Leveraged the capabilities of logistic regression and Support vector machine classification, which can be used to accurately predict user purchase behaviour
This study focuses on helping a magazine company understand the decline in subscriptions over the past year. Despite the increased amount of time people spent at home, the company observed a decrease in readership. The objective is to analyze marketing subscriptions and gain insights into the factors affecting subscription behavior. We will leverage the capabilities of logistic regression and Support vector machine classification, which can be used to accurately predict user purchase behaviour. Both methodologies are exclusive since logistic regression is based on geometrical properties while SVM is based on statiscal properties. In this study, we aim to assist a magazine publisher in determining the reasons for drop in the the subscription last year since they anticipated that a person spending more time at home would increase reading and thereby subscription.
In this study, we will assess two method: Logistic regression and Support vector machines to understand user bhaviour and the accuracy of the model will be the key metric to evaluate the reliability and effectiveness of the model. Evaluating predictvie metrics like Accuracy Rate, Precision, Recall will enable the policymakers generate insights into which factors have a significant influence over user behaviour and improving readershop and are contributing to improved subscription.
Income variables to have outliers present in the data. Therefore, we treated the outliers to range within the IQR (inter quartile range) as adjusted. Assessed the dataset of non unfirm values or characters like ‘?’ or NAN. Due to multiple factors and repetitive variables we paired different sub categories such as in Ecuation, we paired 2n Cycle into “Masters”. We then replaced Together with “Common-Law” and paired YOLO, Alone and Absurb as “Single”. We calculate Age from Birth year by subtracting it from current year. We summed up all Accepted_CMP into Accepted Offers, which were users who has accepted to subscribing for magazine. Similarly, we summed kid home and teen home into Teens and all the purchased items into Total spends.
- Maximum of our users to be married. and common-law/ live-in users to be second highest.
- Graduates could possibly be our target audience who would be interested in our magazine subscription
- There are constant purchase in <20 to 100 days which assesses the purchasing behaviour of the individuals.
- The income distribution of individuals are within the range of 40,000-60,000
- The spending patterns across different product types, where most users are spending on Wine consumption and Meat products and least on Fruits. This could also be interpreted in a way that cost of meat and wine is far more superior than fruits.
- Highest purchase were visibly recorded from stores and web purchases while the lowest were recorded, from deals.
- The income distribution by Marital Status and Education where Married, Divorced and Widow have a higher Income while Singles have a lower income range, similarly PhDs and Graduates have a higher income range than the others indicating purchasing power
- Widow and Single’s have a higher response rate compared to Married , Common Law who often tend to ignore these deals.
So lets get started on data modelling
We conducted Variance Inflation Factor analysis to assess the degree of multicollinearity in our study of regression. Multicollinearity occurs when there is a correlation between numerous independent variables. The findings of the analysis included that all variables had an VIF of <5. A VIF of >5 depict that the variables have higher multicolinearity and hence must be eliminated from the analysis.
Logistics Regression Results Summary From the results below we can interpret the following: Model Information: ● Dependent Variable: Response ● Number of Observations: 1568 ● Model: Logit (logistic regression) ● Method: Maximum Likelihood Estimation (MLE) Model Fit: The Pseudo R-squared: 0.3633, Log-Likelihood: -421.79, Log-Likelihood Ratio (LLR) p-value: 3.246e-82, Converged: False, LL-Null: -662.46. The model consists of approximately 36.33% of the variation from the response variable. Note, that the Pseudo R-squared values must be interpreted with caution and must not be directly compared to R-squared values from a inear regression.
A loglikelihood isa measure of how well a logistic regression fits the observed data where higher the loglikelihood the better, however in this case it is -421.79 which is essential to compare with the results of an alternet model. The LLR P-values is a measure to determine the statistical significance of the model as a whole. While in this case, the LLR p-value is very small 3.246e-82, indicating strong evidence to reject the null hypothesis and be certain that the model significantly improves the fit.
The convergence values is False, which indicates that the p-value and coefficients estimated by the model may not be reliable and we must look at the other predictive indicators and exercise caution while interpreting results for opportunities.
Overall Accuracy = 0.8898809523809523, Precision = 0.7049180327868853, and Recall = 0.434343434343436. Overall accuracy represents the proportion of correctly classified observations from the total number of instances in the data and in this case, the model achieved an overall accuracy of approximately 0.8899, which depicts that the model correctly predicted the outcome for the given percentage. Similarly, the precision suggests that approximately 0.7049 of the instances predicted as positive by the model were true positives, since precision is an indication of the number of positive predictions made by the model are actually correct, predicting a positive subscription behaviour. Recall of 0.4343, indicates that the model correctly identified around 43.43% of the actual positive instances whoa actually subscribed to the magazine.
As we evaluate the confusion matrix, it can be assessed that the predictions were obtained by comparing the predicted probabilities with a threshold of 0.5 which provides a binary classification result (0 or 1) for each instance in the train set and test set.
Support Vector Machine Results Model The SVM model, built with a linear kernel, is taking into account all the features of the data frame. We imposed a 70:30 train-to-test split ratio as a result of the training data set of 70% and testing data set conatins 30% of the split.
The second SVM Model was created only using the significant variables using the SVM Classification Model are ‘MntMeatProducts', 'MntGoldProds', 'NumDealsPurchases', 'NumStorePurchases', 'NumWebVisitsMonth'.
Linear Regression SVM Model 1 SVM Model 2 Models Accuracy Precision Recall Linear regression Support Vector Machine Model comprising All Data Support Vector Machine Model comprising Significant Variables Only Accuracy 0.88 Accuracy 0.86 Accuracy 0.85 Recall 0.43 Precision 0.63, Recall 0.19 Precision 0.55, Recall 0.05
In summary, after a thorough evaluation of the model results, it becomes evident that the logistic regression model outperforms the SVM model in studying user subscription behavior. Through meticulous assessment, we have established that the logistic regression model exhibits higher accuracy, precision, and recall scores, thereby affirming its superior reliability and efficacy in this context. Notably, the logistic regression model demonstrates a greater ability to correctly identify and classify true positives, crucial for understanding the intricacies of customer behavior.
This finding underscores the significance of employing the logistic regression model as a preferred tool for dissecting user subscription patterns. By leveraging its strengths, we can gain deeper insights into the factors influencing subscription decisions and tailor our strategies accordingly. Understanding these nuances is paramount for magazine publishers seeking to optimize subscription rates and mitigate churn.
In essence, the logistic regression model emerges as not only a robust analytical tool but also a strategic asset in the pursuit of enhancing subscription-based revenue streams. Its superior performance underscores its value in informing targeted marketing efforts, personalized engagement strategies, and proactive churn prevention measures. Through continuous refinement and adaptation informed by these insights, publishers can foster stronger connections with their audiences, foster loyalty, and drive sustained growth in subscription revenues.