report Recent site activity shigeyukioba

シェア "report Recent site activity shigeyukioba"

(1)

Report assignment

on matrix factoriza2on

  Analyze the ra2ng data of 1624 movies by 1000 users

  Make a recommenda2on system to oﬀer best movies for

each user

  Evaluate the performance and compare to the random

recommenda2on

Dislike Like

No data

The subject is a part of “EachMovie” dataset

(2)

A simple approach

  Preprocessing

 Transform ordinal ra2ng data to binary [6,5,4]1, [3,2,1]‐1

 Mask the top‐leO 100 x 100 region in the matrix

 Omit movies and users that have few informa2on

  Factoriza2on

 Apply low‐rank matrix factoriza2on with SVD of K=1,2,3,4,5

  ^Predic2on

 Determine +1/‐1 with sign of

  ^Evalua2on

 Calculate predic2on error in the masked region of 100x100

USV

_{= Y}

X

(K )

^{= U}

(K )

^S

(K )

^V

(K ) T

X

_ij

(3)

Advanced approaches

  Preprocessing

 Try alterna2ve transforma2on [6,5,4]3,2,1, [3,2,1]‐1,‐2,‐3

  Factoriza2on

 Apply weighted low‐rank matrix factoriza2on with W_ij=0 if Y_ij=0.

 Apply alterna2ve loss‐func2on such as logis2c model and hinge loss.

report Recent site activity shigeyukioba

Report assignment

on matrix factoriza2on

 Analyze the ra2ng data of 1624 movies by 1000 users

 Make a recommenda2on system to oﬀer best movies for

each user

 Evaluate the performance and compare to the random

recommenda2on

A simple approach

 Preprocessing

 Factoriza2on

 Predic2on

 Evalua2on

USV

= Y

X

= U

S

V

X

Advanced approaches

 Preprocessing

 Factoriza2on

  Analyze the ra2ng data of 1624 movies by 1000 users

  Make a recommenda2on system to oﬀer best movies for

  Evaluate the performance and compare to the random

  Preprocessing

  Factoriza2on

  ^Predic2on

  ^Evalua2on

_{= Y}

^{= U}

^S

^V

  Preprocessing

  Factoriza2on