Summary of content: Needless to say the importance of reasonable diet and balanced nutrition, but how to implement it is not simple. In order to get recipes that are more reasonable, healthier and more in line with people’s tastes, AI has also joined the team of nutritionists.
Keywords: diet recommendation nutritional balance recommendation system
In terms of healthy eating, modern people have an unimaginable enthusiasm for the keywords of health preservation, expert recommendations, and the most complete recipes.
The fitness crowd pursues muscle gain and is obsessed with a low-carb and high-ketogenic diet; the beauty-loving crowd pursues quick-acting weight loss, light fasting, and the Copenhagen weight loss method is the most popular; postoperative patients and pregnant women are brainwashed by word of mouth with millet porridge and trotters ; Young people who yearn for a green life and people who study Buddhism advocate vegetarianism, and try to maintain physical and mental health in the limited choices.
Is it really suitable for everyone?
Faced with a wealth of choices, how to match is scientific and reasonable? Some researchers are using the advantages of algorithms and big data to recommend healthier and more nutritious recipes for everyone.
AI learns to “see people’s dishes”
The biggest problem with the universal recipes mentioned above is that everyone has different physiques, different tastes, different geographical resources, and different consumption levels.
The high-quality recipes of some bloggers often include salmon and avocado, which are expensive and difficult to buy in some regions, which discourages many people.
In order to realize the dietary recommendations for thousands of people, many researchers have also introduced artificial intelligence technology.
Researchers from Rensselaer Polytechnic University in the United States and IBM Research have recently developed a personalized diet recommendation system pFoodReQ, which can recommend tailor-made recipes according to individual users’ preferences and dietary needs.
Personalized food recommendation based on the limited question answering on the large-scale food knowledge graph
Paper address: https://arxiv.org/pdf/2101.01775.pdf
The author pointed out in the paper that the existing diet recommendation methods generally have three main shortcomings:
Does not understand the exact requirements of users;
Does not consider the key factors of allergies and nutritional needs;
There are no customized recipes based on abundant food choices.
In this study, the research team proposed personalized food recommendations and regarded it as an answer to the restricted questions of the Food Knowledge Graph (KG), so as to try to solve the above problems in a unified way.
The team proposed a personalized food recommendation framework based on KBQA (Knowledge Base Question Answering), that is, personalized food recommendation through question and answer.
For different users, the system will conduct different demand analysis
Specifically, the pFoodReQ system will follow the user’s question, such as “What is a good breakfast with bread?”, and then retrieve all recipes that satisfy this query from KG. Then score the applicability of the ingredients in these recipes, and finally recommend the recipes with the highest rankings.
The structure of personalized food recommendation system based on question and answer mechanism
Finally, the verification experiment results show that their proposed method is significantly better than the non-personalized method and can recommend more relevant and healthier recipes.
There are thousands of recipes, AI only recommends dishes that suit you
In general, the team established the system through four steps: creating a data set, generating benchmark questions, compiling health guidelines, and system training.
Step 1: Create a recipe data set
Based on the extensive food knowledge graph FoodKG (FoodKG integrates recipes, food, and nutrition data), the team created a benchmark QA data set (not yet published), which contains more than 1 million recipes, 7.7 million nutrition records, and 730 Thousands of foods also contain corresponding ingredients and nutrients, and refer to the lifestyle guidelines recommended by the ADA American Diabetes Association.
(Recommended) Open food facts data set of food nutrition facts
Collect data on ingredients, allergy information, additives, etc. of various foods around the world
Data set download: https://hyper.ai/datasets/5615
According to the author, this is the first personalized food recommendation data set related to ingredients, nutrition and recipes that supports a question-and-answer system.
Each example in the data set contains a user query, dietary preferences, user-related health guidelines and basic answers (ie recipe recommendations).
Step 2: Generate benchmark questions
In order to get questions that reflect people’s true diet, the author collected more than 200 recipes and diabetes-related issues on the social media Reddit, and found a total of 156 posts from netizens asking for recipes, and everyone focused on four types of issues:
Which ingredients are edible;
Which ingredients are not edible;
What are the recommendations for “low carbohydrate” or “high protein”;
What are the recommendations for “Italian flavor” or “Mediterranean flavor”.
The team analyzes the issues on Reddit
Define the recipes, foods, ingredients, etc. mentioned in it
Based on the questioning methods of these posts, the team summarized 56 different templates and generated some benchmark questions based on them.
Step 3: Compile the health guide
After the preliminary preparations are completed, healthy diet recommendations can be made.
The team selected some food-related guidelines from the ADA lifestyle guidelines, which involve nutrition and trace elements, and used them as additional food recommendations. Therefore, its systematic recommendations are healthy recipes that conform to health guidelines.
Since these criteria are in natural language, they are transformed into a structured representation (for example, a hash table that stores key-value pairs).
Convert natural language dietary recommendations into structured data for representation
Step 4: Training the personalization system
In order to achieve personalization, the team also solved the problems of query expansion and allergy query. The author believes that an effective food recommendation system should respect the individual needs of dietary preferences and health guidelines. Therefore, user queries will be further expanded.
For example, if the user asks the system “please recommend a breakfast with bread”, the system will understand the user’s dietary preferences based on the user’s previous queries, dietary preferences, and diet history logs, and automatically expand the single query into additional personalization Demand query.
The expanded query becomes: “Recommend a good breakfast with bread, no peanuts, and 5g to 30g carbohydrates.”
Because of this, the system can give different recipe suggestions in the face of the same problem raised by different users.
Experimental results: better than other models
The team conducted a human evaluation of the food recommendation by providing 8 evaluators with 50 questions randomly tested and user roles, including ingredient preferences (likes and dislikes) and applicable nutrition guidelines.
For each question, input the four models of BAMnet, P-BOW, P-MatchNN and pFoodReQ in random order, and get the answer. Each answer contains the first three recipes (if there are more than three recipes retrieved), a list of ingredients, and nutritional value.
The scoring range is 1-10 points (the higher the score, the better the result)
pFoodReQ gets the highest score
However, the current personalized diet recommendation system is only the first step of the team’s research. The author said, “There are still many challenges in the future. We need more complex answer benchmarks to deal with implicit user intentions and various special situations.”
AI nutritionist, Chinese people need more
Needless to say, the importance of a reasonable diet, but our country’s eating habits are still very unhealthy. According to a research report released by The Lancet, China ranks 140th among 195 countries in the world.
To fundamentally change unhealthy eating habits is much more than just recommending recipes. And can the intervention of AI bring us more possibilities? The following are some of the advantages of AI nutritionists we have summarized:
Faced with machine recommendations, remove psychological precautions
In the face of human nutritionists and fitness coaches, many people may not disclose their true eating habits and consumption levels due to face or privacy issues.
But changing to an AI dietitian is different. You can put aside your psychological burden, tell the AI of your more real needs, and let it search out the one you are interested in and that is nutritious from thousands of recipes.
Tolerate cheat day, dynamically adjust recipes
Nowadays, popular science articles on various nutritional combinations and balanced diets are flooding the Internet, but for readers, implementation is too difficult and compliance is low.
We always encounter situations where we cannot fully execute the recipes or go out to socialize. AI can make timely adjustments to subsequent recipes in response to these changes and adapt to various changes.
Replenish the gap of nutritionist and improve health awareness
Although people’s health awareness is rising, the profession of nutritionist in our country is facing a huge gap.
According to the survey, in Japan, every 300 people are assigned a nutritionist; in the United States, every 4,000 people are assigned a nutritionist; and in China, there is one nutritionist for every 400,000 people.
If AI has perfected the knowledge of nutrition and health, then everyone can have a private nutritionist who is accompanied 24 hours a day, and can provide dietary guidance anytime and anywhere.
By then, the problem of “how to eat today” will be handed over to the AI!