Öneri Sistemleri
Recommender engines are the one engineering discipline shaping the internet: 80% of Netflix watching, 70% of YouTube consumption, 35% of Amazon revenue come from recommenders. We see the birth, billion-dollar impact, and why now is the moment.
Table of Contents
Module 0: Course Framework & Workshop Setup
- 1
Why Do Recommender Systems Matter? Birth, Present, and Future of a Discipline
Recommender engines are the one engineering discipline shaping the internet: 80% of Netflix watching, 70% of YouTube consumption, 35% of Amazon revenue come from recommenders. We see the birth, billion-dollar impact, and why now is the moment.
- 2
Who Is a Recommender Engineer? Skill Atlas and Junior → Staff Career Map
Recommender Engineer = a specific intersection of Data Engineer + ML Engineer + ML Researcher + Backend Engineer. Full atlas across 8 skill categories, junior → senior → staff path, typical interview questions, and T-shaped specialization strategy.
- 3
Course Philosophy: Math → Manual Code → Library → Benchmark → Production
Why is this course different from a typical 'Coursera style'? We approach every topic in 5 stages: (1) Pen-and-paper math, (2) From-scratch NumPy, (3) Production-style library, (4) Benchmark on the same dataset, (5) 'Production gotcha' note. Why does this order give 3x deeper learning?
- 4
Workshop Setup: Python 3.12, uv, PyTorch, FAISS, Polars and Jupyter Lab
We set up a modern Python environment for recommender systems work from scratch: uv (Rust-based, 80x faster than conda), Python 3.12, PyTorch 2.5+, FAISS CPU+GPU, Polars, implicit, lightfm, surprise, Jupyter Lab. Mac, Windows (WSL2), Linux, and Google Colab options.
- 5
Our Datasets and the Ethics Contract: From MovieLens to H&M, Going to the Field
Full profile of 8 datasets we'll use: MovieLens (3 sizes), Amazon Reviews (2023), RetailRocket, H&M Fashion, MIND News, Spotify MPD, Last.fm, Yelp. Each one's license, size, download steps, suitability and ethics contract.
Module 1: Landscape of Recommender Systems & Problem Formulation
- 1
Where Do Recommenders Run? An Architecture Tour: Netflix, YouTube, Spotify, Amazon, TikTok, Trendyol
A concrete architecture tour of 6 major companies based on published engineering blogs: Netflix retrieval-ranking, YouTube two-stage, Spotify BaRT, Amazon item-CF heritage, TikTok Monolith, Trendyol personalization.
- 2
Problem Typology: Rating Prediction vs. Ranking vs. Top-N Retrieval vs. Sequential
A recommender problem can be formulated in 4 different ways — and choosing the right formulation is often more important than choosing the right algorithm. Each one's mathematical definition, when to choose, which metric, and which real scenarios.
- 3
Explicit and Implicit Feedback: A Complete Guide from 1-5 Stars to Click-Skip Behavior
Two fundamental data types in recommender systems: explicit (intentionally given stars/likes) vs implicit (click, dwell time, completion, skip). Differences, loss function impact, bias sources, hybrid usage, real-world labeling strategies.
- 4
Three Faces of the Cold-Start Problem: User, Item, System — and Practical Solutions
The most annoying problem in recommenders: how do you recommend for a user/item you have no data on? Strategy map for user cold-start, item cold-start, and system cold-start — from Netflix's 5-film screen to TikTok's viral loop.
Module 2: Data, Labels and Ethics
- 1
MovieLens from Zero: Schema, EDA, and Efficient Loading with Polars
File structure of MovieLens-100K, 1M, 25M, row-by-row schema, lazy/streaming load with Polars (10-30x faster than Pandas), sparse matrix conversion, first EDA graphics, and data quality checks.
- 2
Turning Implicit Feedback into Labels: Click, Dwell, and Multi-Signal Aggregation
Raw e-commerce site logs → trainable labeled dataset. Math and NumPy implementation of Hu/Koren confidence weighting, multi-signal weighted aggregation, session reconstruction, preventing label leakage.
- 3
Bias Galaxy: Position, Presentation, Popularity, Exposure and IPS Correction
5 important biases in recommender systems (position, presentation, popularity, exposure, selection), each with mathematical definition, ways to observe in log data, and Inverse Propensity Scoring (IPS) correction derivation + NumPy implementation.
- 4
GDPR, KVKK and the Right to Be Forgotten: Legal Compliance in Recommenders
How does a recommender system comply with data subject rights (access, deletion, portability)? EU AI Act 2024-2026 timeline, KVKK's 2025 update, removing user data from ML models (machine unlearning), audit log requirements.
Module 3: Evaluation Protocols
- 1
Accuracy Metrics: RMSE, MAE, Precision@K, Recall@K, MAP, MRR, NDCG, HR@K — Full Math + NumPy
Full mathematical definition of 8 main accuracy metrics, from-scratch NumPy implementation, comparative run on MovieLens, and which metric to choose when — recommender engineer's metric cheat sheet.
- 2
Beyond-Accuracy: Coverage, Diversity (ILS), Novelty, Serendipity, and Popularity Bias Measurement
Why does a recommender with high NDCG but bored users exist? Because only 'accuracy' was measured. Coverage, intra-list similarity (ILS), novelty, serendipity, gini coefficient measure all faces of the system.
- 3
Data Splitting Strategies: Random, Time, User, Leave-One-Out — Practical Trade-Offs
How you split MovieLens changes NDCG — 0.15 or 0.25. This lesson covers 5 main split strategies, when each is correct, when each leaks, and a comparison from a production realism standpoint.
- 4
Online Evaluation: A/B Test, Interleaving, CUPED and Statistical Power
You raised offline NDCG +2% — verify with A/B test before production deploy that user behavior really changes. A/B test sample size math, interleaving (10x more efficient), CUPED variance reduction, and switchback testing.
- 5
The Offline-Online Gap: The Dacrema Crisis and Correct Protocol Selection
In 2019, Dacrema, Cremonesi, Jannach paper shook the recommender literature: 'Are neural recommenders really better? Most are beaten by even classic k-NN.' This lesson covers the reproducibility crisis, the offline-online correlation problem, and how to select the correct protocol.
Module 4: Content-Based Filtering
- 1
Content-Based Filtering Philosophy: 'What They Watched' vs 'What It's Like'
Collaborative filtering searches for 'similar users', content-based searches for 'similar items'. This philosophical difference dictates technical decisions — cold-start advantage, filter bubble disadvantage, hybrid strategies. Concept + math + industry positioning.
- 2
Item Profiling: TF-IDF, BM25, n-grams, and Categorical Feature Encoding — Math + NumPy
Foundation of content-based recommenders: converting item to numerical vector. Full TF-IDF formula derivation + from-scratch NumPy implementation, BM25 vs TF-IDF difference, n-grams on movie titles, categorical encoding (one-hot, target, frequency).
- 3
From-Scratch NumPy Content-Based Recommender: 150 Lines on MovieLens-100K
The backbone lesson of this module: building a real content-based recommender on MovieLens-100K — pure NumPy, 150 lines, end-to-end. Item profiling, user profile vector, cosine scoring, top-N recommendation, evaluation. Then compare with sklearn and the first row in our baseline table.
- 4
Production Notes: Feature Drift, Multi-Modal Content, and Challenges of Turkish NLP
Module 4 closing: real problems you'll face after keeping a content-based recommender in production for 6 months. Feature distribution drift, multi-modal embeddings (image+text+audio) for cold-start power, CLIP/SBERT modern approaches, Turkish NLP specifics with stemming + BERTurk.
Module 5: Memory-Based Collaborative Filtering (k-NN)
- 1
k-NN Collaborative Filtering: User-User vs Item-Item — When to Use Which?
Recommender's birth paper, GroupLens 1994. After 30 years, still the baseline of every recommender system. This lesson covers philosophical differences between user-user and item-item CF, mathematical formulation of each, and when each wins.
- 2
Similarity Metrics: Pearson, Cosine, Adjusted Cosine, Jaccard — Full Math + NumPy
Foundation of all CF algorithms: 4 main similarity metrics. Pearson correlation (rating bias correction), cosine similarity (vector direction), adjusted cosine (user bias correction), Jaccard (binary implicit). Full mathematical derivation + from-scratch NumPy + MovieLens comparison.
- 3
From-Scratch Item-Item k-NN with NumPy: Production-Grade on MovieLens-1M
Module 5's backbone: production-grade item-item k-NN from scratch on MovieLens-1M. Adjusted cosine + shrinkage, sparse matrix optimizations, offline batch precomputation pattern, top-K neighbor caching, second row in our benchmark.
- 4
Scalability Ceilings: Optimization Strategies Above 100M Ratings
MovieLens-1M is too small — in real world you work with 100M+ ratings, 10M+ items. This lesson: offline batch precomputation pattern, LSH (Locality-Sensitive Hashing), MinHash for approximate Jaccard, distributed computation with MapReduce/Spark, Redis-based serving.