Classification modelling review

IDS 572 DATA MINING FOR BUSINESS
ASSIGNMENT 3 - FUNDRAISING 2
Kaushik Kompella
Jaideep Adusumelli
Sushaanth Srirangapathi

1. Modeling Partitioning - Partition the dataset into 60% training and 40% validation (set the seed
to 12345). In the last assignment, you developed decision tree, logistic regression, naïve Bayes,
random forest and boosted tree models. Now, develop support vector machine models for
classification. Examine different parameter values, as you see suitable. Report on what you
experimented with and what worked best.
Howdo you select the subset ofvariables to include in the model? What methods do you use to select
variables that you feelshouldbe includedinthe model(s)? Doesvariable selectionmake adifference?
Provide a comparative evaluation ofperformance ofyour bestmodels from all techniques (including
those from part 1, i.e. assignment 2) (Be sure NOT to include “TARGET−D” in your analysis.)
Data Selection: Missing Values
In order to focusour effortson the set of variablesthatwouldprovide the mostmeaningful information
we decided to start by looking for variables with excessive amounts of missing values.
We sawan advantage inchoosing to eliminatevariableswith>=85% of all valuesmissing.Thishelpedus
to eliminate anumberof variables.We chose thisroute because we knew that includingthese variables
withgreaternumbersof missingvalueswouldnotbe useful inpredictingdonorresponse,because there
isverylittle informationcontainedwithinthose variables,wellbelow eventhe low responserate of 5.1%.
Certain notable variables were:
Attributes Eliminated Number of Missing values (out of 9999 values)
MAILCODE 9857
PVASTATE 9847
RECP3 9728
RECPGVG 9983
CHILD03 9884
CHILD12 9840
CHILD18 9746
SOLP3 9976
MAJOR 9969
HOMEE 9902
PHOTO 9504
KIDSTUFF 9822
CARDS 9869
PLATES 9933
RDATE_3 9947
RDATE_5 9997
RDATE_6 9896
RAMNT_3 9947
RAMNT_4 9954
RAMNT_5 9997
RAMNT_6 9896
RECSWEEP 9826

These Variables did not even have enough number of values for the actual data which has only 5.1%
response rate. Sothese variableseventhoughcontainveryimportantinformationcannotcontribute well
towards creating a good model. Hence we eliminated all these variables.
We did not eliminate those variables that were coded as X and null values. We transformed
many of those attributes and will discuss those in the ‘transform’ section.
Data Selection: Relevance
Next, we chose to focus on only including variables which wouldhave relevance on predictingdonation
probability.We didthisbyscanningthroughthe variabledescriptionsinthe datadictionarytoonlyretain
those that were relevant.
An example of a variable thatwe chose to eliminate wasDW2. Thisvariable correspondedtothe locale-
specific percentage of single-unit detached structures. Eliminating these variable and other variables
similar to this allowed us to lessen the total number of variables that we analyzed.
In addition to those variables that we found to be irrelevant, we noticedthat there was a great deal of
overlap between certain variables. For instance, the ‘Wealth1’ and ‘Wealth2’ variables had very similar
descriptions in the data dictionary. However, we noticed that Wealth2 contained a more robust
description of the variable –
“Wealth rating uses median family income and population statistics from each area to index relative
wealthwithineachstate.The segmentsare denoted0-9,with9beingthe highestincomegroupandzero
being the lowest. Each rating has a different meaning within each state”.
Whereas “Wealth1” simply contained the description: wealth rating. Since “Wealth1” has a similar
description to “Wealth2” we know that there is a lot of information overlap, which allows us to justify
dropping the less informative of the two: “Wealth1”.
We also found some variables which had no relationship with “TARGET_B” but have a relationship with
“TARGET_D”. Some variables of this kind are:
Data Reduction: PCA
We experimented with creating many different sets of PCAs but found that in many cases we were
including many variable sets that had numerous missing values, or had summary variables that were
alreadyexisting.Therefore,we onlycreated3PCAsthat were valuable PCA_CENSUS_X,PCA_EMP_Xand
PCA_INTERESTS_X.
PCA_CENSUS_X included census variables for neighborhoods
PCA_EMP_X included Employment information of the people i.e. Whether they are army or military
veterans, state government employees or federal government employees, etc.
PCA_INTERESTS_X included the 18 ‘interest/hobby’ based variables.

Data Reduction: Random Forests
We ran random forests and plotted the variable importance graphs for all the variables for all
the remaining attributes after the above processes were carried out. And then we selected all
those variables which had the “Decrease in Gini Index” more than or equal to 0.0064 which
gave us 68 variables. These included PCAs and mostly summary variables.
Most of the variables were not at all included in the split and we were able to eliminate them to
improve the overall performance of our model.
Data Selection: Missing Value Replacement
We utilized two rapid miner operators in our process to ensure that those variables that had missing
valuesbelowourcutoff were transformed.The twooperatorswe usedwere the Replace MissingValues
and the Map operators.
Replace MissingValuesallowedustoimpute the valueswhennecessarywhile the Mapoperatorallowed
us to recode those variables that had encoded characters such as “MDMAUD”. This allowed us to code
the extraneous characters as null values giving us a more accurate picture of which variables actually
contained many more missing values than originally depicted in our first glance.

Data Selection: The Variables Included
CLUSTER DOMAINTYPE WEALTH1
INCOME PEPSTRFL MAXRAMNT
DOMAINSES HIT VIETVETS
WWIIVETS IC1 CARDPROM
MAXADATE CARDPM12 RAMNTALL
NGIFTALL CARDGIFT MINRAMNT
MINRDATE MAXRDATE LASTGIFT
FISTDATE TIMELAG AVGGIFT
CONTROLN TOTALDAYS RECENCYFREQ
LOWINCOME MEDINCOME HIGHINCOME
PCA_NEIGH_PER_1-5 PCA_NEIGH_MMM_1-4 PCA_EMP_1-4

Decision Trees
We chose the J48 version of the decision tree as we were able to obtain the most accurate results.
The parameter that allowed us to build the best model was .4 as a threshold. We selected Laplace
smoothing parameter for the J48 and chose an overall confidence of .35.
We created two main PCA’s which we removed from the model initially. The following confusion matrix
is our J48 without these three PCAs:
Accuracy of Training Data set for the best tree: 94.97
Prediction / Actual Actual Good Actual Bad Class precision
Predicted Good 3796 159 95.98
Predicted Bad 143 1901 93.00
Class recall 96.37 92.28
Accuracy of Test Data set for the best tree: 55.67
Our optimal J48model showed a55.67% accuracy onthe overall testdata.The classprecisionfortrue 1’s
and predicted 1’s was 38.26% and overall class recall was 37.80%. These results were close to training
data accuracy so it implies that we reached optimal model.
Logistic Regression
Once again,we ran the model withandwithoutthe PCAsthatwe generatedforthe J48 model.Inthis
case we alsofoundthatthe inclusionof PCA’swashurting overallmodelperformance. Below isthe
confusionmatrix associatedwithlogisticregression.

We thencomparedthe trainingresultsfromourLogisticRegressiontothe validationdataandfoundour
best LR model. The associated test data confusion matrix is seen above.
We could maintainan overall model accuracyof 48.08%. The class precisionfromthe testdata was very
similartothatof the training data.We obtained a38.83% precisionwithanoverallclassrecall of 77.07%.
We used default parameters for the Logistic regression as these gave us the highest overall accuracy.
Naïve Bayes
After running J48 and Logistic regression we attempted to find better model using Naïve Bayes. We
applied same subset of variables both with and without PCAsto find best Naïve Bayes model. Our best
model again didnot contain any of the PCA’sbut as with the othermodels,containedall of our created
variables described in the initial data preprocessing steps.
The overall model accuracy is 59.35% with a class precision of 43.53% and a class recall of 43.71%
We compared the accuracy of our Naïve Bayes model with the associated testing data. The resulting
confusion matrix is as above.
In this case we see a rise in overall model accuracy on the validation data set. Also, we noticed similar
class precisionandclass recall betweenourtestingand trainingdata suggestingwe hadfound our best,
and most stable Naïve Bayes model.
Support Vector Machines
Initially we tried to deploy SVMmodel on our data but were getting perfect over fit. We were using a
kernel type radial basedfunctionondefaultsettings.Inaddition,we changedthe thresholdto.3, .5, and
.65 but were still getting massive over fit on our training data.

Below is the confusion matrix associated with the SVMtraining data:
Afteranalyzingmany differentkernel typessuchas- polynomial,dot,andGaussiancombination.We also
experimentedwiththe Cpenaltyparametersand determinedthatthe supportvectormachinewasn’tthe
best suited algorithm to our data and did not generalize well.
We could generate an overall accuracyof 59.21% onthismodel,butwithalowerclassprecisionof 35.8%.
The class recall was also worse than all other models with a 23.69% recall on true 1’s/predicted 1’s.
Our testdatahad an overall accuracy of 59.55% butstill withverylow classprecisionof 39.43% and class
recall of 23.21%.
KNN
We usedthe followingparameterstogenerate ourbestKNN model:
K = 150; Measure:Mixed,Mixed-Measure:MixedEucledian
Thisincluded usingak-value of 150, withthresholdof .33 alongwithmixedmeasure of Euclideanasa
distance metric.
Belowisthe confusionmatrix of ourtrainingdatafor KNN:

Using the parameters above we obtained an overall accuracy of 54.3% on the training data. The class
precisionwas lowerthan that of other modelsat 39.57%. The class recall for predictedand true 1’s was
61.5% overall.
Our testing confusionmatrix gave an accuracy of 50.48%, with a class precision of 38.1%, but an overall
class recall of 60.32% on predicted 1’s and true 1’s.
Random Forest
Initially we usedthe same setof attributesthatwe deployedinourSVMmodel andwere onlyable to
generate a~30% Test accuracy but witha ~80% Trainingdata accuracy. We allowedtoomanytreesto
buildandas a resultthismodel wasoverfit.
By limitingdepthof the trees,we could generate amuchhigheraccuracy ontest data.Below isour
RandomForestconfusionmatrix forourtrainingdata:
We obtainedanaccuracy of 60.06% overall on trainingdata,butwitha classprecisionof 44.97% and a
classrecall 72.91% forpredictedandtrue 1’s
On testing data set we obtained a reasonably close accuracy of 53.8%. The class precision was 41.07%
with a class recall close to that of the training data at 65.32%.

Gradient Boosted Trees:
The parameters used were:
Number of trees: 40, Maximal Depth: 5, Minimum Rows: 20, Minimum Split Improvement: 0.0025,
Number of Bins: 25, Learning Rate: 0.1, Sample Rate: 1, Distribution: Auto
From the above matrix we found that gradient boosted trees were the best model we could get with
optimal accuracy and precision and high recall which would help us to maximize the profits.

2. Our overall goal is to identify which individuals to target for maximum donations (profit). We
will try two approaches for this:
(i) using the response model,togetherwith average donation and mailingcosts information,to
identifythe most profitable individualstotarget (Q 2.1 below)
(ii) developa secondmodel on TARGET_D, and combine this with the response model to identifythe
most profitable individualstotarget (Q 2.2 below)
2.1 (a) What is the ‘best’ model for each method in Question 1 for maximizing revenue? Calculate
the net profit for both the training and validation set based on the actual response rate (5.1%). We
can calculate the net profit from given information - the expected donation, given that they are
donors, is $13.00, and the total cost of each mailing is $0.68. Note: to calculate estimated net profit
(on data with the ‘natural’ response rate of5.1%), we will need to “undo” the effects ofthe weighted
sampling, and calculate the net profit that reflects the actual response distribution of 5.1% donors
and 94.9% non-donors.)
Afterchoosingthe bestmodelsfromthe above methods,we coulddeterminethatourbestmodel was
the GradientBoosted Trees.Thisis notjust basedonthe accuracy itself butthe netprofitwe got.
We neededtoundothe effectsof the weightedsampling whichwe didinthe followingway:
Firstly, adjustthe weightedprofit andthe weightedcost.
Weightedprofit:13-$0.68 = $12.32
WeightedCost= ($0.68)
Afterthis,we usedthe followingformulatocalculate maximumprofit:
(12.32*.051)/(3499/9999) = 1.795
Next,we usedthe followingformulatocalculate the adjustedcost:
(($0.68)*0.9491)/(6500/9999) = .992
The liftwe achievedforthe logisticregressionwas 306.679 as max profit.

(b) Summarize the performance of the ‘best’ model from each method, in terms of net profit from
predicting donors in the validation dataset; at what cutoff is the best performance obtained?
Draw profit curves:Draweachmodel’s netcumulative profitcurve forthe validation setontoasingle
graph. Are there any models that dominate?
Best Model: From your answers above, what do you think will be the “best” model to implement?
(What criteria do you use to determine ‘best’?)
Our bestmodel is GradientBoostedTrees.We usedmax liftcriteriabutalsoconsideredthe highestclass
recall (55.39%) and class precision (41.30%). Our next best model (“Logistic Regression”) has a class
precision of 38.83%, and a class recall of 77.07%, but a higher lift curve. It does not make sense from a
businessstandpointtoevaluate ourbestmodelonthe criteriaof liftalone because thenthe J48wouldn’t
generalize well to predict donors accurately.
The performance of all models are as follows:
Model Profit (Performance)
J-48 105.5 $
Logistic 257.63 $
Naïve Bayes 319.58 $
Random Forests 349.10 $
KNN 159.34 $
SVM 90.63 $
The cutoff we usedin GradientBoostedTrees wasa confidence of: 0.496901.
0
50
100
150
200
250
300
1
29
57
85
113
141
169
197
225
253
281
309
337
365
393
421
449
477
505
533
561
589
617
645
673
Cummulative Benefit curveof Decision Tree

The best model that was J-48 according to the lift chart. However, Gradient Boosted Trees had the best
chance to generalize well on unseen data.
-150
-100
-50
0
50
100
150
1
29
57
85
113
141
169
197
225
253
281
309
337
365
393
421
449
477
505
533
561
589
617
645
673
Cummulative benefit curvefor Gradient
Boosted Trees
-50
0
50
100
150
200
1
29
57
85
113
141
169
197
225
253
281
309
337
365
393
421
449
477
505
533
561
589
617
645
673
Cummulative Benefit Curvefor Naive Bayes
-50
0
50
100
150
200
1
29
57
85
113
141
169
197
225
253
281
309
337
365
393
421
449
477
505
533
561
589
617
645
673
Chart Title

2.2. (a) We will next develop a model for the donated amount (TARGET_D). Note that TARGET_D has
valuesonlyforthose individualswhodonors(thatis,TARGET_D valuesare definedonlyforcaseswhere
TARGET_B is 1). What data will you use to developa model for TARGET_D? (Non-donors,obviously,do
not have any donationamount -- shouldyou considerthese as $0.0 donation,or impute missingvalues
here? Should non-donors be included for developing the model to predict donation amount? Also,
should cases with rare very large donation amounts be excluded? [Hah! – leading questions]
Develop a model for TARGET_D. What modeling method do you use (report on any one). Which
variables do you use? What variable selection methods do you use? Report on performance.
For the caseswithrare verylarge donationamounts,we will notexclude thesecasesaswe will lose out
on predictingourmostvaluable donors.Andthesedonationsmightbe large because of asupportive
reasonfromthe valuesof othervariables.Hence we didnotwanttolose ImportantInformation.
We will use linearRegressionmodelforpredictedTARGET_Dwitha splitvalidationof 60-40 on the data
set.
Variable selectionmethods
Afterapplyingthe datacleaningandmissingvalue treatmentasappliedondatasetforTARGET_B, we
lookedatthe scatter plotsof the independentvariablesagainstTARGET_Dandeliminatedthose
variablesthatdid notshowany significanttrendand/orrelationshipwithTARGET_D. Givenbelow are a
fewvariablesthatwe have selectedforregressionmodellingforTarget-Dandtheirscatterplotswith
Target D.
Fig.1 RAMNT_18 vs TARGET_D

Fig.2 RAMNT_8 vs TARGET_D
Fig.3 LASTGIFT vs TARGET_D
Fig.4 RAMNTALL vs TARGET_D

Fig.5 HIT vs TARGET_D
As can be seenfromthe plots,the above variableshadasignificantrelationshipwithTARGET_D,hence
we selectedthese variablesforthe linearregressionmodellingfor the predictionof TARGET_D Variable.
Givenbeloware the scatterplotsof variable withTARGET_D for the oneswhichwe eliminated.
Fig.1 AGE vs TARGET_D

Fig.2 ADI vs TARGET_D
Fig.3 DMA vs TARGET_D
Fig.4 CARDPROMvs TARGET_D
These variables didn’tshowanyparticulartrendwiththe TARGET_D variable hence were eliminated.

AfterScatterplots,we narroweddownthe listof variables from481 to 261. Andthenwe used
correlationmatrix tofinallyselectvariablesformodelling.
The highestcorrelation(R) valuesof independentvariableswithTARGET_Dwas of 0.2 from LASTGIFT
followedbythese variables - NGIFTALL,AVGGIFT,RAMNTALL,MINRAMNT.We choose all the variables
having>= 0.05 as R value withTARGET_D. We finalizedalist of 15 independentvariablesinourdataset
to run the model.Here isthe listof variablesthat we usedin our bestlinearregressionmodelthrough
the T-Testfeature selection.
The variable we gotwere:
HIT WEALTH_1 NUMPROM RAMNT_8 RAMNT_12
RAMNT_14 RAMNT_16 RAMNT_18 RAMNT_22 RAMNTALL
NGIFTALL MINRAMNT MAXRAMNT LATGIFT AVGGIFT
Performance-
Our bestmodel hasgiventhe followingvaluesforRootmeansquarederrorandsquaredcorrelation
Root-MeanSquarederror:9.949 +/- 0.000

(b) How can you use the results from the response model together with results from the donation
amount model to identify targets? (Hint: The response model estimatesthe probability of response
p(y=1|x). The donation amount model estimates the conditional donation amount, E [amt | x, y=1].
The product of these gives …..? )
How do you identify individuals to target, when combining information from both models? [Sorted
values of predicteddonation amount? What thresholdwould you use? Or maybe you prefer to target
all individuals with predicted amounts greater than $0.68 (why?), and/or…..]
Giventhe confidence probabilitiesof donors(TARGET_B=1) fromthe classificationmodel andthe
predicteddonationamounts(TARGET_D) fromthe regressionmodel we multiplythe twovaluesto
combine the twomodels. The multiplicationof the confidence (1) fromthe classificationmodel andthe
probable donationfromthe donorpredictedbythe linearregressionmodelgivesusthe mostprobable
donationbythe responder.
Hence withthe classificationmodel we onlypredictedhow manyindividualsare goingtorespondtoour
promotion,onthe otherhandthe regressionmodel givesusthe probable donationamountbythe
respondersobtainedfromthe classificationmodel.Hence combiningthe twomodelsplaysavery
importantrole.
The modellingtechniquewe usedisGradientBoostedTrees,whichhasanaccuracy of approximately
50%, hence we have selectedall those individualswhoare donating 2$ or more (as the accuracy is 50%,
we assume that outof all the predictedrespondersoutevery2one wouldsurelyrespond,hence we
wantthat one individual tocoverourpromotional chargeswhichwasdownscaledto -0.9928 $ for two
individualsi.e. 1.9856 $.
Hence, inorderto cover ourexpenses,we wouldneedindividualswhowoulddonate aminimum
amountof ~2 $.

5. Testing – chose one model, either the one from 2.1 or 2.2 above,based on performance on the test
data.
The file FutureFundraising.xlscontains the attributes for future mailing candidates. Using your “best”
model from Step 2 which of these candidates do you predict as donors and non-donors? List them in
descending order of probability of being a donor/prediction of donation amount. What cutoff do you
use? Submit this file (xls format), with your best model’s predictions (prob of being a donor).
For testingthe FutureFunraising.xlsfile,the bestmodel whichwe have usedis GradientBoostedTrees.
Afterapplyingthe model,the below are the predicted informationinferred
Numberof Donors : 10857
Numberof Non-Donors : 9143
Cumulative Profit : ~24604 $
The cut off usedto predict donor/non-donorsis the confidence ofpredicting1s = 0.496901
Note:
The listof predictionsusingthe bestmodel is uploadedtothe blackboard.(Future_Data_Prediction.xls).

Appendix-1
Python code used for cleaning the data-
import numpy as np
import pandas as pd
drop_column = ["ODATEDW", "OSOURCE", "TCODE", "ZIP", "MAILCODE", "PVASTATE", "DOB", "NOEXCH", "CLUSTER",
"AGEFLAG", "NUMCHLD", "HIT", "DATASRCE", "MALEMILI", "MALEVET", "VIETVETS", "WWIIVETS", "LOCALGOV",
"STATEGOV", "FEDGOV", "GEOCODE", "HHP1", "HHP2", "DW1", "DW2", "DW3", "DW4", "DW5", "DW6", "DW7", "DW8",
"DW9",
"HV1", "HV2", "HV3", "HV4", "HU1", "HU2", "HU3", "HU4", "HU5", "HHD1", "HHD2", "HHD3", "HHD4", "HHD5",
"HHD6",
"HHD7", "HHD1", "HHD2", "HHD3", "HHD4", "HHD5", "HHD6", "HHD7", "HHD8", "HHD9", "HHD10", "HHD11",
"HHD12", "HUR1",
"HUR2", "RHP1", "RHP2", "RHP3", "RHP4", "HUPA1", "HUPA2", "HUPA3", "HUPA4", "HUPA5", "HUPA6", "HUPA7",
"RP1",
"RP2", "RP3", "RP4", "MSA", "ADI", "DMA", "MC1", "MC2", "MC3", "TPE1", "TPE2", "TPE3", "TPE4", "TPE5",
"TPE6", "TPE7",
"TPE8", "TPE9", "PEC1", "PEC2", "TPE10", "TPE11", "TPE12","TPE13", "ANC1", "ANC2", "ANC3", "ANC4",
"ANC5", "ANC6",
"ANC7", "ANC8", "ANC9", "ANC10", "ANC11", "ANC12", "ANC13", "ANC14", "ANC15", "POBC1", "POBC2", "LSC1",
"LSC2",
"LSC3", "LSC4", "VOC1", "VOC2", "VOC3", "ADATE_2", "ADATE_3", "ADATE_4", "ADATE_5", "ADATE_6", "ADATE_7",
"ADATE_8",
"ADATE_9", "ADATE_10", "ADATE_11", "ADATE_12", "ADATE_13", "ADATE_14", "ADATE_15", "ADATE_16",
"ADATE_17",
"ADATE_18", "ADATE_19", "ADATE_20", "ADATE_21", "ADATE_22", "ADATE_23", "ADATE_24","MAXADATE",
"RDATE_3",
"RDATE_4","RDATE_5", "RDATE_6", "RDATE_7","RDATE_8", "RDATE_9", "RDATE_10", "RDATE_11", "RDATE_12",
"RDATE_13",
"RDATE_14", "RDATE_14", "RDATE_15", "RDATE_16", "RDATE_17", "RDATE_18", "RDATE_19", "RDATE_20",
"RDATE_21",
"RDATE_22", "RDATE_23", "RDATE_24", "MINRDATE","MAXRDATE", "LASTDATE", "FISTDATE","NEXTDATE",
"CONTROLN",
"TARGET_D", "HPHONE_D", "RFA_2R", "RFA_2F", "RFA_2A", "MDMAUD_R", "MDMAUD_F", "MDMAUD_A", "CLUSTER2",
"GEOCODE2", "MDMAUD"]
df = pd.read_csv("C:UserstyrionDocumentsIDS_572_notesassign2pvaBal35Trg.csv", sep=',', na_values=[' '],
low_memory=False)
df.drop(drop_column, axis=1, inplace=True)
list_string = []
non_list_string = []
# filling numeric columns with -1
for c in df.columns:
li = df[c].values.tolist()
a = np.asarray(li)
# print type(a)
flag = 0
for x in np.nditer(a):
si = x.tolist()
if x != "nan":
if type(si) != str:
flag = 1
break;
# filling NAN for all numeric entries
if flag == 1:
df[c].fillna(-1, inplace=True)
# print "was in"
# print c
else:
if df[c].isnull().values.any():
# print type(df[c])
list_string.append(c)
else:
non_list_string.append(c)
# print c
# replacing columns having "X" to 1 and "NaN" to 0
for val in list_string:
str_l = df[val].values.tolist()
a = np.asarray(str_l)
# print type(a)
flag = 0
for x in np.nditer(a):
if x == "X":
flag = 1
break
if x == "M":
flag = 2

break
if x == "Y":
flag = 3
break
if flag == 1:
# print val
df[val] = df[val].replace({'X': 1}, regex=False)
df[val].fillna(0, inplace=True)
if flag == 2:
df[val].fillna(-1, inplace=True)
df[val] = df[val].replace({'M': 1}, regex=False)
df[val] = df[val].replace({'F': 0}, regex=False)
if flag == 3:
df[val] = df[val].replace({'Y': 1}, regex=False)
df[val].fillna(-1, inplace=True)
if val == "HOMEOWNR":
df[val].fillna(0, inplace=True)
df[val] = df[val].replace({'H': 1}, regex=False)
new_attri = []
for val in list_string:
if val.find("RFA",0) == 0:
df[val].fillna("Z5Z", inplace=True)
r = val + "_R"
f = val + "_F"
c = val + "_C"
df[r] = df[val].str.extract('([FNALISZ])',expand = True)
df[f] = df[val].str.extract('(d)', expand=True)
df[c] = df[val].str.extract('[a-zA-Z][d]([a-zA-Z])', expand=True)
"""for che in range(1,9999,1):
if df[val].iloc[che] == "Z5Z": print che
print df[f].iloc[128]
"""
df[r] = df[r].replace({'F': 0,'N': 1,'A': 2,'L': 3,'I': 4,'S': 5,'Z': 6}, regex=False)
df[c] = df[c].replace({'A': 0, 'B': 1, 'C': 2, 'D': 3, 'E': 4, 'F': 5,'G': 6,'Z': 7}, regex=False)
new_attri.append(r)
new_attri.append(f)
new_attri.append(c)
del df[val]
val = "RFA_2"
df[val].fillna("Z5Z", inplace=True)
r = val + "_R"
f = val + "_F"
c = val + "_C"
df[r] = df[val].str.extract('([FNALISZ])',expand = True)
df[f] = df[val].str.extract('(d)', expand=True)
df[c] = df[val].str.extract('[a-zA-Z][d]([a-zA-Z])', expand=True)
"""for che in range(1,9999,1):
if df[val].iloc[che] == "Z5Z": print che
print df[f].iloc[128]
"""
df[r] = df[r].replace({'F': 0,'N': 1,'A': 2,'L': 3,'I': 4,'S': 5,'Z': 6}, regex=False)
df[c] = df[c].replace({'A': 0, 'B': 1, 'C': 2, 'D': 3, 'E': 4, 'F': 5, 'G': 6,'Z':7}, regex=False)
new_attri.append(r)
new_attri.append(f)
new_attri.append(c)
del df[val]
val = "DOMAIN"
df[val].fillna("Z5", inplace=True)
domain_att_1 = val + '_urban_level'
domain_att_2 = val + "_economic_status"
df[domain_att_1] = df[val].str.extract('([UCSTRZ])',expand = True)
df[domain_att_2] = df[val].str.extract('(d)', expand=True)
df[domain_att_1] = df[domain_att_1].replace({'U': 0,'C': 1,'S': 2,'T': 3,'R': 4,'Z':5}, regex=False)
# new_attri.append(domain_att_1)
# new_attri.append(domain_att_2)
del df[val]
# exporting the dataframe to csv
df.to_csv('cleaned_PVA_data.csv')
# NORMALIZE DATA
print np.count_nonzero(df.columns.values)
pca_variables = ["CHILD03", "CHILD07", "CHILD12", "CHILD18", "MBCRAFT", "MBGARDEN", "MBBOOKS", "MBCOLECT", "MAGFAML",
"MAGFEM", "MAGMALE", "PUBGARDN", "PUBCULIN", "PUBHLTH", "PUBDOITY", "PUBNEWFN", "PUBPHOTO", "PUBOPP",
"COLLECT1", "VETERANS", "BIBLE", "CATLG", "HOMEE", "PETS", "CDPLAY", "STEREO", "PCOWNERS", "PHOTO",
"CRAFTS", "FISHER", "GARDENIN", "BOATS", "WALKER", "KIDSTUFF", "CARDS", "PLATES", "LIFESRC",
"PEPSTRFL", "POP901", "POP902", "POP903", "POP90C1", "POP90C2", "POP90C3", "POP90C4", "POP90C5",
"ETH1","ETH2", "ETH3", "ETH4", "ETH5", "ETH6", "ETH7", "ETH8", "ETH9", "ETH10", "ETH11", "ETH12",
"ETH13","ETH14", "ETH15", "ETH16", "AGE901", "AGE902", "AGE903", "AGE904", "AGE905", "AGE906",
"AGE907", "CHIL1", "CHIL2", "CHIL3", "AGEC1", "AGEC2", "AGEC4", "AGEC5", "AGEC6", "AGEC7", "CHILC1",

"CHILC2", "CHILC3", "CHILC4", "CHILC5", "HHAGE1", "HHAGE2", "HHAGE3", "HHN1", "HHN2", "HHN3", "HHN3",
"HHN4", "HHN5", "HHN6", "MARR1", "MARR2", "MARR3", "MARR4", "ETHC1", "ETHC2", "ETHC3", "ETHC4",
"ETHC5", "ETHC6", "HVP1", "HVP2", "HVP3", "HVP4", "HVP5", "HVP6", "IC1", "IC2", "IC3", "IC4", "IC5",
"IC6", "IC7", "IC8", "IC9", "IC9", "IC9", "IC10", "IC11", "IC12", "IC13", "IC14", "IC15", "IC16",
"IC17", "IC18", "IC19", "IC20", "IC21", "IC22", "IC23", "HHAS1", "HHAS2", "HHAS3", "HHAS4", "LFC1",
"LFC2", "LFC3", "LFC4", "LFC5", "LFC6", "LFC7", "LFC8", "LFC9", "LFC10", "OCC1", "OCC2", "OCC3",
"OCC4", "OCC5", "OCC6", "OCC7", "OCC8", "OCC9", "OCC10", "OCC11", "OCC12", "OCC13", "EIC1", "EIC2",
"EIC3", "EIC4", "EIC5", "EIC6", "EIC7", "EIC8", "EIC9", "EIC10", "EIC11", "EIC12", "EIC13", "EIC14",
"EIC15", "EIC16", "OEDC1", "OEDC2", "OEDC3", "OEDC4", "OEDC5", "OEDC6", "OEDC7", "EC1", "EC2", "EC3",
"EC4", "EC5", "EC6", "EC7", "EC8", "SEC1", "SEC2", "SEC3", "SEC4", "AFC1", "AFC2", "AFC3", "AFC4",
"AFC5", "AFC6", "VC1", "VC2", "VC3", "VC4", "HC1", "HC2", "HC3", "HC4", "HC5", "HC6", "HC7", "HC8",
"HC9", "HC10", "HC11", "HC12", "HC13", "HC14", "HC15", "HC16", "HC17", "HC18", "HC19", "HC20", "HC21",
"MHUC1", "MHUC2", "AC1", "AC2", "RAMNT_3", "RAMNT_4", "RAMNT_5","RAMNT_6", "RAMNT_7", "RAMNT_8",
"RAMNT_9", "RAMNT_10", "RAMNT_11", "RAMNT_12", "RAMNT_13", "RAMNT_14","RAMNT_15", "RAMNT_16",
"RAMNT_17", "RAMNT_18", "RAMNT_19", "RAMNT_20", "RAMNT_21", "RAMNT_22", "RAMNT_23", "RAMNT_24",
"RAMNTALL", "NGIFTALL", "CARDGIFT","MINRAMNT", "MAXRAMNT", "TIMELAG","AVGGIFT", new_attri]
print len(pca_variables)
from sklearn import preprocessing
from sklearn.preprocessing import StandardScaler
# std_scale = preprocessing.StandardScaler().fit(df[pca_variables])
# df_std = std_scale.transform(df[pca_variables])
# minmax_scale = preprocessing.MinMaxScaler().fit(df[pca_variables])
# df_minmax = minmax_scale.transform(df[pca_variables])
"""
df1 = pd.DataFrame(index=range(0, 99999), columns=['A'], dtype='int')
for nd_series in df.columns:
print type(nd_series)
if nd_series in pca_variables:
df1.append(df[nd_series])
print df1.column
X_std = StandardScaler().fit_transform(df1)
mean_vec = np.mean(X_std, axis=0)
cov_mat = (X_std - mean_vec).T.dot((X_std - mean_vec)) / (X_std.shape[0]-1)
print('Covariance matrix n%s' %cov_mat)
cov_mat = np.cov(X_std.T)
eig_vals, eig_vecs = np.linalg.eig(cov_mat)
print('Eigenvectors n%s' %eig_vecs)
print('nEigenvalues n%s' %eig_vals)"""

Appendix-2
Process chart:
Fig.1 Classification modelling process Rapidminer.
Fig.2 Regression Modelling Process Rapidminer.

Classification modelling review

More Related Content

What's hot (20)

Similar to Classification modelling review (20)

Recently uploaded (20)

Classification modelling review