.The expression for the maximum margin classi?er will involve?φi,φj?.Next,we de?ne a kernel function K such that K(x i,x j))φ(x i)·φ(x j).This kernel function is required in the training phase to solve the optimization problem,thus getting rid of the need for explicit knowledge of the mapping functionφ.The commonly used kernel functions are as follows:(1)linear kernel, x i t x j;(2)polynomial kernel,(γx i t x j+constant)d,γ>0,d is the degree of the polynomial;(3)radial basis function kernel, e-γ||x i-x j||2;(4)sigmoidal kernel,tanh(γx i t xj+constant).
The appropriate parameters for different kernels are chosen by performing a grid search.For a radial basis function(rbf) kernel,the values ofγand the regularization parameter c are varied and for a polynomial kernelγ,c,and d are varied,while for the linear kernel only the value of c is varied.The usage of a sigmoidal kernel is generally avoided as it does not obey Mercer’s theorem(any kernel function should satisfy this theorem)for all values ofγ.A detailed account on SVMs may be found in the literature.17
SVMs were originally developed for handling only binary class problems.However,most of the practical real world problems are multiclass,and solving these are much tougher than solving binary class problems.Several algorithms have been proposed for solving multiclass problems.18-21
Once classi?cation is achieved,it is important to identify the discriminating features that cause the categorization to enable an in-depth understanding of the disease mechanisms or the path-ways involved in the case of metabolomic data.This can be achieved through feature selection,which involves identifying the optimum subset of the variables in the data set that gives the best separation or classi?cation accuracy.
Feature selection algorithms in machine learning can be broadly classi?ed under two categories:(1)the?lter approach and(2)the wrapper approach.The former is independent of the actual classi?er algorithm and is mainly done on the basis of a ranking system.Univariate correlation scores such as Fisher scores fall under this category.In the wrapper method,feature selection is done in conjunction with the training phase.A subset of the variables is chosen and the performance of the classi?er is evaluated on this subset.The subset of variables that gives the best classi?er performance is chosen for?nal analysis.Detailed reviews on feature selection methods may be found in the literature.22,23
SVM recursive feature elimination(SVM-RFE)is a wrapper approach that uses the norm of the weights w to rank the variables.Initially,all data is taken and a classi?er is computed. The norm of w is then computed for each of the features and the feature with the smallest norm is eliminated.This process is repeated until all the features are ranked.A more elaborate version of this algorithm is described by Guyon et al.11
Variable importance to projection(VIP)is a measure of the importance of the terms in the model(i.e.,variables in the data), with respect to its correlation to the responses(y).VIP metric values are usually computed from all of the extracted components. VIP is computed as follows:
(12)Jaakkola,T.;Diekhans,M.;Haussler,D.Proceedings of the7th International
Conference on Intelligent Systems for Molecular Biology;AAAI Press: Heidelberg,Germany,1999.
(13)Vapnik,V.Estimation of Dependences Based on Empirical Data;Springer
Verlag:New York,1982.
(14)Boser,B.;Guyon,I.;Vapnik,V.Proc.5th Annu.ACM Workshop Comput.
Learn.Theory1992,144–152.
(15)Bottou,L.;Cortes,C.;Denker,J.;Drucker,H.;Guyon,I.;Jackel,L.;LeCun,
Y.;Muller,U.;Sackinger,E.;Simard,P.;Vapnik,V.Proc.12th IAPR Int.
Conf.Pattern Recognition1994,2,77–82.
(16)Cortes,C.;Vapnik,V.Machine Learn.1995,20,273–297.(17)Burges,C.J.Data Min.Knowledge Discovery1998,2,121–167.
(18)Krebel,U.Pairwise Classi?cation and Support Vector Machines.In Advances
in Kernel Methods-Support Vector Learning;Scholkopf,B.,Burges,C.,Smola,
A.,Eds.;MIT Press:Cambridge,MA,1999.
(19)Lee,Y.;Lin,Y.;Wahba.J.Am.Stat.Assoc.2004,99,67–81.
(20)Crammer,K.;Singer,Y.J.Machine Learn.Res.2001,2,265–292.
(21)Weston,J.;Watkins,C.Multi-class support vector machines,1998.
(22)Kohavi,R.;John,G.Artif.Intell.1997,97,273–324.
(23)Chen,Y.;Lin,https://www.doczj.com/doc/0016115356.html,bining SVMs with various feature selection strategies.
In Feature Extraction,Foundations and Applications;Guyon,I.,Gunn,S., Nikravesh,M.,Zadeh,L.,Eds.;Physica-Verlag,Springer:NTU,Taiwan, 2006.
7563 Analytical Chemistry,Vol.80,No.19,October1,2008
VIP
kp
) ∑i)1k(w ip2(SSY i-1-SSY i))p SSY0-SSY k
where k is the number of extracted components,p is the dimension of the system,SSY i is the cumulative sum of squares explained by the i th component,and w ip is the squared PLS weight of that term.
In this work,we will show how SVMs can be used with metabolomic data(using both spectral binning and targeted pro?ling),and how it performs better than the techniques that are being currently used(PCA and PLS-DA).We will also compare the performance of a wrapper approach(SVM-recursive feature elimination)with the VIP(variable importance to projection) approach of PLS-DA.6
METHODS
Populations:Written informed consent was obtained from each subject before entering this study,and the institutional ethics committees approved the protocols outlined below.
Healthy volunteers:For comparison of gender,30male and thirty female subjects aged19-69years self-identi?ed as healthy volunteered to participate in this study as previously described.24 Multiple urine samples were taken either as a?rst void or at 5p.m.,over four nonconsecutive days from each subject to obtain a total of352samples,of which194were female and158were male.It should be noted here that the genders were matched in terms of age and other metadata parameters such as diet and alcohol consumption.Details of age matching statistical analysis are included in the Supporting Information?le.Hence,these metadata parameters do not have any impact on the classi?cation between the two groups.For comparison between healthy volunteers and the pneumonia group,one urine sample obtained from each of59subjects was used.
Pneumonia volunteers:A total of59patients infected with Streptococcus pneumoniae as determined through cultures of blood,sputum,cerebrospinal?uid,bronchoalveolar lavage,en-dotracheal tube secretions,ascites,or a combination of any of these,constituted the pneumococcal https://www.doczj.com/doc/0016115356.html,rmation regarding the metadata parameters for the pneumonia volunteers was not available.However,it is likely that most of the patients were on analgesics and antibiotics.
Sample handling:Upon acquisition of urine samples,sodium azide was added to a?nal concentration of approximately0.02% to prevent bacterial growth.Urine was placed in the freezer and stored at-80°C until ready for preparation and data acquisition. Urine samples were prepared for NMR analysis as described previously.24
NMR Spectral Data.NMR spectra were acquired using standard techniques on a Varian600MHz spectrometer.24 Standard preprocessing steps such as phasing and baseline correction were applied to each spectrum.Standard binning was performed with a bin width of0.025ppm between0.2and10.0 ppm,with the region between 4.65and 4.86, 5.5-6.0,and 7.33-7.39/8.34-8.44ppm excluded to account for water,urea, and imidazole,respectively,and integrals were measured.Each integral region was then normalized to the total area.Standard binning reduced the dimension of the system to360bins (variables).
Concentration Data.Concentration data was obtained by quanti?cation of known urinary components(metabolites)in the NMR spectra by?tting the NMR peaks using Chenomx NMRSuite 4.0(Chenomx Inc.,Edmonton Canada).Peak areas re?ect the concentration of each particular compound.Calibration is done a priori,for all compounds in the database with respect to a known concentration of the reference compound(DSS).Unlike spectral data,this type of data is of a lower dimension,since the number of metabolites that are quanti?ed is of the order of100.The precision and accuracy of the metabolite concentrations thus determined as well as the repeatability of sample preparation has already been demonstrated in an earlier work.24The main advantage of doing the metabolomic analysis on the concentration data is that we are able to preselect the metabolites that are important and remove unwanted metabolites(metabolites related to analgesics and antibiotics were not included in the controls pneumonia case study).However,at the same time it should be noted that this work does not focus on giving a diagnostic for detecting a particular phenotype.The main focus of this work is to introduce support vector machines as an ef?cient tool for doing metabolomic analysis.
An inherent assumption of both analyses was that none of the samples were outliers.
Data Analysis.Multivariate data analysis(PLS-DA,and SVMs) were performed on binned NMR spectral data that had been mean centered and normalized with respect to the total area using MATLAB.The Matlab codes that were used to generate the results are provided in the Supporting Information as a down-loadable zip archive.
Targeted pro?ling data was obtained after measuring the concentration of metabolites in the samples.Mean centering and unit variance scaling was applied for PLS-DA and SVM analysis.
PLS-toolbox4.1of Eigenvector Research was used for PLS-DA.PLS-DA can be considered to be the regression extension of PCA.Details of this algorithm have been described elsewhere.6 Computation of Classi?cation Accuracy.Classi?cation ac-curacy was computed based on k-fold cross validation by dividing each data set comparison into k random subsets.For each computation,one of the k subsets was kept aside as the test set and the remaining k-1subsets were pooled together to form the training set.Each of the k subsets was therefore treated once as the test set,which generated k accuracy values.Classi?cation accuracy was calculated as the mean of k accuracy rates.In order to reduce the variance of the accuracy rate estimate,the process was repeated many times and the mean value was taken.Two types of accuracy rates were reported:
(1)A4-fold cross validation,calculated by randomly dividing the data set into four subsets and calculating the mean of the four accuracy values.This process was repeated100times and the mean accuracy was computed.The estimate calculated in general has a smaller variance than leave-one-out cross validation(LOOCV) but is pessimistically biased.
(2)Leave-one-out cross validation(LOOCV).This is the same as n-fold cross validation,where n is the total number of samples. This method produces a good estimate of the true accuracy rate
(24)Slupsky,C.;Rankin,K.;Wagner,J.;Fu,H.;Chang,D.;Weljie,A.;Saude,
E.;Lix,B.;Adamko,D.;Shah,S.;Greiner,R.;Sykes,B.;Marrie,T.Anal.
Chem.2007,79,6995–7004.
7564Analytical Chemistry,Vol.80,No.19,October1,2008
with less bias.When comparing classi?cation accuracy rates obtained from two different methods,the LOOCV method is preferred since it does not involve formation of random subsets as the k-fold cross-validation approach.Thus,a de?nite measure of the accuracy estimate is obtained.
Kernel parameters were chosen using a3-fold cross-validation approach by dividing the training data set randomly into three subsets.The classi?er was trained on either of the two subsets and was tested on the third subset.The set of parameters which gave the best cross-validation accuracy were chosen for further analysis using LIBSVM interface in MATLAB.25The kernel parameters which corresponds to the maximum cross-validation accuracy is chosen to perform the actual classi?cation.
The same statistics for mean centering and scaling to unit variance was used for both the training and test data.In other words,the mean and the standard deviation of the training data is used to mean-center and scale the test data,since it is assumed that the test data set is fully blinded and thus their statistics is unknown.
Feature Selection.Feature selection using SVM-RFE was done to identify the signi?cant biomarkers causing the classi?ca-tion.For selection of the k best features,feature selection was done on100randomly partitioned sets of training and test data, selecting k features each time generating100subsets,each containing k features.The frequency of all original features appearing in each of the100subsets was calculated and the top k features with the maximum frequency was selected.Further details regarding the kernel and the parameters used for SVM-RFE are included in the Supporting Information?le.The SVM-RFE algorithm was implemented using the Spider feature selection objects.29
Test of Model Over?t.Once the model was trained,it was used to test whether the data was over?t.One way to do this is to have a validation set with known class labels and check whether it gives a comparable accuracy rate to that of the training data. Another method is an R2/Q2validation plot that helps assess the risk that the current model is spurious,i.e.,the model just?ts the training set well but does not predict Y well for new observations.The R2value is the percent variation of the training set that can be explained by the model.The Q2value is a cross-validated measure of R2.The formula for computing the R2/Q2 values are as follows:
R2)1-∑
i)1
n
(y-y?)2∑
i)1
n
(y-y?)2
Q2)1-∑
i)1
n
(y-y?(i))2∑
i)1
n
(y-y?(i))2
In the above equation,n represents total number of samples,y is actual class,y is predicted class when all the samples are used for model building,y(i)is predicted class when all the samples except the i th sample are used for model building.This validation compares the goodness of?t of the original model with the goodness of?t of several models based on data where the order of the Y-observations have been randomly permuted,while the X-matrix has been kept intact.
The criteria for model validity are as follows:(1)All the Q2 values on the permuted data set are lower than the Q2value on the actual data set.If this is not the case,it means that the model is capable of?tting well any kind of data set which is over?tting.
(2)The regression line(line joining the actual Q2point to the centroid of the cluster of permuted Q2values)has a negative value of intercept on the y-axis.
RESULTS
To demonstrate the usefulness of SVMs in classifying both binning and targeted pro?ling metabolomic data,two case studies are presented here:(1)a case where nearly complete distinction may be seen(healthy versus pneumonia);and(2)a case where distinction is not as easily seen(male versus female).
To determine how well SVM and PLS-DA perform to discrimi-nate the data,59healthy subjects were compared to59subjects with pneumococcal pneumonia with a class vector Y of length 118with-1and1representing healthy and pneumonia,respec-tively.The variables consisted of either360bins,or82metabolites derived from the urinary NMR spectral pro?les.
Figure1shows a visual comparison of the controls-pneumonia spectral and concentration data separated using PLS-DA and SVMs.Traditionally,most classi?cation methods for biological data involve visualization of the observations,and where they lie with respect to one another.This is because the majority of methods involve projection of the data set to a lower dimensional space(usually2D or3D)as in PCA or PLS.However,in SVM, separation of the data set involves mapping to a higher dimen-sional feature space.Hence visualization of the resulting data is dif?cult.This problem may be overcome by plotting the decision function of the sample,where the sign of the decision function determines the class of the test sample.Here,we plotted the decision function against a dummy variable.In this case,a random Gaussian variable was chosen as the dummy variable.Figure1a is a plot of the separation achieved between control and pneu-monia data based on spectral binning using SVM,and Figure1b is a plot of the separation achieved with PLS-DA.Table1shows the classi?cation accuracy rates for both methods using4-fold cross validation as well as LOOCV.While the visual separation appears slightly better for the SVM,the classi?cation accuracy rates are similar,with SVMs being marginally better.
Classi?cation using targeted pro?ling data produces slightly different results.Figure1c shows a visual comparison of the raw targeted pro?ling data using SVMs,and Figure1d shows the visual comparison using PLS-DA.Table1shows the classi?cation accuracy rates for both methods.It is clear that visual separation on the raw data by PLS-DA is not as good as the separation using SVMs.Moreover,the classi?cation accuracy is poor compared to SVMs for both4-fold cross validation and LOOCV.
To determine if SVMs perform better in a case where separation is not as apparent,we analyzed the urine pro?les of
(25)Chang,C.;Lin,C.Software available at https://www.doczj.com/doc/0016115356.html,.tw/cjlin/
libsvm,2001.
7565 Analytical Chemistry,Vol.80,No.19,October1,2008
352individuals to see if we could separate male from female.In this case,healthy human urinary pro?les were investigated to assess the differences associated with gender.The data set used here is same as previously published 24and consists of processed NMR spectra from 158male and 194female samples.Standard binning was done on the spectra to reduce the dimension of the system to 360.A class vector Y of length 352with -1and 1representing males and females,respectively,was assigned.Metabolite concentration data was obtained by quanti?cation of 72known urinary components (metabolites)in the NMR spectrum.The concentration data consisted of 352samples with 158males and 194females.Each observation has associated variables of the concentration value (measured in μmolar)of 72metabolites.This class vector Y had a length of 352with -1and 1representing males and females respectively.
Table 1shows classi?cation accuracies for gender separation (both spectral and concentration data).As with the normal-pneumonia case,SVM appears to perform marginally better than PLS-DA for the concentration data,however for spectral data PLS-DA appeared slightly better.Figure 2shows the visualization plot for both binned and concentration data.The visualization plot for SVM concentration data (Figure 2a)is better than that of PLS-DA (Figure 2b);however,that of SVM spectral data (Figure 2c)is only as good that of PLS-DA spectral data (Figure 2d).It should be noted that the actual PLS-DA model for this spectral gender data was built using 10components;however,in the visualization plot only two components were used.This explains the very poor performance of PLS-DA visualization.On the other hand,SVM visualization plots are independent of the dimension of the data set or the kernel used to build the model.
In order to show that none of the models generated over?t the data,the R 2-Q 2?gures were plotted along with the separation plots for all the data sets as in Figures 1and 2.This plot can be explained as follows:the vertical axis represents the R 2/Q 2
value
Figure https://www.doczj.com/doc/0016115356.html,parison of PLS-DA and SVM multivariate analysis techniques on controls-pneumonia data.Control data (solid circles,n )59);pneumococcal data (hollow circles,n )59).(a)SVM spectral data;(b)PLS-DA spectral data;(c)SVM concentration data;(d)PLS-DA concentration data.
Table 1.Case Studies:Classi?cation Accuracy Rates
case study
method parameters 4-fold CV (mean %)
LOOCV (mean %)
case study 1a:control-pneumonia spectral data SVM rbf γ)2-13,c )21195.6596.61PLS-DA k )394.4495.76case study 1b:control-pneumonia concentration data SVM rbf γ)2-9,c )21093.0295.76PLS-DA k )291.8092.37case study 2a:male -female spectral data SVM rbf γ)2-9,c )21183.6586.93PLS-DA k )1084.4487.50case study 2b:male -female concentration data
SVM rbf γ)2-6.6,c )211
90.9092.04PLS-DA
k )5
90.13
91.19
7566
Analytical Chemistry,Vol.80,No.19,October 1,2008
for the original model (far to the right)and of the cluster of Y -permuted models further to the left.The horizontal axis shows the correlation between the permuted Y -vectors and the original Y -vector.The original Y has the correlation 1.0with itself,de?ning the high point on the horizontal axis.In all the plots (Figures 1and 2),the R 2and Q 2values for the permuted models are lower than that of the original model.The Q 2regression line has a negative intercept in all the cases.Hence it can be concluded that neither of the models over?t the data.Moreover,the R 2/Q 2values and the R 2/Q 2regression line slopes of SVM-RFE are higher than that of the PLS-DA based VIP method.
Orthogonal partial least-squares-discriminant analysis (OPLS-DA)is a major development of the PLS-DA technique which was proposed to handle the class orthogonal variation in the data matrix.27,28OPLS-DA augments the classi?cation performance only in cases where individual classes exhibit divergence in within-class variation.In terms of regression prediction results,OPLS-DA is nearly equivalent to PLS-DA.27However,the main advantage of the OPLS-DA method lies in the fact that there is more transparency in the generated model.This is because OPLS-DA can separate the predictive from nonpredictive (orthogonal)variation.In order to verify these statements,O-PLSDA algorithm
was coded using MATLAB and was applied to the male -female concentration data set.Summary of O-PLSDA results have been included in the Supporting Information ?le.It was observed that the classi?cation performance was very similar to that of PLS-DA.The top 10features (metabolites)that contributed the most to the separation in PLS-DA and OPLS-DA was also compared.It was observed that the list of features and their ordering were also very similar.Hence,it can be concluded that the OPLS-DA method performance was very similar to that of PLS-DA.This similarity in performance may be because of the absence of divergence in within-class variation between the two classes.
Feature Selection.In order to extract meaningful information about the features (bins or metabolites)giving rise to the separation,the PLS-DA based VIP method and the SVM-RFE method were implemented and compared.The goal was to ?nd the optimum number of features that would result in the best classi?cation.For the PLS-DA based feature selection,this was done as follows:based on the PLS loadings,the VIP values were calculated for each feature.These features were then arranged in decreasing order of VIP (or decreasing order of importance).Starting from the most important feature,the features were included sequentially and the corresponding accuracy rate for that set of features was calculated.In the case of SVM-RFE,the algorithm is a backward elimination procedure,where the entire data set is used at the beginning.Depending on the weights of the individual features,each feature was eliminated sequentially and the corresponding accuracy rate for that set of features was calculated.
(26)Ambroise,C.;McLachlan,G.Proc.Natl.Acad.Sci.U.S.A.2002,99,6562–
6566.
(27)Bylesjo,M.;Rantalainen,M.;Cloarec,O.;Nicholson,J.;Holmes,E.;Trygg,
J.J.Chemom.2006,20,341–351.
(28)Trygg,J.;Wold,S.J.Chemom.2002,16,119–128.
(29)Weston,J.;Elisseeff,A.;BakIr,G.;Sinz,F.Software available at http://
www.kyb.tuebingen.mpg.de/bs/people/spider/main.html,
2006.Figure https://www.doczj.com/doc/0016115356.html,parison of PLS-DA and SVM multivariate analysis techniques on gender separation data.Male (solid circles,n )158);female data (hollow circles,n )194).(a)SVM spectral data;(b)PLS-DA spectral data;(c)SVM concentration data;(d)PLS-DA concentration data.
7567
Analytical Chemistry,Vol.80,No.19,October 1,2008
For the case of healthy versus pneumococcal pneumonia spectral data,classi?cation accuracy rate was plotted against the corresponding number of features (Figure 3a).From Figure 3a,it can be seen that the feature selection curve initially rises steeply,attains a maximum,then becomes stable,and decreases slightly at the end.This is because the ?rst few features are most important in terms of classi?cation.The point at which the curve reaches a maximum corresponds to the optimum number of metabolites.Further addition of features to the classi?er slightly degrades the performance because of the addition of redundant “noise”to the system.SVM-RFE signi?cantly outperforms PLS-DA as SVM-RFE gives a classi?cation accuracy rate of 99.4%for as low as 30features,whereas the maximum accuracy rate achieved by PLS-DA is 98.4%requiring 50features.A few of the signi?cant bins (ppm regions)identi?ed by both the methods are listed in Table 2in the order of importance.Table 2compares
the signi?cant features (metabolites for concentration data and bin regions for spectral data)identi?ed by PLS-DA and SVM-RFE for each of the four case studies.For the case of healthy versus pneumococcal pneumonia concentration data (Figure 3b),it can be observed that SVM-RFE signi?cantly outperforms PLS-DA.SVM-RFE gives a classi?cation accuracy rate of 98.5%for as low as eight features,whereas the maximum accuracy rate achieved by PLS-DA is 95%.A few of the signi?cant metabolites for this concentration data identi?ed by both methods are listed in Table 2with their rank order of importance included in parentheses.However,both methods perform better when compared to the case where no feature selection is done.Hence it can be concluded that feature selection reduces the dimension of the system as well as increases the separation between the two classes.
In the case of healthy versus pneumococcal pneumonia,very good separation was achieved.However,in the case of data
that
Figure https://www.doczj.com/doc/0016115356.html,parison of feature selection results using SVM-RFE (solid line)and PLS-DA VIP (dashed line):(a)controls-pneumonia spectral data;(b)controls-pneumonia concentration data;(c)gender spectral data;(d)gender concentration data.Table 2.Signi?cant Features (CP )Controls Pneumonia Data Set,MF )Gender Data Set)
CP
MF
SVM-RFE
PLS-DA
SVM-RFE PLS-DA spectral (bin regions)
3.025-3.05 3.025-3.05
4.025-4.05 3.9-3.9252.675-2.70 2.675-2.70 6.95-6.975 2.525-2.554.025-4.05 1.25-1.2757.15-7.175 2.675-2.79.1-9.125 2.525-2.553-3.025 2.5-2.5252.7-2.725 2.5-2.525 3.9-3.9253-3.025concentration (metabolites)
citrate (1)
citrate (1)
carnitine (1)carnitine (1)O -acetylcarnitine (3)O -acetylcarnitine (2)citrate (2)citrate (4)trigonelline (5)trigonelline (6)succinate (6)succinate (5)fumarate (4)
fumarate (3)
creatinine (8)creatinine (-)1-methylnicotinamide (2)
1-methylnicotinamide (7)
fumarate (9)
fumarate (11)
7568
Analytical Chemistry,Vol.80,No.19,October 1,2008
is more overlapped,the differences in the performance and feature selection are more pronounced.Figure3c shows that SVM-RFE outperforms PLS-DA when classifying gender spectral data.For the case when all the features are used for classi?cation,the accuracy rate is84%.This low accuracy rate is likely due to the presence of irrelevant features or“noise”that can potentially distort the classi?cation.As these irrelevant features are removed sequentially,the accuracy improves until it attains the maximum of89%at40bins.Further removal of features causes a reduction in the accuracy rate due to a lack of suf?cient information for the model to classify the data.In the case of PLS-DA,the maximum accuracy rate achieved is approximately86%and requires almost 70features to attain this maximum.
For the gender spectral data,some of the signi?cant bin
regions identi?ed by both PLS-DA and SVM-RFE have been listed in Table2.Bin regions3.9-3.925ppm(creatine),3.0-3.025,and a few others were found to be important by both SVM-RFE and PLS-DA based methods.However,their rank order of importance was found to be different.The metabolites to which these bin regions correspond can be identi?ed from the literature.Never-theless,for some of the bin regions,this mapping is not unique, due to severe overlapping of these metabolite bin regions.Several metabolites such as creatinine,citrate,carnitine,and citrate were found to be signi?cant for classi?cation in both concentration and spectral data.Further investigation of these bin regions may lead to important?ndings of potentially signi?cant biomarkers which may be missed in the concentration data.
For the case of gender separation concentration data,Figure 3d shows that both the methods give a similar maximum classi?cation accuracy rate of91%.Nevertheless,on careful observation,it can be seen that SVM-RFE achieves this maximum at30features whereas PLS-DA requires almost40features to achieve this maximum.Moreover,when the number of features selected is low(initial phase of the graph),SVM-RFE performs better than its counterpart.A few of the signi?cant metabolites have been listed in Table2with their rank order of importance included in parentheses.Further study has to be done to investigate the biological signi?cance of these metabolites to get a better understanding as to why these metabolites were classi?ed as important for the classi?cation.
DISCUSSION
The tabulated results suggest that SVM performs marginally better than PLS-DA in terms of the leave-one-out classi?cation accuracy rate on the entire data set.Cross validation for calculating accuracy rates is preferred over computing the same using a single blinded data set since the latter case is a biased approach.The reason for this is because it cannot be guaranteed that the test set has been sampled from the same population distribution as the training set.Moreover,if a different blinded set is randomly chosen from the original data set,the same accuracy rate may not be obtained every time.Thus,comparison of classi?cation accuracy rates based on a cross-validation approach is preferred over comparison to a single blinded data set.In general,the LOOCV method generates a better estimate than the4-fold cross validation method because the latter gives a pessimistically biased estimate(the estimate is always lower than the true value).
In terms of classi?cation on the full data set,SVM performs only marginally better than PLS-DA.However,the main advantage of SVM over PLS-DA is that,after doing feature selection,SVM generates a better predictive model than PLS-DA from the subset of features describing the phenotype.The SVM-RFE method gives a much better classi?cation accuracy rate with fewer features when compared to the PLS-DA based VIP method.This is important when a number of markers are being used for disease diagnostics. The summary of feature selection results for the two different case studies is tabulated in Table3.Table2shows that the order of importance of metabolites ranked by SVM and PLS-DA is very different.These metabolite subsets are crucial for the development of a predictive model for disease diagnostics.With better clas-si?cation accuracy using the least number of features possible means SVMs are more ef?cient than PLS-DA for diagnostic purposes.
The other advantage of SVMs over PLS-DA is that the latter is essentially a linear classi?er whereas the former can be used for nonlinear cases with the help of appropriate kernel functions. Moreover,SVMs have an excellent generalization performance because of the usage of the regularization parameter c.Further-more,in PLS-DA there is an inherent assumption that the predicted class(y)values of each class are normally distributed. This assumption is made to calculate the prediction probabilities of different classes of data and the classi?cation threshold according to the Bayesian method.A normal distribution is?tted to the predicted class(y)values,and the value at which the probability of a sample belonging to either of the class(in a binary class problem)is equal is chosen as the classi?cation threshold. In order to verify this normality assumption,the histogram of of predicted class values for the controls class in case study1is plotted as in Figure4.It can be clearly seen that the assumption of normality is not valid.In the case of SVMs,there are no such assumptions on the data set distribution.
Clearly,further study has to be done to investigate the biological signi?cance of these metabolites.However,it is of interest that the ordering of the metabolites is different in both PLS-DA and SVMs.Moreover,the PLS-DA based method may have missed one or more important distinguishing factors.One such factor is creatinine.It has been shown24that creatinine is related to body mass,and since males in general have a higher body mass than females,it would seem appropriate to assume the creatinine is a key differentiator between males and females. PLS-DA did not indicate creatinine as an important indicator whereas SVMs did.A detailed analysis of all the signi?cant metabolites identi?ed in either of the methods is beyond the scope Table3.Case Studies:Summary of Feature Selection Results
case study method
optimum no.
of features
accuracy
rate case study1a:SVM3099.4 control-pneumoniae
spectral data PLS-DA5098.4 case study1b:SVM898 control-pneumoniae
concentration data PLS-DA795 case study2a:SVM4089 male-female spectral
data PLS-DA7086 case study2b:SVM3091 male-female
concentration data PLS-DA4591
7569 Analytical Chemistry,Vol.80,No.19,October1,2008
of this work and will be described elsewhere.24It should be noted that this work does not focus on giving a diagnostic for detecting a particular phenotype.The main focus of this work is to introduce support vector machines as an ef?cient tool for doing metabolomic analysis.
CONCLUSIONS
In this work,we have shown that support vector machines (SVMs)are an effective tool to analyze metabolomic data.Its performance and ability to ?nd fewer features is better than the traditionally used PLS-DA technique.Moreover,SVMs are able to generate better predictive models after doing the appropriate feature selection.This is done to identify the potential biomarkers or the signi?cant distinguishing features which causes the classi?cation.With fewer number of features selected,SVM-RFE gives a much better classi?cation accuracy rate when compared to the PLS-DA based VIP method.The superior performance of
SVM-RFE over PLS-DA based method was demonstrated with the help of two case studies.Unlike PLS-DA,SVM does not require any implicit assumptions on the distribution of the data set.However,in the case of metabolomic data sets where there is a signi?cant divergence in the within-class variation between the two classes,OPLS-DA might perform better than PLS-DA.Further work needs to be done to compare the performances of SVM and OPLS-DA in such cases.
ACKNOWLEDGMENT
The authors thank Ms.Shana Regush for preparing and acquiring the NMR data and the Canadian National High Field NMR Centre (NANUC)for use of the facilities for collection of the NMR data.The authors also want to express their gratitude to Chenomx Inc.and Varian Inc.for their support.The authors also acknowledge Allison McGeer for providing the S.pneumoniae samples,Kathryn Rankin and Hao Fu for analyzing the same,and NSERC for providing ?nancial support.Part of this work was supported by an establishment grant from Alberta Heritage Foundation for Medical Research,and from The Lung Association of Alberta and the Northwest Territories.This work was also supported by funds from Western Economic Development,and Alberta Advanced Education and Technology.The authors wish to thank all volunteers and patients for providing urine samples for this study.
SUPPORTING INFORMATION AVAILABLE
Details regarding the kernel and the parameters used for SVM-RFE;the Matlab codes that were used to generate the results;summary of O-PLSDA results and details of age matching statistical analysis.This material is available free of charge via the Internet at https://www.doczj.com/doc/0016115356.html,.
Received for review June 4,2008.Accepted July 22,2008.
AC800954C
Figure 4.Histogram of PLS-DA y predicted for normal class in normal-pneumonia case study.PLS-DA model is built on the controls-pneumonia data and the predicted controls class (y )values is computed and its histogram is plotted.
7570Analytical Chemistry,Vol.80,No.19,October 1,2008
例谈高考地理试题数据分析的方法和技巧 地理数据是地理事物和现象空间位置、属性特征及其动态变化等的数量化表示。通过地理数据的分析，可以解析其所表达的地理事物的分布、特征及其运动变化的地理过程，进而了解自然和人文地理环境特征以及人类活动与地理环 境之间的关系。全国各地高考试题近年来出现了不同类型的地理数据分析类试题，其解析方法和技巧各不相同。 一、比较法 比较法是高考试题中地理数据分析的常用方法，包括纵比法、横比法和联系比较法等。纵比法是将不同历史阶段的地理数据加以比较，借以揭示地理现象在历史发展过程各阶段的共性与个性。横比法是将同一属性的不同地理事物加以比较，找出其在同一发展阶段在地理特征、发展趋势等方面的差异。联系比较法是联系相关地理事物进行比较，以利于探求地理规律，解决地理问题。 例题1：图1示意某城市20世纪80年代和90年代平均人口年变化率，当前，该城市总人口约1300万。据此完成（1）～（2）题。 （1）20世纪90年代和80年代相比，该城市 A.总人口增长速度加快 B.总人口减少 C.人口自然增长率降低
D.人口净迁入量减少 （2）该城市所在的国家可能是 A.美国 B.日本 C.俄罗斯 D.德国 解析：该题涉及了人口地理学的相关地理数据，主要有人口的自然增长率、迁移率、总人口增长率。第（1）题的解题方法是典型的纵比法，可以将“1981～1990”和“1991～2000”两个不同历史阶段的同类地理数据进行比较，排除A、C，由于总人口增长率一直为正值，所以B选项“总人口减少”是错误的，所以选D。 技巧：纵比法主要用于同一地理实体不同历史阶段气候资料、水文信息、自然资源、人口数量、农业分布、工业产值等的比较，可以揭示同一地理事物属性特征的时间变化过程，利用纵比法进行比较时，一是要注意比较实体和属性数据的同一性，必须是同一地理实体不同时期的同类属性数据的比较。二是要注意将属性数据变化值与时间尺度联系起来分析地理实体的变化特征，不能夸大或缩小地理实体特征的变化幅度。横比法主要用于不同地理实体间同一历史阶段属性数据的比较，可以是国家间的作物面积的比较，可以是河流间水文特征的比较，也可以是功能区间属性的比较等等。利用横比法进行比较时，一是要注意所比较的地理实体间的同质性和层次性，例如，不能将城市功能区与农业规划区进行比较，也不能将国家的工业产值与城市的工业产值进行比
经营数据分析报告一、确定分析目标分析目标主要包括以下三个方面分析目的。分析范围。分析时间。如下图所示，分析目标除了主要包括三个方面外，还有备注一栏，这里备注的是计算周期问题。强调一点，我们做运营数据分析的时候通常都会拿更新前和更新后的数据进行比较，因此我们的设定的分析周期一般都会跟着游戏实际的更新情况走。二、分析综述分析综述主要包括两方面的内容1、上周本周充值数据对比充值总额充值人数服务器数服务器平均充值服务器平均充值人数针对上述内容进行差额对比以及增减率对比，如游戏有特殊要求，可以适当增加其它数据内容。2、上周本周更新内容对比主要陈列两周内分别更新的活动内容或一些重大调整。三、一周运营数据分析1、本周收入概况日均充值金额，环比上周日均充值金额用户值，环比上周值简述与上周或之前的充值情况的比较，如上升还是下降、影响充值的较大的因素。2、新用户概况新用户就是新进游戏的玩家，这里主要介绍这些新玩家的动态数据，一般以两个月为总时长进行陈列比较，具体周期数据仍以周为单位。新用户数据主要包括安装下载数、创建角色数、安装→角色转化率、付费人数、创建角色→付费转化率、值、次日留存、三日留存、七日留存等，可根据游戏实际情况进行添加。3、活跃用户概况活跃用户概况主要包括三部分内容日均在线人数，环比上周实时在线人数，提升下降百分比日均付费用户登陆人数，环比上周付费登陆数，提升
下降百分比日均活跃玩家数，环比日均活跃玩家数，提升下降百分比4、道具消费概况道具方面的消费概况主要包括产出活动类别道具分类单类道具消费元宝，消费占比，环比上周日均消费元宝，总消费元宝，环比上周下降上升简述活动效果较好较差的道具分类5、当前元宝库存当前元宝库存是指玩家充了元宝还没花出去的存量，以及游戏中额外获得的元宝存量。例如，我充了1000块，拿了1元宝，花了8，我造成的存量是2，当平台各服的元宝存量不断上涨，就代表消费点不够了，要不补新消费系统，要不上消费类的运营活动。6、重点商业活动付费玩家参与情况活动参与情况主要考虑以下几点付费群体类别，活跃付费玩家数付费玩家的参与比例付费玩家在活动中消费的元宝数付费玩家在活动中消费的元宝占周消费元宝总数的比例付费玩家的人均消费元宝数根据活动的这些付费玩家的相关数据，判断该活动产生的效益以及玩家的接受程度。如果数据不佳，则代表该活动不行，需深究其存在的问题，看看问题是出现在活动难度、活动的奖励不吸引、还是活动本身的可玩性太差。根据分析的原因在下次更新活动时判断是需要进行调整玩法设定还是替换成新活动。另外，同一时期可能会推出多个活动，在进行单个活动数据分析时，也要横向比较各个活动的效果，对于下次运营其它产品，有个经验借鉴。注付费玩家数活动期间登陆过游戏的玩家数;消费占比=活动道具总消费元宝当周总消费元宝四、游戏运营数据总分析在简单分析完一
数据分析中常用得10种图表 1折线图 折线图可以显示随时间(根据常用比例设置)而变化得连续数据,因此非常适用于显示在相等时间间隔下数据得趋势。 表1家用电器前半年销售量 月份冰箱电视电脑平均销售量合计 1月68 45 139 84 252 2月33 66 166 88 265 3月43 79 160 94 282 4月61 18 115 65 194 5月29 19 78 42 126 6月22 49 118 63 189 图1 数点折线图 图2堆积折线图
图3百分比堆积折线图 2柱型图 柱状图主要用来表示各组数据之间得差别。主要有二维柱形图、三维柱形图、圆柱图、圆锥图与棱锥图。 图4二维圆柱图 3堆积柱形图 堆积柱形图不仅可以显示同类别中每种数据得大小还可以显示总量得大小。 图5堆积柱形图
图6百分比堆积柱形图 百分比堆积柱形图主要用于比较类别柱上每个数值占总数得百分比,该图得目得就是强调每个数据系列得比例。 4线-柱图 图7线-柱图 这种类型得图不仅可以显示出同类别得比较,更可以显示出平均销售量得趋势情况。 5两轴线-柱图 月份工资收 入(元) 其她收入 (元) 工资占其她收入得百分 比 1月5850 12000 48、75% 2月5840 15000 38、93% 3月4450 20000 22、25%
4月6500 10000 65、00% 5月5200 18000 28、89% 6月5500 30000 18、33% 图8两轴线-柱图 操作步骤:01 绘制成一样得柱形图,如下表所示: 图1 操作步骤02: 左键单击要更改得数据,划红线部分所示,单击右键选择【设置数据系列格式】,打开盖对话框,将【系列选项】中得【系统绘制在】更改为“次坐标轴”,得到图4得展示结果。
实验空间分析基本操作 一、实验目的 1. 了解基于矢量数据和栅格数据基本空间分析的原理和操作。 2. 掌握矢量数据与栅格数据间的相互转换、栅格重分类(Raster Reclassify)、栅格计算－查询符合条件的栅格(Raster Calculator)、面积制表（Tabulate Area）、分区统计(Zonal Statistic)、缓冲区分析(Buffer) 、采样数据的空间内插(Interpolate)、栅格单元统计（Cell Statistic）、邻域统计（Neighborhood）等空间分析基本操作和用途。 3. 为选择合适的空间分析工具求解复杂的实际问题打下基础。 二、实验准备 预备知识： 空间数据及其表达 空间数据（也称地理数据）是地理信息系统的一个主要组成部分。空间数据是指以地球表面空间位置为参照的自然、社会和人文经济景观数据，可以是图形、图像、文字、表格和数字等。它是GIS所表达的现实世界经过模型抽象后的内容，一般通过扫描仪、键盘、光盘或其它通讯系统输入GIS。 在某一尺度下，可以用点、线、面、体来表示各类地理空间要素。 有两种基本方法来表示空间数据：一是栅格表达; 一是矢量表达。两种数据格式间可以进行转换。 空间分析 空间分析是基于地理对象的位置和形态的空间数据的分析技术，其目的在于提取空间信息或者从现有的数据派生出新的数据，是将空间数据转变为信息的过程。 空间分析是地理信息系统的主要特征。空间分析能力（特别是对空间隐含信息的提取和传输能力）是地理信息系统区别与一般信息系统的主要方面，也是评价一个地理信息系统的主要指标。 空间分析赖以进行的基础是地理空间数据库。 空间分析运用的手段包括各种几何的逻辑运算、数理统计分析，代数运算等数学手段。 空间分析可以基于矢量数据或栅格数据进行，具体是情况要根据实际需要确定。 空间分析步骤 根据要进行的空间分析类型的不同，空间分析的步骤会有所不同。通常，所有的空间分析都涉及以下的基本步骤，具体在某个分析中，可以作相应的变化。 空间分析的基本步骤: a)确定问题并建立分析的目标和要满足的条件 b)针对空间问题选择合适的分析工具 c)准备空间操作中要用到的数据。 d)定制一个分析计划然后执行分析操作。 e)显示并评价分析结果
一、企业生产经营的基本情况 （一）企业主营业务范围和附属其他业务，纳入年度会计决算报表合并范围内企业从事业务的行业分布情况；未纳入合并的应明确说明原因；企业人员、职工数量和专业素质的情况；报表编报口径说明。 （二）本年度生产经营情况，包括主要产品的产量、主营业务量、销售量（出口额、进口额）及同比增减量，在所处行业中的地位，如按销售额排列的名次；经营环境变化对企业生产销售（经营）的影响；营业范围的调整情况；新产品、新技术、新工艺开发及投入情况。 （三）开发、在建项目的预期进度及工程竣工决算情况。 （四）经营中出现的问题与困难，以及需要披露的其他业务情况与事项等。 二、利润实现、分配及企业亏损情况 （一）主营业务收入的同比增减额及主要影响因素，包括销售量、销售价格、销售结构变动和新产品销售，以及影响销售量的滞销产品种类、库存数量等。 （二）成本费用变动的主要因素，包括原材料费用、能源费用、工资性支出、借款利率调整对利润增减的影响。 （三）其他业务收入、支出的增减变化，若其收入占主营业务收入10%（含10%）以上的，则应按类别披露有关数据。
（四）同比影响其他收益的主要事项，包括投资收益，特别是长期投资损失的金额及原因；补贴收入各款项来源、金额、以及扣除补贴收入的利润情况；影响营业外收支的主要事项、金额。 （五）利润分配情况。 （六）利润表中的项目，如两个期间的数据变动幅度达30%（含30%）以上，且占报告期利润总额10%（含10%）以上的，应明确说明原因。 （七）会计政策变更的原因及其对利润总额的影响数额，会计估计变更对利润总额的影响数额。 （八）其他。 三、资金增减和周转情况 （一）各项资产所占比重，应收账款、其他应收款、存货、长期投资等变化是否正常，增减原因；长期投资占所有者权益的比率及同比增减情况、原因、购买和处臵子公司及其他营业单位的情况。 （二）资产损失情况，包括待处理财产损益主要内容及其处理情况，按账龄分析三年以上的应收账款和其他应收款未收回原因及坏账处理办法，长期积压商品物资、不良长期投资等产生的原因及影响。 （三）流动负债与长期负债的比重，长期借款、短期借款、应付账款、其他应付款同比增加金额及原因；企业尝还
空间分析复习提纲 一、基本概念(要求：基本掌握其原理及含义，能做名词解释) 1、空间分析：是基于地理对象的位置和形态的空间数据的分析技术，其目的在于提取和传输空间信息。 2、空间数据模型：以计算机能够接受和处理的数据形式，为了反映空间实体的某些结构特性和行为功能，按一定的方案建立起来的数据逻辑组织方式，是对现实世界的抽象表达。分为概念模型、逻辑模型、物理模型。 3、叠置分析：是指在同一地区、同一比例尺、同一数学基础、不同信息表达的两组或多组专题要素的图形或数据文件进行叠加，根据各类要素与多边形边界的交点或多边形属性建立多重属性组合的新图层，并对那些结构和属性上既互相重叠，又互相联系的多种现象要素进行综合分析和评价；或者对反映不同时期同一地理现象的多边形图形进行多时相系列分析，从而深入揭示各种现象要素的内在联系及其发展规律的一种空间分析方法。 4、网络分析：网络分析是通过研究网络的状态以及模拟和分析资源在网络上的流动和分配情况，对网络结构及其资源等的优化问题进行研究的一种空间分析方法。 5、缓冲区分析：即根据分析对象的点、线、面实体，自动建立它们周围一定距离的带状区，用以识别这些实体或主体对邻近对象的辐射范围或影响度，以便为某项分析或决策提供依据。其中包括点缓冲区、线缓冲区、面缓冲区等。 6、最佳路径分析：也称最优路径分析，以最短路径分析为主，一直是计算机科学、运筹学、交通工程学、地理信息科学等学科的研究热点。这里“最佳”包含很多含义，不仅指一般地理意义上的距离最短，还可以是成本最少、耗费时间最短、资源流量（容量）最大、线路利用率最高等标准。 7、空间插值：空间插值是指在为采样点估计一个变量值的过程，常用于将离散点的测量数据转换为连续的数据曲面，它包括内插和外推两种算法。，前者是通过已知点的数据计算同一区域内其他未知点的数据，后者则是通过已知区域的数据，求未知区域的数据。 8、空间量算：即空间量测与计算，是指对GIS数据库中各种空间目标的基本参数进行量算与分析，如空间目标的位置、距离、周长、面积、体积、曲率、空间形态以及空间分布等，空间量算是GIS获取地理空间信息的基本手段，所获得的基本空间参数是进行复杂空间分析、模拟与决策制定的基础。 9、克里金插值法：克里金插值法是空间统计分析方法的重要内容之一，它是建立在半变异函数理论分析基础上，对有限区域内的区域变化量取值进行无偏最优估计的一种方法，不仅考虑了待估点与参估点之间的空间相关性，还考虑了各参估点间的空间相关性，根据样本空间位置不同、样本间相关程度的不同，对每个参估点赋予不同的权，进行滑动加权平均，以估计待估点的属性值。 二、分析类(要求：重点掌握其原理及含义，能结合本专业研究方向做比较详细的阐述) 1、空间数据模型的分类？ 答：分为三类： ①场模型：用于表述二维或三维空间中被看作是连续变化的现象； ②要素模型：有时也称对象模型，用于描述各种空间地物； ③网络模型：一种某一数据记录可与任意其他多个数据记录建立联系的有向图结构的数据模型，可 以模拟现实世界中的各种网络。
PB级大数据存储与分析解析 部门： xxx 时间： xxx 制作人：xxx 整理范文，仅供参考，可下载自行修改
PB级大数据存储技术与分析技术解读 2018年12月2日 目录 一、PB级大数据存储技术解读2 二、大数据分析系统应规避的问题5 三、剖析Hadoop和大数据的七误解8 四、6个优秀的开源文件系统助力大数据分析13 五、大数据与关系型数据库是否水火不容？NO (17) 六、大数据探讨：如何整理1700亿条Twitter发布信息？21 七、畅谈阿里巴巴的大数据梦26 八、Twitter利用Storm系统处理实时大数据35 一、PB级大数据存储技术解读 对于存储管理人员来说，大数据应该分为大数据存储和大数据分析，这两者的关系是——大数据存储是用于大数据分析的。然而，到目前为止这是两种截然不同的计算机技术领域。本文就重点解读一下PB级大数据存储技术，希望对您有所帮助。b5E2RGbCAP
越来越多的存储产品都在融入大数据的概念和功能，并使之成为产品的一大卖点。但对于从事存储管理的专业人员来说，对“大数据”在具体应用场景中的特点和区别有所了解。p1EanqFDPw 大数据存储致力于研发可以扩展至PB甚至EB级别的数据存储平台；大数据分析关注在最短时间内处理大量不同类型的数据集。DXDiTa9E3d 在快速变化的技术趋势中有两个特点需要存储管理人员重视起来。 第一，大数据分析流程和传统的数据仓库的方式完全不同，其已经变成了业务部门级别和数据中心级别的关键应用。这也是存储管理员的切入点。随着基础平台(分布式计算或其它架构>变得业务关键化，用户群较以往更加地依赖这一平台，这也使得其成为企业安全性、数据保护和数据管理策略的关键课题。RTCrpUDGiT
第一章 地理数据的初步整理 第一节 地理数据的类型、特征及其采集 一、地理数据的类型 根据地理学的研究对象可将地理数据分为空间数据和属性数据。 （一）空间数据 空间数据，主要用于描述地理实体、地理要素、地理现象、地理事件及地理过程产生、存在和发展的地理位置、区域范围及空间联系。空间数据的表达，可以将其归纳为点、线、面三种几何实体以及描述它们之间联系的拓扑关系。 点：由一个独立的坐标点),(y x 定位，可以表示精确的地理坐标点，也可以是一些地理实体的抽象，如道路交叉点、河流汇聚点以及小比例尺地图上的城镇、村庄等。 线：由两个以上坐标点i i y x i i ,2,1),,( 定义，有一定的长度和走向，表示线状地物或点实体之间的联系。如交通线、河流及各种地理区域的界线等，都是线实体。 面：表示在空间上连续分布的地理景观或区域。如居民区、工业区、行政区等都是面实体。 点、线、面三种地理几何实体，按照一定的拓扑关系组合、排列，就可以形成更为复杂的地理几何实体。如点、线组合形成网络；线、面组合形成地带；点、面组合形成地域类型；点、线、面组合形成地理区。 （二）属性数据 属性数据主要用于描述地理实体、地理要素、地理现象、地理事件及地理过程的有关属性特征，如海拔高度、气温、植被覆盖率、人口数量等。属性数据可以分为两种类型：即数量标志数据和品质标志数据。 1.数量标志数据 根据测度标准，可以将数量标志数据分为以下两类： ⑴ 间隔尺度数据。是以有量纲的数据形式表示测度对象在某种量纲下的绝对量。如摄氏温标表示气温，以面积量纲表示土地面积，以时间量纲表示地理事件、地理现象发生的时间等，如表1.1。 表1.1 间隔尺度数据 区域 年平均气温（℃） 年降水量（mm ） 土地面积（hm 2） 人口（人） 国内生产总值（万元） 1 8.0 500.2 1245.6 1210 2678.28 2 7.6 498.6 1064 1023 2015.47 3 6.5 550.9 894.3 848 1754.56 4 8.5 586.4 668.7 654 1365.46 ⑵ 比例尺度数据。是以无量纲的数据形式表示测度对象的相对量。这种数据要求事先规定一个基点，然后将其它同类数据与基点数据相比较，换算为基点数据的比例。因此这类数据常常又称为指数或比例数。如耕地指数、工业发展指数、舒适度指数等，如表1.2。 表1.2 比例尺度数据（某地区耕地复种指数及农业发展指数） 年份 1996 1997 1998 1999 2000 耕地复种指数① 120.40 113.56 126.54 132.76 121.43 农业发展指数 ② 100 115.68 124.50 135.69 129.56
数据存储分析和设计 第一步：收集各种表格 由某企业物资管理系统的供应计划管理部分的数据流程图和数据字典得到了下面的数据存储表： 第二步：确定各种表格需要存储的内容 ?根据系统功能确定是否有必要增加新表，对已有表，是否增加新的属性 ?去除多余的数据元素 表5中的生产需要量可由计量单位*消耗定额得到 ?增加必要的代码项 如加上材料码 第三步：列出各种表格存储的1NF数据元素 生产计划 1NF关系：部门码+部门名+产品码+产品名+计划产量 材料消耗定额表 1NF关系：产品码+产品名+材料码+材料名+型号+规格+计量单位+消耗定额 材料计划价格表 1NF关系：材料码+材料名+型号+规格+计量单位+单价 维修用材计划 1NF关系：部门码+部门名+材料码+材料名+型号+规格+计量单位+维修用量 生产用材计划 1NF关系：部门码+部门名+产品码+产品名+计划产量+材料码+材料名+型号+规格+计量单位+消耗定额
第四步：1NF关系的规范化 生产计划 1NF关系：部门码+部门名+产品码+产品名+计划产量 3NF关系：①*部门码+部门名 ②*产品码+产品名 ③*部门码+*产品码+计划产量 材料消耗定额表 1NF关系：产品码+产品名+材料码+材料名+型号+规格+计量单位+消耗定额 3NF关系：④*产品码+产品名 ⑤*材料码+材料名+型号+规格+计量单位 ⑥*产品码+*材料码+消耗定额 材料计划价格表 1NF关系：材料码+材料名+型号+规格+计量单位+单价 3NF关系：⑦*材料码+材料名+型号+规格+计量单位+单价 维修用材计划 1NF关系：部门码+部门名+材料码+材料名+型号+规格+计量单位+维修用量 3NF关系：⑧*部门码+部门名 ⑨*材料码+材料名+型号+规格+计量单位 ⑩*部门码+*材料码+维修用量 生产用材计划 1NF关系：部门码+部门名+产品码+产品名+计划产量+材料码+材料名+型号+规格+计量单位+消耗定额 3NF关系：⑾*部门码+部门名 ⑿*产品码+产品名 ⒀*部门码+*产品码+计划产量 ⒁*材料码+材料名+型号+规格+计量单位 ⒂*产品码+*材料码+消耗定额 第五步：3NF关系的归纳和合并 对以上15个表按照相同的关键字进行归纳与合并，最后得到供应计划管理的六个3NF关系： ①部门=*部门码+部门名 ②产品=*产品码+产品名 ③计划=*部门码+*产品码+计划产量 ④材料=*材料码+材料名+型号+规格+计量单位+单价 ⑤维修=*产品码+*材料码+维修用量 ⑥消耗=*产品码+*材料码+消耗定额
店长经营数据分析 Document number：WTWYT-WYWY-BTGTT-YTTYU-2018GT
店长必学：店长必须要会的数据分析 店长定期进行科学的数据分析，是店长掌握门店经营方向的重要手段。在日常工作中还有一些数据需要总部、门店分析，但无论哪方面数据，分析只是一个开始，关键是能够找出门店存在的问题及可以挖掘的能力，指导如何开始下一步工作才是重要的。店长需要每周或者每月开会，做以上各种数据分析，总结过去，找出差距。 一、门店经营指标数据分析 1）销售指标分析：主要分析本月销售情况，本月销售指标完成情况，与去年同期对比情况，通过这组数据的分析可以知道同比销售趋势，实际销售与计划的差距。 2）毛利分析：主要分析本月毛利率、毛利率情况，与去年同期对比情况。通过这组数据的分析可以知道同比毛利率状况，以及是否在商品毛利方面存在不足。 3）营运可控费用分析：主要是本月各项费用明细分析，与去年同期对比情况，有无节约控制成本费用，这里的各项费用是指：员工成本、能耗、物料及办公用品费用，维修费用，房租，存货损耗，日常营运费用（电话费、交通费、卫生费、税收、工商费），通过这组数据的分析，可清楚地知道门店营运可控费用后的列支，是否有同比异常的费用发生，有无可以节约的费用空间。 4）评效：主要是本月评效情况，与去年同期对比“日均评效”是指“日均单位面积销售额”，即日均销售额/门店营业面积。 5）人均劳效：主要是本月人均劳效情况，与去年同期对比，“本月人均劳效”计算方法：本月销售额/本月工资人数
6)盘点损耗率分析：主要是门店盘点结果简要分析，通过分析，及时发现门店在进、销、存各个环节存在的问题。 7)门店商品库存分析：主要是本月平均商品库存、周转天数，与去年同期对比分析。通过这组数据分析，看门店库存是否出现异常，特别是否有库存积压现象。 二、商品经营数据分析 1)经营商品目录执行情况总结分析：主要是本店执行商品目录情况与经营业态主力商品情况及新品引进情况、淘汰商品是否进行及时请退，总部每月1号将最新目录主力商品货号、目录新引进商品货号、目录淘汰商品货号发至各门店，门店根据相关货号查询出经营情况，特别是热销商品、新品商品经营情况，以及淘汰产品有没有及时请退，通过这组数据，可以了解门店是否按照商品目录的调整进行了门店的商品结构调整。 2)商品动销率分析：主要是本月商品动销品种统计，动销率分析，与上月对比情况，商品动销率计算公式：动销品种/门店经营总品种数*100，滞销品种数：门店经营总品种数-动销品种数。通过此组数据及具体单品的分析，可以看出门店在商品经营中存在的问题及潜力。 3)商品品类分析：主要是本店本月各品类销售比重及与去年同期对比情况，门店本月各品种类毛利比重及与去年同期对比情况，门店需对本月所有品类销售与毛利情况，特别是所有销售下降及毛利下降的品类进行全面分析，并通过分析找出差距，同时提出改进方案。 4)本月商品引进分析：主要是引进商品产生销售、毛利分析，这时的引进商品需要门店日常对新引进商品建档，并跟踪分析引进商品的动销率、适销率、销售额以及毛利
分区耕地坡度结构图的制作 姓名： 学号： 年级： 专业： 学院： 指导老师： 华北水利水电学院 年月日
1 目的意义 对耕地坡度进行分级是对耕地管理的重要前提。耕地坡地分级赋值是一个繁杂的人机交互过程，人为地判断分割面积可能造成多次反复，还极可能产生误差，特别是矢量化和分割图斑会带来拓扑错误及属性丢失的现象。通过运用ArcGIS 的可进行耕地坡度分级赋值，该方法在使用时需提供经过拓扑错误检查后带属性数据的土地利用图斑数据和坡度分级图。 2材料方法 2.1数据 耕地坡度结构图制作需要用到的数据有：地类图斑.shp、线状地物.shp、注记点.shp、整饰线.shp、等高线.shp、村界.shp、遥感影像H50G044024DOM.tif，每个数据包含的内容见表1。 表1 耕地坡度结构图制作需要用到的数据 2.2数据分析方法(把数据来源、数据处理与分析方法说清楚) 2.2.1数据裁切方法(研究区域边界的确定) （1）矢量数据的裁切(clip) 矢量数据的剪切用clip，该工具在 （2）栅格数据的裁切(extract by mask) 栅格数据的剪切用extraction，该工具在 2.2.2影像的地理配准方法(定义数据的坐标系) 影像的地理配准用georeferencing，该工具在 2.2.3图形的矢量化方法(把多用到的矢量化方法尽量全部列出) 图形的矢量化运用Editor工具，该工具在点击Editor——Start Editing，在
Target中选择图层，进行图形矢量化。 为了方便绘制，可以打开效果Effect工具条，将绘制层设置为透明(也可将地类图斑层设置为无填充) ①一般面状轮廓的画法使用Sketch Tool，沿面状地物的边界进行描绘，双击完成。 ②面状边界的跟踪矢量画法 ③岛的画法 ④面要素边界转换为线要素 利用已经录入完毕的面状要素，使用面转线工具直接将所绘所有面要素边界转换为线要素。具体方法为：打开ArcToolbox，选择Data Management Tools菜单下的Features，双击Polygon To Line，即打开Polygon To Line对话框。在Input Feature中输入已画好的待转换的面层，在Output Feature Class中输入转换后的线层确定后即完成面转线。 ⑤线要素转换为面要素 利用已有的线要素生成面要素 2.2.4 拓扑分析方法 拓扑分析所要用到的工具topology 2.2.5矢量数据属性值的录入方法 2.2.6符号库的建立方法 （1）面状符号的制作方法 （2）线状符号的制作方法
一、数据库知识 单项选择题 1. 数据库系统的核心是（B） A、数据模型 B、数据库管理系统 C、软件工具 D、数据库 2. 下列叙述中正确的是(C)。 A、数据库是一个独立的系统，不需要操作系统的支持 B、数据库设计是指设计数据库管理系统 C、数据库技术的根本目标是要解决数据共享的问题 D、数据库系统中，数据的物理结构必须与逻辑结构一致 3. 下列模式中，能够给出数据库物理存储结构与物理存取方法的是( A )。 A、内模式 B、外模式 C、概念模式 D、逻辑模式 4. SQL语句中修改表结构的命令是(C )。 A、MODIFY TABLE B、MODIFY STRUCTURE C、ALTER TABLE D、ALTER STRUCTURE 5. SELECT-SQL语句是(B ) 。 A、选择工作区语句 B、数据查询语句 C、选择标准语句 D、数据修改语句 6. SQL语言是( C )语言。 A、层次数据库 B、网络数据库 C、关系数据库 D、非数据库 7. 如果要创建一个数据组分组报表，第一个分组表达式是"部门"，第二个分组表达式是"性别"，第三个分组表达式是"基本工资
"，当前索引的索引表达式应当是( B )。 A、部门+性别+基本工资 B、部门+性别+STR(基本工资) C、STR(基本工资)+性别+部门 D、性别+部门+STR(基本工资) 8. 数据库DB、数据库系统DBS、数据库管理系统DBMS三者之间的关系是( A )。 A、DBS包括DB和BMS B、DBMS包括DB和DBS C、DB包括DBS和DBMS D、DBS就是DB，也就是DBMS 9. 下列有关数据库的描述，正确的是( C )。 A、数据库是一个DBF文件 B、数据库是一个关系 C、数据库是一个结构化的数据集合 D、数据库是一组文件 10. 下列说法中，不属于数据模型所描述的内容的是( C )。 A、数据结构 B、数据操作 C、数据查询 D、数据约束 11. 数据库管理系统能实现对数据库中数据的查询、插入、修改和删除等操作，这种功能称为( C ) 。 A．数据定义功能 B．数据管理功能 C．数据操纵功能 D．数据控制功能 12. 数据库管理系统是( B ) 。 A．操作系统的一部分 B．在操作系统支持下的系统软件 C．一种编译程序
2003年一季度经营情况分析报告
新奥燃气控股有限公司 2003年4月
前言 03年度一季度已经匆匆过去。继02年度成功的市场开拓之后，控股公司有28个成员企业投入运作，从而使新奥燃气的覆盖人口从02年度的685万人迅速地增大到935万人。市场的扩展也使控股公司的经营收入比去年同期增长71.27%，达到11745.25万元，首次实现了季度收入过亿元。民用户的市场发展量和安装量、工商户的发展量和安装量比去年同期也有较大幅度的提高。一季度，控股公司成功的完成了部分A类物资的招标采购，实现了物资采购的质的飞跃并有效的降低了物资采购的成本；针对公司规模的迅速扩张，成立控股公司的安全管理委员会和安全管理办公室，为实施有效的安全管理打下了基础；工程管理迅速的开展了对成员企业的技术指导和流程支持，有效的支撑了企业的场站建设和基建工程建设。 也应该看到，随着新公司的增加，市场容量的迅速增大，销售收入并未实现同比的增长。老公司市场容量的日益减少、新公司市场培育尚未完成，给控股公司业绩的迅速提升带来巨大的压力。同口径相比，虽然老公司的业绩比去年同期增长14%，但新公司的市场增量依旧给控股公司一季度完成情况的差距。一季度，控股公司销售收入仅完成季度计划的82.69%，完成年度计划的11.51%。总体经营情况依旧没有摆脱严峻的形势，这就要求控股公司努力探求迅速提升市场发展的有效途径，寻找降低成本、提升业绩的有效手段，给投资者以信心。
一、总体经营情况 一季度，控股公司共实现销售收入11745.25万元，虽比去年同期增长
71.27%，但仅完成年度计划的11.51%，年度计划完成比比去年同期下降2.59个百分点；实现回款12456.45万元，比去年同期增长78.5%，回款率为106.06%。 经营收入与回款状况见附表一：03年一季度经营情况。 在销售收入的排名中，廊坊燃气、淮安燃气和蚌埠燃气分别以2603.71万元、1948.69万元和1715.91万元位居前三位，新乡燃气以1247.64万元位居第四。 在生产情况中，民用户发展完成34679户，完成季度计划的96.02%，完成年度计划的14.45%，比去年同期增长132.15%；在与去年同期老公司的数据对比分析中，今年老公司完成发展18019户，比去年同期的17073户增长了5.54%，显示出老公司的市场发展情况基本稳定；去年下半年度及今年成立的新公司的市场发展尽管也完成了16660户，但未能显现出市场发展的强劲势头来。工商业户发展完成19992.63方/日，完成年度计划的15.38%，完成季度计划的160.27%，比去年同期增长336.29%。 一季度，民用户安装完成10758户，虽比去年同期增长94.43%，但仅完成季度计划的73.36%，完成年度计划的5.57%；在与去年同期老公司的数据对比分析中，今年老公司完成安装仅完成4671户，比去年同期的5669户降低了17.6%，依旧显示出老公司注重房地产开发商的发展，对老户的开发仍然缺乏有效措施；工商业户安装完成7964.66方/日，完成季度计划的94.81%，完成年度计划的4.42%，完成量虽比去年同期增长65.22%，但年度完成率比去年同期下降1.52个百分点；总体的安装形势依旧呈现低迷状态。 在财务状况中，可控费用支出2856.63万元，比季度计划超支13.79%，比去年同期比例增高了1.46个百分点； 详见一季度经营指标完成情况表。 燃气集团2003年第一季度其它业务指标完成情况
大数据可视化分析平台 一、背景与目标 基于邳州市电子政务建设的基础支撑环境，以基础信息资源库(人口库、法人库、宏观经济、地理库)为基础，建设融合业务展示系统，提供综合信息查询展示、信息简报呈现、数据分析、数据开放等资源服务应用。实现市府领导及相关委办的融合数据资源视角，实现数据信息资源融合服务与创新服务，通过系统达到及时了解本市发展的综合情况，及时掌握发展动态,为政策拟定提供依据。 充分运用云计算、大数据等信息技术，建设融合分析平台、展示平台，整合现有数据资源,结合政务大数据的分析能力与业务编排展示能力，以人口、法人、地理，人口与地理,法人与地理，实现基础展示与分析,融合公安、交通、工业、教育、旅游等重点行业的数据综合分析，为城市管理、产业升级、民生保障提供有效支撑。 二、政务大数据平台 1、数据采集和交换需求:通过对各个委办局的指定业务数据进行汇聚,将分散的数据进行物理集中和整合管理，为实现对数据的分析提供数据支撑。将为跨机构的各类业务系统之间的业务协同，提供统一和集中的数据交互共享服务。包括数据交换、共享和EＴL等功能。 ２、海量数据存储管理需求：大数据平台从各个委办局的业务系统里抽取的数据量巨大,数据类型繁杂,数据需要持久化的存储和访问。不论是结构化数据、半结构化数据，还是非结构化数据,经过数据存储引擎进行建模后，持久化保存在存储系统上。存储系统要具备
高可靠性、快速查询能力。 3、数据计算分析需求：包括海量数据的离线计算能力、高效即席数据查询需求和低时延的实时计算能力。随着数据量的不断增加，需要数据平台具备线性扩展能力和强大的分析能力,支撑不断增长的数据量，满足未来政务各类业务工作的发展需要，确保业务系统的不间断且有效地工作。 4、数据关联集中需求：对集中存储在数据管理平台的数据，通过正确的技术手段将这些离散的数据进行数据关联,即:通过分析数据间的业务关系，建立关键数据之间的关联关系，将离散的数据串联起来形成能表达更多含义信息集合，以形成基础库、业务库、知识库等数据集。 5、应用开发需求:依靠集中数据集,快速开发创新应用，支撑实际分析业务需要。 6、大数据分析挖掘需求:通过对海量的政务业务大数据进行分析与挖掘,辅助政务决策,提供资源配置分析优化等辅助决策功能， 促进民生的发展。
不同大数据分析的存储选择 目前市场上有两种类型的大数据分析方式——同步的和异步的，两种都有各自在存储容量和特性上的要求。 近来大数据分析这个词正逐渐成为IT界流行的一个术语，以代指有关大数据本身的猜想，通俗说来即成堆数据背后问题的答案。然而，如果我们能够从足够的数据点入手比对及交叉分析，或许能帮助我们找到一些有用的数据，甚至可能帮助避免灾难。 问题是显而易见的，所有的分析都需要大量甚至海量的数据，这便给当今的IT管理人员带来了更新的挑战，即如何捕获、存取、以及分析这些数据并将从中得到的分析用于后续任务的执行? 大数据分析应用通常会使用例如网络流量、金融交易记录以及敏感数据来替代传统形式的内容。数据本身的价值在于数据间的比对、关联或者引用。对大数据的分析通常会意味着与大量的小数据对象打交道，而这些小数据对象往往对响应延时要求非常之高。 当前业界主要有两种大数据分析场景，而它们通常是根据数据处理的形式而区分：在实时使用场景下，响应效率是最为关键的，因此大数据存储架构本身的设计需要满足最小延时的功能。 同步，即实时的或者近乎于实时的；另外一种就是异步的方式，这种方式下，数据首先会被获取，记录下来然后再用批处理进程进行分析。 同步分析 可以想到的近乎于实时的大数据分析的最早的例子就是超级市场里的工作人员是如何统计消费者行为习惯以便于提供相应的优惠促销券的。事实上是，消费者购买行为计算很可能在用户收银前就已经完成，但是概念本身是非常类似的。另外一个相关的例子是在线社交
网站可以通过访问用户的行为建立属于他们的行为数据库，这样就可以根据各自不同的消费习惯提供不同的点对点广告植入。 在零售行业，一些大型商铺正开始在停车场对前来购物的消费者使用面部识别技术，这样一旦他们路过或者经过对应的商铺与之相应的促销信息便随之而来。因此，在这样一类的实时大数据分析场景中，速度是第一要素，故而大数据存储架构需要建设成为低延时的场景。 针对同步大数据分析的存储 实时分析应用通常会运行在例如NoSQL之类的数据库上，通常都能支持海量可扩展的商用硬件上。Hadoop，从另一角度考虑，非常适合批量的数据处理，这种技术非常合适于异步大数据分析。由于在很多场合下，存储本身会成为延时问题的瓶颈，那么固态存储设备对于实时数据分析是很有帮助的。闪存存储可以以多种形式进行部署：作为传统存储磁盘阵列的一层，以NAS系统的方式，再或者以应用服务器本身的方式都可以实现。 这种服务器端的闪存实施方式广受用户欢迎，之所以这样是由于它能够实现最低程度的延时（因该方式下的存储最为接近CPU），并且提供了很灵活的容量选择，几百GB容量就可以实现。SAS/SATA接口的固态硬盘本身就是个选择，但是近来我们看到PCIe板卡为接口的固态设备逐渐成了性能应用（比如实时分析）的标准，因为相对于前者，其延时更低。 如今，业界有许多提供PCIe闪存存储的公司，包括Fusion-io、LSI、Micron Technology、SanDisk、sTec（现在是HGST的一部分，作为Western Digital的一个部门）、Violin Memory 以及Virident (也被Western Digital收购)。其它所有主流服务器及存储厂商们也都提供PCIe 解决方案，大多数是与这些公司通过了OEM协议。 尽管PCIe卡最大容量已经近乎于10 TB，但仍无法满足用户的需求，因此一个共享的存储资源池也是需要考虑的。一个解决方案是使用Virident的FlashMAX Connect software，这种软件可以实现将PCIe卡的资源通过服务器上的InfiniBand，进行资源池化。
2016年数据处理和存储服务行业简析 一、行业主管部门及监管体制 (2) 二、行业主要法律、法规及政策 (2) 三、行业发展现状和未来趋势 (3) 四、进入本行业的主要障碍 (6) 1、技术壁垒 (6) 2、人才壁垒 (7) 3、市场与客户壁垒 (7) 五、因素行业发展的因素 (8) 1、有利因素 (8) （1）互联网技术的发展对行业的促进 (8) （2）国家信息化建设趋势 (8) （3）IT基础科技的快速发展 (8) 2、不利因素 (9) （1）技术人才流失风险 (9) （2）资金风险 (9) 六、行业主要企业简况 (9) 1、上海天玑科技股份有限公司 (9) 2、南京斯坦德云科技股份有限公司 (10) 3、上海爱可生信息技术股份有限公司 (10) 4、南京云创大数据科技股份有限公司 (11)
一、行业主管部门及监管体制 数据处理和存储服务行业从属于软件和信息技术服务业，行政主管部门是工业和信息化部以及各地的信息产业主管部门。工业和信息化部负责制订我国软件和信息技术服务业的产业政策、产业规划和行业规则制度，制订行业的技术政策和技术标准等，对行业的发展方面进行宏观调控。 此外，国家发改委、科技部等部门分别从产业发展、科技发展等方面对行业进行宏观指导，国家版权局负责本行业知识产权相关保护工作。 我国软件和信息技术服务业自律机构为中国软件行业协会。中国软件行业协会主要负责产业和市场研究、行业协调、为会员企业提供公共服务、行业自律管理；受工信部委托对各地软件企业认定机构的认定工作进行业务指导、监督和检查，负责软件产品登记认证和软件企业资质认证工作；代表会员企业与相关政府部门进行行业信息的交流与协调，向政府部门提出产业发展建议等。 数据处理和存储服务等业务，涉及的监管部门还包括公安部门，相关的行业协会主要有中国安全防范产品行业协会、国家计算机行业协会等。 二、行业主要法律、法规及政策 行业涉及的国家及地方的相关法律、法规及政策如下：
怎么从公司财务报表中分析一个公司经营状况 假如有两家公司在某一会计年度实现的利润总额正好相同，但这是否意味着它们具有相同的获利能力呢？答案是否定的，因为这两家公司的资产总额可能并不一样，甚至还可能相当悬殊。再如，某公司2000年度实现税后利润100万元。很显然，光有这样—个会计数据只能说明该公司在特定会计期间的盈利水平，对报表使用者来说还无法做出最有效的经济决策。但是，如果我们将该公司1999年度实现的税后利润60万元和1998年度实现的税后利润30万元加以比较，就可能得出该公司近几年的利润发展趋势，使财务报表使用者从中获得更有效的经济信息。如果我们再将该公司近三年的资产总额和销售收入等会计数据综合起来进行分析，就会有更多隐含在财务报表中的重要信息清晰地显示出来。可见，财务报表的作用是有一定局限性的，它仅能够反映一定期间内企业的盈利水平、财务状况及资金流动情况。报表使用者要想获取更多的对经济决策有用的信息，必须以财务报表和其它财务资料为依据，运用系统的分析方法来评价企业过去和现在的经营成果、财务状况及资金流动情况。据以预测企业未来的经营前景，从而制定未来的战略目标和作出最优的经济决策。 为了能够正确揭示各种会计数据之间存在着的重要关系，全面反映企业经营业绩和财务状况，可将财务报表分析技巧概括为以下四类：横向分析；纵向分析；趋势百分率分析；财务比率分析。 一、财务报表分析技巧之一：横向分析 横向分析的前提，就是采用前后期对比的方式编制比较会计报表，即将企业连续几年的会计报表数据并行排列在一起，设置“绝对金额增减”和“百分率增减”两栏，以揭示各个会计项目在比较期内所发生的绝对金额和百分率的增减变化情况。下面，以ABC公司为例进行分析（见下表）。 比较利润及利润分配表分析： ABC公司比较利润表及利润分配表金额单位：元 项目2001年度2002年度绝对增减额百分率增减额（％） 销售收入7655000 9864000 2209000 28.9