Data Mining Exercise 1

Problem: Predict "Which customers are likely to change their provider next month?"

1. Data Selection

We don't need to pull all the data in the company's universe. We only select data that is likely to affect the decision to change providers.

Selected:

• Call history
• Billing amount
• Calls to Call Center
• Promotion type

Not selected:

• First and last name (unrelated to the decision)
• ID card number
• Favorite color

Simulated Raw Data Extracted:

| ID | Monthly_Bill (THB) | Data_Usage (GB) | Calls_to_Support (Times) | Contract | Last_Month_Usage | Churn (Label) | | --- | ---

--------------- | --------------- | ---------------------- | --------------- | ---------------- | ------------- | | 101 | 899 | 50 | 0 | 1 Year | 50 GB | No | | 102 | 1200 | 10 | 5 | Monthly | NULL | Yes | | 103 | 399 | 2 | 0 | Monthly | 2 GB | No | | 104 | 2500 | 100 | 1 | 2 Years | 200 GB | Yes |

2. Data Preprocessing

Raw data is often dirty. We need to handle it first.

🔍 Missing Values

• Problem: Customer 102's Last_Month_Usage data is missing (NULL).
• Solution: Mean Imputation or drop it.
  • Here, let's assume we fill in the mean = 40 GB.

⚡ Noise/Outlier

• Problem: Suppose Customer 104's internet usage last month was 200GB (abnormally high).
• Solution: Check if it's an error or actual usage.
  • If actual usage → Keep it, because they might be a VIP group.

Result after Cleaning:

• ID 102: Last_Month_Usage changed from NULL → 40 GB

3. Data Transformation

This step involves transforming data to make it easier for Algorithms to understand and learn.

🔨 Feature Construction

Create Usage_Drop_Rate:

Usage_Drop_Rate = (Last month - This month) / Last month

Reason: A sharp drop in usage might mean they've started using another provider's SIM.

📊 Normalization

Problem: Monthly_Bill (thousands) and Calls_to_Support (single digits) have scales that are too different.

Solution: Scale everything to a range of 0-1 (Min-Max Scaling).

🔢 Encoding

| ID | Norm_Bill | Norm_Support_Calls | Usage_Drop_Rate | Is_Monthly_Contract | Churn (Target) | | --- | ---

------ | ------------------ | ------------------ | ------------------- | -------------- | | 101 | 0.3 | 0.0 | 0.0 (Constant) | 0 | No | | 102 | 0.5 | 1.0 (Frequent calls)| 0.6 (Large drop) | 1 | Yes | | 103 | 0.1 | 0.0 | 0.0 | 1 | No | | 104 | 1.0 | 0.2 | 0.5 (Drop) | 0 | Yes |

4. Data Mining

Send the table data above into an Algorithm (like Decision Tree or Logistic Regression).

Goal: Find the relationship between $X$ (Bill, Calls, Drop_rate) and $Y$ (Churn).

Simulating Algorithm Operation (Decision Tree Learning)

5. Pattern Evaluation

We have the "rules" now, we must see if they are reliable.

🧪 Test

Try applying the rules to guess a new set of customer data (Test Set).

Result:

• ✅ Guessed correctly 90 people
• ❌ Guessed incorrectly 10 people (e.g., guessed they would churn, but they didn't)
• Accuracy = 90%

💡 Interpretation

We found an interesting Pattern (Knowledge):

Simulating Algorithm Operation (Decision Tree Learning)

The model will try to create rules like this:

• First Node: Check Calls_to_Support first → If > 0.8 (calls often) → High chance of Churn.
• Second Node: Check Usage_Drop_Rate → If drops more than 40% → High chance of Churn.

5. Pattern Evaluation

We have the "rules" now, we must see if they are reliable.

🧪 Test

Try applying the rules to guess a new set of customer data (Test Set).

Result:

• ✅ Guessed correctly 90 people
• ❌ Guessed incorrectly 10 people (e.g., guessed they would churn, but they didn't)
• Accuracy = 90%

💡 Interpretation

We found an interesting Pattern (Knowledge):

6. Knowledge Presentation & Action

This step is the destination of KDD.

🎯 Action

The system will automatically pull a list of customers who fall into this Pattern every morning.

🎁 Retention Campaign

Send an SMS or have an employee call to offer:

• "50% discount for 6 months" immediately.
• Before the customer walks over to a competitor's service center.

📝 Takeaway

| Step | Details | | ------

-------- | ------------------------------------------------------------------------------------------------- | | Input | Millions of rows of customer data | | KDD Process| Filter → Clean → Transform numbers → Create model → Select rules | | Output | "List of at-risk customers" and "Logic" indicating why they are at risk (e.g. frequent Call center calls) |

🎯 Goal: Allow the marketing team to solve the problem directly. KDD Process: Filter -> Clean -> Transform numbers -> Create model -> Select rules

Output: "List of at-risk customers" and "Logic" indicating why they are at risk (e.g. frequent Call center calls) to allow the marketing team to solve the problem directly.