Framing a Prediction Problem

Prediction Task

• Type: Binary classification
• Goal: Predict whether a recipe will receive high or low ratings based on its nutritional health informations and complexity (high rating represented by having an average rating ≥ 4.5 stars).

Response Variable

• high_rating (Boolean) — True if the recipe’s avg_rating ≥ 4.5; otherwise False.
• Why:
  • Converting recipes’ average rating to high/low rating classes can make our predicting result more intuitive.
  • Since ratings are mainly concentrated on 5 points, we think it is reasonable to define that recipes with high overall ratings have an average ratings of at least 4.5 points.

Feature Selection Rationale

We began with the full suite of complexity and nutrition metrics but pared down to those with the strongest, most interpretable signals:

• Complexity: originally considered minutes, n_steps, and n_ingredients, but found that n_steps and n_ingredients tracked very closely with minutes (similar distributions and correlations). To avoid redundancy, we retained minutes alone.
• Nutrition: Whether containing healthy-related tags provides some health information about the recipes. From the nutrition data, we selected features that, as observed in earlier analysis, appear to have certain patterns with average rating (listed below). ‘total_fat’, ‘protein’, and carbohydrates’ do not directly provide nutritional health information, since total fat may contain saturated fat (healthy) or unsaturated fat (unhealthy). Protein may come from fish that meet the omega-3 requirement, or from unhealthy processed meat or high-fat red meat. And carbohydrates are divided into high-quality carbohydrates and low-quality carbohydrates. Thus, they were omitted to at the prediction stage.

Chosen Features Known at Prediction Time

• Complexity
  • minutes (Preparation time)
• Nutrition
  • calories
  • sugar, sodium (PDV)
• Meta Tag
  • is_healthy (Boolean derived from health-related tags)

Evaluation Metric

F₁-score

• Class imbalance: Only ~20 % of recipes cross the 4.5-star line; accuracy would be inflated by the majority class.
• Balanced error cost: We care about both precision (avoid over-hyping average dishes) and recall (don’t miss hidden gems). F₁ is the harmonic mean of the two and captures that trade-off.

Time-of-Prediction Justification

All chosen features exist the moment a recipe is published: preparation time, nutrition breakdown, and author-assigned tags. Excluding any data that arrives later ensures the prediction is feasible in real time and avoids future information leakage.

---