Cookpadのレシピ検索 — 学習to-rank実装

Cookpad, as one of Japan's largest-class recipe posting services, processes several million recipe searches per day. To rank the recipes that most satisfy users thinking about "tonight's menu" at the top, simple keyword matching alone is insufficient. This article organizes the overall picture of Cookpad's search ranking.

The features fed into the ranking model are diverse — ingredient match degree, recipe popularity (number of tsukurepo "cooked reports" and saves), seasonal compatibility, the user's past browsing history, author trust level, time elapsed since posting, and more. Dozens to hundreds of features are combined.

Japanese search queries are not separated by spaces between words like English. To process the query "pork cabbage easy," it must be decomposed into words by a morphological analyzer (mecab or kuromoji). Cookpad maintains a dictionary specialized for the cooking domain so that ingredient names and cooking terms can be correctly extracted.

The same ingredient has abundant notation variations — "人参 / ニンジン / にんじん" for carrot, "豚バラ / 豚ばら肉" for pork belly. By normalizing these at indexing time and performing synonym expansion on the query side, the same recipes will hit regardless of how the user writes them.

Rather than weighting features manually, models that learn from user clicks and tsukurepo postings as ground truth are used. Gradient boosting tree models like LambdaMART and XGBoost are representative, automatically learning the importance of each feature. The strength is that they can evaluate non-linearly through combinations with ingredient match, rather than as a simple linear relationship like "high popularity is good."

Japanese food culture is strongly tied to seasons — chilled Chinese noodles and somen are popular in summer, hot pots and stews in winter. Cookpad adds features linked to ingredient peak seasons and temperature, designed so that even the same "easy recipe" query yields different results by time period. Regional differences (such as Kansai vs Kanto seasoning) are also taken into consideration.

Before deploying a new model or feature to production, the difference from the existing model is always measured via A/B test. Evaluation metrics emphasize not only click rate but also downstream actions that show the dish was actually cooked — "was a tsukurepo posted?" "was it saved?" — because clicking without cooking cannot be called true satisfaction.

Even with the same query "curry," personalization considering an individual's past behavior is performed so that meat-eating users see chicken curry on top, and vegetable-centric users see vegetable curry on top. However, overly strong personalization creates a filter bubble where "only the same recipes ever appear," so a device for mixing in new recipes at a certain ratio is needed.

To instantly return dozens of results from millions of recipes, two-stage search is the basic approach. The first stage narrows candidates down to hundreds via an inverted index, and the second stage applies a heavy learning-to-rank model for final ranking. Applying the model to all items would explode response time, so devices for producing accuracy at realistic computational cost are applied.

Cookpad's search is a mechanism that, in addition to the basics of "ingredient matches" and "high popularity," combines Japanese-specific notation variations, seasonal and regional variation, and individual preferences to deliver "the recipe I want to cook tonight." Only when both machine learning and cooking domain knowledge come together is a satisfying experience realized.