01 · Real data, features, and ItemCF
Trace real temporal histories, ItemCF contributions, and seen-item filtering.
17 LESSONS · ONE CONNECTED PATH
Recommendation
Start from user behavior and study retrieval, sequences, interactions, ranking and evaluation.
Before you begin: Start with 31–34. Real-data experiments require Retailrocket; XMRec is an advanced track.
THE LEARNING PATH
Build one idea on the last.
01
Start · data, baselines and training
Real interaction logs, two-tower retrieval, gradients and evaluation. These are recommendation-specific examples.
02 · Two towers: source code, tensors, gradients, and real training
Inspect two-tower tensors and gradients, then save a new real training run.
03 · Foundations, metrics, and A/B testing
Distinguish metric denominators, candidate pools, and simulated A/B assignment.
02
Mechanisms · one algorithm at a time
Retrieval → multiple objectives → feature interactions → sequences → diversity → cold start → engineering. Constructed tensors and simulated traffic are labeled.
04 · Retrieval, samples, and indexes
Study collaborative retrieval, negative sampling, vector indexes, and filtering.
05 · Multiple objectives, MMoE, and pre-ranking
Inspect conditional objectives, MMoE gates, duration weighting, and cached towers.
06 · FM, DCN, LHUC, and FiBiNET
Verify feature-interaction operators and gradients before scaling model capacity.
07 · LastN, DIN, and SIM
Compare mean pooling, candidate-dependent attention, and retrieval before activation.
08 · Diversity, rules, and DPP
Inspect diversity objectives, infeasible rules, and DPP numerical equivalence.
09 · Cold start and traffic allocation
Separate category-based cold-start mechanisms from simulated traffic guarantees.
10 · System improvement and engineering experiments
Study retrieval quotas, distillation, drift, and the limits of social simulations.
03
Practice · XMRec competition core
Task and data → features → graph models → ranking → evaluation. A teaching implementation of selected methods, not the full winning ensemble.
11 · A real competition: from the task to one prediction
Understand the full official XMRec task, candidate protocol, and user-isolated splits.
12 · A winning core method: collaborative-filtering scores as features
Reproduce ItemCF formulas and inspect real cross-market ranking features.
13 · Negative sampling, BPR, and LightGCN
Sample real interaction triples, check LightGCN operators, and run small training.
14 · Retrain a fusion model yourself
Retrain LightGBM and preserve feature and negative-sampling ablation results.
15 · Is evaluation trustworthy, and what should change next?
Audit nDCG, cross-market weighting, paired uncertainty, and reproduction limits.
04
Synthesis · controlled comparisons
Compare negative sampling, sequence models and indexing under a common budget. Preserve negative results.
16 · Deeper recommendation experiments: negatives, sequences, and indexes
Compare negative proposals, sequence encoders, and approximate retrieval on real data.
05
Reference · original recommendation map
Chapter 00 describes the recommendation course. Use this website for the current cross-domain map.
00 · Recommendation systems · Notebook course map
An entry map to recommendation mechanisms and the XMRec competition track.
No lessons match. Try another concept or clear the search.
TRACE THE IDEAS
Sources & lecture notes
Follow the original material alongside the Lab. Source scope, original language and the boundaries of each teaching unit are explicit.
Shusen Wang · Recommender Systems
45 lecture references mapped to practice. An exercise on a topic is not a complete paper reproduction. Originals retain their language and authors’ rights.
WSDM Cup 2022 · Official XMRec Task
Official cross-market recommendation task. Only local public-split results are reported, not hidden leaderboard scores.
OPDAI · XMRec Winning Solution
Selected sparse ItemCF, graph retrieval and ranking methods. Not the complete winning ensemble.
Chapter-to-source map 8 units
R01
Behavioral data and baselines
Run the linked chapters and explain inputs, outputs, controls and the limits of the evidence.
- Shusen Wang · Recommender Systems Corresponding author-course module; see individual lecture links below
Boundary: Real interaction logs; clicks, impressions and unobserved preferences are not interchangeable.
R02
Retrieval and representation learning
Run the linked chapters and explain inputs, outputs, controls and the limits of the evidence.
02 · Two towers: source code, tensors, gradients, and real training04 · Retrieval, samples, and indexes
- Shusen Wang · Recommender Systems Corresponding author-course module; see individual lecture links below
Boundary: Real two-tower training plus constructed retrieval mechanisms; self-supervision and Deep Retrieval are not fully reproduced.
R03
Ranking, multiple objectives and interactions
Run the linked chapters and explain inputs, outputs, controls and the limits of the evidence.
- Shusen Wang · Recommender Systems Corresponding author-course module; see individual lecture links below
Boundary: Constructed tensors and simulated tasks; no claim of real impression CTR or production MMoE results.
R04
Sequence interests
Run the linked chapters and explain inputs, outputs, controls and the limits of the evidence.
- Shusen Wang · Recommender Systems Corresponding author-course module; see individual lecture links below
Boundary: Candidate-dependent attention and long-sequence mechanisms, not complete DIN/SIM recipes.
R05
Diversity and reranking
Run the linked chapters and explain inputs, outputs, controls and the limits of the evidence.
- Shusen Wang · Recommender Systems Corresponding author-course module; see individual lecture links below
Boundary: Small-set and real-vector mechanisms, with no claim of online diversity gains.
R06
Cold start and engineering
Run the linked chapters and explain inputs, outputs, controls and the limits of the evidence.
- Shusen Wang · Recommender Systems Corresponding author-course module; see individual lecture links below
Boundary: Controlled traffic and diagnostics; real new-item/segment outcomes require additional data.
R07
XMRec practice
Run the linked chapters and explain inputs, outputs, controls and the limits of the evidence.
11 · A real competition: from the task to one prediction12 · A winning core method: collaborative-filtering scores as features13 · Negative sampling, BPR, and LightGCN14 · Retrain a fusion model yourself15 · Is evaluation trustworthy, and what should change next?
- WSDM Cup 2022 · Official XMRec Task Official task, data and evaluation protocol
- OPDAI · XMRec Winning Solution Core methods from the authors’ solution
Boundary: Public local splits and selected winning methods, not the full ensemble or hidden leaderboard scores.
R08
Controlled experiments and synthesis
Run the linked chapters and explain inputs, outputs, controls and the limits of the evidence.
- Shusen Wang · Recommender Systems Corresponding author-course module; see individual lecture links below
Boundary: Controlled comparisons on real samples, retaining budgets and negative results; not industrial scale.
Original lecture library 45 entries
These links open the original author or publisher material, in its original language and under its own terms.
- Recommender-system basics
01_Basics_01.pdf · 10 pages · Chinese
Partial method or mechanism exercise; full pipelines and scale are not covered.
- Recommendation pipeline
01_Basics_02.pdf · 15 pages · Chinese
Partial method or mechanism exercise; full pipelines and scale are not covered.
- A/B testing
01_Basics_03.pdf · 23 pages · Chinese
A related teaching mechanism is readable; the full method in the lecture remains an advanced objective.
- Item-based collaborative filtering
02_Retrieval_01.pdf · 39 pages · Chinese
A corresponding method experiment is available; this does not establish complete paper reproduction or industrial outcomes.
- Swing
02_Retrieval_02.pdf · 12 pages · Chinese
A related teaching mechanism is readable; the full method in the lecture remains an advanced objective.
- User-based collaborative filtering
02_Retrieval_03.pdf · 37 pages · Chinese
A related teaching mechanism is readable; the full method in the lecture remains an advanced objective.
- Categorical features
02_Retrieval_04.pdf · 17 pages · Chinese
A corresponding method experiment is available; this does not establish complete paper reproduction or industrial outcomes.
- Matrix completion and ANN
02_Retrieval_05.pdf · 34 pages · Chinese
A related teaching mechanism is readable; the full method in the lecture remains an advanced objective.
- Two towers: architecture and training
02_Retrieval_06.pdf · 31 pages · Chinese
Partial method or mechanism exercise; full pipelines and scale are not covered.
- Two towers: positive and negative samples
02_Retrieval_07.pdf · 22 pages · Chinese
Partial method or mechanism exercise; full pipelines and scale are not covered.
- Two towers: serving and updates
02_Retrieval_08.pdf · 21 pages · Chinese
Partial method or mechanism exercise; full pipelines and scale are not covered.
- Two towers and self-supervised learning
02_Retrieval_09.pdf · 38 pages · Chinese
A related teaching mechanism is readable; the full method in the lecture remains an advanced objective.
- Deep Retrieval
02_Retrieval_10.pdf · 48 pages · Chinese
A related teaching mechanism is readable; the full method in the lecture remains an advanced objective.
- Other retrieval channels
02_Retrieval_11.pdf · 16 pages · Chinese
A mechanism is readable; the real business task needs additional data and labels.
- Exposure filtering and Bloom filters
02_Retrieval_12.pdf · 25 pages · Chinese
Partial method or mechanism exercise; full pipelines and scale are not covered.
- Multiple objectives and sampling calibration
03_Rank_01.pdf · 18 pages · Chinese
A related teaching mechanism is readable; the full method in the lecture remains an advanced objective.
- Multi-gate Mixture-of-Experts
03_Rank_02.pdf · 18 pages · Chinese
A related teaching mechanism is readable; the full method in the lecture remains an advanced objective.
- Prediction-score fusion
03_Rank_03.pdf · 8 pages · Chinese
A related teaching mechanism is readable; the full method in the lecture remains an advanced objective.
- Video watch-time modeling
03_Rank_04.pdf · 15 pages · Chinese
A mechanism is readable; the real business task needs additional data and labels.
- Ranking features and serving
03_Rank_05.pdf · 14 pages · Chinese
Partial method or mechanism exercise; full pipelines and scale are not covered.
- Three-tower coarse ranking and COLD
03_Rank_06.pdf · 17 pages · Chinese
A related teaching mechanism is readable; the full method in the lecture remains an advanced objective.
- Factorization machines
04_Cross_01.pdf · 11 pages · Chinese
A related teaching mechanism is readable; the full method in the lecture remains an advanced objective.
- DCN V2
04_Cross_02.pdf · 24 pages · Chinese
A related teaching mechanism is readable; the full method in the lecture remains an advanced objective.
- LHUC / PPNet
04_Cross_03.pdf · 13 pages · Chinese
A related teaching mechanism is readable; the full method in the lecture remains an advanced objective.
- SENet and FiBiNET
04_Cross_04.pdf · 22 pages · Chinese
A related teaching mechanism is readable; the full method in the lecture remains an advanced objective.
- Last-N mean features
05_LastN_01.pdf · 7 pages · Chinese
A corresponding method experiment is available; this does not establish complete paper reproduction or industrial outcomes.
- Deep Interest Network
05_LastN_02.pdf · 19 pages · Chinese
A related teaching mechanism is readable; the full method in the lecture remains an advanced objective.
- Search-based Interest Model
05_LastN_03.pdf · 16 pages · Chinese
A related teaching mechanism is readable; the full method in the lecture remains an advanced objective.
- Diversity metrics
06_Rerank_01.pdf · 15 pages · Chinese
Partial method or mechanism exercise; full pipelines and scale are not covered.
- Maximal Marginal Relevance
06_Rerank_02.pdf · 12 pages · Chinese
A corresponding method experiment is available; this does not establish complete paper reproduction or industrial outcomes.
- Reranking rules
06_Rerank_03.pdf · 8 pages · Chinese
A related teaching mechanism is readable; the full method in the lecture remains an advanced objective.
- DPP mathematics
06_Rerank_04.pdf · 17 pages · Chinese
A related teaching mechanism is readable; the full method in the lecture remains an advanced objective.
- DPP diversification
06_Rerank_05.pdf · 18 pages · Chinese
A related teaching mechanism is readable; the full method in the lecture remains an advanced objective.
- Cold-start metrics
07_ColdStart_01.pdf · 18 pages · Chinese
A mechanism is readable; the real business task needs additional data and labels.
- Simple cold-start retrieval
07_ColdStart_02.pdf · 21 pages · Chinese
A related teaching mechanism is readable; the full method in the lecture remains an advanced objective.
- Cluster retrieval
07_ColdStart_03.pdf · 20 pages · Chinese
A mechanism is readable; the real business task needs additional data and labels.
- Look-alike audience expansion
07_ColdStart_04.pdf · 17 pages · Chinese
A related teaching mechanism is readable; the full method in the lecture remains an advanced objective.
- Traffic control and guaranteed delivery
07_ColdStart_05.pdf · 22 pages · Chinese
A related teaching mechanism is readable; the full method in the lecture remains an advanced objective.
- Cold-start A/B testing
07_ColdStart_06.pdf · 20 pages · Chinese
A mechanism is readable; the real business task needs additional data and labels.
- Improving metrics: overview
08_Improvement_01.pdf · 7 pages · Chinese
Partial method or mechanism exercise; full pipelines and scale are not covered.
- Retrieval improvements
08_Improvement_02.pdf · 20 pages · Chinese
Partial method or mechanism exercise; full pipelines and scale are not covered.
- Ranking-model improvements
08_Improvement_03.pdf · 35 pages · Chinese
A related teaching mechanism is readable; the full method in the lecture remains an advanced objective.
- Diversity and exploration
08_Improvement_04.pdf · 15 pages · Chinese
Partial method or mechanism exercise; full pipelines and scale are not covered.
- Special user segments
08_Improvement_05.pdf · 23 pages · Chinese
A related teaching mechanism is readable; the full method in the lecture remains an advanced objective.
- Interactions and social value
08_Improvement_06.pdf · 24 pages · Chinese
A mechanism is readable; the real business task needs additional data and labels.