Recommendation System

Netflix এর next show, YouTube এর next video, Amazon এর next product — সব Recommendation System এর কাজ। মূল প্রশ্ন: এই user এর কোনটা পছন্দ হবে?

• Content-based — item এর feature ও user এর past pref এর মিল।
• Collaborative Filtering — অন্য similar user যা পছন্দ করেছে।
• Hybrid — দুটোই (Netflix, Spotify)।

• User-based — তোমার মতো user যা পছন্দ করেছে।
• Item-based — যে item তুমি পছন্দ করেছ, তার similar item।
• Matrix Factorization (SVD, ALS) — latent factor।
• Neural CF — deep learning দিয়ে interaction।

from surprise import SVD, Dataset, Reader
from surprise.model_selection import cross_validate
import pandas as pd

df = pd.read_csv("ratings.csv")  # userId, movieId, rating
reader = Reader(rating_scale=(1, 5))
data = Dataset.load_from_df(df[["userId","movieId","rating"]], reader)

algo = SVD(n_factors=64, n_epochs=20, lr_all=0.005, reg_all=0.02)
print(cross_validate(algo, data, measures=["RMSE","MAE"], cv=5))

Content-based (Cosine Similarity)

from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.metrics.pairwise import cosine_similarity

tfidf = TfidfVectorizer(stop_words="english")
mat = tfidf.fit_transform(movies["description"])
sim = cosine_similarity(mat)

def recommend(title, top=5):
    i = movies.index[movies["title"] == title][0]
    scores = sorted(enumerate(sim[i]), key=lambda x: -x[1])[1:top+1]
    return movies.iloc[[s[0] for s in scores]]["title"]

• RMSE / MAE — rating prediction।
• Precision@K, Recall@K — top-K list।
• NDCG — ranking quality।
• Coverage, Diversity, Novelty — beyond accuracy।

• Sparsity — অধিকাংশ user-item interaction নেই।
• Popularity bias — popular item বেশি recommend হয়।
• Feedback loop — algorithm নিজেই future data shape করে।

• Two-Tower Neural Network (YouTube)।
• Sequential — SASRec, BERT4Rec।
• Graph Neural Network (PinSage)।
• LLM-based recommendation।