3.1 The Challenge of Size Recommendation

Size recommendation is a complex problem that goes beyond simple measurement matching. Different brands use different sizing standards, fabrics behave differently, and individuals have varying fit preferences. A machine learning-based size recommendation system must account for all these variables while providing recommendations with high confidence.

Traditional size charts provide ranges (e.g., "Medium: chest 96-101cm"), but real-world sizing is more nuanced. A person with a 98cm chest might prefer Medium in one brand but need Large in another due to differences in cut, fabric stretch, or design philosophy. Effective size recommendation requires understanding these subtleties through data-driven modeling.

3.2 Recommendation Algorithm Architecture

The WIA-IND-001 size recommendation system uses a multi-model ensemble approach that combines several specialized algorithms:

Measurement-to-Size Model

Maps body measurements directly to garment sizes using neural networks trained on millions of fit outcomes. This model learns the relationship between measurements and size labels across different brands and garment types. It uses a deep learning architecture with layers that progressively learn higher-level patterns from raw measurements.

Collaborative Filtering

Leverages purchase and return data from similar users. If many users with similar measurements purchased size M and kept it, the system recommends M with higher confidence. This addresses the "cold start" problem for new brands by learning from collective user behavior.

Brand-Specific Sizing Models

Each major brand has unique sizing characteristics. The system maintains brand-specific models that capture these differences. For example, luxury brands often size smaller than fast-fashion brands. These models are continuously updated as new data becomes available.

Fabric and Garment Type Adjustments

Stretchy fabrics allow for tighter fits, while rigid materials require more room. The algorithm applies adjustments based on fabric composition (cotton, polyester, spandex percentage, etc.) and garment type (t-shirt, dress shirt, jacket, jeans, etc.).

Personal Preference Learning

As users purchase and provide feedback, the system learns individual preferences. Some users consistently prefer looser fits, while others like tight fits. The model personalizes recommendations based on this learned preference profile.

3.3 Training Data and Model Development

Effective size recommendation requires extensive training data:

Data Sources

• Body Measurements: Millions of 3D body scans with precise measurements
• Purchase Data: Historical purchase records with size selections
• Return Data: Return reasons and replacement sizes chosen
• Fit Feedback: User ratings on how well items fit (too small, perfect, too large)
• Garment Specifications: Detailed measurements of actual garments from brands
• Fabric Data: Material composition and stretch characteristics

Data Privacy Considerations

All training data must be anonymized and aggregated to protect individual privacy. Differential privacy techniques add statistical noise to prevent identification of specific individuals while maintaining overall pattern accuracy. Users consent to data usage with full transparency about how their information contributes to model improvement.

Model Training Pipeline

1. Data Collection and Cleaning
   - Remove outliers and errors
   - Normalize measurements
   - Encode categorical variables

2. Feature Engineering
   - Calculate body shape indices
   - Extract brand sizing patterns
   - Generate interaction features

3. Model Training
   - Split data: 70% train, 15% validation, 15% test
   - Train ensemble of models
   - Hyperparameter optimization
   - Cross-validation

4. Evaluation and Deployment
   - Measure accuracy on test set
   - A/B testing in production
   - Monitor performance metrics
   - Continuous retraining

3.4 Confidence Scoring

Every size recommendation includes a confidence score indicating the algorithm's certainty:

Confidence Levels

• 95-100%: Very High Confidence - Standard measurements, common size, extensive brand data
• 85-94%: High Confidence - Good measurement quality, moderate brand data
• 70-84%: Moderate Confidence - Some uncertainty in measurements or limited brand data
• Below 70%: Low Confidence - Unusual body proportions, new brand, or measurement quality issues

Factors Affecting Confidence

• Measurement accuracy and method used
• Availability of brand-specific data
• Number of similar users in training data
• Fabric stretch and garment type
• How close measurements are to size boundaries
• Consistency of user's fit preference history

When confidence is below 85%, the system may recommend trying multiple sizes or suggest consulting detailed size charts.

3.5 Multi-Size Recommendations

Rather than recommending a single size, the system provides a ranked list with probabilities:

{
  "primaryRecommendation": {
    "size": "M",
    "confidence": 0.92,
    "fitQuality": "perfect",
    "reasoning": "Based on chest 92cm, waist 78cm"
  },
  "alternativeRecommendations": [
    {
      "size": "L",
      "confidence": 0.65,
      "fitQuality": "slightly loose",
      "reasoning": "Consider if you prefer looser fit"
    },
    {
      "size": "S",
      "confidence": 0.28,
      "fitQuality": "tight",
      "reasoning": "May fit if fabric is very stretchy"
    }
  ],
  "philosophy": "弘益人間"
}

This probabilistic approach gives users more information to make informed decisions and reduces the likelihood of returns.

3.6 Handling Edge Cases

Certain scenarios require special handling:

Between Sizes

When measurements fall exactly between two sizes, the system recommends based on fit preference and garment type. For fitted garments, it may suggest the smaller size; for comfort-focused items, the larger size.

Unusual Proportions

Bodies with proportions outside typical ranges (e.g., very tall with narrow build) may not fit standard sizes well. The system identifies these cases and suggests custom tailoring or brands known for accommodating specific body types.

Limited Brand Data

For new or small brands with insufficient data, the system uses transfer learning from similar brands and applies conservative confidence scoring. It may request user feedback to improve future recommendations.

Size Availability

When the recommended size is out of stock, the system evaluates alternatives considering fit tolerance. It might suggest a nearby size if the garment has stretch, or recommend waiting for restock if the fit would be significantly compromised.

3.7 Regional Sizing Variations

The standard supports international size systems:

US Sizing

Numeric (2, 4, 6, 8, etc.) for women, letter (XS, S, M, L, XL) for both genders. The system maps measurements to US sizes using region-specific charts.

European Sizing

Numeric system based on body measurements (38, 40, 42, etc.). Generally more consistent across brands than US sizing.

UK Sizing

Similar to US but offset (UK 10 = US 6 approximately). The system handles conversions accurately.

Asian Sizing

Often runs smaller than Western sizing. The algorithm applies regional adjustment factors and learns brand-specific patterns for Asian market brands.

3.8 Fit Preference Profiles

Users can define explicit fit preferences that influence recommendations:

• Tight Fit: Prefers form-fitting garments, may size down
• Slim Fit: Fitted but not restrictive
• Regular Fit: Standard fit with comfortable room
• Relaxed Fit: Looser, comfortable fit
• Oversized: Deliberately large, trending style

These preferences can be set globally or per garment category (tight shirts but relaxed pants, for example).

3.9 Real-Time Learning and Adaptation

The recommendation system continuously improves through feedback loops:

Implicit Feedback

• Purchase signals interest and likely good fit
• Returns indicate sizing issues
• Replacement size reveals correct fit
• Time to purchase shows decision confidence

Explicit Feedback

• Post-purchase fit ratings
• Review comments mentioning sizing
• Direct size accuracy feedback
• Fit preference adjustments

Models are retrained weekly using the latest data, ensuring recommendations improve over time. A/B testing validates changes before full deployment.

3.10 Integration with Virtual Try-On

Size recommendations are enhanced when combined with virtual try-on visualization:

Users can see how recommended sizes will actually look on their body shape. The virtual fitting shows differences between sizes visually, making the choice more intuitive. This combination of data-driven recommendation and visual confirmation creates the most confident purchasing experience.

The system can highlight fit issues (tight shoulders, loose waist) and suggest size adjustments or alternative products that better match body shape.