機械学習の先進的な事例 マーケティング (S3-109)

Main ideas, concepts, and lessons

• Purpose of the session (series “9th installment”)

  • Learn about advanced case studies of machine learning in marketing.
  • The presenter outlines a structured flow: 1) what “big data” B2C companies hold, 2) challenges companies face, 3) AI solutions, 4) real examples of AI in marketing, 5) a specific predictive model example from their company.

• What counts as “big data” in B2C

  • B2C companies target general consumers (e.g., department stores, convenience stores, mass retailers).
  • Data sources mentioned:
    • True POS / barcode scanning → purchase/transaction data.
    • IC cards for train travel → boarding/alighting data and shopping-related behavior.
    • E-commerce → purchase data, clicks, website visits.
  • Key idea: the combined data reveals customer attributes and behavior patterns linked to profit/engagement.

• Challenges in marketing that require AI

  • Marketers historically design scenarios/promotions using statistical methods applied to newly obtained customer data.
  • Prior approaches have limitations.
  • Reasons AI/machine learning are needed:
    • Consumers are increasingly diverse
    • Population is declining
  • Lesson: because of these shifts, companies need analysis methods that can learn over time, not rely only on static statistical planning.

• How AI is used in marketing operations (system landscape)

  • Data and operational systems described:

    • DMP (Data Management Platform): consolidates separately managed data (e.g., visitor behavior and attribute information for magazines and websites).

    • CRM (Customer Relationship Management): manages customer information.

    • Automated distribution systems: sends email newsletters and coupons.
    • MA (Marketing Automation): delivers information to consumers via mobile apps.
      • Core AI lesson:
    • AI should analyze and predict targets and recommendations (e.g., who to recommend what to).
    • Then human marketers apply optimal settings based on the AI outputs.

Methodology / instructional content (predictive model example, step-by-step)

Annual visit prediction model (example from the presenter’s company)

Goal: Predict how many days a customer will visit a store per year, based on their product purchase history.

• Step 1: Define input features from purchased products

  • Take the customer’s purchased products and represent them numerically.
  • For each product item:
    • Assign 1 if the customer purchased it
    • Assign 0 if the customer did not purchase it

• Step 2: Input the feature vector into the AI prediction model

  • The model uses the “combination of purchased products” as the input feature set.

• Step 3: Generate prediction output

  • The model outputs a predicted value for annual visiting days.
  • The example includes relative/visualized predictions (exact numeric labels are unclear due to subtitle artifacts), described as categories/values.

Example scenario A

• Purchase history assumption for “customer A”:
  • Purchased: wine game products
  • Not purchased: SD cards, cosmetics
• Representation idea: purchased items → 1, unpurchased items → 0
• Actual number of visits (example context): 3
• Interpretation:
  • If the model predicts an increase in annual purchases, it implies an increase in annual visits.
  • Conclusion stated: If customer A purchases cosmetics, annual visits increase (as described in the narrative).

Example scenario B (illustrative prediction outcomes)

• Sub-cases mentioned in the subtitles:
  • If customer A purchases product A → AI predicts annual visits “black” (subtitle label)
  • If customer B purchases product A → AI predicts annual visits “black triangles”
  • If customer B purchases product B → AI predicts annual visits “# b” (subtitle garbling)

Note: Some predicted labels are garbled in the subtitles (e.g., “black,” “black triangles,” “# b”), but the underlying instructional method is clear: encode purchases as features, run the model, interpret the prediction as expected visit frequency.

Speakers / sources featured

• Wada (introduced as: “Today, I’m Wada from Company C…”)