Recurrent Recommender NetworksChao-Yuan Wu1, Amr Ahmed2, Alex Beutel2

Alex Smola3, How Jing4

21 3 4

1

Traditional recommender systems

4

3 5 1

3

2

1 ?

2 5

Predict missing values

Observe interactions

2

Traditional recommender systemsassume stationary states

4

3 5 1

3

2

1 ?

2 5

f(vu, vm)u

mUser states

Movie states3

However…, user & movie states should be time-dependent.

4

User preference changes over time

5

?10 years ago

now

Movie reception changes over time

So bad that it’s great to watch

Bad movie6

Exogenous effects“La La Land” won big at Golden Globes

7

Seasonal effects

Only watch during Christmas

8

Traditional MF

Modeling temporal dynamics within each user and movieTraditional MF

10

Also…, we consider real prediction instead of interpolation.

11

Kid started using

Consider a user profile….

12

?

Traditional random split

??

13

Real scenario

?? ??

14

Real scenario

?? ?

In this paper we design and evaluate our model with this real scenario.

?

15

Recurrent Recommender Networks

User RNN

Movie RNN

user

movie16

User Recurrent Neural Network (User RNN)

17

User Recurrent Neural Network (User RNN)

18

Rating

Movie RNN

19

Rating

Recurrent Recommender Networksuser

movie 20

Experiments

21

Rating prediction accuracy

PMF: Salakhutdinov & Mnih NIPS ‘07T-STD: Koren KDD ‘09U-AR & I-AR: Sedhain et al. WWW ‘15

(RMSE)

22

How does the model react to the temporal effects?

23

24

?

Automatically captures exogenous effects

Oscar & Golden globe

25

Automatically captures system-wise effects

Netflix changed the Likert scale

26

Automatically models effects that used to be captured by hand-crafted features

Movie age effects

27

Large improvement when movies have large fluctuations

28

Improvement

Fluctuation

SummaryNovel model

SummaryNovel model Future prediction

? ? ??

SummaryNovel model Future prediction

Accurate prediction

? ? ??

32

SummaryNovel model Future prediction

Accurate prediction Temporal dynamics

? ? ??