Case-Based Reasoning & Data Mining in DSS/ES: Techniques, Tools & Applications

Key Concepts: Case-Based Reasoning (CBR) in DSS/ES

CBR solves new problems by adapting solutions from similar past cases (stored in a case base). Unlike rule-based systems, it avoids explicit knowledge engineering by leveraging memory and analogy.

1. CBR Cycle: Retrieve → Reuse → Revise → Retain

How it works:

• Retrieve: Find k most similar cases using metrics like Euclidean distance or cosine similarity.
• Reuse: Apply the solution from the top-k cases (e.g., majority vote for classification).
• Revise: Adjust the solution if it fails (e.g., tweak loan approval rules).
• Retain: Store the new case for future use.

2. Worked Example: Daraz Customer Complaint Resolution

Scenario: Daraz receives a complaint about a delayed order. The CBR system matches it to past cases:

• Case 1: Delayed by 3 days → Refunded 50%.
• Case 2: Delayed by 5 days → Full refund + discount coupon.
• Case 3: Delayed by 2 days → No refund, but expedited shipping.

Steps:

1. Feature Vector: [delay_days=4, order_value=12000, customer_rating=4.2].
2. Similarity Metric: Euclidean distance to stored cases.
3. Retrieve Top-3: Cases 1, 2, and 3 (distances: 0.8, 1.2, 1.5).
4. Reuse: Majority vote suggests partial refund + discount (Case 1 + Case 2).
5. Revise: Agent overrides to full refund (customer is premium user).
6. Retain: New case added: [delay_days=4, action=full_refund, customer_type=premium].

3. Data Mining in DSS/ES: Extracting Patterns

Data mining discovers hidden patterns in DSS/ES data to support decisions. Key techniques:

4. Worked Example: NEPSE Stock Trend Prediction (Regression)

Problem: Predict tomorrow’s stock price for Nepal Bank Limited (NBL) using past 30 days of data.

Steps:

1. Data Collection: Features = [opening_price, volume, moving_avg_7day]; Target = next_day_price.
2. Model: Linear Regression (simplified): $${\text{Price}}_{\text{t+1}}=2.1\times {\text{Volume}}_t+0.8\times {\text{MovingAvg}}_t+150$$
3. Training: Fit model on 25 days of data.
4. Prediction: For Day 30:
   • Volume = 500,000; MovingAvg = 1,200 → $${\text{Price}}_{\text{31}}=2.1\times 500,000+0.8\times 1,200+150=1,050,990\text{ (NPR)}$$
5. DSS Use: Trader uses this to decide whether to buy/sell/hold.

In the Real World

1. eSewa’s Fraud Detection

   • CBR: Matches new transactions to past fraud cases (e.g., "user in Kathmandu paying for a service in Pokhara").
   • Data Mining: Uses clustering to flag unusual spending patterns (e.g., sudden high-value transfers).

2. Pathao’s Driver Routing

   • Association Rules: "Drivers in Lalitpur with >3 trips/day earn 20% more."
   • Regression: Predicts demand spikes during Dashain (e.g., +40% trips on Day 1).

3. NTC’s Network Outage Prediction

   • Classification: Trained on past outages (e.g., "rain + old cables → 80% chance of outage").
   • CBR: New outage in Chitwan? Matches to similar outages in Makwanpur.

5. Data Mining Tools for DSS/ES

6. Case-Based Reasoning vs. Data Mining: Key Differences

7. Challenges and Limitations

CBR Pitfalls

• Cold Start Problem: No cases for new scenarios (e.g., first COVID-19 lockdown).
• Similarity Metric Bias: Euclidean distance may fail for high-dimensional data.
• Maintenance Overhead: Case base must be updated regularly.

Data Mining Challenges

• Garbage In, Garbage Out (GIGO): Poor data quality → useless patterns.
• Overfitting: Model works on training data but fails in real-world (e.g., NEPSE model trained only on bull markets).
• Interpretability: Black-box models (e.g., deep learning) are hard to explain to stakeholders.

Exam Tip

1. For CBR:

   • Always explain the 4-step cycle (Retrieve-Reuse-Revise-Retain) with a real example (e.g., Daraz, eSewa).
   • Compare CBR with rule-based systems (CBR uses past cases; rule-based uses IF-THEN).
   • Graph the similarity metric (e.g., Euclidean distance formula for 2D/3D cases).

2. For Data Mining:

   • Define the technique (classification/clustering/association) and give a Nepalese example.
   • Show the math for simple cases (e.g., linear regression equation with real numbers).
   • Tool names matter: Weka, RapidMiner, and Orange are high-yield for short-answer questions.

3. Common Exam Traps:

   • ❌ Saying "data mining is the same as big data."
   • ❌ Forgetting to link DSS/ES (e.g., "This clustering helps the bank approve loans faster").
   • ❌ Using vague examples (e.g., "Amazon uses data mining" → specify association rules for recommendations).

Visual Summary:

mindmap
  root((Case-Based Reasoning & Data Mining))
    CBR
      Retrieve["Find Similar Cases\n(Euclidean/Cosine Similarity)"]
      Reuse["Apply Past Solutions"]
      Revise["Adjust for New Context"]
      Retain["Update Case Base"]
    Data Mining
      Classification["Predict Categories\n(e.g., Loan Approval)"]
      Clustering["Group Data\n(e.g., NEPSE Stock Segments)"]
      Association["Find Patterns\n(e.g., Khalti Transaction Rules)"]
      Regression["Predict Values\n(e.g., Pathao Earnings)"]
    Tools
      Weka["Open-Source ML"]
      RapidMiner["Drag-and-Drop"]
      Orange["Visual Workflows"]
    Real World
      eSewa["Fraud Detection via CBR"]
      Daraz["Customer Complaints via CBR"]
      NTC["Outage Prediction via Clustering"]