Recommender Systems & Cross-Selling/Upselling: Algorithms, Strategies & E-Commerce Impact

Core Concepts: What Are Recommender Systems?

Recommender systems are AI-driven tools that predict user preferences to suggest products/services. They are the backbone of cross-selling (selling complementary items, e.g., phone + case) and upselling (pushing premium versions, e.g., basic → premium Daraz membership).

How They Work: The Two Pillars

graph TD
    A["Recommender System"] --> B["Collaborative Filtering"]
    A --> C["Content-Based Filtering"]
    B --> D["User-User CF\n(‘People like you bought…’)"]
    B --> E["Item-Item CF\n(‘Customers who bought X also bought Y’)"]
    C --> F["User Profile + Item Features\n(‘You liked X, so here’s Y with similar tags’)"]
    D -->|"Example"| G["Netflix: ‘Because you watched *The Dark Knight*, we recommend *Inception*’"]
    E -->|"Example"| H["Amazon: ‘Frequently bought together’"]
    F -->|"Example"| I["Spotify: ‘Discover Weekly’ based on your genre preferences"]

Key Definitions:

• Collaborative Filtering (CF): Relies on user-item interactions (ratings, purchases, clicks) to find patterns.
  • User-User CF: "Users like you also liked..."
  • Item-Item CF: "Users who bought X also bought Y."
• Content-Based Filtering: Uses item features (keywords, categories) and user profiles to match preferences.
  • Example: If you buy organic tea, the system suggests other organic products.

Visual: The Recommender System Pipeline

flowchart TD
    A["User Data\n(Purchases, clicks, ratings)"] --> B["Preprocessing\n(Cleaning, normalization)"]
    B --> C["Collaborative Filtering\nor\nContent-Based Model"]
    C --> D["Ranking Algorithm\n(Cosine similarity, Pearson correlation)"]
    D --> E["Post-Processing\n(Diversity, novelty, business rules)"]
    E --> F["Recommendations\n(‘You may also like…’)"]
    F --> G["Feedback Loop\n(User clicks/ignores updates model)"]

Why This Matters:

• 80% of what users watch on Netflix comes from recommendations.
• 35% of Amazon’s revenue is driven by "Frequently bought together" suggestions.

Real-World Applications in Nepal

1. Khalti & E-Sewa: Personalized Discounts

How it works: Khalti’s "My Offers" section uses collaborative filtering to show discounts on products you’ve previously viewed or bought from linked merchants (e.g., if you bought a laptop charger, it suggests a laptop skin). Visual:

2. Daraz: "Frequently Bought Together"

How it works: Daraz’s "Customers who bought this also bought" uses item-item collaborative filtering. For example:

• Buy a DSLR camera → System suggests a tripod, memory card, and lens cleaner.
• Worked Example:
  • Input: User buys a Logitech mouse (₹2,500).
  • Recommendation: Mouse pad (₹500) + USB cable (₹300) (based on 60% of buyers who purchased the mouse also bought these).
  • Revenue Impact: Upsells average order value by 15–20%.

3. Pathao: Dynamic Pricing + Location-Based Upsells

How it works: Pathao’s "Pathao Plus" membership uses content-based filtering to upsell:

• If you frequently order biryani at night, it promotes "Night Biryani Combo" or "Late-Night Discounts."
• Location-based: If you’re near Thamel, it suggests "Tourist Packs" (water bottle, umbrella, SIM card).

Cross-Selling vs. Upselling: Key Differences

How Recommender Systems Drive Sales

1. Collaborative Filtering in Action: Amazon’s "Frequently Bought Together"

Step-by-Step Trace:

1. Data Collection: Amazon tracks which items are bought together (e.g., diapers + beer).
2. Association Rule Mining: Uses Apriori algorithm to find rules like:
   • If X → Y, then confidence = 70%, support = 5%.
3. Recommendation: If you add diapers to cart, Amazon suggests beer (even if you’ve never bought it before).
4. Result: 35% of Amazon’s revenue comes from these recommendations.

2. Content-Based Filtering: Spotify’s "Discover Weekly"

How It Works:

• User Profile: Your top genres (e.g., hip-hop, Nepali rock).
• Item Features: Songs/artists tagged with similar genres.
• Algorithm: Cosine similarity between your profile and new tracks.
• Example:
  • You listen to Ariana Grande → System suggests Billie Eilish (similar pop/R&B tags).
  • You listen to Indra Joshi → System suggests Anil Chitrakar (Nepali rock).

Challenges & Ethical Considerations

1. The "Cold Start" Problem

• Issue: New users/items have no interaction data.
• Solutions:
  • Demographic-based recommendations (e.g., "New users in Kathmandu may like...").
  • Hybrid models (combine CF + content-based).
• Example: When you first open Daraz, it asks for preferences before showing recommendations.

2. Filter Bubbles & Echo Chambers

• Problem: Over-recommending similar items can limit user exposure.
• Example: If you only buy organic products, the system may never suggest non-organic items, reinforcing bias.
• Solution: Diversity-aware algorithms (e.g., YouTube’s "Mixed Recommendations").

3. Privacy Concerns

• Issue: Recommender systems rely on user data, raising privacy risks.
• Nepalese Context:
  • E-Sewa collects transaction history to personalize offers.
  • Ncell uses location + call data for targeted ads.
• Mitigation:
  • Federated learning (train models on-device without raw data).
  • Anonymization (e.g., Khalti’s aggregated purchase trends).

Exam Tip: How to Score Full Marks

1. Define Clearly:

   • "Collaborative filtering is a recommender technique that predicts user preferences by analyzing patterns in user-item interactions (e.g., ratings, purchases)."
   • "Cross-selling is the practice of suggesting complementary products during checkout (e.g., phone + case)."

2. Use Nepalese Examples:

   • For collaborative filtering, cite Daraz’s "Frequently Bought Together" or Khalti’s personalized discounts.
   • For content-based, cite Spotify’s Discover Weekly or Ncell’s plan upgrades.

3. Compare CF vs. Content-Based:

   Criteria	Collaborative Filtering	Content-Based Filtering
   Data Used	User-item interactions (ratings, clicks)	Item features + user profile
   Cold Start	Poor for new users/items	Works better for new items
   Example	Netflix recommendations	Spotify’s genre-based suggestions

4. Explain the Business Impact:

   • "Amazon’s recommender system increases average order value by 30% by suggesting add-ons."
   • "Pathao’s location-based upsells boost revenue during peak hours (e.g., 7–9 PM in Thamel)."

5. Avoid Vague Statements:

   • ❌ "Recommender systems are important." → ✅ "Recommender systems account for 35% of Amazon’s revenue by leveraging item-item collaborative filtering."

Worked Example: Calculating Recommendation Confidence

Scenario: Daraz wants to recommend a wireless charger (₹1,200) to users who bought a smartphone (₹45,000). Given:

• Support (X → Y): 10% of smartphone buyers also bought the charger.
• Confidence (X → Y): 60% (60% of smartphone buyers who bought the charger also bought it together).

Steps:

1. Market Basket Analysis:

   • Total smartphone buyers = 1,000.
   • Buyers who bought both smartphone + charger = 600 (60% confidence × 1,000).
   • Buyers who bought only smartphone = 400.

2. Recommendation Logic:

   • If a new user adds a smartphone to cart, Daraz shows the charger with:
     • Confidence: 60% (high likelihood of purchase).
     • Upsell Potential: ₹1,200 extra per transaction.

Real-World Tie-In:

• Daraz’s "Complete the Look" section uses this exact logic.
• Result: 15% increase in average order value for smartphone categories.

Key Takeaways for Exam Success

1. Recommender systems use collaborative filtering (CF) or content-based methods to predict preferences.
2. Cross-selling = complementary items (e.g., phone + case); upselling = premium versions (e.g., basic → premium Ncell plan).
3. Collaborative filtering excels with existing data but struggles with new users/items (cold start).
4. Content-based filtering works well for new items but may create filter bubbles.
5. Nepalese examples:
   • Daraz: Item-item CF for "Frequently Bought Together."
   • Khalti: User-user CF for personalized discounts.
   • Pathao: Location-based upselling for "Night Combos."
6. Ethical challenges: Privacy, filter bubbles, and explainability (why was this recommended?).

Quick Revision Table