TELECOMMUNICATION-DIRECT MARKETING
Prepared by,
Anish Dixit,
Aumkar Gadekar,
Aditi Kandoi
The following document contains a detailed summary of the solution to the problem statement 'Telecommunication- Direct Marketing' prepared by team 'BADM_074'.
Table of Contents
[Topics]{.underline}
1**.** Introduction 3
2. Data Preparation 4
3. Exploratory Data Analysis 7
4. Proposed Approach 12
5. Report 15
6. Conclusion 18
Direct marketing is a form of communicating an offer, where organizations communicate directly to a pre selected customer and supply a method for a direct response. Among practitioners, it is also known as direct response marketing.
This marketing technique is particularly prevalent in the telecommunication sector where direct calling is often among the most relevant methods of converting customers.
The given business problem gives a dataset containing various parameters of instances of this methodology which can be used to predict if an existing customer will subscribe to a data pack or not.
The objective of this study is to implement a statistical model to give the most accurate prediction validated on the hold out set.
The objective of this document is to explain the proposed solution, the process of obtaining the solution and establishing how the proposed model is the most befitting solution for the given business case.
DATA PREPARATION
It is a data mining technique that transforms raw data into an understandable format. Raw data(real world data) is always incomplete and that data cannot be sent through a model. That would cause certain errors. That is why we need to preprocess data before sending through a model.
[Steps in data preparation:]{.underline}
- Import libraries
{width="6.5in"
height="1.2638888888888888in"}
- Read data
{width="7.135416666666667in"
height="2.276042213473316in"}
- Checking for Missing Data
We first check for null values, none of the entries were found to have Nan values. However, some entries for education had unknown as the value specified. To reduce ambiguous data, we drop the rows with education= "unknown".
{width="6.5in"
height="3.611111111111111in"}
{width="6.5in"
height="0.6388888888888888in"}
- Checking for Categorical Data
Since our model cannot accept categorical data or strings as input; we have to convert it into numerical values. We use a label Encoder to give a numerical value to the columns having string/category data.
{width="7.057292213473316in"
height="2.3125in"}
- Normalize the Data
We scale down the features, without distorting their range. We split the data frame into input features X and dependent variable target(output) ie. Y
{width="6.989583333333333in"
height="1.3802088801399826in"}
EXPLORATORY DATA ANALYSIS
Arguably the most underrated yet pivotal phase of formulating a statistical model, EDA involves carefully scrutinizing the data, familiarizing the data and then trying to derive insights about overpowering features, irrelevant features, etc. from it before actually moving on to modelling.
- Graphing parameter last month against target shows us that the
customers with packs expiring in May are most likely to buy new packs.
{width="6.125in"
height="3.276042213473316in"}
- Graphing parameter job against target shows us that the customers
with management jobs are most likely to buy new packs.
{width="6.21875in"
height="3.651042213473316in"}
- Graphing parameter education against target shows us that the
customers with secondary education are most likely to buy new packs.
{width="6.270833333333333in"
height="3.3489588801399823in"}
- Graphing parameter marital status against target shows us that the
customers with management jobs are most likely to buy new packs.
{width="6.380208880139983in"
height="3.3020833333333335in"}
- Graphing parameter duration against target shows us that the
customers whose calls lasted for 0-500 seconds are most likely to buy new packs.
{width="6.09375in"
height="3.463542213473316in"}
- Graphing parameter smart against target shows us that the customers
who did not have smartphones are most likely to buy new packs.
{width="5.692708880139983in"
height="3.3541666666666665in"}
[Final step in EDA: Feature Extraction]{.underline}
{width="7.302083333333333in"
height="1.609375546806649in"}
{width="7.229166666666667in"
height="2.588542213473316in"}
We tried various feature extraction techniques to find the dependence of the output on the features and remove redundant attributes. The extra trees classifier is a reliable feature selector and based on its results, the attribute "connect" was found to have negligible impact on the output. Hence this redundant feature was removed. Our validation set too gave higher accuracy with this feature removed.
{width="6.5in"
height="0.6822922134733158in"}
PROPOSED APPROACH
Once the data preprocessing and exploration is done, we have to actually build a Machine Learning model for prediction and analysis.
In the given dataset, as the output variable(target) is a binary variable(yes/no), the given problem falls under the category of [Binary Classification.]{.underline}
The algorithm we have chosen for this is a [Random Forest classifier]{.underline}.
Random forest algorithm is a supervised classification algorithm. As the name suggests, this algorithm creates the forest with a number of trees.
In general, more trees in the forest, more robust the forest looks like. In the same way in the random forest classifier, the higher the number of trees in the forest, the higher the accuracy results.
{width="6.5in"
height="1.203125546806649in"}
Now that the classifier has been trained, we use the validation data to test its performance and compare with other algorithms.
{width="6.5in"
height="5.473958880139983in"}
The validation data has been loaded and processed. We now predict output for the validation data.
{width="6.5in"
height="3.9427088801399823in"}
As it is observed that the given dataset is imbalanced, in the training set only 4000 positive examples are observed as opposed to 35000 negative examples. Therefore, overall accuracy isn't the best indicator of the model's performance. Hence, we use F1 score that takes into consideration the precision and recall values, to evaluate the performance of the classifier. An overall accuracy of 98.7% is observed. An F1 score of 0.99/1 and 0.94/1 is observed for negative and positive examples, respectively. This indicates high accuracy and performance.
REPORT
As we have seen in the previous step, the algorithm proposed in this solution, i.e. a Random Forest Classifier gives an [accuracy of 98.7%]{.underline}, and an [F1 score of 0.99]{.underline} which matches almost perfectly with the real world observations. However, to further validate our approach, we have implemented the same algorithm using R language and libraries.
{width="6.479166666666667in"
height="5.015625546806649in"}
{width="6.46875in"
height="4.203125546806649in"}
We obtain an accuracy of around 90% using this, which successfully cross validates our technique.
To further strengthen our proposed methodology, we crossed the Random Forest Algorithm with another popular classification algorithm, the Naive Bayes classifier, and ran a ROC AUC curve (Receiver Operating Characteristic, Area Under the Curve) and the results are as follows:
{width="7.25in"
height="3.276042213473316in"}
ROC plots :
{width="3.2911439195100614in"
height="2.244792213473316in"}
{width="2.8906255468066493in"
height="2.0865430883639546in"}
Random Forest Naive Bayes
This also proves that the algorithm that has been used is better than other classification algorithms for this problem.
CONCLUSION
The given business problem about Digital Marketing in Telecommunication was a classical statistical modelling problem. The solution proposed in this document, while of a high degree of accuracy, is not without its flaws, however, from a practicality purpose, it is a viable, and therefore a sustainable solution.