Recommendation Systems (Online)

Summary

This is an internal-only course for our VIP lab, focusing on the fundamentals of recommendation systems. This course covers basic ideas of traditional collaborative filtering and content-based recommendations, with videos as the main domain. It is assumed that the students already have taken basic machine learning courses, but some fundamentals are reviewed in the context of recommendation systems. Topics in collaborative filtering (CF) include * Memory-based CF, * Matrix factorization, * Ranking, * Deep-learning-based CF, and * Local models. In content-based recommendations, we cover * Video content analysis, * Metric learning, and * Multimodal video retrieval. Lastly, we cover recent trends in * Graph-based, * Sequential, and * LLM-based recommendation models.

Logistics

Textbook:
- Charu C. Aggarwal, Recommender Systems: The Textbook, 2016, Springer. [Link]
- Additional reading materials and papers will be provided.
Prerequisite:
- Intermediate+ Python programming: you should be able to code what you think in Python.
- Machine learning basics: took this course or equivalent
- Basic calculus, linear algebra, datastructures and algorithms

Syllabus

Topic	Recording	Reading
1. Introduction to Recommendation Systems Definition of recommender systems Examples of recommendations Difference between recommendation and advertisement		Survey 1, Survey 2
2. Understanding Recommendation Systems History of recommendation systems Goals of recommendation systems Types of recommendation systems		Tapestry, GroupLens
Collaborative Filtering
3. Memory-based (Neighborhood) Collaborative Filtering Problem formulation in matrix view User-based collaborative filtering Item-based collaborative filtering		User CF, Item CF
4. Matrix Factorization I Structure of the rating matrix Rank-k Matrix factorization Incomplete singular value decomposition (SVD)		SVD
5. Matrix Factorization II Overfitting, Regularized SVD Optimization: Stochastic gradient descent, Alternating least squares SVD++		SVD++
6. Advanced Matrix Factorization Non-negative Matrix Factorization (NMF) Probabilistic Matrix Factorization (PMF) Bayesian PMF		NMF, PMF, BPMF
7. Collaborative Ranking Learning to rank Collaborative ranking, ordering function		CofiRank
8. Deep Collaborative Filtering I Deep learning revolution Neural networks review Neural collaborative filtering		Lecture note Neural CF
9. Deep Collaborative Filtering II Autoencoders, Denoising autoencoders AutoRec, Collaborative Denoising Autoencoders (CDAE) Recurrent Neural Networks (RNN) GRU4Rec		Lecture note DAE, AutoRec, CDAE, GRU4Rec
10. Deep Collaborative Filtering III Transformers (Encoder, Decoder) BERT BERT4Rec		Transformer, BERT, BERT4Rec
11. Local Approaches for Collaborative Filtering I CF ensembles Local Low-Rank Matrix Approximation (LLORMA) Local Collaborative Ranking (LCR)		CF Ensemble, LLORMA, LCR
12. Local Approaches for Collaborative Filtering II Limitations of previous local models Local Collaborative Autoencoders (LOCA)		LOCA
Content-based Recommendations
13. Introduction to Content-based Video Recommendations Why content-based recommendations? Basics on video content understanding Challenges in video understanding
14. Video Content Analysis I AlexNet, VGGNet, GoogLeNet, ResNet Initial ideas in video understanding Temporal pooling		AlexNet, VGG, GoogleNet, ResNet, Temporal pooling
15. Video Content Analysis II Two-stream models C3D, R3D, R2+1D, I3D SlowFast Networks		Two-stream model, Two-stream fusion, C3D, R3D (ArXiv, CVPR), I3D, SlowFast
16. Content-based Metric Learning I Triplet loss Semi-hard negative mining		P 16.2 Triplet, Contrastive learning
17. Content-based Metric Learning II Collaborative Deep Metric Learning (CDML) Graph Clustering Metric Learning (GCML)		CDML, GCML
18. Multimodal Video Retrieval Multimodality Video-text models: VideoBERT, Hammer		VideoBERT, Hammer
19. Evaluation of Recommendation Systems Evaluation metrics Online and offline evaluations		Survey
20. Cold-start Problems Cold-start users and items Tree-based initial profiling		FMF
21. Case Study: YouTube Video Recommendations Challenges in YouTube video recommendation Multi-stage Approach: Candidate generation, Ranking		YouTube
Recent Trends
1. Graph-based Recommendation Top-N recommendation Graph Convolution Networks (GCN) Neural Graph Collaborative Filtering, LightGCN DGCF, SGL, HCCF		GCN, NGCF, LightGCN, DGCF, SGL, HCCF
2. Sequential Recommendation Transformer-based approaches: SASRec, BERT4Rec Content-based approaches: FDSA, SASRec++, NOVA, DIF-SR		SASRec, BERT4Rec FDSA, NOVA, DIF-SR
3. Content-based Multimedia Recommendation Warm-start methods: MMGCN, SLMRec Cold-start methods: CDML, CLCRec, CCFCRec		MMGCN, SLMRec CDML, CLCRec, CCFCRec
4. Recommendation for Local Communities Local approaches review LLORMA, LCR, GLSLIM, LOCA		LLORMA, LCR, GLSLIM, LOCA
5. Recommendation with Large Language Models (LLM) Large Language Models (LLMs) UniSRec, Prompt4NR, ChatGPT for recommendation, P5		Survey, UniSRec, Prompt4NR, ChatGPT, P5