1. Introduction

I have choosing Marketing domain as it is inline to my profession. I have choosen my datset from Kaggle website. The link to the data source is https://www.kaggle.com/datasets/jackdaoud/marketing-data. This data source has de- tails about customer profiles and their product preferences which can be analyzed on the attributes like campaign performance across channels.

The problem I am trying to analyze is to figure out how campaign performance across channels impact product preferences for different customer profiles. This is important in real world as it will analyze the impact of campaigns and channels on customer preferences for different products and how it can impact transformation in consumer behaviour. [1]

Below are the steps or phases of your project implementation:

1. Finalizing Dataset and preparing Github Repo

2. Identifying the key parameters for analysis

3. Data preparation and cleansing

4. Planning and building the model

5. Operationalizing the model and capturing the resulting analytics

6. Submitting finalized latex and pdf report using Overleaf

Key components of my approach are Data cleansing and Machine Learning model which will lead me to my analysis.

1.1. Sections of This Project

- Data Collection

- Data Cleansing

- Exploratory Data Analysis

- Predictive Analytics

- Models Evaluation Results

- Conclusion

2. Data Collection

2.1. Data Source

My Data source is https://www.kaggle.com/datasets/jackdaoud/marketing-data.

2.2. Data Format

The format of the data source is CSV and I imported the data source as it is, in CSV format.

2.3. Data Scraping

The data was available in a structured manner and I did not have to use any Data scraping technique to extract information. I imported the CSV in python code and use panda library to read the attribute values.

2.4. Data Attributes to be Used for Analysis

Few of the important attributes which I used as a part of this analysis is Income, NumStorePurchases, NumCatalogPurchases, NumWebPurchases, NumWebVis- itsMonth, NumDealsPurchases, AcceptedCmp5, AcceptedCmp4, AcceptedCmp3, AcceptedCmp2, AcceptedCmp1, AcceptedCmpOverall, Age, educationbasic, educationGraduation, educationMaster, educationPhD to list a few.

2.5. Data Extraction

We did a one time import of the full data set using Panda libraries directly from the source - Kaggle site. As the data is already structured and transformed to a huge extent, we used the data as it is for our analysis.

3. Data Cleaning

3.1. Data Cleaning Process

Although I have chosen a structured data set but I have to undergo the process of Data Cleaning and Transformation.

Figure 1. Data Cleaning Process [4]

3.2. Tools and Techniques to be Used for Data Cleaning

- For removing duplicates we used drop duplicate()

- For checking Null values we used isnull.any() functioning for dropping null rows we will be using dropna(thresh=thresh, axis = 1).shape

- For removing extreme outliers we used Quantile function with operator

- We used INT() function to convert Income values from float to INT

3.3. Missing Value Strategy

Isnull.any() will check for missing values and dropna() will remove the row with the null value.

3.4. Data Attributes and Record Count after Cleaning Process

My data data attributes remain the same I.e 2206 Rows and 39 columns as I choose a structured set and none of my attribute had missing values.

3.5. Definitions of Important Data Attributes

We will be using selected attributes from our dataset in our analysis to derive findings. Below are the list of attributes which will be important for our EDA and our results:

- Income: This is one of the most important attribute which we used for our analysis in consumer behavior.

- Age: Age is used to determine the category of products consumed by different age groups.

- NumOfWebVisitispermonth: This attribute helps us analyze if the website influenced any purchase behavior.

- MntTotal: This is the total products consumed and this attribute will be used as a denominator for other products.

- AcceptedCmpOverall: Total number of campaign accepted by the consumer which might influence purchase behavior.

- MnRegularProducts: This is the total number of regular products consumed.

- Education basic: This identifies if a consumer has basic level education which might influence purchase behavior.

- Education Master: This identifies if a consumer has master level education which might influence purchase behavior.

- Education Graduation: This identifies if a consumer has Graduation level education which might influence purchase behavior.

- Education Phd: This identifies if a consumer has PHD level education which might influence purchase behavior.

- Marital Single: This identifies if a consumer’s marital status is single which might influence purchase behavior.

- Marital Divorced: This identifies if a consumer’s marital status is divorced which might influence purchase behavior.

- Customer Days: This attribute defines if loyalty of consumer impacts purchase behavior.

3.6. Dependent and Independent Variables for the Analysis

MntTotal is a dependent variables related to consumer behavior analysis. Income, AcceptedCmpOverall, Education and Marital status are independent variables for our analysis.

4. Exploratory Data Analysis

Exploratory Data Analysis (EDA) is a methodology which helps identify the understand the data model and identify co-relations and pattern between the data attributes. It is very essential in a data science or data analytics project as it reveals essential characteristics of the dataset which eventually helps in holistic data insights.

4.1. Exploratory Data Analysis Technique

Below are primary types of EDA techniques:

- Univariate nongraphical

- Multivariate nongraphical

- Univariate graphical

- Multivariate graphical

Histograms and Box plots are going to help comprehend the relationships and identify outliers. So Univariate graphical technique will work for my project. I will also leverage Multivariate graphical technique when and where required.

4.2. Preliminary Results of Exploratory Analysis

This technique helped me identify the characteristics of these important data attributes. For ex:

- The Average age range doing the transactions

- Average customer days

- Income range of the surveyed customer

- Graduated vs basic profiles

- Marital status

Exploratory Data Analysis phase helps understand the various data attributes and what kind of relationship I can generate between Customer Profile, Education and Marital Status. It also helps identify the campaign acceptance based on profiles which is key to this analysis. This phase gave me an idea about the ideal range of values for each profile attribute. I will continue with my analysis for deeper insights on purchase behavior and link them to the profile attribute.

4.3. Specifics of Processing

Below are the sequence of activities which we performed as a part of EDA:

- Converted Income column from Float to Int to deal with whole numbers

- Removed the Outliers for teh accuracy of the analysis

- Identified customer profile data attributes and created histogram for each attribute - Income, Age, Dt Customer, Recency, Kidhome, Teenhome

- Identified Education related and created histogram for each attribute - education Basic, education Graduation, education Master, education PhD

- Identified Marital status attributes and created histogram for each attribute - marital Divorced, marital Married, marital Single, marital Together, marital Widow

4.4. Code Snapshot

Removing Null values:

thresh = len(data _frame)*0.6 data _frame.dropna(thresh=thresh, axis = 0).shape

(Converting Income from Float to Int:

data _frame.Income.value _counts() data frame[’Income’] = data _frame[’Income’].apply(int) data _frame.Income.value _counts()

Removing Extreme Outliers:

data _frame _outliers = data _frame[(data _frame.Income ¡ data _frame.Income.quantile(0.995)) (data _frame.Income ¿ data _frame.Income.quantile(0.005))] data _frame _outliers.hist(’Income’)

Sum of food purchased:

data _frame[’MntTotal’] = data _frame.loc[:,[’MntWines’, ’MntFruits’, ’MntMeat-

Products’, ’MntFishProducts’, ’MntSweetProducts’]].sum(axis=1) data _frame.MntTotal.head()

Histograms:

data _frame.hist(column = [’Age’, ’Customer Days’, ’Income’, ’Kidhome’, ’Teen- home’], figsize=(45,25))

Histograms showcase the mean values for each attribute of the customer profile - Age, Customer Days, Income, Kidshome and Teenhome. We will have boxplots for these customer profiles in Section 6 of the report, where we will explain our analysis. For detailed code, please visit my My Github Repo.

Figure 2. Histograms

4.5. Exploratory Data Analysis Findings

Exploratory Data Analysis phase helps us understand the various data attributes and what kind of relationship I can generate between Customer Profile, Education and Marital Status. It also helps me identify the campaign acceptance based on profiles which will be the key to this analysis. This phase gave me an idea about the ideal range of values for each profile attribute. We continue with our analysis for deeper insights on purchase behavior in the following sections and link them to the profile attribute.

5. Predictive Analytics

5.1. Analysis Technique

Creating a pipeline for predictive application [5] is all about automating the data ingestion and processing the data models for desired outcome. This helps in maintenance of the overall system with reduced effort. The model I am trying to follow can be depicted with the below diagram:

Figure 3. ML Automation process [4]

1. Data Import and Validation - We are not continuously streaming data so we will do a one time import of the data which will be validated before processing.

2. Data Exploration: Essential information was extracted from the ingested data which will help us understand and predict the data.

3. ML Model Training: A model is selected and data is transformed into test and train data sets.

4. ML Model Validation and Deployment: After the predictions are verified for the targeted values the Model is deployed.

5. ML model evaluation and refinement: Predictions are constantly monitored to improve the ML Model.

5.2. Machine Learning Algorithms

We are using multiple ML algorithms to analyze the data and eventually select the best performing one to make predictions:

1. Linear Regression

2. Random Forest

3. Neural Network

5.3. Training and Testing Process

For training and testing phase of the data, We had to split the data 80-20 using train test split library. As it is a one time imported data, we will do this partition once. Training set will be used for building the predictive model and testing set will be used for actual predictions. We will be applying multiple ML algorithms to choose the best model amongst all of them.

Figure 4. 80-20 train-test split of dataset

5.4. ML Algorithms Implementation

We have identified 3 different ML algorithms for building my model on the training set.

Linear Regression: We will evaluate this supervise learning method to determine any linear relationships between Income and Purchase considering Campaign acceptance.

Random Forest: This algorithm is a type of ensemble learning which generally has less bias and variance as it is an extension of classification and regression decision and combines multiple decision trees.

Neural Network: This deep learning technique will give us result like hu- man brain processing which will be different than other ML learning models.

Once we have defined the models, we are going to evaluate the accuracy of these 3 models by comparing the Bias and Variance for Linear Regression. For Neural Network and Random Forest we will be comparing Accuracy and Precision to determine the preferred model.

Figure 5. Linear Regression Test and Train model with resulting Bias, Coefficients, Score, Mean Absolute Error and Mean Squared Error

6. Models Evaluation Results

Basic results for our classification model to predict Campaign Acceptance based on Income and Total Purchase

Model— Training Features— Acc Train— F1 Train— Acc Test— F1 Test Neural Network— ’Income’, ’MntTotal’ — 78—69—80.74—72.13

Random Forest— ’Income’, ’MntTotal’ —100—100—77.28—74.72

Random Forest looks to be overfitting with the Training statistics as com- pared to the Neural Network model.

Results on Linear Regression:

Bias is -0.3651311628340533

Coefficients [1.29545462e-05] Score 0.15410723940666216

Data type — MAE — RMSE — MSE — R2 Train data — 0.43 — 0.62 — 0.39 — 0.15

Test data — 0.42 — 0.63 — 0.40 — 0.12

We prefer Linear regression over classification models as the error rate is minimal. This Linear regression model shows the Campaign acceptance percent- age based on Income variable. Below screenshot shows the linear slope of the equation.

The linear regression prediction model is y = Coefficient (X) + Bias (C). So our model will have a prediction equation y = [1.29545462e-05]X + [-0.365], where Y is the total amount of purchase and X is the Income value. Based on this equation, household purchase predictions can be made for other customers using their income.

For detailed code, please visit my My Github Repo.

Figure 6. Random Forest Test and Train model with resulting scores of Accuracy, Preci- sion, Recall and F1

Figure 7. Neural Network Test and Train model with resulting scores for Accuracy, Pre- cision, Recall and F1

7. Conclusions

The problem we are trying to analyze is to figure out how campaign performance across channels impact product preferences for different customer profiles. This is important in real world as it will analyze the acceptance of campaigns impacts customer preferences for different products with their income. Our data source has details about customer profiles and their product preferences which can be analyzed on the attributes like campaign performance across channels. Our Linear Regression slope shows a direct correlation between Income and Number of Accepted campaigns. Surveyed people with higher income tend to accept more campaigns resulting in higher purchase.

Figure 8. Slope resulting from the equation of the linear regression model

Figure 9. Linear Slope

Figure 10 bar graph shows the relationship of the number of Web Visits to the number of deals purchased. It cannot be concluded that high number of web visits is resulting in higher number of deals purchased.

Figure 10. Web visits vs deals purchased

Figure 11 Bar graph shows the relationship between number of dependents (Kids and Teens) in an household with the number of stores purchased. House- holds with Teens and Kids do tend to make frequent store purchases.

Figure 11. Households with Kids and Teens

For detailed code, please visit my My Github Repo.

7.1. Conclusions Based on Statistical Evidence

Based on the statistical evidence across all channels [2], Store Purchases outper- form all other channels of purchase. Below seaborn plot in figure 13 compares the Store purchases with the Web Purchases, Catalog Purchases and Deals Purchases. Store Purchases is a clear winner followed by Webpurchases. Please see the code snippet used to create the Horizon Bar plots.

Based on the statistical evidence For campaigns, newer campaigns have higher acceptance as compared to the older campaigns executed previously. Below seaborn plot compares all the campaigns vis a vis total number of accepted campaigns. Barplot on figure 15 clearly shows that Campaign no. 4 has highest acceptance and campaign no. 2 has the least acceptance. Please see the code snippet used to create the Horizon Bar plots.

Figure 12. Channel Performance Comparison plotting

Figure 13. Channels Performance

Figure 14. Campaign Performance comparison plotting

Figure 15. Campaign Performance

For detailed code, please visit my My Github Repo.

7.2. Limitations

Limitations of this project:

- Data collected is only for Food items

- Data has only 2206 observations

7.3. Future Work Recommendations

“As customers go through a series of touch points across media, channels and devices on their paths to purchase understanding the effectiveness of each touch point and their complementary roles in leading to overall conversions is becoming very important” [3]. As future work I recommend to analyze how each of the campaign is impacting the purchase decisions or campaign acceptance, which will be a deeper analysis for combination of the below graphs:

Figure 16. Channels and Campaigns

Ethical Declaration

Not applicable.