How AI Knows What You're Thinking (And Why It Feels So Accurate)

import numpy as np
from sklearn.metrics.pairwise import cosine_similarity

# Simplified collaborative filtering example
# Real systems use embeddings trained on billions of interactions

# User-item interaction matrix
# Rows = users, Columns = items (movies, songs, products)
# Values = implicit score (watch time, clicks, purchases)
user_item_matrix = np.array([
    [5, 3, 0, 1, 0],   # User A: liked action movies
    [4, 0, 4, 1, 2],   # User B: liked action + comedy
    [1, 1, 0, 5, 4],   # User C: liked romance movies
    [0, 0, 5, 4, 5],   # User D: liked comedy + romance
])

def find_similar_users(user_id: int, matrix: np.ndarray, top_n: int = 2):
    """Find users most similar to target user using cosine similarity."""
    similarities = cosine_similarity(matrix)
    user_similarities = similarities[user_id]
    user_similarities[user_id] = -1  # exclude self
    similar_users = np.argsort(user_similarities)[::-1][:top_n]
    return similar_users, user_similarities[similar_users]

def recommend_items(user_id: int, matrix: np.ndarray, top_n: int = 3):
    """
    Recommend items user hasn't seen based on similar users' behavior.
    This is the core of collaborative filtering.
    """
    similar_users, similarity_scores = find_similar_users(user_id, matrix)
    unseen_items = np.where(matrix[user_id] == 0)[0]

    item_scores = {}
    for item_id in unseen_items:
        score = sum(
            similarity * matrix[sim_user, item_id]
            for sim_user, similarity in zip(similar_users, similarity_scores)
        )
        if score > 0:
            item_scores[item_id] = score

    ranked = sorted(item_scores.items(), key=lambda x: x[1], reverse=True)
    return [item_id for item_id, _ in ranked[:top_n]]

recommendations = recommend_items(user_id=0, matrix=user_item_matrix)
print(f"Recommended items for User A: {recommendations}")
# Output: items that User B (similar taste) liked but User A hasn't seen