IntelliMeal - Enhancing Creativity by Reusing Domain Knowledge in the Adaptation Process

Kari Skjold, Marthe Øynes, Kerstin Bach, Agnar AamodtICCBR (Workshops) 2017 · OpenReview

TL;DR

This paper introduces IntelliMeal, a Case-Based Reasoning (CBR) system that enhances recipe recommendation by creatively adapting existing recipes using domain knowledge. By intelligently reusing ingredient hierarchies and adaptation rules, IntelliMeal can generate novel recipe variations that better satisfy user preferences, even with a limited initial recipe database.

What Problem Does This Paper Address?

Recommending recipes that perfectly match a user's desires can be challenging. Traditional recommendation systems often struggle to go beyond simple matching or substitution, leading to a lack of creativity and personalization. The 2017 Computer Cooking Contest (CCC) highlighted this challenge, prompting research into more sophisticated approaches. This paper addresses the gap in creating truly personalized and creative recipe recommendations by focusing on how to adapt existing recipes using a rich understanding of ingredients and cooking principles. The goal is to move beyond just finding a recipe that fits to one that is tailored and potentially improved based on user input.

Methods

IntelliMeal employs a modified Case-Based Reasoning (CBR) cycle. The core innovation lies in its adaptation process, which leverages domain knowledge in the form of hierarchical taxonomies of ingredients and a set of adaptation rules.

Here's a breakdown of the key methods:

  • Modified CBR Cycle: IntelliMeal incorporates a second retrieval step using an "ephemeral case base." This allows the system to evaluate adapted recipes before permanently adding them to the main case base.
  • Hierarchical Taxonomies: Ingredient categories are organized hierarchically, defining similarity relationships between ingredients. This allows for more nuanced substitutions and additions (e.g., any type of "meat" can substitute for any type of "fish").
  • Adaptation Rules: The system uses two main types of rules:
  • Simple Rules: These handle additions, deletions, and title modifications based on specified conditions.
  • Substitution Rules: These are more complex, allowing for the replacement of one ingredient with another, potentially based on user preferences or similarity.
  • Creativity through Adaptation: The system focuses on enhancing creativity by using these rules to perform addition, creation, substitution, suitability adjustments, and title customization. This allows for more radical modifications than simple ingredient swaps.

Datasets

  • Case Base: The system was initially trained and evaluated on a limited case base of 21 sandwich recipes.
  • Taxonomies: The taxonomies used were extended versions of those employed in previous CCC participants' systems (specifically, CookIIS).
  • Evaluation Data: For user evaluation, the system gathered 3414 responses over 42 recipes in an online quiz where users guessed if a recipe was human or computer-generated.

Key Insights and Ideas

  • Reusing Domain Knowledge is Key: The paper strongly suggests that effectively reusing structured domain knowledge (like ingredient hierarchies) is crucial for enabling creative adaptation in recipe recommendation.
  • Beyond Simple Substitutions: IntelliMeal demonstrates that moving beyond basic ingredient substitutions to more complex rule-based adaptations (additions, deletions, similarity-based replacements) leads to more personalized and potentially novel recipe outcomes.
  • Limited Case Base, High Potential: The success with a small initial case base highlights the power of intelligent adaptation. The system can generate a wide variety of recommendations by creatively modifying a few core recipes.
  • Human-like Creativity is Achievable: The evaluation results suggest that users found it difficult to distinguish between computer-adapted and human-created recipes, indicating a high level of perceived creativity.

Key Results

  • Improved Similarity Scores: Adaptation of recipes led to an average increase of 0.32 in the similarity score for the top five suggested recipes per query.
  • User Perception: In an online quiz, users incorrectly identified computer-adapted recipes as human-created in 53.43% of cases, and original (non-adapted) recipes as computer-created in 50.09% of cases. This indicates that users struggled to differentiate between the two.
  • Ranking Accuracy: The system's recipe rankings showed a discrepancy of 5.52 out of 25 compared to human test subjects' rankings, and they shared 3.57 out of 5 recipes in their top 5.

Limitations and Future Work

  • Limited Case Base: While the system performed well with a small case base, its full potential might be realized with a larger and more diverse set of initial recipes.
  • Subjectivity of Taste: The evaluation relied on user feedback regarding "tastiness" and perceived creativity. Taste is inherently subjective, and further research could explore more objective measures or a wider range of user preferences.
  • Computational Cost: The paper doesn't extensively discuss the computational cost of the adaptation process, which could be a factor for real-time applications with very large case bases.
  • Future Work: The authors suggest further exploration of rule generation and more sophisticated adaptation strategies to enhance creativity and user satisfaction. They also hint at the potential for integrating more complex cooking workflows.

Summary updated

BibTeX
@inproceedings{skjold2017intellimeal,
  author = {Kari Skjold and Marthe Øynes and Kerstin Bach and Agnar Aamodt},
  title = {{IntelliMeal - Enhancing Creativity by Reusing Domain Knowledge in the Adaptation Process}},
  booktitle = {ICCBR (Workshops)},
  year = {2017},
  url = {https://openreview.net/forum?id=RsYHihxSfc},
}
Figure