Course: MBA563 ************************************************************************
# load needed packages
library(tidyverse)## ── Attaching core tidyverse packages ──────────────────────── tidyverse 2.0.0 ──
## ✔ dplyr 1.1.4 ✔ readr 2.1.5
## ✔ forcats 1.0.0 ✔ stringr 1.5.1
## ✔ ggplot2 3.5.1 ✔ tibble 3.2.1
## ✔ lubridate 1.9.4 ✔ tidyr 1.3.1
## ✔ purrr 1.0.4
## ── Conflicts ────────────────────────────────────────── tidyverse_conflicts() ──
## ✖ dplyr::filter() masks stats::filter()
## ✖ dplyr::lag() masks stats::lag()
## ℹ Use the conflicted package (<http://conflicted.r-lib.org/>) to force all conflicts to become errors#install.packages('e1071')
library(e1071)
# Confusion Matrix function
my_confusion_matrix <- function(cf_table) {
true_positive <- cf_table[4]
true_negative <- cf_table[1]
false_positive <- cf_table[2]
false_negative <- cf_table[3]
accuracy <- (true_positive + true_negative) / (true_positive + true_negative + false_positive + false_negative)
sensitivity_recall <- true_positive / (true_positive + false_negative)
specificity_selectivity <- true_negative / (true_negative + false_positive)
precision <- true_positive / (true_positive + false_positive)
neg_pred_value <- true_negative/(true_negative + false_negative)
print(cf_table)
my_list <- list(sprintf("%1.0f = True Positive (TP), Hit", true_positive),
sprintf("%1.0f = True Negative (TN), Rejection", true_negative),
sprintf("%1.0f = False Positive (FP), Type 1 Error", false_positive),
sprintf("%1.0f = False Negative (FN), Type 2 Error", false_negative),
sprintf("%1.4f = Accuracy (TP+TN/(TP+TN+FP+FN))", accuracy),
sprintf("%1.4f = Sensitivity, Recall, Hit Rate, True Positive Rate (How many positives did the model get right? TP/(TP+FN))", sensitivity_recall),
sprintf("%1.4f = Specificity, Selectivity, True Negative Rate (How many negatives did the model get right? TN/(TN+FP))", specificity_selectivity),
sprintf("%1.4f = Precision, Positive Predictive Value (How good are the model's positive predictions? TP/(TP+FP))", precision),
sprintf("%1.4f = Negative Predictive Value (How good are the model's negative predictions? TN/(TN+FN)", neg_pred_value)
)
return(my_list)
}We will now us the KNN algorithm to address our business problem.
TECA is planning for the future and would like to set up their business so they are not so reliant on selling gas. One way to do that is to increase the sales of profitable products. Thus, TECA would like to predict when a transaction is a going to be a sale of a high gross profit margin product. Profit margin is the percentage of gross profit to revenue, or revenue minus costs divided by revenue. It is a percentage of how much profit each product makes or what percentage of profit is earned for each dollar of revenue.
The first thing we need to do is bring in the dataset we are going to
work with. This is TECA data that has been transformed to be used in our
KNN analysis. Let’s load it and examine it a bit. We first notice that
the dataset has 20,000 rows and 38 features. This data is aggregated at
the level of a purchase. Thus, each example or row is the purchase of
one item. The first feature is the target feature or the dependent
variable. This is the variable we are trying to predict. It is called
high_gpm. It is a factor that is either low or high. Next,
we see revenue. This variable is the amount of revenue TECA
makes for each purchase. Revenue is a continuous variable. Next, we have
four variables related to quarter. These are one-hot encoded or dummy
variables for the quarter of the year. For example, if the purchase
happened during the first quarter of the year, quarter.1
would have a 1 value and the other three quarters’ variables would have
a 0. In using the KNN algorithm we must all of the features, except the
target feature, as numbers. Thus, to we created these four dummy
variables from a one feature that called quarter that had
four different types of entries–‘quarter 1’, ‘quarter 2’, ‘quarter 3’,
and ‘quarter 4’. Next, income,
bachelors_degree, and population list these
averages for the location of the store where the purchase took place.
Next, we have another set of dummy variables–for each of the states TECA
operates in. Next, we have num_trans, which indicates how
many purchases were made by this same person as part of this
transaction. basket.no and basket.yes measure
whether the purchase was part of a multi-part transaction.
refill.no and refill.yes indicate whether or
not the purchase was a refill of fountain soda. The next 11 dummy
variables that all start with area, indicate the area of
the store the purchase was made, as follows: *
area.alcohol: products in this area include, for example,
wine and beer; * area.cooler: products in this area
include, for example, energy drinks, canned soda, and juice; *
area.dispensed: products in this area include, for example,
cold and hot dispensed (fountain) drinks; * area.fresh:
products in this area include, for example, pizza, hot sandwiches,
salads, and roller grill items; * area.fuel: products in
this area include, for example, gas; * area.grocery:
products in this area include, for example, milk, eggs, and cheese; *
area.lottery: products in this area include, for example,
lottery tickets; * area.miscellaneous: products in this
area include, for example, store services and coupons; *
area.nongrocery: products in this area include, for
example, clothing, magazines, medicine, and newspapers; *
area.snacks: products in this area include, for example,
candy, gum, chips, and salty snacks; and * area.tobacco:
products in this area include, for example, cigarettes and chewing
tobacco.
Next, items_sold indicate how many of these items were
purchased. Finally, the two loyalty dummies indicate whether the
purchase was made by a loyalty customer (a customer who scanned their
loyalty card) or not. Note, that the data was adjusted so that about
half of the purchases in this dataset were made by loyalty customers,
since TECA is particularly interested in these customers.
knn_input <- read_rds('knn_input.rds')
str(knn_input)## 'data.frame': 20000 obs. of 38 variables:
## $ high_gpm : Factor w/ 2 levels "low","high": 2 1 1 2 2 2 1 1 1 1 ...
## $ revenue : num 0.99 27.25 2.49 2.26 3 ...
## $ quarter.1 : int 1 0 1 1 0 0 0 0 0 0 ...
## $ quarter.2 : int 0 1 0 0 1 1 1 0 0 0 ...
## $ quarter.3 : int 0 0 0 0 0 0 0 0 1 0 ...
## $ quarter.4 : int 0 0 0 0 0 0 0 1 0 1 ...
## $ income : int 65268 58854 61115 45874 43547 41827 45874 66250 70963 61115 ...
## $ bachelors_degree : int 118 68 264 2488 502 169 2488 22 1959 264 ...
## $ population : int 1003 734 2048 30327 10615 2379 30327 567 16052 2048 ...
## $ state_province.Alabama : int 0 0 0 0 1 0 0 0 1 0 ...
## $ state_province.Arkansas : int 0 0 0 1 0 0 1 0 0 0 ...
## $ state_province.Colorado : int 0 0 0 0 0 0 0 0 0 0 ...
## $ state_province.Iowa : int 1 0 0 0 0 0 0 1 0 0 ...
## $ state_province.Minnesota : int 0 0 0 0 0 0 0 0 0 0 ...
## $ state_province.Missouri : int 0 0 0 0 0 0 0 0 0 0 ...
## $ state_province.Nebraska : int 0 0 0 0 0 0 0 0 0 0 ...
## $ state_province.Oklahoma : int 0 0 1 0 0 1 0 0 0 1 ...
## $ state_province.South Dakota: int 0 1 0 0 0 0 0 0 0 0 ...
## $ state_province.Wyoming : int 0 0 0 0 0 0 0 0 0 0 ...
## $ num_trans : int 1 1 1 1 1 1 1 1 1 1 ...
## $ basket.no : int 1 1 1 1 1 1 1 1 1 1 ...
## $ basket.yes : int 0 0 0 0 0 0 0 0 0 0 ...
## $ refill.no : int 1 1 1 1 1 1 1 1 1 1 ...
## $ refill.yes : int 0 0 0 0 0 0 0 0 0 0 ...
## $ area.alcohol : int 0 0 0 0 0 0 0 0 0 0 ...
## $ area.cooler : int 0 0 1 0 0 0 0 0 0 0 ...
## $ area.dispensed : int 0 0 0 0 0 1 0 0 0 0 ...
## $ area.fresh : int 1 0 0 1 1 0 0 0 0 0 ...
## $ area.fuel : int 0 1 0 0 0 0 1 1 0 1 ...
## $ area.grocery : int 0 0 0 0 0 0 0 0 1 0 ...
## $ area.lottery : int 0 0 0 0 0 0 0 0 0 0 ...
## $ area.miscellaneous : int 0 0 0 0 0 0 0 0 0 0 ...
## $ area.nongrocery : int 0 0 0 0 0 0 0 0 0 0 ...
## $ area.snacks : int 0 0 0 0 0 0 0 0 0 0 ...
## $ area.tobacco : int 0 0 0 0 0 0 0 0 0 0 ...
## $ items_sold : num 1 1 1 1 3 1 1 1 1 1 ...
## $ loyalty2.not loyal : int 1 1 1 1 1 1 1 1 1 1 ...
## $ loyalty2.loyal : int 0 0 0 0 0 0 0 0 0 0 ...
## - attr(*, "dummies")=List of 6
## ..$ quarter : int [1:4] 3 4 5 6
## ..$ state_province: int [1:10] 10 11 12 13 14 15 16 17 18 19
## ..$ basket : int [1:2] 21 22
## ..$ refill : int [1:2] 23 24
## ..$ area : int [1:11] 25 26 27 28 29 30 31 32 33 34 ...
## ..$ loyalty2 : int [1:2] 37 38slice_sample(knn_input, n=10)## high_gpm revenue quarter.1 quarter.2 quarter.3 quarter.4 income
## 1402640 low 8.00 1 0 0 0 45278
## 81954 low 1.69 0 0 1 0 43438
## 381626 high 1.00 1 0 0 0 58594
## 393438 low 1.79 1 0 0 0 43547
## 116217 low 1.00 0 0 0 1 37917
## 450690 low 54.34 0 0 0 1 89196
## 964038 low 5.79 0 0 1 0 37130
## 857079 low 3.59 0 0 1 0 70963
## 353406 low 5.98 0 0 0 1 58386
## 296850 high 2.00 0 1 0 0 60350
## bachelors_degree population state_province.Alabama
## 1402640 162 1741 0
## 81954 51 279 0
## 381626 1167 8624 0
## 393438 502 10615 1
## 116217 10 546 0
## 450690 90 423 0
## 964038 1238 15153 0
## 857079 1959 16052 1
## 353406 7602 39333 0
## 296850 4076 55532 0
## state_province.Arkansas state_province.Colorado state_province.Iowa
## 1402640 0 0 0
## 81954 0 1 0
## 381626 0 0 0
## 393438 0 0 0
## 116217 1 0 0
## 450690 0 0 0
## 964038 0 1 0
## 857079 0 0 0
## 353406 0 1 0
## 296850 0 1 0
## state_province.Minnesota state_province.Missouri
## 1402640 0 0
## 81954 0 0
## 381626 0 1
## 393438 0 0
## 116217 0 0
## 450690 0 0
## 964038 0 0
## 857079 0 0
## 353406 0 0
## 296850 0 0
## state_province.Nebraska state_province.Oklahoma
## 1402640 0 1
## 81954 0 0
## 381626 0 0
## 393438 0 0
## 116217 0 0
## 450690 0 1
## 964038 0 0
## 857079 0 0
## 353406 0 0
## 296850 0 0
## state_province.South Dakota state_province.Wyoming num_trans basket.no
## 1402640 0 0 1 1
## 81954 0 0 1 1
## 381626 0 0 1 1
## 393438 0 0 1 1
## 116217 0 0 1 1
## 450690 0 0 1 1
## 964038 0 0 1 1
## 857079 0 0 1 1
## 353406 0 0 1 1
## 296850 0 0 1 1
## basket.yes refill.no refill.yes area.alcohol area.cooler area.dispensed
## 1402640 0 1 0 0 0 0
## 81954 0 1 0 0 0 0
## 381626 0 1 0 0 0 1
## 393438 0 1 0 0 1 0
## 116217 0 1 0 0 0 0
## 450690 0 1 0 0 0 0
## 964038 0 1 0 0 0 0
## 857079 0 1 0 0 0 0
## 353406 0 1 0 0 1 0
## 296850 0 1 0 0 0 0
## area.fresh area.fuel area.grocery area.lottery area.miscellaneous
## 1402640 0 1 0 0 0
## 81954 0 0 0 0 0
## 381626 0 0 0 0 0
## 393438 0 0 0 0 0
## 116217 0 0 0 1 0
## 450690 0 1 0 0 0
## 964038 0 0 0 0 0
## 857079 0 0 0 0 0
## 353406 0 0 0 0 0
## 296850 1 0 0 0 0
## area.nongrocery area.snacks area.tobacco items_sold loyalty2.not loyal
## 1402640 0 0 0 1 1
## 81954 0 1 0 1 0
## 381626 0 0 0 1 0
## 393438 0 0 0 1 1
## 116217 0 0 0 1 0
## 450690 0 0 0 1 0
## 964038 0 0 1 1 1
## 857079 0 0 1 1 1
## 353406 0 0 0 2 0
## 296850 0 0 0 2 0
## loyalty2.loyal
## 1402640 0
## 81954 1
## 381626 1
## 393438 0
## 116217 1
## 450690 1
## 964038 0
## 857079 0
## 353406 1
## 296850 1Let’s look more in depth at the target feature. For KNN analysis, the target feature is a categorical variable. In this implementation we can leave it as a factor. About 44% of these purchases are for high gross profit items.
freq <- table(knn_input$high_gpm)
freq[2]/(freq[1]+freq[2])## high
## 0.43695contrasts(knn_input$high_gpm)## high
## low 0
## high 1Before using the algorithm, we need to prepare the data. The first
line below loads the caret package. The next line sets the
seed for the randomization that will be used for the algorithm. The
caret::createDataPartition() function uses the caret
package to split the data. Basically, it creates a number for
p amount, in this case 0.75, of the target feature and
lists these numbers in a matrix (since list is
FALSE). Next, the training data and testing data are
created. These use the numbers from the partition matrix
that we just created. data_train retains each of the rows
with numbers in partition while data_test
takes the numbers not in partition, i.e., -partition.
library(caret)## Loading required package: lattice##
## Attaching package: 'caret'## The following object is masked from 'package:purrr':
##
## liftset.seed(77)
partition <- caret::createDataPartition(y=knn_input$high_gpm, p=.75, list=FALSE)
data_train <- knn_input[partition, ]
data_test <- knn_input[-partition, ]Next, we need to remove the dependent variable, high_gpm
to create testing data and training data without the target variable,
y and training and testing data that is just the dependent
variable.
X_train <- data_train %>% select(-high_gpm)
X_test <- data_test %>% select(-high_gpm)
y_train <- data_train$high_gpm
y_test <- data_test$high_gpmNext, let’s standardize our data. Recall, that KNN uses a distance function to determine its nearest neighbor. Thus, we need all of the variables to be on similar scales. This function creates matrices that have variables that have been standardized with z-score standardization. The mean of each feature is subtracted from each individual feature and then divided by that feature’s standard deviation. This transformation rescales the feature such it has a mean of zero and a standard deviation of one. Thus, it is measured in how many standard deviations it falls above or below its mean.
X_train <- scale(X_train)
X_test <- scale(X_test)Let’s just double check that our training data is the correct percentage and that the size of the training input variables and the training output variables are the same.
nrow(X_train)/(nrow(X_test)+nrow(X_train))## [1] 0.75005dim(X_train)## [1] 15001 37length(y_train)## [1] 15001Finally, let’s run the model. Please remember that this algorithm works differently from some of the others you will see. It is what we call lazy because it does not create a model, but rather stores all of the training data and then uses it on the test data all in one step. Thus, the training step and the prediction step are combined into one step.
To implement KNN we use one function that accepts four arguments: *
The training data with the label’s column/target variable removed * The
testing data with the label’s column/target variable removed * The class
labels (cl)/target variable for the training data (but not
the testing data, since that is what is being predicted), and * The K
that we select.
How did I select this k? There is no one right way to do this, but I just took the square root of the total number of rows.
knn1 is a feature vector of predicted labels for each of
the rows/examples in the test data.
library(class)
knn1 = class::knn(train=X_train, test=X_test, cl=y_train, k=141)Finally, let’s check to see how accurate our model is. We will use the function created at the beginning of the notebook. Overall, our model is very accurate. The model makes the correct prediction about 83% of the time! Recall, that KNN has used the training data to predict on the testing data. Thus, there is some assurance that the model would predict well on new data.
Let’s explore some details of the accuracy of this model. The table shows the following output: * When a transaction is actually/truthfully a high gross profit margin transaction, the model correctly classifies it as such-by saying “high”, 1700 times. This is called a True Positive (TP), Hit.
When a transaction is truthfully a low gross profit margin transaction, the model correctly classifies it as such–by saying “low”, 2448 times. This is a True Negative (TN), Rejection.
On the other hand, the model makes two kinds of errors. When the model transaction is not a high gross profit margin transaction, but the model incorrectly says it is, this is called a False Positive (FP), Type 1 Error. It happens 367 times.
Finally, when a transaction is a high gross profit margin transaction and the model says it is not, which here happens 484, it is called a False Negative (FN), Type 2 Error.
These numbers can then be manipulated to create different measures of accuracy, as follows: * Overall accuracy (TP+TN/(TP+TN+FP+FN)) is 0.8298.
Sensitivity, Recall, Hit Rate, True Positive Rate (How many positives did the model get right? TP/(TP+FN)), is 0.7784.
Specificity, Selectivity, True Negative Rate (How many negatives did the model get right? TN/(TN+FP)) is 0.8696.
Precision, Positive Predictive Value (How good are the model’s positive predictions? TP/(TP+FP)) is 0.8224.
Negative Predictive Value (How good are the model’s negative predictions? TN/(TN+FN) is 0.8349.
But what does this all tell us? Overall, our model does a great job at predicting when a purchase will be high profit margin versus low profit margin and gets it right 83% of the time. The aspect of the model that is the least effective is its sensitivity and the aspect that is the most effective is the specificity. This means that the models is very good at classifying the low margin purchases successfully (specificity), but slightly worse at classifying the high margin purchases correctly (sensitivity). Thus, while the model is quite good overall, it particularly excels at identifying bad transactions.
table2 <- table(knn1, y_test) #prediction on left and truth on top
my_confusion_matrix(table2)## y_test
## knn1 low high
## low 2448 484
## high 367 1700## [[1]]
## [1] "1700 = True Positive (TP), Hit"
##
## [[2]]
## [1] "2448 = True Negative (TN), Rejection"
##
## [[3]]
## [1] "367 = False Positive (FP), Type 1 Error"
##
## [[4]]
## [1] "484 = False Negative (FN), Type 2 Error"
##
## [[5]]
## [1] "0.8298 = Accuracy (TP+TN/(TP+TN+FP+FN))"
##
## [[6]]
## [1] "0.7784 = Sensitivity, Recall, Hit Rate, True Positive Rate (How many positives did the model get right? TP/(TP+FN))"
##
## [[7]]
## [1] "0.8696 = Specificity, Selectivity, True Negative Rate (How many negatives did the model get right? TN/(TN+FP))"
##
## [[8]]
## [1] "0.8224 = Precision, Positive Predictive Value (How good are the model's positive predictions? TP/(TP+FP))"
##
## [[9]]
## [1] "0.8349 = Negative Predictive Value (How good are the model's negative predictions? TN/(TN+FN)"The above confusion matrix was made from our function above. It is
helpful because I put in some extra text that helps interpret the
results. R, of course, has several packages that will build the
confusion matrix for you. Here is one. Note, that we did need to specify
which of our levels is the “Positive Class”, since this normally takes
the first level of the variable, which is our case is low,
which is the “negative” class, or the wrong thing, rather than the thing
we are predicting. One thing that is provided here is a significance
test of accuracy. As you can see from the very low P-Value
and Mcnemar's Test P-Value, our model is significant.
caret::confusionMatrix(knn1, y_test, positive='high')## Confusion Matrix and Statistics
##
## Reference
## Prediction low high
## low 2448 484
## high 367 1700
##
## Accuracy : 0.8298
## 95% CI : (0.8191, 0.8401)
## No Information Rate : 0.5631
## P-Value [Acc > NIR] : < 2.2e-16
##
## Kappa : 0.6519
##
## Mcnemar's Test P-Value : 6.996e-05
##
## Sensitivity : 0.7784
## Specificity : 0.8696
## Pos Pred Value : 0.8224
## Neg Pred Value : 0.8349
## Prevalence : 0.4369
## Detection Rate : 0.3401
## Detection Prevalence : 0.4135
## Balanced Accuracy : 0.8240
##
## 'Positive' Class : high
## Finally, we can put the prediction back into the test data and
compare when our model is and is not accurate. The first line of the
code below adds the predicted low or high gross profit margin back to
the test data. The second line creates a new variable called
correct that takes on the value TRUE when the
model was correct and FALSE when the model was incorrect.
Printing out a sample of this dataframe and scrolling through it gives
us the ability to investigate where the model went wrong.
It also gives us some insight into TECA’s problem. Just doing a quick scan, it seems like promoting certain areas, such as fresh and dispensed might help increase the sale of high profit margin products, while other areas, such as lottery, might not help. However, are efforts here were to make a prediction model. To examine the relevance of the factors we would need to employee another method, such as logistic regression or decision trees.
data_test$prediction <- knn1
data_test <- data_test %>% mutate(correct = high_gpm==prediction)
slice_sample(data_test, n=20)## high_gpm revenue quarter.1 quarter.2 quarter.3 quarter.4 income
## 209931 high 1.79 0 0 1 0 34210
## 1403545 low 26.91 0 0 0 1 29091
## 1226812 low 10.00 1 0 0 0 42772
## 428400 low 5.00 1 0 0 0 66250
## 417907 low 21.00 1 0 0 0 76071
## 38464 high 0.99 0 1 0 0 52500
## 1612610 high 1.00 0 0 0 1 29091
## 1224527 low 45.25 0 0 0 1 70557
## 879815 high 1.98 0 0 1 0 66160
## 318760 low 7.03 0 1 0 0 37130
## 1242817 low 26.90 0 0 0 1 50819
## 382899 high 1.69 0 0 1 0 66250
## 118951 low 13.98 1 0 0 0 41827
## 855741 low 4.75 0 1 0 0 70963
## 1551687 high 2.69 0 0 1 0 66250
## 1201431 low 20.00 0 0 0 1 51461
## 471531 high 7.56 0 0 1 0 63673
## 15110 low 2.13 0 0 0 1 53026
## 1555083 high 1.30 1 0 0 0 132230
## 417764 low 44.64 0 0 1 0 73148
## bachelors_degree population state_province.Alabama
## 209931 1632 25999 1
## 1403545 3 181 0
## 1226812 160 1876 0
## 428400 22 567 0
## 417907 93 493 0
## 38464 95 1192 0
## 1612610 3 181 0
## 1224527 4048 51863 0
## 879815 10178 57364 0
## 318760 1238 15153 0
## 1242817 958 8778 0
## 382899 22 567 0
## 118951 169 2379 0
## 855741 1959 16052 1
## 1551687 22 567 0
## 1201431 2922 52172 0
## 471531 1172 9046 0
## 15110 19 557 0
## 1555083 974 3294 0
## 417764 139 852 0
## state_province.Arkansas state_province.Colorado state_province.Iowa
## 209931 0 0 0
## 1403545 0 0 0
## 1226812 0 0 0
## 428400 0 0 1
## 417907 0 0 0
## 38464 0 0 0
## 1612610 0 0 0
## 1224527 0 1 0
## 879815 0 0 0
## 318760 0 1 0
## 1242817 0 0 0
## 382899 0 0 1
## 118951 0 0 0
## 855741 0 0 0
## 1551687 0 0 1
## 1201431 0 1 0
## 471531 0 0 1
## 15110 0 0 1
## 1555083 1 0 0
## 417764 0 0 1
## state_province.Minnesota state_province.Missouri
## 209931 0 0
## 1403545 0 0
## 1226812 0 0
## 428400 0 0
## 417907 0 0
## 38464 0 0
## 1612610 0 0
## 1224527 0 0
## 879815 0 1
## 318760 0 0
## 1242817 1 0
## 382899 0 0
## 118951 0 0
## 855741 0 0
## 1551687 0 0
## 1201431 0 0
## 471531 0 0
## 15110 0 0
## 1555083 0 0
## 417764 0 0
## state_province.Nebraska state_province.Oklahoma
## 209931 0 0
## 1403545 0 1
## 1226812 0 1
## 428400 0 0
## 417907 1 0
## 38464 1 0
## 1612610 0 1
## 1224527 0 0
## 879815 0 0
## 318760 0 0
## 1242817 0 0
## 382899 0 0
## 118951 0 1
## 855741 0 0
## 1551687 0 0
## 1201431 0 0
## 471531 0 0
## 15110 0 0
## 1555083 0 0
## 417764 0 0
## state_province.South Dakota state_province.Wyoming num_trans basket.no
## 209931 0 0 1 1
## 1403545 0 0 1 1
## 1226812 0 0 1 1
## 428400 0 0 1 1
## 417907 0 0 1 1
## 38464 0 0 1 1
## 1612610 0 0 1 1
## 1224527 0 0 1 1
## 879815 0 0 1 1
## 318760 0 0 1 1
## 1242817 0 0 1 1
## 382899 0 0 1 1
## 118951 0 0 1 1
## 855741 0 0 1 1
## 1551687 0 0 1 1
## 1201431 0 0 1 1
## 471531 0 0 1 1
## 15110 0 0 1 1
## 1555083 0 0 1 1
## 417764 0 0 1 1
## basket.yes refill.no refill.yes area.alcohol area.cooler area.dispensed
## 209931 0 1 0 0 0 0
## 1403545 0 1 0 0 0 0
## 1226812 0 1 0 0 0 0
## 428400 0 1 0 0 0 0
## 417907 0 1 0 0 0 0
## 38464 0 1 0 0 1 0
## 1612610 0 1 0 0 0 1
## 1224527 0 1 0 0 0 0
## 879815 0 1 0 0 0 0
## 318760 0 1 0 0 0 0
## 1242817 0 1 0 0 0 0
## 382899 0 1 0 0 0 1
## 118951 0 1 0 0 0 0
## 855741 0 1 0 0 0 0
## 1551687 0 1 0 0 0 0
## 1201431 0 1 0 0 0 0
## 471531 0 1 0 0 1 0
## 15110 0 1 0 0 1 0
## 1555083 0 1 0 0 0 0
## 417764 0 1 0 0 0 0
## area.fresh area.fuel area.grocery area.lottery area.miscellaneous
## 209931 1 0 0 0 0
## 1403545 0 1 0 0 0
## 1226812 0 1 0 0 0
## 428400 0 1 0 0 0
## 417907 0 1 0 0 0
## 38464 0 0 0 0 0
## 1612610 0 0 0 0 0
## 1224527 0 1 0 0 0
## 879815 0 0 0 0 0
## 318760 0 0 0 0 0
## 1242817 0 1 0 0 0
## 382899 0 0 0 0 0
## 118951 0 0 0 0 0
## 855741 0 0 0 0 0
## 1551687 1 0 0 0 0
## 1201431 0 1 0 0 0
## 471531 0 0 0 0 0
## 15110 0 0 0 0 0
## 1555083 1 0 0 0 0
## 417764 0 1 0 0 0
## area.nongrocery area.snacks area.tobacco items_sold loyalty2.not loyal
## 209931 0 0 0 1 1
## 1403545 0 0 0 1 1
## 1226812 0 0 0 1 1
## 428400 0 0 0 1 1
## 417907 0 0 0 1 0
## 38464 0 0 0 1 0
## 1612610 0 0 0 1 1
## 1224527 0 0 0 1 1
## 879815 0 1 0 2 1
## 318760 0 0 1 1 0
## 1242817 0 0 0 1 1
## 382899 0 0 0 1 0
## 118951 0 0 1 2 1
## 855741 0 0 1 1 1
## 1551687 0 0 0 2 1
## 1201431 0 0 0 1 1
## 471531 0 0 0 4 1
## 15110 0 0 0 1 0
## 1555083 0 0 0 1 1
## 417764 0 0 0 1 0
## loyalty2.loyal prediction correct
## 209931 0 high TRUE
## 1403545 0 low TRUE
## 1226812 0 low TRUE
## 428400 0 low TRUE
## 417907 1 low TRUE
## 38464 1 low FALSE
## 1612610 0 high TRUE
## 1224527 0 low TRUE
## 879815 0 high TRUE
## 318760 1 low TRUE
## 1242817 0 low TRUE
## 382899 1 high TRUE
## 118951 0 low TRUE
## 855741 0 low TRUE
## 1551687 0 high TRUE
## 1201431 0 low TRUE
## 471531 0 low FALSE
## 15110 1 low TRUE
## 1555083 0 high TRUE
## 417764 1 low TRUEFinally, note that we are not covering advanced topics such as
hyperparameter tuning. When you gain more experience, you might want to
examine methods for picking k, cross validation, etc. Our
goal is to provide you the framework to understand the algorithm and
start working with it.