main.py

import torch 
import nltk
import benepar
import pandas as pd 
import numpy as np
from sklearn.metrics.pairwise import cosine_similarity 

from utils.clean_text import clean_text 
from utils.semantic_similarity import Encoder 
from utils.syntactic_similarity import Parser 
from utils.tfidf_similarity import TFIDF_Vectorizer 

# Set default device to CUDA if available, otherwise CPU 
if torch.cuda.is_available(): 
    torch.set_default_device("cuda") 
else: 
    torch.set_default_device("cpu")

# Download models/data
nltk.download('punkt')
nltk.download('punkt_tab')
benepar.download('benepar_en3_large')

# Load dataset 
data = pd.read_csv("data/toy_data_aggregated_embeddings.csv") 

# Load precomputed TF-IDF features 
restaurant_tfidf_features = np.load("data/toy_data_tfidf_features.npz") 

# Extract embeddings 
all_desc_embeddings = np.vstack(data["embedding"].values) 

# Initialize encoder 
encoder = Encoder() 

# Initialize syntactic parser 
parser = Parser() 

# Initialize TF-IDF vectorizer 
tfidf_vectorizer = TFIDF_Vectorizer(load_vectorizer=True) 

def retrieve_candidates(query: str, n_candidates: int): 
    # Encode query 
    query_emb = encoder.encode([query]).cpu().numpy() 
    
    # Semantic similarities 
    desc_sem_sim = cosine_similarity(query_emb, all_desc_embeddings)[0] 
    
    # TF-IDF similarities 
    tfidf_sim = tfidf_vectorizer.compute_tfidf_scores(query, restaurant_tfidf_features) 
    
    # Syntactic similarities 
    parsed_query = parser.parse_text(query)
    parsed_query = parser.subtree_set(parsed_query)

    syn_sims = []
    for trees_list in data["syntactic_tree"]:
        review_sims = []
        for review_tree_subs in trees_list:
            if review_tree_subs is None:
                review_tree_subs = set()
            sim = parser.compute_syntactic_similarity(parsed_query, review_tree_subs)
            review_sims.append(sim)
        syn_sims.append(np.mean(review_sims))

    # Combined Stage 1 score
    syn_sims = np.array(syn_sims)
    combined_stage1_scores = 0.8*desc_sem_sim + 0.1*syn_sims + 0.1*tfidf_sim

    # Get top N candidates for Stage 2 reranking
    candidates_idx = np.argsort(combined_stage1_scores)[-n_candidates:][::-1]

    return candidates_idx


def rerank(candidates_idx: np.ndarray, n_rec: int = 10, ) -> list:
    
    # Get popularity scores for stage 1 candidates 
    rerank_scores = data.loc[candidates_idx, "pop_score"].values

    # Retrieve n_rec restaurant based on pop_score 
    topN_reranked_local_idx = np.argsort(rerank_scores)[-n_rec:][::-1] 
    topN_reranked_global_idx = candidates_idx[topN_reranked_local_idx]
    
    # Get restaurant_id for final recommendations 
    restaurant_ids = data.loc[topN_reranked_global_idx, "id"].tolist() 
    
    return restaurant_ids

def get_recommendations(query: str, n_candidates: int = 100, n_rec: int = 30):
    query_clean = clean_text(query) 
    candidates_idx = retrieve_candidates(query_clean, n_candidates) 
    restaurant_ids = rerank(candidates_idx, n_rec) 
    return restaurant_ids