Eat System - Machine Learning Model

File: cogs/eat/train/train.py

The Train class encapsulates the machine learning component of the "Eat What" system. It is responsible for training a personalized recommendation model for each server and using that model to predict food preferences.

Train Class

__init__(self, db: DB, ...)

Initializes the training manager.

• Parameters:
  • db (DB): An instance of the database class, used to fetch training data.
  • Other parameters control the model's architecture and training hyperparameters (embedding dimensions, hidden layers, learning rate, etc.).

genModel(self, discord_id: str)

This is the core training method. It builds and trains a recommendation model based on a server's historical data.

• Parameters:
  • discord_id (str): The ID of the server for which to train the model.
• Process:
  1. Data Loading: It uses a DataLoader to fetch all search records for the given discord_id from the database.
  2. Data Processing: It processes the raw data, extracting features like restaurant titles, tags, and user-provided keywords.
  3. Vocabulary Creation: It builds a vocabulary of all unique words from the processed data.
  4. Tensor Conversion: The data is converted into numerical tensors that the model can understand.
  5. Model Training: It initializes a Net model (a neural network defined in cogs/eat/train/model.py) and trains it on the tensor data for a set number of epochs. The user's ratings (self_rate from the database) implicitly influence the training data distribution.
  6. Saving: After training, the model's state (.model file) and the vocabulary mapping (.pickle file) are saved to the models/ directory, named after the discord_id.

predict(self, discord_id: str)

Uses a pre-trained model to generate a food recommendation.

• Parameters:
  • discord_id (str): The ID of the server to generate a prediction for.
• Process:
  1. It loads the saved model and vocabulary files for the specified discord_id.
  2. It feeds a random seed from the server's vocabulary into the model.
  3. The model outputs a probability distribution over the entire vocabulary.
  4. It uses this distribution to randomly select and return a predicted keyword.
• Returns: A string containing the predicted food keyword (e.g., "ramen"). Returns None if no model exists for the server.