Enabling Next-Generation Smart Homes Through Bert Personalized Food Recommendations - RecipeBERT

This paper introduces RecipeBERT, a BERT-based NLP model for personalized meal recommendations in the context of smart homes research. As a content-based recommendation system, it ranks recipes based on semantic similarity of ingredients and procedures. Recipes have similar ingredients and procedures are considered similar recipes. The proposed model effectively curates a personalized list of top recipes, taking into account the household's eating patterns over an extended period. Unlike previous BERT-based models, which lacked comprehensive fine-tuning to adopt to food recommendation domain, our approach enhances efficiency through fine-tuning BERT to better deliver tailored recommendations. Our evaluation included testing different transformer models: Sentence BERT and it's distilled version along with Sentence-distilled RoBERTa, which we optimized through both full and various partial fine-tuning methods. This comparative analysis helped us determine the most suitable model and fine-tuning process for accurately adapting BERT model to this task. Here we hypothesize that carefully fine-tuned LLM's can demonstrate improved performance in the recipe recommendation domain compared to pretrained models. In this work, we utilized a semi-automated labeling method combining distillation and manual data review process for fine-tuning data needs. The complete code and dataset for our project are available at https://github.com/DivyaMereddy007/RecipeBert/tree/mainlRecipeBert.