collaborative_user.py

import pandas as pd
import numpy as np
import streamlit as st
from pathlib import Path
import sys
import os
import time
from scipy.sparse import coo_matrix, csr_matrix
from sklearn.decomposition import TruncatedSVD, IncrementalPCA
from sklearn.preprocessing import normalize
from sklearn.metrics.pairwise import cosine_similarity
import joblib

# ====================== Streamlit Setup ======================
st.set_page_config(layout="wide")
st.title("🎬 Movie Recommender System User-Based")

# ====================== Paths ======================
sys.path.append(os.path.dirname(__file__))
data_dir = Path(__file__).parent / 'data'
cache_dir = data_dir / 'cache'
cache_dir.mkdir(exist_ok=True)

movies_path = 'movies_final.csv'
ratings_path = data_dir / 'ratings.csv'

# ====================== Load Data ======================

# used to **cache the output of a function that loads or processes data** so that it doesn’t get recomputed every time the app reruns. 
# This is especially useful when loading large datasets or performing expensive computations.Since our data is huge

@st.cache_data
def load_data():
    movies = pd.read_csv(movies_path)
    ratings = st.session_state['ratings_df']
    # ratings = ratings.drop(columns=['timestamp'], errors='ignore')
    ratings = ratings.drop_duplicates(subset=['userId', 'movieId'], keep='first')
    ratings['userId'] = ratings['userId'].astype(int)
    ratings['movieId'] = ratings['movieId'].astype(int)
    return movies, ratings

movies, ratings = load_data()

ratings = ratings[ratings['movieId'].isin(movies['movieId'].unique().tolist())]

st.text(f"Initial ratings shape: {ratings.shape}")

# ====================== Data Filtering ======================
user_counts = ratings['userId'].value_counts()
active_users = user_counts[user_counts >= 450].index
ratings_filtered = ratings[ratings['userId'].isin(active_users)]

movie_counts = ratings_filtered['movieId'].value_counts()
popular_movies = movie_counts[movie_counts >= 450].index
ratings_filtered = ratings_filtered[ratings_filtered['movieId'].isin(popular_movies)]
ratings = ratings_filtered.copy()
del ratings_filtered

st.text(f"Filtered ratings shape: {ratings.shape}")

st.markdown("#### 🎯 Dataset Filtering Summary")
st.markdown("""
We filtered the dataset in **two steps** to focus on active users and popular movies:

1. **Active Users**  
   - Kept users who have rated **at least 450 movies**.  
   - Ensures we focus on users who are actively engaged.

2. **Popular Movies**  
   - Kept movies that have **at least 450 ratings**.  
   - Ensures we focus on movies with enough feedback to be meaningful.

 **Result:**  
The final dataset contains ratings from active users on popular movies, improving the quality of recommendations and decreasing the size of the dataset.
""")

formatted_num = "{:,}".format(ratings.shape[0])
print(formatted_num)

# Optional: Show numbers dynamically
st.markdown(f"**📊 Ratings after filtering:** {formatted_num} ratings from {ratings['userId'].nunique()} users on {ratings['movieId'].nunique()} movies.")


# ====================== Index Mapping ======================
# For each unique userId, assigns a unique integer code starting from 0.
ratings['user_idx'] = ratings['userId'].astype('category').cat.codes
ratings['movie_idx'] = ratings['movieId'].astype('category').cat.codes

num_users = ratings['user_idx'].nunique()
num_movies = ratings['movie_idx'].nunique()

# Map indices back to original IDs
# {0: 100, 1: 101, 2: 105}
user_idx_to_id = dict(enumerate(ratings['userId'].astype('category').cat.categories))
movie_idx_to_id = dict(enumerate(ratings['movieId'].astype('category').cat.categories))

# ====================== Sparse Matrix ======================

# Rows → users
# Columns → movies
# Values → ratings
# Most entries are **0 (or missing)** because not all users rate all movies.

user_item_matrix = coo_matrix(
    (ratings['rating'], (ratings['user_idx'], ratings['movie_idx'])),
    shape=(num_users, num_movies)
).tocsr()

# CSR format allows fast operations for matrix factorization or nearest-neighbor search
# The full matrix would have 50 million cells. But maybe only 1 million ratings exist.
# Dense matrix: stores all 50 million entries → huge memory usage.Sparse matrix: stores only the non-zero ratings → huge memory savings.
# There are different sparse formats, e.g., COO, CSR, CSC.

# ====================== Incremental SVD Caching ======================
SVD_CACHE_PATH = cache_dir / "svd_incremental_cache.joblib"

def compute_incremental_svd(matrix: csr_matrix, n_components=100, batch_size=1000):
    """
    Memory-efficient incremental SVD using IncrementalPCA.
    Works even when matrix cannot fit entirely in memory.
    """
    if SVD_CACHE_PATH.exists():
        # st.text("✅ Loading cached SVD decomposition...")
        U, Sigma, VT = joblib.load(SVD_CACHE_PATH)
        return U, Sigma, VT

    st.text("⚙️ Computing Incremental SVD... (this may take several minutes)")
    ipca = IncrementalPCA(n_components=n_components)
    n_samples = matrix.shape[0]

    # Fit in chunks
    for start_idx in range(0, n_samples, batch_size):
        end_idx = min(start_idx + batch_size, n_samples)
        batch = matrix[start_idx:end_idx].toarray()
        ipca.partial_fit(batch)
        st.text(f"Processed rows {start_idx} to {end_idx}")

    # Transform full data
    U = []
    for start_idx in range(0, n_samples, batch_size):
        end_idx = min(start_idx + batch_size, n_samples)
        batch = matrix[start_idx:end_idx].toarray()
        U.append(ipca.transform(batch))
    U = np.vstack(U)

    VT = ipca.components_
    Sigma = np.ones(ipca.n_components_)

    joblib.dump((U, Sigma, VT), SVD_CACHE_PATH)
    st.text("✅ SVD computation cached successfully.")
    return U, Sigma, VT

U, Sigma, VT = compute_incremental_svd(user_item_matrix, n_components=100, batch_size=100000)
U_normalized = normalize(U)

# ====================== Recommendation Function ======================
def recommend_for_user_svd(user_id, top_n=5, n_similar_users=30):
    start = time.time()
    if user_id not in ratings['userId'].values:
        return "User not found.", 0

    user_idx = ratings.loc[ratings['userId'] == user_id, 'user_idx'].iloc[0]

    # Cosine similarity in latent space
    user_vector = U_normalized[user_idx]
    similarities = np.dot(U_normalized, user_vector)

    # Select top similar users
    similar_user_indices = similarities.argsort()[::-1][1:n_similar_users+1]
    similar_scores = similarities[similar_user_indices]

    user_rated_movie_indices = user_item_matrix.getrow(user_idx).nonzero()[1]
    user_rated_movie_set = set(user_rated_movie_indices)

    # Aggregate weighted ratings
    scores = {}
    for sim_user_idx, sim_score in zip(similar_user_indices, similar_scores):
        sim_user_ratings = user_item_matrix.getrow(sim_user_idx)
        movie_indices = sim_user_ratings.nonzero()[1]
        sim_ratings = sim_user_ratings.data
        for m_idx, rating in zip(movie_indices, sim_ratings):
            if m_idx not in user_rated_movie_set:
                scores[m_idx] = scores.get(m_idx, 0) + sim_score * rating

    # Sort and recommend
    recommended_movie_indices = sorted(scores, key=scores.get, reverse=True)[:top_n]
    recommended_movie_ids = [movie_idx_to_id[idx] for idx in recommended_movie_indices]

    end = time.time()
    recommendations = movies[movies['movieId'].isin(recommended_movie_ids)][['title']]
    return recommendations, end - start


# 1. Works in latent space → captures hidden user preferences.
# 2. Uses top similar users → collaborative filtering logic.
# 3. Avoids recommending already rated movies.
# 4. Weighted aggregation ensures more similar users influence recommendations more.
# 5. Returns top N movies along with computation time.

# ====================== Streamlit UI ===========================================

# st.dataframe(ratings.head(1000))
# st.text('unique')
# st.text(ratings['user_idx'].nunique())
# st.text(ratings['userId'].nunique())

user_id_input = st.number_input(
    "Enter User ID",
    min_value=int(ratings['user_idx'].min()),
    max_value=int(ratings['user_idx'].max()),
    step=1
)


user_id_input = ratings[ratings['user_idx'] == user_id_input]['userId'].unique()[0]

# st.dataframe(ratings)

# st.text(user_id_input)


if st.button("🎥 Recommend Movies"):    

    col1, col2 = st.columns(2)
    with col1:
        st.text('User Activity')
        current_user = ratings[ratings['userId']==user_id_input]
        merge_results = pd.merge(current_user,movies, how='left', on = 'movieId')
        merge_results = merge_results[['title','rating']]
        st.dataframe(merge_results)
    with col2:
        st.text('Recommended Movies')
        recs, duration = recommend_for_user_svd(user_id_input, top_n=5)
        st.dataframe(recs)
    st.text(f"Computed in {duration:.2f} seconds.")