Recommendation system mind map

Get Started. It's Free
or sign up with your email address
Recommendation system mind map by Mind Map: Recommendation system mind map

1. SVD-based algorithm

1.1. PureSVD

1.2. TruncatedSVD

2. T-trial payoff

2.1. Regret minimization

3. Replace missing entries with 0

4. Eigendecompostion algorithms

5. top-N recomendations

6. ranking mechnism

6.1. pointwise

6.1.1. NDCG

6.2. listwise

6.2.1. Directly optimize metrics

6.2.1.1. MAP

6.2.1.2. MRR

6.3. pairwise

6.3.1. Bayessian Personalized Ranking

6.3.2. Weighted approximate-rank pairwise

7. Feature based exploration/explotation problem

7.1. Exploit

7.2. Exploration

7.2.1. Can increaes short-term regret

7.2.2. Can reduce long-term regret

7.3. Set of arms (actions)

8. epsilon-greddy

9. Weighted matrix factorization

10. Switching

11. Biases

11.1. Baseline predictions

12. Content-based systems

13. Knowledge-based approach

14. Hybrid systems

14.1. Algorithms

14.1.1. Regulization

14.1.2. Weighted

14.1.3. Mixed

14.1.4. Meta-learned

14.1.5. Feature augmentation

14.2. Problems

14.2.1. Gray sheep users

14.2.2. Black sheep users

15. User experience

16. Matrix factorization

16.1. No the best choise for implicit case

16.1.1. Deep Learning based

16.1.1.1. Neural Collaborative Filtering

16.1.1.2. DeepMF

16.1.1.3. Variational Autoencoders for Collaborative Filtering (Mult-VAE)

16.2. Embedings

16.3. Loss function

17. Learning-to-rank

18. Bandits

18.1. Exploration strategy

18.1.1. EXP3

18.1.2. UCB1

18.1.3. Thompson sampling

18.2. Contextual bandit

18.2.1. Traits

18.2.2. Context

18.2.2.1. Current user

18.2.2.2. Feature vector

18.2.3. PayOff

19. Sacrifice the buisness metrics

20. Candidate generation strategies

20.1. TrainItems

21. RelPlusN

22. Unexepectedness metrics

22.1. Anonymys

22.1.1. Scalability

22.2. Problems

23. Intra-list diversity

24. Off-policy

25. Require existence of large community of users

26. Non-probabalistic

26.1. Item-based

26.2. User based

27. Pearson correlation

28. Memory-based

28.1. Matrix completion

28.2. Algorithms

28.3. Behaviour types

28.3.1. Items

28.3.2. Actions

28.3.3. Users

28.4. Behavioral similarity function

28.4.1. Spearman correlation

28.4.2. Cosine similarity

28.4.3. Jaccard index

28.4.4. Dice coefficient

28.5. Neighborhood formation

28.5.1. Select fixed number of nearest neighbors

29. Paralle ALS

30. Value-based

31. Problems

31.1. Extremely large action space

32. Percentage of users with at least one correct recommendations

33. Parallel SGD

33.1. Shared-memory approach

34. Lack of personalization

35. Offline Metrics

35.1. Precision-based

35.1.1. Precision@k

35.1.2. Recall@k

35.1.3. RMSE

35.1.4. MAE

35.1.5. Ranking metrics

35.1.5.1. NDCG

35.1.6. MAP

35.1.7. MRR

35.1.8. F-measure

35.1.9. AUC

35.2. Hit rate

35.2.1. Pairwise similarity

35.3. Development-based

35.3.1. CPU per cost

35.3.2. Diversity-based

35.4. Cost of re-training

35.5. Matrix factorization

35.5.1. Storage cost

36. SGD

37. Pre-filtering

37.1. Item splitting

37.2. User splittiing

37.3. Optimal splitting criteria

38. Collaborative filtering

38.1. Features

38.1.1. Cold start problem

38.1.2. Don't rely on metadata

38.1.2.1. Model-based

38.1.2.1.1. Markov models

38.1.2.1.2. Session-based recommendations

38.1.2.1.3. Algorithmns

38.1.2.1.4. User

38.1.3. Don't rely on content features

39. Evaluation

39.1. Online metrics

39.1.1. CR

39.1.2. Business-oriented metrics

39.1.2.1. CTR

39.1.2.2. GMV

39.1.3. Serendipity

39.1.4. Novelty

39.1.5. Coverage

39.2. Train/eval split

39.2.1. Time based

39.3. UserTest

40. Optimization techniques

40.1. ALS

40.2. CD

41. Parallel computation

42. Context-aware recommendations

42.1. Post-filtering

42.1.1. Deviation

42.2. Contextual modeling

42.2.1. Deviation based modeling

42.2.2. Similarity-based modeling

43. Reinforcement learning

43.1. Model-based

43.2. Policy-based

43.3. Model-free

43.3.1. On-policy

43.4. Evaluation