Recommender Systems

Edit Me

Recommender systems are one of the most important AI applications in industry. From traditional collaborative filtering to modern large-model recommendations, this section covers the complete learning path for recommender systems.

8.1 Recommender Systems Learning Path (WIP)

8.1.1 Understanding the Business

Search, advertising, and recommendation have different scopes:

Search: Focuses on query-based retrieval, understanding user search intent
Advertising: Balances three-way interests among advertisers, ad platforms, and user experience
Recommendation: Focuses on long-term ecosystem, maximizing long-term user value

Core Differences:

Recommendation only cares about whether the ranking order of item fusion scores is accurate
Advertising must also ensure score distances (related to bidding mechanisms)
Generally, item scores in recommender systems can be summarized as: recall → ctr × cvr
Ad systems additionally multiply by bid and deep_cvr scores (e.g., payment amount)

Business Understanding Advice: Whether it's search, advertising, or recommendation, it's recommended to read "Computational Advertising" to understand the business fundamentals.

8.1.2 Learning Models

Evolution of recommender system models:

Traditional Methods: Collaborative filtering, matrix factorization
Deep Learning: DeepFM, Wide&Deep, DIN, etc.
Pre-trained Models: BERT4Rec, SASRec, etc.
LLM Era: LLM4Rec, ChatRec, etc.

8.2 Recommender Learning Resources

8.2.1 Wang Shusen's Recommender Systems Course

Video: Bilibili Recommender Systems Course
Highlights: Balanced theory and practice, rich industry experience
Content: From basic collaborative filtering to deep learning recommendation models

8.2.2 Datawhale Large Model Recommender Systems Study Group

Tutorial: GitHub Repository
Website: https://datawhalechina.github.io/fun-rec/
Highlights: Collaborative learning format, suitable for beginners

8.2.3 "Internet Giant Recommender Algorithm in Practice"

Resource Link: Zhihu Article Introduction
Highlights: Real-world industry experience sharing
Content: Practical application cases of recommender algorithms at major companies

8.3 Beginner Projects

8.3.1 Alibaba Tianchi News Recommendation System

Competition: Tianchi News Recommendation
Highlights: Real business scenarios, high data quality
Learning Value: Complete recommender system development workflow

Recommender System Project Collection: A curated collection of more hands-on project resources

8.4 Recommender Systems Study Notes

Recommender Systems Study Notes Table: View the complete organized study notes

Core Content Includes:

Classic recommendation algorithm principles
Deep learning recommendation models
Industry deployment experience
Evaluation metrics and optimization strategies

8.5 Recommender Systems Papers (To Be Curated)

GitHub paper collection, continuously updated. Content is closer to practical business, without much focus on generative recommendation.

Recommender Systems Paper Curation Table: View the complete paper curation list

Curation Status:

Not Started: Pending curation
Under Review: Currently being evaluated
Completed: Review finished
Not Recommended: Quality below standard
Not Open-Source: Code not publicly available

LLM4REC

Applications of large models in recommender systems, including:

Pre-trained Recommendation Models: Pre-training on large-scale data
Generative Recommendation: Framing recommendation as a generation task
Multimodal Recommendation: Combining text, images, and other multimodal information
Conversational Recommendation: Natural language interaction-based recommendation

Tencent Advertising Algorithm Competition

Website: https://algo.qq.com
Problem: Participants predict users' next likely ad interactions based on desensitized multi-modal historical behavior data (collaborative, textual, visual)
Technical Requirements: Each interaction includes ID-type ad features and multimodal information (images, text, etc.)
Innovation Direction: Via baseline models and solution review stages, the competition encourages participants to break out of traditional discriminative recommendation frameworks and explore generative recommendation

Multimodal Short Video Click Prediction — Malanshan Cup

Website: https://challenge.ai.mgtv.com/
Problem: Predict videos users watch and completion rate based on Mango TV's multimodal features, combined with user features and behavior data
Technical Highlights: Full baseline sharing for 2025 MGTV Multimodal Video Recommendation (0.256+)

Kaggle Playground — FlightRank2025

Problem: 2025 personalized flight recommendation for passengers
Goal: Build an intelligent flight ranking model to predict which flight option business travelers will choose from search results
Reference Solution: CatBoost Ranker Baseline

Technology Trends

Traditional Recommendation → LLM Recommendation

Representation Learning: From sparse features to dense embeddings
Sequence Modeling: From static features to dynamic sequences
Multimodal Fusion: From single modality to multi-modal information
Generative Recommendation: From discriminative models to generative models

Industry Deployment Considerations

Latency Requirements: Millisecond-level response time
Throughput: High-concurrency request handling
Storage Optimization: Model compression and quantization
A/B Testing: Online performance evaluation

Evaluation Metric Framework

Offline Metrics:

Accuracy-based: Precision, Recall, F1
Ranking-based: NDCG, MAP, MRR
Diversity: Coverage, Diversity

Online Metrics:

Click-Through Rate (CTR)
Conversion Rate (CVR)
Dwell Time
User Retention

Learning Suggestions

Business Understanding First: Deeply understand the nature and goals of recommendation business
Progressive Algorithm Learning: Start from classic algorithms, gradually go deeper into deep learning methods
Practice-Oriented: Accumulate hands-on experience through projects and competitions
Engineering Skills: Emphasize system design and engineering implementation
Continuous Learning: Follow frontier technologies like large models applied to recommendation

Other Summaries

LLM4REC (Large Model Recommendation)
Learning Path (WIP): Business understanding, model learning
Recommender Learning Resources: Wang Shusen, Datawhale, hands-on projects

贡献者

这篇文章有帮助吗？