🥗 Cold-Start Recipe Recommendation via Template-Based Language Inference with LLM-Generated Summaries

Official Implementation of the paper submitted to Expert Systems with Applications (ESWA).

📝 Introduction

Recipe-sharing platforms face the Strict Cold-Start (SCS) problem: newly uploaded recipes lack user interactions (ratings, clicks), making them invisible to traditional algorithms (CF, GNNs) that rely on interaction history.

In this work, we propose a novel framework that bridges Large Language Models (LLMs) and Pre-trained Language Models (PLMs) to solve this. Instead of relying on collaborative filtering, we reformulate recommendation as a Sentence Pair Classification (SPC) task.

By utilizing GPT-4o-mini to distill complex user and recipe attributes into coherent natural language summaries, we enable a BERT-based model to predict compatibility via Natural Language Inference (NLI). Our approach achieves state-of-the-art performance on SCS-Food.com and SCS-AllRecipes.com benchmarks.

---

📂 Project Structure

Ensure your directory is organized as follows:

coldreciperec/
├── data/                       # Main data directory
│   ├── foodcom/                # Food.com dataset files
│   │   ├── metadata.pkl
│   │   ├── user_descriptions_RAW_ID.pkl
│   │   ├── RAW_recipes.csv     # From Kaggle
│   │   └── ...
│   └── allrecipe/              # AllRecipes dataset files
│       ├── metadata.pkl
│       ├── raw-data_recipe.csv # From Kaggle
│       └── ...
├── make_train_data.py          # Generates training/validation pairs
├── make_test_data.py           # Generates ranking test data
├── finetune.py                 # Fine-tunes the BERT model
├── evaluate.py                 # Calculates NDCG & Recall metrics
└── README.md

📥 Data Setup

Before running the scripts, please download the required datasets from the sources below.

1. Download Datasets

Dataset	Source	Description
Summary Data (Required)	Download via Figshare	Contains the foodcom and allrecipe summary folders.
Food.com Raw Data	Download via Kaggle	Original recipe attributes and interaction data.
AllRecipes Raw Data	Download via Kaggle	Original recipe attributes and interaction data.

2. File Placement

After downloading the Summary Data from Figshare, extract the contents and move the foodcom and allrecipe folders into your project directory here:

coldreciperec/data/

💻 Usage Guide

Step 1: Generate Training Data

Create the training and validation datasets. This script handles negative sampling and template injection:

For Food.com

python make_train_data.py --dataset foodcom

For AllRecipes

python make_train_data.py --dataset allrecipe

Step 2: Generate Test Data (Ranking)

Create the inference dataset. This generates a ranking list (User x All Items) formatted into natural language prompts.

python make_test_data.py --dataset foodcom 

Output: Saves scp_test_data_{dataset}.pkl.

Step 3: Fine-tune the Model

Fine-tune the PLM (default: bert-base-uncased) on the sentence pairs.

python finetune.py \
    --dataset foodcom 

Outputs: Models are saved to ./data/{dataset}/models/.

Step 4: Evaluation

Evaluate the model using NDCG@K and Recall@K .

python evaluate.py \
    --dataset foodcom