Skip to content

Latest commit

 

History

History
170 lines (142 loc) · 5.35 KB

PITCHME.md

File metadata and controls

170 lines (142 loc) · 5.35 KB

---?image=https://source.unsplash.com/Oaqk7qqNh_c

@color[black](Latent Factor Analysis)

@color[black](aka Matrix Factorization)

+++ @snap[north]

Overview

@snapend @snap[west] @ul

  • Matrix Factorization
  • SVD
  • SGD
  • surprise
  • Our Implementation @ulend @snapend

---?image=https://source.unsplash.com/jgKgekpnmCI

Matrix Factorization

+++?image=https://source.unsplash.com/jgKgekpnmCI

Matrix Factorization

mat factorization image

+++?image=https://source.unsplash.com/jgKgekpnmCI

Singular Vector Decomposition

+++?image=https://source.unsplash.com/jgKgekpnmCI

Singular Vector Decomposition

---?image=https://source.unsplash.com/tMvuB9se2uQ

Stochastic Gradient Descent

+++?image=https://source.unsplash.com/tMvuB9se2uQ ### Stochastic Gradient Descent @ul * Randomly initialize `P` and `Q` * For a given `epoch`, minimize:

sgd loss function +++?image=https://source.unsplash.com/tMvuB9se2uQ

Stochastic Gradient Descent

  • adjust p(u) and q(i) at each epoch according to:

sgd derivatives @ulend

---?image=https://source.unsplash.com/iVVBVb2RqLc

surprise


![image of surprise library](latent_factor_analysis/pitch_assets/surprise_logo.png) +++?image=https://source.unsplash.com/iVVBVb2RqLc ### SVD vs SVD++ @div[top-50 fragment]

@divend @div[bottom-50 fragment]

@divend +++?image=https://source.unsplash.com/iVVBVb2RqLc ### SVD vs SVD++ Recommendations for user `276847` using...

@div[left-50 fragment] surpise-svd: svdpp results @divend

@div[right-50 fragment] surprise-svd++: fake our model results @divend ---?image=https://source.unsplash.com/JFeOy62yjXk

Our implementation

+++?image=https://source.unsplash.com/JFeOy62yjXk

Our implementation

Biases

our fun +++?image=https://source.unsplash.com/JFeOy62yjXk

Our implementation

Evaluation

our rmse +++?image=https://source.unsplash.com/JFeOy62yjXk

Our implementation

Recommendation

our recommendation ---?image=https://source.unsplash.com/i0K3-IHiXYI

Conclusion

@div[left-50 fragment] Pros @ul

  • Users or Books features not necessary
  • easy to evaluate and understand @ulend @divend

@div[right-50 fragment] Cons @ul

Further Development

@ul

  • dataset with time stamps
  • confidence interval
  • Grid Search
  • KFold Evaluation @ulend