Santander_Transaction_Prediction
Viditya
22/12/2019
1. Introduction
1.1 Problem Statement
This is a Project Report for the Santander Customer Transaction Prediction Project. The objective of this Case is to help people and businesses prosper. We are always looking for ways to help our customers understand their financial health and identify which products and services might help them achieve their monetary goals
1.2 Data
The goal is to build regression models which will predict whether a customer is satisfied? Will a customer buy the product? Can a customer pay the loan?
## [[1]]
## [1] TRUE
##
## [[2]]
## [1] TRUE
##
## [[3]]
## [1] TRUE
##
## [[4]]
## [1] TRUE
##
## [[5]]
## [1] TRUE
##
## [[6]]
## [1] TRUE
##
## [[7]]
## [1] TRUE
##
## [[8]]
## [1] TRUE
##
## [[9]]
## [1] TRUE
##
## [[10]]
## [1] TRUE## [1] 200000 202## [1] 200000 201## 'data.frame': 200000 obs. of 202 variables:
## $ ID_code: Factor w/ 200000 levels "train_0","train_1",..: 1 2 111113 122224 133335 144446 155557 166668 177779 188890 ...
## $ target : int 0 0 0 0 0 0 0 0 0 0 ...
## $ var_0 : num 8.93 11.5 8.61 11.06 9.84 ...
## $ var_1 : num -6.79 -4.15 -2.75 -2.15 -1.48 ...
## $ var_2 : num 11.91 13.86 12.08 8.95 12.87 ...
## $ var_3 : num 5.09 5.39 7.89 7.2 6.64 ...
## $ var_4 : num 11.5 12.4 10.6 12.6 12.3 ...
## $ var_5 : num -9.28 7.04 -9.08 -1.84 2.45 ...
## $ var_6 : num 5.12 5.62 6.94 5.84 5.94 ...
## $ var_7 : num 18.6 16.5 14.6 14.9 19.3 ...
## $ var_8 : num -4.92 3.15 -4.92 -5.86 6.27 ...
## $ var_9 : num 5.75 8.09 5.95 8.24 7.68 ...
## $ var_10 : num 2.925 -0.403 -0.325 2.306 -9.446 ...
## $ var_11 : num 3.18 8.06 -11.26 2.81 -12.14 ...
## $ var_12 : num 14 14 14.2 13.8 13.8 ...
## $ var_13 : num 0.575 8.414 7.312 11.97 7.889 ...
## $ var_14 : num 8.8 5.43 7.52 6.46 7.79 ...
## $ var_15 : num 14.6 13.7 14.6 14.8 15.1 ...
## $ var_16 : num 5.75 13.83 7.68 10.74 8.49 ...
## $ var_17 : num -7.24 -15.58 -1.74 -0.43 -3.07 ...
## $ var_18 : num 4.28 7.8 4.7 15.94 6.53 ...
## $ var_19 : num 30.7 28.6 20.5 13.7 11.3 ...
## $ var_20 : num 10.54 3.43 17.76 20.3 21.42 ...
## $ var_21 : num 16.22 2.74 18.14 12.56 18.96 ...
## $ var_22 : num 2.58 8.55 1.21 6.82 10.11 ...
## $ var_23 : num 2.47 3.37 3.51 2.72 2.71 ...
## $ var_24 : num 14.38 6.98 5.68 12.14 14.21 ...
## $ var_25 : num 13.4 13.9 13.2 13.7 13.5 ...
## $ var_26 : num -5.149 -11.768 -7.994 0.814 3.174 ...
## $ var_27 : num -0.407 -2.559 -2.903 -0.906 -3.342 ...
## $ var_28 : num 4.93 5.05 5.85 5.91 5.9 ...
## $ var_29 : num 5.997 0.548 6.144 2.841 7.935 ...
## $ var_30 : num -0.308 -9.299 -11.102 -15.24 -3.158 ...
## $ var_31 : num 12.9 7.88 12.49 10.44 9.47 ...
## $ var_32 : num -3.8766 1.2859 -2.2871 -2.5731 -0.0083 ...
## $ var_33 : num 16.89 19.37 19.04 6.18 19.32 ...
## $ var_34 : num 11.2 11.4 11 10.6 12.4 ...
## $ var_35 : num 10.579 0.74 4.109 -5.916 0.633 ...
## $ var_36 : num 0.676 2.8 4.697 8.172 2.792 ...
## $ var_37 : num 7.89 5.84 6.93 2.85 5.82 ...
## $ var_38 : num 4.67 10.82 10.89 9.17 19.3 ...
## $ var_39 : num 3.874 3.678 0.9 0.666 1.445 ...
## $ var_40 : num -5.24 -11.11 -13.52 -3.83 -5.6 ...
## $ var_41 : num 7.37 1.87 2.24 -1.04 14.07 ...
## $ var_42 : num 11.58 9.88 11.53 11.78 11.92 ...
## $ var_43 : num 12 11.8 12 11.3 11.5 ...
## $ var_44 : num 11.64 1.24 4.1 8.05 6.91 ...
## $ var_45 : num -7.02 -47.38 -7.91 -24.68 -65.49 ...
## $ var_46 : num 5.92 7.37 11.14 12.74 13.87 ...
## $ var_47 : num -14.2136 0.1948 -5.7864 -35.1659 0.0444 ...
## $ var_48 : num 16.028 34.401 20.748 0.761 -0.135 ...
## $ var_49 : num 5.33 25.7 6.89 8.38 14.43 ...
## $ var_50 : num 12.9 11.8 12.9 12.7 13.3 ...
## $ var_51 : num 29.05 13.23 19.59 9.55 10.49 ...
## $ var_52 : num -0.694 -4.108 0.727 1.79 -1.437 ...
## $ var_53 : num 5.17 6.69 6.41 5.21 5.76 ...
## $ var_54 : num -0.747 -8.095 9.312 8.091 -8.541 ...
## $ var_55 : num 14.83 18.6 6.28 12.4 14.15 ...
## $ var_56 : num 11.3 19.3 15.6 14.5 17 ...
## $ var_57 : num 5.38 7.01 5.82 6.58 6.18 ...
## $ var_58 : num 2.02 1.92 1.1 3.32 1.95 ...
## $ var_59 : num 10.12 8.87 9.19 9.46 9.2 ...
## $ var_60 : num 16.18 8.01 12.6 15.78 8.66 ...
## $ var_61 : num 4.96 -7.24 -10.37 -25.02 -27.74 ...
## $ var_62 : num 2.077 1.794 0.875 3.442 -0.495 ...
## $ var_63 : num -0.215 -1.315 5.804 -4.392 -1.784 ...
## $ var_64 : num 8.67 8.1 3.72 8.65 5.27 ...
## $ var_65 : num 9.53 1.54 -1.1 6.31 -4.32 ...
## $ var_66 : num 5.81 5.4 7.37 5.62 6.99 ...
## $ var_67 : num 22.43 7.93 9.86 23.61 1.62 ...
## $ var_68 : num 5.01 5.02 5.02 5.02 5.03 ...
## $ var_69 : num -4.7 2.23 -5.78 -4 -3.24 ...
## $ var_70 : num 21.64 40.56 2.36 4.05 40.12 ...
## $ var_71 : num 0.566 0.513 0.852 0.25 0.774 ...
## $ var_72 : num 5.2 3.17 6.358 1.252 -0.726 ...
## $ var_73 : num 8.86 20.11 12.17 24.42 4.59 ...
## $ var_74 : num 43.11 7.78 19.73 4.53 -4.53 ...
## $ var_75 : num 18.38 7.05 19.45 15.42 23.35 ...
## $ var_76 : num -2.34 3.27 4.5 11.69 1.03 ...
## $ var_77 : num 23.4 23.5 23.2 23.6 19.2 ...
## $ var_78 : num 6.52 5.51 6.32 4.08 7.17 ...
## $ var_79 : num 12.2 13.8 12.8 15.3 14.4 ...
## $ var_80 : num 13.647 2.546 7.473 0.784 2.96 ...
## $ var_81 : num 13.8 18.2 15.8 10.5 13.3 ...
## $ var_82 : num 1.367 0.368 13.353 1.621 -9.259 ...
## $ var_83 : num 2.94 -4.82 10.19 -5.29 -6.71 ...
## $ var_84 : num -4.52 -5.49 5.46 1.6 7.9 ...
## $ var_85 : num 21.5 13.8 19.1 18 14.5 ...
## $ var_86 : num 9.32 -13.59 -4.46 -2.32 7.08 ...
## $ var_87 : num 16.46 11.1 9.54 15.63 20.17 ...
## $ var_88 : num 8 7.9 11.91 4.55 8.01 ...
## $ var_89 : num -1.71 12.23 2.14 7.55 3.8 ...
## $ var_90 : num -21.449 0.477 -22.404 -7.587 -39.8 ...
## $ var_91 : num 6.78 6.89 7.09 7.04 7.01 ...
## $ var_92 : num 11.09 8.09 14.16 14.4 9.36 ...
## $ var_93 : num 9.99 10.96 10.51 10.78 10.43 ...
## $ var_94 : num 14.84 11.76 14.26 7.29 14.06 ...
## $ var_95 : num 0.1812 -1.2722 0.2647 -1.093 0.0213 ...
## $ var_96 : num 8.96 24.79 20.4 11.36 14.72 ...
## [list output truncated]2. Methods used
2.1 Pre processing
Any predictive modeling requires that we look at the data before we start modeling. However, in data mining terms looking at data refers to so much more than just looking. Looking at data refers to exploring the data, cleaning the data as well as visualizing the data through graphs and plots. This is often called as Exploratory Data Analysis. To start this process we will first try and look at all the probability distributions of the variables. Most analysis like regression, require the data to be normally distributed. We can visualize that in a glance by looking at the probability distributions or probability density functions of the variable.
## [1] 02.2 Distribution of all types of variables
We check for the distribution shown by various variables. For this, we draw various plots like Bargraph, Histogram, Scatterplot an Boxplot etc.
2.3 Check for outliers
The next step is to check for outliers. Outliers are those points in the data that are not following the pattern shown by other data poins. These outliers need to be removed from the data.
##
## 0 1
## 179902 20098##
## 0 1
## 89.951 10.049
#Distribution of feature variables
From the above layout, we can select any of the independennt variables and see corresponding relations between the train and train datasets
Distribution of mean values per row and column in train and test dataset
Distribution of standard deviation values per row and column in train and test dataset
Distribution of skewness values per row and column in train and test dataset
Correlations Analysis
Correlations in train data
correlation between the train attributes is very small.
Correlations in train data
correlation between the test attributes is also very small.
###############################################Feature Engineering############################################## ## 2.4 Feature Selection
Before performing any type of modeling we need to assess the importance of each predictor variable in our analysis. There is a possibility that many variables in our analysis are not important at all to the problem of class prediction. There are several methods of doing that. Below we have used Random Forests to perform features selection.
As we could not get any relevant informatkion from correlation analysis and our objective is to find out the variables that can stand out so that we can concentrate on those variables only. So, we will take a sample of data and try to find out value importance.
## [1] 150000 202## [1] 50000 202Random forest classifier
Feature importance by random forest
Here the features are sorted by their importance and we can observe that the most significant variable is var_81
We can also visualise using the plot
Impact of the main features which are discovered in the previous section by using pdp package.
Performing models on data
3. Model Selection
The dependent variable can fall in either of the four categories:
1. Nominal
2. Ordinal
3. Interval
4. Ratio
If the dependent variable, in our case Quality, is Nominal the only predictive analysis that we can perform is Classification, and if the dependent variable is Interval or Ratio the normal method is to do a Regression analysis, or classification after binning.
#Splitting the data
## [1] 160000 202## [1] 40000 202Training and validation dataset
#Logistic Regression
## Length Class Mode
## a0 61 -none- numeric
## beta 12200 dgCMatrix S4
## df 61 -none- numeric
## dim 2 -none- numeric
## lambda 61 -none- numeric
## dev.ratio 61 -none- numeric
## nulldev 1 -none- numeric
## npasses 1 -none- numeric
## jerr 1 -none- numeric
## offset 1 -none- logical
## classnames 2 -none- character
## call 4 -none- call
## nobs 1 -none- numericCross validation prediction
##
## Call: cv.glmnet(x = X_t, y = y_t, type.measure = "class", family = "binomial")
##
## Measure: Misclassification Error
##
## Lambda Measure SE Nonzero
## min 0.0003163 0.08590 0.0005722 195
## 1se 0.0008800 0.08632 0.0006442 184Plotting the missclassification error vs log(lambda) where lambda is regularization parameter
## [1] 0.000316272
# Observation ## Miss classification error increases as increasing the log(Lambda).
Model performance on validation dataset
Creating Confusion Matrix for Logistic Regression
## Confusion Matrix and Statistics
##
## Reference
## Prediction 1 2
## 1 35618 2973
## 2 434 975
##
## Accuracy : 0.9148
## 95% CI : (0.912, 0.9175)
## No Information Rate : 0.9013
## P-Value [Acc > NIR] : < 2.2e-16
##
## Kappa : 0.3292
##
## Mcnemar's Test P-Value : < 2.2e-16
##
## Sensitivity : 0.9880
## Specificity : 0.2470
## Pos Pred Value : 0.9230
## Neg Pred Value : 0.6920
## Prevalence : 0.9013
## Detection Rate : 0.8904
## Detection Prevalence : 0.9648
## Balanced Accuracy : 0.6175
##
## 'Positive' Class : 1
##