20
Ef cient Retrieval of Recommendations in a Matrix Factorization Framework Noam Koenigstein Parikshit Ram Yuval Shavitt School of Electrical Engineering Tel Aviv University Computational Science & Engineering Georgia Institute of Technology School of Electrical Engineering Tel Aviv University

Efficient Retrieval of Recommendations in a Matrix Factorization Framework

  • Upload
    gitel

  • View
    49

  • Download
    0

Embed Size (px)

DESCRIPTION

Efficient Retrieval of Recommendations in a Matrix Factorization Framework. Motivation. In the field of Recommender System , Matrix Factorization (MF) models have shown superior accuracy for recommendation tasks. E.g., The Netflix Prize, KDD-Cup’11, etc. - PowerPoint PPT Presentation

Citation preview

Page 1: Efficient Retrieval of Recommendations in a Matrix Factorization Framework

Efficient Retrieval of Recommendationsin a Matrix Factorization Framework

Noam Koenigstein Parikshit Ram Yuval ShavittSchool of Electrical

Engineering

Tel Aviv University

Computational Science &Engineering

Georgia Institute ofTechnology

School of ElectricalEngineering

Tel Aviv University

Page 2: Efficient Retrieval of Recommendations in a Matrix Factorization Framework

Motivation• In the field of Recommender System, Matrix Factorization

(MF) models have shown superior accuracy for recommendation tasks.E.g., The Netflix Prize, KDD-Cup’11, etc.

• Training is fast. Computing test scores is fast.But… Retrieval of Recommendations (RoR) is s--l--o--w !

• This problem is well known in the industry, yet never been approached before in academia!

Page 3: Efficient Retrieval of Recommendations in a Matrix Factorization Framework

22

45

3 32

. . . .

4 12

25

I T E M S

USERS

Yahoo! Music:

1M Users625K Items

6 Tera elements ~300 multiplications ~5 days CPU

Naïve Multithreading: High latency + wasteful

Yahoo! Music:

1M Users625K Items

6 Tera elements ~300 multiplications ~5 days CPU

Page 4: Efficient Retrieval of Recommendations in a Matrix Factorization Framework

Reduction to Inner Product

.

.

.

1

.

.

.

�̂�𝑢𝑖=𝐩𝑢𝑇 𝐪𝑖=‖𝐩𝑢

❑‖‖𝐪𝑖‖cos𝜃𝑢𝑖

Core problem:

Given a user vector and a set of item, find an item vector that will maximize

Page 5: Efficient Retrieval of Recommendations in a Matrix Factorization Framework
Page 6: Efficient Retrieval of Recommendations in a Matrix Factorization Framework

Best Matches Algorithms

• Metric Space

• Cosine Similarity

• Locality Sensitive Hashing

Page 7: Efficient Retrieval of Recommendations in a Matrix Factorization Framework

Metric TreesR

R

Page 8: Efficient Retrieval of Recommendations in a Matrix Factorization Framework
Page 9: Efficient Retrieval of Recommendations in a Matrix Factorization Framework

Branch-and-bound Algorithm

Page 10: Efficient Retrieval of Recommendations in a Matrix Factorization Framework

Bounding Inner Product Similarity

Page 11: Efficient Retrieval of Recommendations in a Matrix Factorization Framework
Page 12: Efficient Retrieval of Recommendations in a Matrix Factorization Framework

Approximate Solution

Users vectors can be normalized Users can be clustered based on their spherical angle!

Page 13: Efficient Retrieval of Recommendations in a Matrix Factorization Framework

Relative Error Bound

What is the error when recommendations are retrieved based on an approximate user vector?

|𝑒𝑟𝑟𝑜𝑝𝑡|=|𝑝𝑐❑𝑇𝑞𝑖

❑−𝑝𝑢❑𝑇𝑞𝑖

❑||𝑝𝑐❑

𝑇 𝑞𝑖❑|

≤1−cos (𝜃𝑝𝑐𝑞 𝑖

+Δ )cos (𝜃𝑝𝑐𝑞 𝑖 )

Page 14: Efficient Retrieval of Recommendations in a Matrix Factorization Framework

Adaptive Approximate Solution

Page 15: Efficient Retrieval of Recommendations in a Matrix Factorization Framework

Experimentations Set-up MovieLens Netflix Yahoo!

MusicRatings 1,000,206 100,480,507 252,800,275

Users 6,040 480,189 1,000,990

Items 3,952 17,770 624,961

Sparsity 95.81% 98.82% 99.96%

Yahoo! Music Recommendations: Modeling Music Ratings with Temporal Dynamics and Item TaxonomyGideon Dror, Noam Koenigstein, Yehuda Koren(RecSys-11`)

Page 16: Efficient Retrieval of Recommendations in a Matrix Factorization Framework

Exact Alg. Speedup

Page 17: Efficient Retrieval of Recommendations in a Matrix Factorization Framework

Approximate Alg. Speedup

Page 18: Efficient Retrieval of Recommendations in a Matrix Factorization Framework

Speedup vs. Precision

Page 19: Efficient Retrieval of Recommendations in a Matrix Factorization Framework

Speedup vs. MedianRank

Page 20: Efficient Retrieval of Recommendations in a Matrix Factorization Framework

Conclusions

• We introduce a new and relevant research problem

• An exact solution with limited speedup

• An approximate solution with a tradeoff between accuracy and speedup

• Much room for further research…