Business Problem

Companies often spend significant resources on marketing campaigns without fully understanding which customers are most likely to respond. Sending the same promotions to all customers can result in low response rates and inefficient use of marketing budgets.

This project seeks to identify customer segments based on demographics and purchasing behavior and determine which characteristics are associated with successful campaign responses.

Objective

The goal of this project is to use customer analytics to better understand customer behavior, specifically their spending patterns to improve market decision-making.

Research Questions

What demographic characteristics are associated with higher customer spending?
What item(s) do they typically buy?
Can customers be segmented into meaningful groups based on purchasing behavior?
Which customer characteristics are associated with positive campaign responses?
How can marketers improve campaign targeting using customer segmentation?

Managerial Questions

Which customers should receive future marketing promotions?
Tailoring Marketing Campaigns
1. Is there a way to tailor marketing to customer needs/spending behavior?
2. How can we allocate marketing resources efficiently?
Customer Groups

Which customer groups generate the highest revenue?
Which customer group has the highest potential for growth?

Analytics Methods

Descriptive Analytics

Summary Statistics
Customer Demographics
Spending Behavior *** ### Predictive Analytics
Predict campaign response likelihood

Introduction

Customer relationship management and targeted marketing have become increasingly important in today’s competitive business environment. Organizations that understand customer behavior can better allocate marketing resources and improve campaign effectiveness. The Customer Personality Analysis dataset provides demographic, purchasing, and campaign response information that can be used to better understand customer preferences and marketing outcomes.

The purpose of this project is to analyze customer characteristics and purchasing behavior to identify meaningful customer segments and evaluate factors associated with marketing campaign success. Using customer segmentation and predictive analytics techniques, this study seeks to provide actionable recommendations that improve marketing targeting and increase customer engagement.

Exploratory Data Analysis

The goal of this exploratory data analysis is to understand customer demographics, income, spending behavior, purchase channels, and marketing campaign responses.

This EDA will generate summary statistics, visualizations, trends, patterns, anomalies, and business insights for the Customer Personality Analysis dataset.

Pre-Processing

## ── Attaching core tidyverse packages ──────────────────────── tidyverse 2.0.0 ──
## ✔ dplyr     1.2.1     ✔ readr     2.2.0
## ✔ forcats   1.0.1     ✔ stringr   1.6.0
## ✔ ggplot2   4.0.3     ✔ tibble    3.3.1
## ✔ lubridate 1.9.5     ✔ tidyr     1.3.2
## ✔ purrr     1.2.2     
## ── 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
## 
## Attaching package: 'janitor'
## 
## 
## The following objects are masked from 'package:stats':
## 
##     chisq.test, fisher.test
## 
## 
## 
## Attaching package: 'naniar'
## 
## 
## The following object is masked from 'package:skimr':
## 
##     n_complete
## 
## 
## corrplot 0.95 loaded
## 
## 
## Attaching package: 'kableExtra'
## 
## 
## The following object is masked from 'package:dplyr':
## 
##     group_rows
## 
## 
## 
## Attaching package: 'scales'
## 
## 
## The following object is masked from 'package:purrr':
## 
##     discard
## 
## 
## The following object is masked from 'package:readr':
## 
##     col_factor
## 
## 
## Loading required package: lattice
## 
## 
## Attaching package: 'caret'
## 
## 
## The following object is masked from 'package:purrr':
## 
##     lift
## 
## 
## Type 'citation("pROC")' for a citation.
## 
## 
## Attaching package: 'pROC'
## 
## 
## The following objects are masked from 'package:stats':
## 
##     cov, smooth, var

Data Cleanup and Exploratory Analysis

Dataset Overview
Item	Value
Rows	2240
Columns	29

Variable Summary
	Variable	Type	Missing_Values
id	id	integer	0
year_birth	year_birth	integer	0
education	education	character	0
marital_status	marital_status	character	0
income	income	integer	24
kidhome	kidhome	integer	0
teenhome	teenhome	integer	0
dt_customer	dt_customer	character	0
recency	recency	integer	0
mnt_wines	mnt_wines	integer	0
mnt_fruits	mnt_fruits	integer	0
mnt_meat_products	mnt_meat_products	integer	0
mnt_fish_products	mnt_fish_products	integer	0
mnt_sweet_products	mnt_sweet_products	integer	0
mnt_gold_prods	mnt_gold_prods	integer	0
num_deals_purchases	num_deals_purchases	integer	0
num_web_purchases	num_web_purchases	integer	0
num_catalog_purchases	num_catalog_purchases	integer	0
num_store_purchases	num_store_purchases	integer	0
num_web_visits_month	num_web_visits_month	integer	0
accepted_cmp3	accepted_cmp3	integer	0
accepted_cmp4	accepted_cmp4	integer	0
accepted_cmp5	accepted_cmp5	integer	0
accepted_cmp1	accepted_cmp1	integer	0
accepted_cmp2	accepted_cmp2	integer	0
complain	complain	integer	0
z_cost_contact	z_cost_contact	integer	0
z_revenue	z_revenue	integer	0
response	response	integer	0

Preview of First 10 Rows
id	year_birth	education	marital_status	income	recency
5524	1957	Graduation	Single	58138	58
2174	1954	Graduation	Single	46344	38
4141	1965	Graduation	Together	71613	26
6182	1984	Graduation	Together	26646	26
5324	1981	PhD	Married	58293	94
7446	1967	Master	Together	62513	16
965	1971	Graduation	Divorced	55635	34
6177	1985	PhD	Married	33454	32
4855	1974	PhD	Together	30351	19
5899	1950	PhD	Together	5648	68

Missing Values Analysis

Missing Values by Variable
Variable	Missing_Values
income	24

One of the main data quality issues we found was that the Income variable contained 24 missing values, while all other variables were complete. Because income is an important factor in understanding customer purchasing behavior, we couldn’t simply ignore these missing values. Instead of deleting those records, we replaced the missing income values with the median income. We chose the median because it is less affected by extreme values than the mean, making it a better representation of a typical customer’s income

Data Cleaning

Duplicate Row Check
Cleaning_Check	Count
Duplicate Rows	0

Income Outliers Identified Using IQR
id	income	education	marital_status
8475	157243	PhD	Married
1503	162397	PhD	Together
5555	153924	Graduation	Divorced
1501	160803	PhD	Married
5336	157733	Master	Together
4931	157146	Graduation	Together
11181	156924	PhD	Married
9432	666666	Graduation	Together

Next, we checked for duplicate records and found zero duplicate rows, so no records needed to be removed for duplication.

We also converted the customer enrollment date into the proper date format so it could be used correctly in later analyses. After that, we examined the income variable for unusual values using the Interquartile Range, or IQR, method. This identified several extremely high-income outliers, including one value of 666,666, which was much higher than the rest of the dataset. These extreme outliers were removed because they could distort summary statistics and influence our models.

Feature Engineering for EDA

Finally, we prepared the data for exploratory analysis by creating several new variables through feature engineering, including customer age, total spending, total purchases, number of children, and total campaigns accepted. These new variables allowed us to better analyze customer behavior and served as important inputs for the clustering and logistic regression models later in the project.

Overall, these preprocessing steps improved the quality of the dataset and ensured that our analyses and business recommendations were based on clean, consistent, and reliable data.

Summary Statistics

Summary Statistics for Key Variables
Statistic	Value
age_Mean	57.21
age_Median	56.00
age_SD	11.99
age_Min	30.00
age_Max	133.00
income_Mean	51630.93
income_Median	51381.50
income_SD	20601.68
income_Min	1730.00
income_Max	113734.00
recency_Mean	49.11
recency_Median	49.00
recency_SD	28.95
recency_Min	0.00
recency_Max	99.00
children_Mean	0.95
children_Median	1.00
children_SD	0.75
children_Min	0.00
children_Max	3.00
total_spent_Mean	605.60
total_spent_Median	396.50
total_spent_SD	601.44
total_spent_Min	5.00
total_spent_Max	2525.00
total_purchases_Mean	12.54
total_purchases_Median	12.00
total_purchases_SD	7.18
total_purchases_Min	0.00
total_purchases_Max	32.00
total_campaigns_accepted_Mean	0.30
total_campaigns_accepted_Median	0.00
total_campaigns_accepted_SD	0.68
total_campaigns_accepted_Min	0.00
total_campaigns_accepted_Max	4.00

Customer Age Distribution

The age distribution helps identify the main age groups in the customer base.

Income Distribution

Income Boxplot

The income distribution and boxplot help identify variation and possible high-income outliers.

Total Spending Distribution

Total Spending Boxplot

Total spending helps identify high-value customers.

Product Category Spending

Total Spending by Product Category
Product	Spending
Wine	680604
Fruits	58881
Meat	368993
Fish	84023
Sweets	60611
Gold	98579

This chart shows which product categories generate the most spending.

Purchase Channel Analysis

Purchases by Channel
Channel	Purchases
Web	9146
Catalog	5884
Store	12964

This chart compares customer purchasing activity across web, catalog, and store channels.

Education Distribution

Customer Education Distribution
education	n	Percentage
2n Cycle	203	9.09
Basic	54	2.42
Graduation	1124	50.36
Master	369	16.53
PhD	482	21.59

Marital Status Distribution

Customer Marital Status Distribution
marital_status	n	Percentage
Absurd	2	0.09
Alone	3	0.13
Divorced	231	10.35
Married	861	38.58
Single	480	21.51
Together	576	25.81
Widow	77	3.45
YOLO	2	0.09

Campaign Acceptance

Campaign Acceptance by Campaign
Campaign	Accepted
Campaign 1	144
Campaign 2	30
Campaign 3	163
Campaign 4	167
Campaign 5	163

Response Rate

Response Rate for Most Recent Campaign
response	n	Percentage
0	1898	85.04
1	334	14.96

Recency Distribution

Income and Total Spending Relationship

## `geom_smooth()` using formula = 'y ~ x'

Spending by Campaign Response

Correlation Analysis

Anomaly Detection

Oldest Customers

Oldest Customers
id	year_birth	age	income	total_spent
11004	1893	133	60182	22
1150	1899	127	83532	1853
7829	1900	126	36640	65
6663	1940	86	51141	157
6932	1941	85	93027	2119
2968	1943	83	48948	902
6142	1943	83	65073	900
7106	1943	83	75865	1242
8800	1943	83	48948	902
1453	1943	83	57513	1060

Highest-Income Customers

Highest-Income Customers
id	income	age	total_spent
4619	113734	81	277
4611	105471	56	1724
10089	102692	52	1112
2798	102160	49	1240
7215	101970	43	1135
4248	98777	66	2008
7451	98777	66	2008
500	96876	49	1941
2109	96843	36	1544
6815	96547	46	809

Highest-Spending Customers

Highest-Spending Customers
id	income	age	total_spent
5735	90638	35	2525
5350	90638	35	2525
1763	87679	38	2524
4580	75759	57	2486
4475	69098	77	2440
5453	90226	70	2352
10133	93790	56	2349
9010	83151	54	2346
5386	94384	73	2302
6024	94384	73	2302

Export Clean Dataset

EDA Findings

The exploratory analysis shows that customer income and total spending vary substantially across the dataset. This suggests that the customer base contains different customer segments, including lower-spending customers and high-value customers.

Product spending is not evenly distributed across categories. Some product categories generate more total spending than others, which may help identify important product preferences.

Purchase channel analysis shows how customers interact with the company through web, catalog, and store purchases.

Campaign acceptance appears to be relatively low, meaning most customers do not accept marketing campaigns.

Business Recommendation

Based on the EDA, the modeling team should pay close attention to income, total spending, purchase channels, recency, and previous campaign acceptance.

High-spending customers and recently active customers may represent valuable segments for future marketing campaigns.

K-Means Clustering Model

Can customers be segmented into meaningful groups based on purchasing behavior?

We used the k-means model to identify customer segments based on age, income, number of children, recency, total spending and purchase activity.

The Elbow plot helps justify how many customer segments to use. In this analysis, the elbow begins at about 2 clusters and forms the elbow at between 3 and 4. We elect to view 4 clusters for the analysis.

What demographic characteristics are associated with higher customer spending? Which customer characteristics are associated with positive campaign responses?

Customer Segment Summary by Cluster
cluster	Average Age	Average Income	Average Children	Average Total Spent	Average Total Purchases	Average Campaigns Accepted
1	55.47	81175.71	0.26	1612.07	19.94	2.53
2	59.03	74210.67	0.21	1323.66	19.56	0.29
3	54.68	34417.29	1.25	98.86	5.85	0.08
4	60.48	58306.43	1.12	736.39	17.22	0.22

Logistic Regression Model

The data is split into training and testing sets. The model is trained on the training data and evaluated on the testing data.

## [1] 1563

## [1] 669

The dependent variable is: response: whether the customer responded to the campaign

The independent variables include customer demographics, spending behavior, purchasing channels, campaign history, and recency.

## 
## Call:
## glm(formula = response ~ income + age + children + recency + 
##     total_spent + total_purchases + total_campaigns_accepted + 
##     num_web_purchases + num_catalog_purchases + num_store_purchases + 
##     num_web_visits_month + num_deals_purchases + complain + education + 
##     marital_status, family = binomial(link = "logit"), data = train_data)
## 
## Coefficients: (1 not defined because of singularities)
##                            Estimate Std. Error z value Pr(>|z|)    
## (Intercept)              -1.114e+00  7.731e-01  -1.441 0.149664    
## income                    5.020e-06  9.494e-06   0.529 0.596975    
## age                      -1.400e-02  7.728e-03  -1.811 0.070112 .  
## children                 -4.598e-01  1.778e-01  -2.586 0.009714 ** 
## recency                  -3.011e-02  3.466e-03  -8.688  < 2e-16 ***
## total_spent               6.445e-04  3.074e-04   2.097 0.036019 *  
## total_purchases          -2.117e-01  4.066e-02  -5.207 1.92e-07 ***
## total_campaigns_accepted  1.375e+00  1.306e-01  10.531  < 2e-16 ***
## num_web_purchases         2.639e-01  5.763e-02   4.579 4.68e-06 ***
## num_catalog_purchases     3.124e-01  5.960e-02   5.242 1.59e-07 ***
## num_store_purchases              NA         NA      NA       NA    
## num_web_visits_month      1.714e-01  5.242e-02   3.270 0.001074 ** 
## num_deals_purchases       1.270e-01  5.541e-02   2.292 0.021910 *  
## complain                  1.417e+00  1.038e+00   1.364 0.172418    
## educationBasic           -1.196e+00  1.088e+00  -1.099 0.271719    
## educationGraduation       2.247e-01  3.538e-01   0.635 0.525314    
## educationMaster           6.476e-01  3.881e-01   1.668 0.095240 .  
## educationPhD              9.431e-01  3.749e-01   2.516 0.011877 *  
## marital_statusMarried    -1.070e+00  2.865e-01  -3.733 0.000189 ***
## marital_statusSingle      3.688e-02  2.902e-01   0.127 0.898876    
## marital_statusTogether   -1.064e+00  3.063e-01  -3.475 0.000511 ***
## marital_statusWidow       2.485e-01  4.705e-01   0.528 0.597446    
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## (Dispersion parameter for binomial family taken to be 1)
## 
##     Null deviance: 1319.83  on 1562  degrees of freedom
## Residual deviance:  873.04  on 1542  degrees of freedom
## AIC: 915.04
## 
## Number of Fisher Scoring iterations: 6

Odds Ratios

An odds ratio greater than 1 means the variable is associated with higher odds of campaign response. An odds ratio below 1 means the variable is associated with lower odds of campaign response.

Logistic Regression Odds Ratios
term	estimate	conf.low	conf.high	p.value
total_campaigns_accepted	3.9550	3.0800	5.1407	0.0000
recency	0.9703	0.9636	0.9768	0.0000
num_catalog_purchases	1.3667	1.2191	1.5404	0.0000
total_purchases	0.8092	0.7458	0.8749	0.0000
num_web_purchases	1.3020	1.1632	1.4595	0.0000
marital_statusMarried	0.3432	0.1962	0.6050	0.0002
marital_statusTogether	0.3449	0.1890	0.6301	0.0005
num_web_visits_month	1.1870	1.0727	1.3180	0.0011
children	0.6314	0.4415	0.8873	0.0097
educationPhD	2.5680	1.2583	5.4950	0.0119
num_deals_purchases	1.1354	1.0198	1.2675	0.0219
total_spent	1.0006	1.0000	1.0012	0.0360
age	0.9861	0.9711	1.0010	0.0701
educationMaster	1.9109	0.9095	4.1867	0.0952
(Intercept)	0.3283	0.0705	1.4663	0.1497
complain	4.1242	0.4925	28.4763	0.1724
educationBasic	0.3025	0.0158	1.7584	0.2717
educationGraduation	1.2520	0.6408	2.5768	0.5253
income	1.0000	1.0000	1.0000	0.5970
marital_statusWidow	1.2820	0.4957	3.1672	0.5974
marital_statusSingle	1.0376	0.5909	1.8474	0.8989
num_store_purchases	NA	NA	NA	NA

Model Evaluation

The confusion matrix evaluates how well the model classifies campaign responders and non-responders at a 50% probability threshold.

## Confusion Matrix and Statistics
## 
##           Reference
## Prediction  No Yes
##        No  553  66
##        Yes  16  34
##                                           
##                Accuracy : 0.8774          
##                  95% CI : (0.8502, 0.9013)
##     No Information Rate : 0.8505          
##     P-Value [Acc > NIR] : 0.02656         
##                                           
##                   Kappa : 0.3928          
##                                           
##  Mcnemar's Test P-Value : 6.262e-08       
##                                           
##             Sensitivity : 0.34000         
##             Specificity : 0.97188         
##          Pos Pred Value : 0.68000         
##          Neg Pred Value : 0.89338         
##              Prevalence : 0.14948         
##          Detection Rate : 0.05082         
##    Detection Prevalence : 0.07474         
##       Balanced Accuracy : 0.65594         
##                                           
##        'Positive' Class : Yes             
##

ROC Curve

The ROC curve and AUC score evaluate how well the model separates likely responders from unlikely responders across all possible thresholds.

## Setting levels: control = 0, case = 1

## Setting direction: controls < cases

## Area under the curve: 0.8658

Advertising Budget Recommendation

The model can help determine how to spend advertising dollars by ranking customers according to predicted response probability. Customers with the highest predicted probability should receive the most advertising attention.

Advertising Priority Tiers
advertising_tier	customers	avg_predicted_response	actual_response_rate	avg_total_spent	avg_income
High Priority	134	0.4730	0.4925	1032.3657	61779.81
Medium Priority	201	0.1033	0.1194	580.6866	50780.73
Low Priority	334	0.0229	0.0299	436.6497	48542.03

How can marketers improve campaign targeting using customer segmentation?

Customer segmentation helps marketers improve campaign targeting by grouping customers with similar behaviors, preferences, and spending patterns. Instead of sending the same campaign to everyone, marketers can tailor promotions to the segments most likely to respond. For example, high-spending customers may receive premium product offers, while lower-spending or less-engaged customers may receive discounts, reactivation campaigns, or introductory offers. This makes campaigns more relevant, improves response rates, and reduces wasted marketing spend.

Comparison of Analyses

Kaggle

Focuses primarily on exploratory data analysis (EDA) and basic customer segmentation.
Provides general insights into customer demographics, spending, and purchasing behavior.
Uses K-means clustering but offers limited validation of cluster quality.
Emphasizes descriptive analytics over predictive modeling.

Our Analysis

Includes comprehensive EDA with additional business-focused metrics
- Purchase channels
- Campaign response
- Recency
- Anomaly detection
Performs feature engineering to improve customer profiling:
- Total Spending
- Total Purchases
- Customer Age
- Connects analytical findings to managerial recommendations for targeted marketing.
Uses customer segmentation to support business decision-making rather than simply identifying clusters.

Limitations

Predictive Analytics
- Marketing recommendations are based on observed patterns
- Customer purchasing behavior varies widely across individuals, making spending patterns more difficult to predict
Cluster validation could be strengthened using a Silhouette Score
Analysis is limited by available dataset variables
- No customer lifetime value, geographic data, or longitudinal purchase history

Managerial Conclusion

High-spending segments (1 & 2) should be prioritized for retention through loyalty offers, premium promotions, and personalized recommendations.

Product spend patterns also show that customers prefer different categories, such as wine, meat, fruits, fish, sweets, or gold products. These differences allow marketers to create more targeted campaigns instead of using the same message for every customer.

Overall, segmentation can improve campaign effectiveness by helping marketers send the right offer to the right customer group. This can increase response rates, improve customer lifetime value, and reduce wasted marketing spend.

Group Project

Roci Barnes, Brandon Chin, Leilani De La Cruz, Jaason Loucks

2026-06-19