PaperlayerView exact revision on arXiv

Source-linked AI summary

Generating Personalized Recipes from Historical User Preferences

Bodhisattwa Prasad Majumder, Shuyang Li, Jianmo Ni, Julian McAuley

arXiv:1909.00105v1cs.CLcs.AIcs.LG

TL;DR

The paper addresses recipe generation for cooks who know a dish and some ingredients but lack complete ingredient knowledge and want recipes reflecting their tastes. It generates personalized instructions from incomplete inputs and historical preferences, and reports plausible, personalized, coherent recipes preferred by human evaluators.

  • Problem

    Existing recipe-generation approaches do not address users who have incomplete ingredient knowledge and culinary preferences that should shape a recipe.

    Source: §1 · p. 1

  • Method

    The paper generates personalized recipes from incomplete input specifications by leveraging preferences learned from users' recipe histories.

    Source: §1 · p. 1, §6 · p. 5

  • Results

    Personalized models generated plausible, personalized, coherent recipes preferred by human evaluators; evaluators preferred them to baseline 63% of the time.

    Source: §6 · p. 5, §5 · p. 5

  • Takeaways & Limitations

    The paper introduces coherence measures for instructional text and personalization metrics to assess personalized recipe generation.

    Source: §6 · p. 5

    BLEU-1/4 are not correlated with subjective recipe quality because valid recipes can use the same main entities while being written in many ways.

    Source: §5 · p. 4

Abstract

from arXiv · show

Existing approaches to recipe generation are unable to create recipes for users with culinary preferences but incomplete knowledge of ingredients in specific dishes. We propose a new task of personalized recipe generation to help these users: expanding a name and incomplete ingredient details into complete natural-text instructions aligned with the user's historical preferences. We attend on technique- and recipe-level representations of a user's previously consumed recipes, fusing these 'user-aware' representations in an attention fusion layer to control recipe text generation. Experiments on a new dataset of 180K recipes and 700K interactions show our model's ability to generate plausible and personalized recipes compared to non-personalized baselines.

1 Introduction

The paper introduces personalized recipe generation for cooks with incomplete ingredient knowledge, using historical preferences to produce recipes tailored to individual tastes. It also presents a dataset and evaluation strategies for assessing generated instructional text.

1 Introduction

  • The model attends to latent profiles learned from a user's previously consumed recipes and fuses user-aware representations with decoder output to generate personalized instructions.Inputs include a dish name, key ingredients, and calorie level.Source: §1 · p. 1
  • The paper is the first to study personalized recipe generation, where output quality depends heavily on content such as ingredients and cooking techniques.Source: §1 · p. 1
  • The task expands incomplete recipe inputs into plausible, personalized recipes by leveraging historical user preferences.The motivating users may know a dish and a few basics but not all its ingredients, a gap existing recipe-generation approaches do not address.Source: §1 · p. 1
  • The work releases a dataset of 180K+ recipes and 700K+ user reviews and introduces coherence-centered evaluation strategies for instructional text.It also reports qualitative and quantitative evidence that personalized models generate high-quality, specific recipes aligned with historical user preferences.Source: §1 · p. 1

2 Related Work

Prior recipe-generation work models ingredients and cooking actions or uses attention to guide ingredient use, while personalized generation has focused on writing style. This paper instead uses historical preferences to guide recipe content.

2 Related Work

  • Recipe-generation approaches represent ingredients and cooking actions or use checklist and ingredient-reference attention to guide generation.Kiddon et al. (2015) and Bosselut et al. (2018b) model recipes as ingredient entities acted on by cooking actions; Kiddon et al. (2016) and Yang et al. (2017) use attention over ingredients.Source: §2 · p. 2
  • Inferring a full ingredient list to constrain generation is infeasible when recipes may draw from a space of 10K+ ingredients, so the paper uses historical preferences instead.Source: §2 · p. 2
  • Prior personalized generation addressed review writing styles, whereas this work attends to historical recipe preferences to generate content.Recipe personalization must address content such as ingredients and cooking techniques.Source: §2 · p. 2

3 Approach

The model expands a dish name, partial ingredients, and calorie level into recipe instructions, conditioning generation on encoded inputs and a user’s prior recipes and techniques. It combines these contexts with decoder output to produce personalized text.

3 Approach

  • A two-layer GRU decoder generates recipe instructions from the recipe name, partial ingredient list, and calorie-level input.The output is a token sequence; the decoder is initialized by projecting concatenated encoder outputs.Source: §3 · p. 2
  • Name, ingredient, and calorie inputs are encoded separately, with attention over encoded ingredients providing a context at each decoding step.Name and ingredient sequences use two-layer bidirectional GRUs; the calorie level is embedded and projected.Source: §3 · p. 2
  • User preference comes from attention over previously reviewed recipes and techniques, with the k most recent recipes used to account for preference drift.Recipe histories can be represented by recipe embeddings or averaged name-token embeddings.Source: §3 · pp. 2–3
  • An attention-fusion layer combines user-aware contexts with decoder output to determine the generated token distribution.Training maximizes the log-likelihood of the generated sequence conditioned on input specifications and user preferences.Source: §3 · p. 3
  • In the illustrated Prior Name example, attention to previously consumed savory recipes suggests adding cilantro.The figure emphasizes prior-recipe attention scores with darker shading; ingredient attention is omitted.Source: §3 · p. 3, Figure 1 · p. 3

4 Recipe Dataset: Food.com

The dataset is drawn from Food.com recipes and reviews spanning 18 years, then filtered to support recipe generation and user-history modeling. Evaluation holds out each user’s two most recent reviews and estimates technique adherence by string matching.

4 Recipe Dataset: Food.com

  • Filtering for recipes with at least 3 steps and 4–20 ingredients, and users with at least 4 reviews, leaves 180K+ recipes and 700K+ reviews.The source collection contains 230K+ recipe texts and 1M+ reviews from 2000–2018.Source: §4 · p. 3
  • The latest review per user is the test set, the second-latest is validation, and earlier reviews are training; evaluation uses recipes absent from training.This is sequential leave-one-out evaluation.Source: §4 · p. 3
  • The authors manually select 58 techniques from 384 cooking actions and approximate technique adherence by string matching recipe text to the list.Bake, combine, pour, and boil together account for 36.5% of technique mentions.Source: §4 · p. 3

5 Experiments and Results

The experiments find that personalization improves several measures of recipe generation, including user matching, recipe-level coherence, and human preference. The results also show a diversity–overlap trade-off and use coherence measures suited to instructional text.

5 Experiments and Results

  • Personalized models outperform the baseline in BPE perplexity and generate more diverse recipes, while BLEU-1/4 and ROUGE-L are comparable overall.Greater diversity accompanies better BLEU-1 with more key entities but worse BLEU-4 as phrasing diverges from gold.Source: §5 · p. 4
  • BLEU-1/4 are limited proxies for subjective recipe quality because valid recipes can share ingredients while differing in wording.The paper notes that n-gram matching does not correlate with subjective recipe quality.Source: §5 · p. 4
  • In cocktail samples, Prior Name adds lemon and mint associated with the user's previously consumed coconut mousse and pork skewers.The qualitative comparison presents outputs ranging from a generic baseline to a blended cocktail.Source: §5 · p. 4, Table 3 · p. 4
  • All personalized models beat baselines in UMA and MRR; Prior Name leads both by a large margin, with prior recipe names strong personalization signals.The evaluation ranks each generated recipe against ten user profiles, including the gold user.Source: §5 · p. 5
  • 1.78–1.82 versus 1.77 for the baseline: personalized models score higher on recipe-level coherence.The metric scores step ordering by comparing generated recipes with forward and reversed orderings.Source: §5 · p. 5, Table 4 · p. 5
  • Local coherence is also measured by BERT-predicted entailment across consecutive steps, averaged over generated recipes.The entailment model is trained on consecutive positive step pairs and non-consecutive negative pairs.Source: §5 · p. 5
  • Human evaluators preferred personalized outputs to baseline 63% of the time; in a 90-recipe survey, 60% found them more coherent and preferable.The pairwise evaluation covered 310 recipe pairs.Source: Table 2 · p. 4, §5 · p. 5

6 Conclusion

The paper concludes that personalized recipes can be generated from incomplete specifications and user histories. It reports plausible, coherent outputs preferred by human evaluators, while identifying structured recipe representations as future work.

6 Conclusion

  • The paper frames personalized recipe generation from incomplete input specifications and user histories as a new task.Its conclusion describes a dataset of 180K recipes and 700K reviews.Source: §6 · p. 5
  • On that dataset, personalized models generate plausible and coherent recipes that human evaluators prefer for consumption.The paper introduces automatic coherence measures for instructional text and personalization metrics to support these claims.Source: §6 · p. 5
  • Future work includes structured recipe representations for ingredient properties and references to ingredient collections such as “dry mix.”Source: §6 · p. 5

Appendix

The appendix reports the scale and coverage period of the raw Food.com data and points to interaction statistics and sample generated recipes.

Appendix

  • The raw Food.com data contains 270K recipes and 1.4M user-recipe interactions spanning January 2000 to December 2018.Source: Food.com: Dataset Details · p. 8
  • Table 5 presents Food.com interaction statistics before and after data processing.Source: Table 5 · p. 8
  • Tables 7 and 8 provide sample generated recipes for chicken chili and sweet waffles, respectively.Source: Generated Examples · p. 8

Human Evaluation

The human evaluation compares baseline and personalized recipe outputs using partial recipe specifications. The supplied figures and tables identify survey examples and sample recipes or interaction data.

Human Evaluation

Related papers

Loading 1909.00105v1…