FlavorGraph Serves Up Food Pairings with AI, Molecular Science
It’s not just gourmet chefs who can discover new flavor combinations— a new ingredient mapping tool by Sony AI and Korea University uses molecular science and recipe data to predict how two ingredients will pair together and suggest new mash-ups.
不仅仅是美食家可以发现新的口味组合——
索尼人工智能
和
高丽大学
开发了一种新的
配料映射工具
，它利用
分子科学
和
食谱数据
来预测两种配料如何搭配，并提出新的
混搭建议
。
Dubbed FlavorGraph, the graph embedding model was trained on a million recipes and chemical structure data from more than 1,500 flavor molecules.
The researchers used PyTorch, CUDA and an NVIDIA TITAN GPU to train and test their large-scale food graph.
这个被称为
FlavorGraph
的图嵌入模型是在来自
1500多个风味分子
的
100万份食谱
和
化学结构数据
上进行训练的。研究人员使用
PyTorch
、
CUDA
和
NVIDIA TITAN
GPU
来训练和测试他们的
大规模食物图
。
Researchers have previously used molecular science to explain classic flavor pairings such as garlic and ginger, cheese and tomato, or pork and apple — determining that ingredients with common dominant flavor molecules combine well. In the FlavorGraph database, flavor molecule information was grouped into profiles such as bitter, fruity, and sweet.
研究人员以前用分子科学来解释
经典的味道搭配
，比如
大蒜和姜
，
奶酪和番茄
，或者
猪肉和苹果
——确定具有
共同主导风味分子
的成分结合得很好。在
FlavorGraph
数据库中，风味分子信息被分为
苦味
、
果味
和
甜味
等。
But other ingredient pairings have different chemical makeups, prompting the team to incorporate recipes into the database as well, giving the model insight into ways flavors have been combined in the past.
但其他成分的组合有不同的
化学组成
，这促使研究小组也将食谱纳入数据库，让模型了解过去
各种风味
的
组合方式
。
FlavorGraph. (A) Ingredient-ingredient relation. The relations between ingredients are shown; two ingredients are a “good pair” if they are used together in a large number of food recipes. The relations were obtained from Recipe1M. (B) Ingredient-compound relation. The relations between ingredients and chemical compounds are shown. These relations were obtained from FlavorDB and HyperFoods. (C) A partial view of FlavorGraph. Only 160 out of 6653 ingredients, 154 out of 1646 compounds, and their relations are shown in Fig. 1 for better illustration. Note that the whole graph was used for model training.
FlavorGraph。(A)成分-成分关系。说明了各成分之间的关系;如果两种食材在大量的食物食谱中一起使用，那么它们就是一对“好搭档”。关系式由Recipe1M求得。(B)成分-化合物关系。指出了成分与化合物之间的关系。这些关系从FlavorDB和HyperFoods中得到。(C) FlavorGraph的局部视图。6653种成分中只有160种，1646种化合物中只有154种，它们的关系如图1所示，以便更好地说明。注意，整个图用于模型训练。
“The outcome is pairing suggestions that achieve better results than ever before,” wrote Korea University researcher Donghyeon Park and Fred Gifford, strategy and partnerships manager at Sony. “These suggestions can be used to predict relationships between compounds and foods, hinting at new and exciting recipe techniques and driving new perspectives on food science in general.”
高丽大学研究员
朴东铉
(
Donghyeon Park
)和索尼战略与合作经理
弗雷德·吉福德
(
Fred Gifford
)写道:“结果是，配对建议取得了比以往任何时候都好的效果。”“这些建议可以用来
预测化合物和食物之间的关系
，
暗示新的和令人兴奋的食谱技术
，并
推动食品科学的新视角
。”
Featuring in
Scientific Reports
, FlavorGraph shows the connections between flavor profiles and the underlying chemical compounds in specific foods. It’s based on the
metapath2vec
model, and outperforms other baseline methods for food clustering.
在《
Scientific Reports
》中，FlavorGraph展示了
风味特征与特定食物中潜在化合物
之间的联系。它基于
metapath2vec
模型，并且优于其他
食物聚类
的基准方法。
Predicting compound-food relations When our modified version of metapath2vec is trained on all the nodes of FlavorGraph, the method trains not only the relations between foods but also the relations between foods and chemical compounds. Therefore, we believe that it is possible to predict pre-existing compound-food relations and undiscovered ones through a similarity search of our learned food representation vectors. We demonstrated a toy example task to predict relationship between compounds and foods. Figure 5) shows the prediction results of
network. To build this relation network, we first picked the 5 most frequently appeared flavor profiles (out of 582) in FlavorDB. We then randomly sampled ten of each corresponding random flavor compounds (out of 1561) upon the picked flavor profiles. Lastly, we randomly sampled 20 food ingredients (out of 6653) for each of the flavor compounds.
当我们改进的metapath2vec在FlavorGraph的所有节点上进行训练时，该方法不仅训练了食物之间的关系，还训练了食物与化合物之间的关系。因此，我们相信，通过我们学习的食物表示向量的相似性搜索，可以预测预先存在的复合食物关系和未发现的关系。我们演示了一个简单的任务来预测化合物和食物之间的关系。图5)显示了
网络的预测结果。为了建立这个关系网络，我们首先选择了FlavorDB中出现频率最高的5个风味配置文件(从582个中)。然后，我们随机抽取了10种相应的随机风味化合物(从1561种中)。最后，我们从6653种食品成分中随机抽取了20种风味化合物。
The researchers hope the project will lead to the discovery of new recipes, more interesting flavor combinations, and potential substitutes for unhealthy or unsustainable ingredients.
研究人员希望该项目能够发现
