An analysis of customer behavior is vital for maintaining active customers, reducing churn, and correctly designing marketing campaigns. In this project, I explore the demographics, spending habits, and campaign engagement of a store's customers. The objective is to predict the success of future marketing strategies to retain or increase customer engagement. All data is downloaded from the Marketing Campaign data from Kaggle.
The first step in this analysis was to capture the distribution of the demographic information available: age, education, marital status, and income. This breakdown will help understand customer diversity and identify dominant segments.
To understand spending patterns, we analyzed the products each of these groups purchased. Because this data does not contain the quantity purchased and only the dollar amount spent, it is difficult to identify appropriate analyses of favorable products.
Continuing:
Across all demographics, I visualized the distribution of total spending using violin plots that show the concentration of spending amounts and the variability within and across each group. The results are almost as expected: those with lower education or lower income are spending less, while the other groups see a nice distribution in spending. Surprisingly, the '18-24' and '65+' groups have similar spending amounts. This could be due to strict or fixed incomes that lead to more informed spending habits.
In order to understand any correlation between each of the products purchased, as well as the platform in which they were purchased, I created a heatmap in Python. There are not necessarily strong correlations but the ones that jump out are: 1) the correlation between meat and wine - as previously noticed, these are popular items, 2) the correlation between fish/sweets and fruit, and 3) the meat purchases attributed to purchasing via catalog. There appears to be a very weak correlation between the "deals" (aka discounts) purchases for each of the products. This proves the store could do a better job offering coupons and discounts to drive in more customers.
Below are the four graphs that represent products purchased with a discount, via web, via catalog, and via in-store. Because the first three graphs show a sharp decline over time, the store should work to promote products via these avenues for consistency. Web and catalog could increase due to a revised interface or online-exclusive coupons.
Campaigns generally saw the same success rate, around 7%. However, the second campaign performed poorly at just 1.4%.
The following heatmap effectively suggests the likelihood of customers who respond to one campaign would respond to another. It also displays the contribution the campaign made on total spent amount. Campaign 1 and campaign 5 stand out as the most successful in terms of customers reacting to both campaigns and total spending because of the campaigns. The store should analyze each campaign to determine why they were successful and replicate/improve upon them.
In this analysis, I've highlighted appropriate steps this company should take in order to make informed, data-driven decisions for customer satisfaction. Lastly, I applied a Random Forest classifier to assess its predictive performance for a campaign. This machine learning algorithm spits out the overall accuracy of its performance to help with future predictions. The test achieved an overall accuracy of 87%, indicating it correctly identified 87% of the samples within our data. Specifically, Class 0 had an 89% precision and Class 1 had a 62% precision; this means Class 1 was not correctly identified amongst true positives. Additional steps should be made to correct any imbalances to improve the model.