Due to fast-paced and long work environments, people are struggling to pay due to attention to their meals.
Most of these meals are ready-to-eat either delivered by restaurants or packaged meals from convinience stores.
Due to the purpose of these meals they cannot be a substitute for a balance diet; Especially if they are consumed without dietary planning.
This can be contributed to high amounts of calories, fat, sugar, and sodium and lack of low in nutrients, such as fiber, vitamins, and minerals.
Developing a service that integrates with E-Carts of food order, convenience store platforms and provide nutritional insights, statistics and recommendations, to help consumers make healthier choices, and enhance the user experience in this trillion dollar market.
Datasets: Limited item options from particular convenience stores or restaurants limit the datasets.
Bypassing Variance: Reliance on user participation can be bypassed using content based filtering.
Parameter Tuning: Weighing the priority of a general healthy diet with user specialized data in forming recommendations.
Considerations for palette and budget
Time Series Component: Consideration of all meals consumed by the user on a daily, weekly, or monthly scale to form recommendations.
- The USP of the product is it aims for a more structured and healthier consumption of ready-to-eat, order-to-eat meals for regular consumers to promote a healthier lifestyle.
- Therefore the outcome of the project is hopeful towards developing a model that can recommend items based on user utility in this domain.
- Dataset generation: An intermediate method of using language models to judge the nutrition content and calorific count for each dish based on its description. Data is created for 1 zipcode in Dallas and 1 in Austin, texas. attributes included are:
- nutrient contents
- estimated calories
- possible allergens
- User calories: Based on the user’s height,weight,age,sex and activeness an in-house developed service (YouCal)[https://github.com/FoodStats/YouCal] is being utillised; which is trained in reference to the Mayo clinic researched parameters!
- User utility generation: formulated as follows meals
- Explict utility
- Nutritional content (Calories + Nutrients)
- Price
- User ratings
- Implicit utility
- user flavour profile
- user vested interest (time spent on an item)
- Utility function:
- An IU-SLSQP (Implicit utility with Sequential Least square optimizer for non linear programming) for user preferences
- Explicit utility
- calculated for Nutrition based on user goals
- price
- Customer ratings
- combined over a Multi Atribute Utility Theory function (MAUTF).
- IU-GA (Genetic algorithm also explored)
- Evaluation: base results run, with acceptible outcomes.
- GUI Achieved.
- Backend API developed.
- Refining app/web interface.
- API destribution.
- using the said interface for crawlers to collect implicit utility data, ie. frequency of attribute,browse time.
- domain expertise oriented tokenisation of estimation of nutrition content for language model being used. Lot of scope in the fine tuning of LLM based data generation.
[!Tip]
An approach would be to leverage a USDA datasets for individual ingredients to leverage an NLP encoder such as BERT to tokenise nutritional content based of dish discriptions/ recipes.
- Suggestion on the model itself. Domain experts based utility functions finetuning.
- Introducing a long term time series component
- Further testing and implementation of results with real users.
https://github.com/FoodStats/EasyEats
https://www.kaggle.com/darshagarwal41/datasets
{{ gettext('Login to leave a comment') }}
{{ gettext('Post a comment…') }}{{ errorMsg }}
{{ gettext('No comments posted yet') }}