discrete collabora ve filtering
play

Discrete Collabora.ve Filtering Hanwang Zhang 1 , Fumin Shen 2 , Wei - PowerPoint PPT Presentation

Discrete Collabora.ve Filtering Hanwang Zhang 1 , Fumin Shen 2 , Wei Liu 3 , Xiangnan He 1 , Huanbo Luan 4 , Tat-Seng Chua 1 Presented by Xiangnan He 1. Na>onal University of Singapore 2. University of Electronic Science and Technology of China


  1. Discrete Collabora.ve Filtering Hanwang Zhang 1 , Fumin Shen 2 , Wei Liu 3 , Xiangnan He 1 , Huanbo Luan 4 , Tat-Seng Chua 1 Presented by Xiangnan He 1. Na>onal University of Singapore 2. University of Electronic Science and Technology of China 3. Tencent Research 4. Tsinghua University 19 July 2016

  2. Online Recommenda-on • An Efficient Recommender System • Latent Model: Binary Representa>on for Users and Items • Recommenda>on as Search with Binary Codes Offline Training • End-to-end binary op>miza>on • Balanced and Decorrelated Constraint • Small SVD + Discrete Coordinate Descent 2

  3. Latent Factor Approach [Koren et al. 2009] User-Item Matrix Latent Space 3

  4. Recommenda>on is Search Ranking by <user vector, item vector> Search in Euclidean space is slow Requires float opera>ons & linear scan of the data Search in Hamming Space is fast. Only requires XOR opera>on & constant-.me lookup User-Item Database Hash Table Query Code 4

  5. [Zhang et al, SIGIR’14; Zhou et al, KDD’12] • Stage 1: Relaxed Real-Valued Problem {B, D} ß Con-nuous CF Methods • Stage 2: Binariza>on B ß sgn (B), D ß sgn (D) Code learning and CF are isolated Quan-za-on Loss 5

  6. 1. A,B , a,b are close but they are separated into different quadrants 2. C, d should be far but they are assigned to the same quadrant 6

  7. 7

  8. Observed ra>ng User code Item code Binary Constraint Ra>ng Predic>on However, it may lead to non-informa>ve codes, e.g.: 1. Unbalanced Codes à each bit should have split the dataset evenly 2. Correlated Codes à each bit should be as independent as possible 8

  9. Illustra>on of the effec>veness of the two constraints in DCF Balanced: Decorrelated: Without any constraints: Separated in the Well separated 3 points are (-1, -1) and 1 1 st & 3 rd quadrant point is (+1, -1), which is not discrimina>ve. 9

  10. However, the hard constraints of zero-mean and orthogonality may not be sa>sfied in Hamming space! 10

  11. Objec>ve func>on: Constraint Trade-off Ra>ng Predic>on Binary Constraint Decorrelated Constraint Balanced Constraint Delegate Code Quality Constraint Mixed-Integer Programming NP-Hard [Hastad 2001] 11

  12. Alterna>ve Procedure • B-Subproblem • D-Subproblem • X-Subproblem • Y-Subproblem 12

  13. Objec.ve Func.on For each user code b i , op>mize bit by bit Parallel for loop over m users Usually converges in 5 itera>ons for loop over r bits D-Subproblem can be solved in a similar way 13

  14. #bits #bit-by-bit itera>ons #compu>ng threads #training ra>ngs Linear to data size! 14

  15. Objec.ve Func.on r x m row-centered user code matrix Small SVD r x m Orthogonaliza>on Y-Subproblem can be solved in a similar way 15

  16. #bits #users Linear to data size! 16

  17. • Recommenda>on is search • We can accelerate search by hashing • Unlike previous erroneous two-stage hashing, DCF is an end-to-end hashing method • Fast O(n) discrete op-miza-on for DCF 17

  18. • Dataset (filtering threshold at 10) : • Random split: 50% training and 50% tes>ng. • Metric: NDCG@K • Search Protocol: Hamming ranking or hash table lookup 18

  19. • MF: M atrix F actoriza>on [Koren et al 2009] Classific MF, Euclidean space baseline • BCCF: B inary C ode learning for C ollabora>ve F iltering [Zhou&Zha, KDD 2012] MF+balance+binariza6on • PPH: P reference P reserving H ashing [Zhang et al. SIGIR 2014] Cosine MF + norm&phase binariza6on • CH: C ollabora>ve H ashing [Liu et al. CVPR 2014] Full SVD MF + balance + binariza6on 19

  20. Performance of NDCG@10. 1. DCF learns compact and informa>ve codes. 2. DCF’s performance is most close to the real-valued MF. 3. End-to-end > Two stage 20

  21. Training: full histories of 50% users Tes>ng: the other 50% users that have no histories in training Evalua>on: simulate online learning scenario. 21

  22. MF: original MF MFB: MF+Binariza>on DCFinit: the variant of DCF that discards the two constraints. 22

  23. • D iscrete C ollabora>ve F iltering: an end-to-end hashing method for efficient CF • A fast algorithm for DCF • DCF is a general framework. It can be extended to any popular CF variants, such as SVD++ and factoriza>on machines. 23

  24. Code available: hups://github.com/hanwangzhang/Discrete-Collabora>ve-Filtering 24

Recommend


More recommend