💻 computer science

Quantifying Recipe Difficulty Based on Recipe Features

This paper proposes a method for quantifying recipe difficulty by integrating six normalized features—cooking action level, text length, utensils, seasonings, ingredients, and steps—using regression coefficients to avoid double-counting, with its validity confirmed through a survey and statistical tests to enable fine-grained recipe recommendations.

Original authors: Sae Okatani, Naoko Yoshii, Masami Takata

Published 2026-11-12

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Sae Okatani, Naoko Yoshii, Masami Takata

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the modern kitchen, the barrier to cooking a healthy meal is rarely a lack of ingredients or a shortage of time, but rather a mismatch between a cook's skill and the complexity of the dish they attempt. While the internet has made millions of recipes available at the click of a button, these digital libraries often lack a crucial piece of information: a reliable measure of how hard a recipe actually is. Without this, a beginner might accidentally select a dish requiring advanced techniques, leading to frustration or a ruined meal, while an expert might find the available options too simplistic. The field of food computing has long focused on matching recipes to what people like to eat or what is nutritionally sound, but it has largely overlooked the physical and mental effort required to prepare them. To bridge this gap, researchers need a way to translate the text of a recipe into a single, clear number that reflects its true difficulty, moving beyond vague labels like "easy" or "hard" to a precise scale that can guide users to the right challenge.

A team of researchers at Nara Women's University set out to solve this problem by developing a method to quantify recipe difficulty using nothing more than the text of the recipe itself. They began by analyzing a massive collection of approximately 605,000 recipes from Cookpad, Japan's largest recipe-sharing website. Instead of relying on human judges to taste and rate every dish, the team designed a system to extract six specific details from the writing of each recipe. These details included the most complex cooking action mentioned, the total number of characters in the instructions, the variety of kitchen tools required, the number of different seasonings used, the count of main ingredients, and the total number of steps in the process. The researchers recognized that these factors are not independent; for instance, a recipe with many steps naturally tends to have more characters, and a dish with many ingredients often requires more tools. If they simply added these numbers together, they would be counting the same information twice, skewing the final result.

To handle this overlap, the team used a statistical approach to understand how these features relate to one another. They discovered that the number of characters and the number of steps were so closely linked that they essentially shared the same information. Rather than treating them as separate burdens, the researchers combined them into a single, adjusted value that represented the overall verbosity and complexity of the instructions without double-counting. They applied similar logic to the other features, using regression analysis to blend the data in a way that respected the natural connections between the ingredients, tools, and steps. This process resulted in four different mathematical models for calculating difficulty, each weighing the factors slightly differently. The goal was to produce a continuous score, a smooth scale where every recipe has a unique position, rather than forcing them into a few rigid categories.

To see which of these four models worked best, the researchers conducted a survey with 30 participants, all women in their twenties. They asked these participants to rank nine different recipes from easiest to most difficult based on their own judgment. The researchers then compared the human rankings against the rankings generated by their four computer models. The results showed that one specific approach, which they called Method 2, aligned most closely with human perception. This method calculated difficulty by averaging the complexity of the cooking actions with the counts of cooking utensils, seasonings, and ingredients, while carefully integrating the information about steps and text length to avoid overestimating the effort. It successfully predicted that a recipe with a high number of seasonings and ingredients would feel more difficult to a human cook, even if the number of steps was moderate.

The study confirmed that a recipe's difficulty is not just about the hardest single action, such as deep-frying, but is a cumulative effect of the entire process. The researchers found that simply counting the number of steps or the length of the text was not enough, as these factors often overlap. By using their integrated method, they could create a strict ordering of recipes where no two dishes had the exact same difficulty score, allowing for a much finer level of recommendation than previous systems. While the study was limited to a small group of young women and a specific set of recipes, the findings suggest that it is possible to automatically generate a reliable difficulty score for hundreds of thousands of dishes. This advancement could eventually help recipe websites guide users to meals that match their actual abilities, turning the overwhelming choice of the internet into a manageable and successful cooking experience.

Technical Summary: Quantifying Recipe Difficulty Based on Recipe Features

Problem Statement

Despite the proliferation of recipe-sharing platforms like Cookpad, which host millions of recipes, there is a lack of standardized, computable metrics for recipe difficulty. Existing search and recommendation systems primarily rely on user preferences, ingredient relationships, or nutritional health, often neglecting the skill level and effort required to execute a recipe. Consequently, users frequently select recipes that exceed their capabilities, leading to failed cooking attempts or abandonment. While previous studies have attempted to quantify difficulty, they often rely solely on cooking actions (resulting in coarse, discrete levels) or simply sum multiple features without accounting for information overlap, failing to provide the fine-grained, continuous difficulty values necessary for precise recipe recommendation.

Methodology

The authors propose a top-down computational method to quantify recipe difficulty as a continuous value based on six automatically extracted features from recipe text. The methodology proceeds through the following stages:

1. Feature Extraction and Quantification

Six features are extracted from approximately 605,000 Cookpad recipes (excluding entries with missing data):

(a) Cooking Action Level: Derived from a dictionary of 118 cooking verbs classified into four difficulty levels (1–4) based on the Japanese Home Cooking Skills Certification. The level for a recipe is defined as the maximum difficulty among its extracted actions.
(b) Number of Characters: The total character count of the recipe instructions (excluding spaces and line breaks).
(c) Number of Cooking Utensils: The count of unique utensil types mentioned.
(d) Number of Seasonings: The count of unique seasoning types mentioned.
(e) Number of Ingredients: The count of unique ingredients, excluding those classified as seasonings.
(f) Number of Steps: The total count of cooking steps.

2. Normalization

All features are normalized to a [0, 1] scale using Min-Max normalization to ensure comparability across different units and magnitudes.

3. Correlation and Regression Analysis

Pearson correlation analysis revealed that the cooking action level (a) is largely independent of other features. However, the number of characters (b) showed strong correlations with the number of steps (f) ( $r=0.86$ ), cooking utensils (c) ( $r=0.58$ ), and ingredients (e) ( $r=0.36$ ). To prevent double-counting shared information (e.g., a long text often implies many steps), the authors employed simple linear regression.

4. Integration via Regression Slopes

Instead of simple summation, the method integrates correlated features using regression slopes ( $\alpha$ ) to create intermediate variables. For two correlated features $x$ and $y$ (where $y$ is regressed on $x$ with slope $\alpha$ ), an integrated variable $z$ is defined as a weighted average: $z = \frac{\alpha x + y}{1 + \alpha}$ .

Stage 1: The number of steps (f) is integrated with the number of characters (b) to form $\tilde{b}$ .
Stage 2: The remaining features (utensils, seasonings, ingredients) are integrated with $\tilde{b}$ to form $\tilde{c}$ , $\tilde{d}$ , and $\tilde{e}$ .

5. Proposed Difficulty Formulas

Four integration methods were constructed:

Method 1: Averages the action level (a) and all intermediate variables ( $\tilde{b}, \tilde{c}, \tilde{d}, \tilde{e}$ ).
Method 2: Averages the action level (a) and the mean of $\tilde{c}, \tilde{d}, \tilde{e}$ . This method does not include $\tilde{b}$ directly but incorporates it indirectly through the calculation of $\tilde{c}, \tilde{d},$ and $\tilde{e}$ .
Method 3: Averages the action level (a) and the mean of all intermediate variables.
Method 4: Averages the action level (a) and the raw counts of utensils, seasonings, and ingredients, ignoring characters and steps.

Experimental Evaluation

The validity of the four methods was evaluated through a questionnaire survey involving 30 female respondents in their 20s. Participants ranked nine selected recipes by perceived difficulty.

Statistical Significance: A Friedman test confirmed that the nine recipes were perceived as having significantly different difficulty levels ( $p < 0.001$ ).
Rank-Distance Metrics: The rankings generated by the four methods were compared against the survey rankings using four metrics: Spearman's footrule distance, Kendall distance, Cayley distance, and Ulam distance.
Results: Method 2 demonstrated the highest similarity to the survey results, achieving the lowest Spearman's footrule distance (12) and low values in other metrics.
- Method 1 and Method 3 showed slightly higher deviations, attributed to the overestimation of character count influence.
- Method 4 showed the largest deviation (Footrule distance 18), particularly failing to rank a recipe with many steps (Recipe I) as difficult, suggesting that ignoring step count leads to significant errors.
Single-Feature Baseline: Single-feature rankings (e.g., difficulty based only on ingredients) resulted in many tied ranks and lower similarity to survey results compared to the proposed integrated methods.

Key Contributions

Automated Feature Extraction: The study extracts six distinct features from approximately 605,000 recipes, moving beyond simple action-based classification to a multi-faceted approach.
Correlation-Aware Integration: By utilizing regression slopes to integrate correlated features, the method avoids double-counting overlapping information (specifically between character count and step count), a limitation in previous additive models.
Continuous Total Ordering: The proposed method generates a strict total order (continuous values without ties), enabling fine-grained comparison of recipes, which is a property not provided by single-feature baselines or discrete-level systems.
Empirical Validation: The study validates the computational model against human perception using rank-distance metrics, identifying a specific integration strategy (Method 2) that best aligns with user perception.

Significance and Claims

The paper claims to provide a necessary "recipe-side" difficulty measure that complements existing "person-side" measures of cooking skills and food literacy. By offering a computable, fine-grained difficulty score, the method addresses a gap in food recommender systems, which currently lack explicit modeling of the effort and skill required for preparation.

The authors modestly conclude that while the proposed method (specifically Method 2) successfully produces a ranking consistent with human perception, the study is limited by the small sample size (30 respondents) and demographic homogeneity (women in their 20s). They state that the generalizability of the results to users of different ages, genders, and experience levels requires further verification. The primary significance lies in establishing a framework for automated, continuous difficulty quantification that can support systems designed to help users select recipes matching their specific skill levels.