Analysing Portugese Bank Telemarketing Dataset

— By Rohit Jalani, Adhithyan Sivakumar, Atif Rehman

In this project, we took the task of analysing a Marketing campaign dataset of a Portuguese banking institution. The code and report can also be found on our GitHub. The classification goal is to predict if the client will subscribe to a term deposit. Analyzing marketing campaign datasets of banks is crucial for banks to stay competitive, improve their marketing strategies, and ultimately, better serve their customers. Analyzing marketing campaign datasets of banks is significant for several reasons:

• Understanding Customer Behavior: Analyzing the marketing campaign datasets of banks can help in understanding customer behavior, such as their preferences, interests, and demographics. By understanding customer behavior, banks can tailor their marketing campaigns to better target their intended audience and increase the chances of success.
• Improving Campaign Effectiveness: Analyzing the data from previous marketing campaigns can help banks identify what worked and what didn’t. This information can be used to improve the effectiveness of future campaigns by optimizing the channels, messaging, and timing of the campaigns.
• Measuring ROI(Rate of Interests): By analyzing the marketing campaign datasets, banks can track the ROI of their campaigns. This information is important to determine whether the investment in the campaign was worthwhile or not.
• Identifying Opportunities: Analyzing marketing campaign datasets can help banks identify opportunities for new products or services. By identifying patterns in customer behavior, banks can develop new products or services that meet customer needs and preferences.
• Benchmarking Performance: Analyzing marketing campaign datasets can help banks benchmark their performance against competitors. This information is important to determine how well the bank is performing in the market and what improvements can be made to remain competitive.

The dataset had many columns pertaining to different aspects of the campaigning journey. Following were the classification of the features :

• Client Data : age, job, marital status, education, previous housing loan and previous personal loan.
• Related to last contact of the current campaign : Last communication type, Last contact month, Day of the week contacted and duration.
• Miscellaneous attributes : No of contacts in this campaign, Days since last contact, No of contacts in previous campaign and Outcome of previous campaign
• Social and Economics Attributes : Employee Variation Rate, Consumer Price Index, Consumer Confidence Index, Euribor 3 month rate and Number of Employees

EDA

Exploratory Data Analysis for the dataset, provided with the following insights

• The subscription depends on continuous and discrete features such as age, job, education, housing loan, etc.
• Andrews curve was plotted for the given data to get a broader view regarding the correlations in the data. Following conclusion were drawn with it:
• Highly overlapping suggest that there is a lot of noise in the current state of the dataset.
• Similar shape and size conclude that neither yes nor no is exclusively discriminated by the data and it cannot separate the two groups well.
• However, visible and distinguishable curves suggest that there is some degree of separation, and thus proper data cleaning, feature selection, and handling outliers would solve the underlying problems.

• Heatmap between all the continuous numeric variables shows there existed high correlations between (nr.employed , emp.var.rate) (nr.employed , euribor3m) (emp.var.rate , euribor3m) (emp.var.rate , cons.price.idx). These variables were concluded to be insignificant and were removed.

• Plotting heatmap of various variables shows that feature ‘Default’ could be removed to avoid any fake data in the analysis.

• While the client’s age also dramatically affects the classification, around 50 percentile of the people in the age category of 35 to 45 are likely to subscribe to the term deposit.

• From the plot, we have concluded that people in admin-related jobs have 12%, i.e., the highest chance of subscribing to a term deposit. Similarly, we found that singles with university degrees have a high chance of subscribing to a term deposit.

• Plotting count plots category-wise for most of the categorical features like marital, education, housing, loan, contact etc gave no significant anomaly and followed regular trends without any discrimination or partiality. Thus these features were directly fitted after encoding to the model.

• On analysing continuous variables in the same way using density plots, symmetric and nearly normal trends were obtained for most of these features.

• The feature pdays had significant amount of 999(i.e. Unknown values). But since it was a significant factor, it was encoded with yes or no.
• ‘Poutcome’ was dropped due to high count of ‘nonexistent’ category.
• While checking the data about the day of the week, it doesn’t significantly affect the subscription. Still, clients subscribing in December, March, October, and September have very high chances of subscribing to the term deposit.
• On doing EDA, it was obtained that the data had high class imbalance, with NO being the dominating class. Since ML models doesnt work well with this distribution of data, The data was balanced using SMOTE algorithm, which also increased the size of data for better learning for the model.

Outlier analysis

Features ‘age’ and ‘campaign’ were found to have significant amount of outliers. To counter this, age was scaled using StandardScaler and campaign by Log Transform, looking at their range of values. The effect of outliers were vanished.

From the exploratory data analysis, we have come to the above intuitions. The conclusions drawn from above steps regarding the dataset were checked and suitable pre processing techniques were applied to obtain flawless dataset for further processing.

Modeling

Taking insights from the EDA, removing unnecessary features and scaling and normalizing for proper symmetry and weighing of all features, some of them turned out to be important and determining factors for telemarketing offer-related analysis. Data containing these features were then used for further processing and modeling.

Before modeling, SMOTE algorithm was applied to tackle the class-imbalance problem. Earlier, the data contained 41188 rows with majority being of ‘no’ class. After SMOTE, it produced synthetic data to make the data containing 73096 rows.

We preprocessed the available data and filled in the unknown values using the mode, but further, we have tried implementing machine learning techniques to replace them more appropriately. Many features were found to be insignificant in the classification, like the day of the week. Similarly, the features which were very much correlated were removed not to overfit the data. The final training data contained 73076 rows, 11 features.

At first, we modeled the data using DNN, gradient boosting, KNN, and naive byes. While using the KNN model had a good accuracy of 83.6% in predicting the subscription, but models like naive byes didn’t fit the model very greatly with an accuracy of 71.3%.

The data had a significant amount of outliers and quite a few missing values. Additionally, since the data was a mixture of numerical and categorical variables, it was expected that Tree based models would work well on the dataset.

On applying Tree-Based models, Random Forest, Gradient Boost and Decision Trees, using a pipeline along with GridSearchCV and 3-folds Cross validation, for tuning the hyper-parameters, it was found that Gradient Boost technique was working best with the data. Following were the results obtained :

The auc-roc score obtained for these parameters in the test data was :

The test data was fit into model, along with the obtained hyper-parameters, gave highly accurate and satisfactory scores.

Looking at training and testing scores, it was inferred that since the scores were close to each other and the test accuracy score is well below 97%(which is usually the threshold limit for overfitting model), the model had the best parameters possible that fits the dataset.

Following was the Confusion Matrix obtained :

The Classification Report shows that the model was appropriately trained, without any overfitting issues.

Conclusion

In conclusion, this project has demonstrated the power and potential of machine learning in analyzing complex datasets. By conducting exploratory data analysis and implementing various algorithms, we were able to gain valuable insights into the telemarketing dataset and develop a predictive model with a high level of accuracy.

After the whole analyzing process it was found that certain factors some of the most significant parameters for successful analysis and forecasting of a telemarketing campaign for offers like term-deposit. These parameters are

• ‘Age’
• ‘Job’
• ‘Previous Loan’
• ‘Previous campaigning data’
• ‘Marital Status’
• ‘Employment Variation Rate’
• ‘Consumer Confidence Index’

This project has highlighted the importance of data preprocessing, feature selection, and hyperparameter tuning in achieving optimal results. The model developed in this project can be used to optimize telemarketing campaigns and improve the overall success rate. Overall, this project has showcased the practical applications of machine learning in business and marketing and serves as a foundation for future research and development in this field.