Review-Based Cross-Domain Collaborative Filtering: a Neural - - PowerPoint PPT Presentation

β–Ά
review based cross domain collaborative filtering a
SMART_READER_LITE
LIVE PREVIEW

Review-Based Cross-Domain Collaborative Filtering: a Neural - - PowerPoint PPT Presentation

Dec 10, 2022 293 likes β€’525 views

COMPLEXREC 2019 Co-located with RecSys19, Copenhagen Review-Based Cross-Domain Collaborative Filtering: a Neural Framework Thanh-Nam Doan, Sherry Sahebi University at Albany, SUNY Albany, NY 1 Cold-start scenario Music User ? 2

slide-1
SLIDE 1

Review-Based Cross-Domain Collaborative Filtering: a Neural Framework

Thanh-Nam Doan, Sherry Sahebi

University at Albany, SUNY Albany, NY

1

COMPLEXREC 2019 Co-located with RecSys’19, Copenhagen

slide-2
SLIDE 2

Cold-start scenario

2

Music User

?

slide-3
SLIDE 3

Music Book User

?

Cross-domain recommenders

3

  • To address problems

such as cold-start and sparsity

  • Information transfer
  • Mostly collaborative

filtering

slide-4
SLIDE 4

Music Book User

?

Problem: hard to justify!

4

  • We propose:
  • using both ratings and reviews

(hybrid and cross-domain)

  • to generate reviews across

domains

Why?

slide-5
SLIDE 5

Music Book User

?

Problem: hard to justify!

5

  • We propose:
  • using both ratings and reviews

(hybrid and cross-domain)

  • to generate reviews across

domains

  • First step towards cross-

domain hybrid recommendation and review generation

slide-6
SLIDE 6

Deep Hybrid Cross Domain (DHCD)

6

slide-7
SLIDE 7

Rating Regression Component

  • We concatenate latent representations of

user and item

𝑦"#

$ = [𝑀"; 𝑀# $]

  • We put it through Q layers

* 𝑧,

$ = β„Ž, $

β„Ž,./

$

… β„Ž/

$ 𝑦"# $

  • The prediction is

Μ‚ 𝑠"#

$ = π‘₯4 $ *

𝑧,

$ + 𝑐4 $

  • The regression loss is

𝑀8 = 9

$∈;

9

"∈<,#∈>?

𝑠"#

$ βˆ’ Μ‚

𝑠"#

$ A

7

slide-8
SLIDE 8

Review Generation Component

  • We concatenate latent

representations of user and item

π‘ž 𝑒D 𝑒ED, Ξ¦$) = πœ€(J β„ŽD

$)

  • The review generation loss is

𝑀K = βˆ’ 9

$∈;

9

"∈<,#∈>?

9

DL/ MNO

log π‘ž(𝑒D|𝑒ED, Ξ¦T)

8

slide-9
SLIDE 9

Joint Model Learning

𝑀 = πœ‡8𝑀8 + πœ‡K𝑀K + πœ‡ π‘Š

" A A +

π‘Š

# A A +

Ξ¦

A A

Where

  • WX

WY controls the trade off

between RRC and RGC

  • πœ‡ is to avoid overfitting

9

slide-10
SLIDE 10

The evaluation of DHCD Model

  • Performance in rating prediction
  • Cold and hot-start
  • Performance in review generation
  • Training convergence performance
  • Trade-off between review generation and rating prediction

10

slide-11
SLIDE 11

Dataset

  • Amazon dataset from 1996 to 2004
  • Three categories: Book, Digital Music and Office Products
  • First 80% of user ratings for training and last 20% for testing

11

slide-12
SLIDE 12

Experiment Setup - Baselines

Model Input Design Generate Reviews MF-based NN-based Ratings Reviews Single- domain Cross- Domain Matrix Factorization (MF) βœ“ βœ“ βœ“ Neural Collaborative Filtering (NCF) βœ“ βœ“ βœ“ Collaborative Deep Learning (CDL) βœ“ βœ“ βœ“ βœ“ βœ“ Collaborative Filtering with Generative Concatenative Net- works (CF-GCN) βœ“ βœ“ βœ“ βœ“ βœ“ Cross-domain neural network (CDN) βœ“ βœ“ βœ“ Our Model (DHCD) βœ“ βœ“ βœ“ βœ“ βœ“

Two setups for each algorithm: single and cross-domain (e.g., cdCDL)

slide-13
SLIDE 13
  • DHCD outperforms single-domain baselines, in each separate domain

13

Performance in rating prediction

slide-14
SLIDE 14

Performance in rating prediction

  • DHCD outperforms cross-domain baselines in mixed-domains

14

slide-15
SLIDE 15

Cold-start Prediction

  • DHCD outperforms the best baseline in cold-start setting (users with 5
  • r less ratings)

15

slide-16
SLIDE 16

Review Generation Analysis

  • Compared to
  • character LSTM, word LSTM, CF-GCN
  • DHCD has better perplexity in review generation

16

slide-17
SLIDE 17

Examples of Generated Reviews: Digital Music and Book

Negative Review Positive Review Real Review

This album has terrible sound. Its very tinny and distant. Nothing like vol 1. I was very disappointed with it. Lightfoot should re release this after firing his producer as there are several great songs on it. Another superb album by Herb Alpert and the Tijuana

  • Brass. Their music is so happy and full of fun. Love them.

DHCD

It not good, awful. Just poor quality that means it bring after the purchase. Nothing like before. should not good. the simplistic. nice jazz of the band, happy with this purchase and enjoy the story, what a sweet voice for me.

CF-GCN

The song is terrible and need to be better, some every dissatisfied and undevelopment. not like it and enjoy song. tribes followers see the awesome song and the lyrics is as

  • always. great recommendation to buy and enjoy the song.

17

slide-18
SLIDE 18

The Effect of Reviews in Training

  • Compare the rating regression training loss of CDN and DHCD through

epochs

  • DHCD has a faster convergence

18

slide-19
SLIDE 19

Trade-off between Rating Prediction and Review Generation

  • πœ‡8/πœ‡K controls the trade off between rating prediction and review generation tasks
  • πœ‡K = 1 and use various values of πœ‡8 for training
  • Increasing πœ‡8/πœ‡K leads to better RMSE but worse perplexity

19

slide-20
SLIDE 20

Conclusion

  • Deep Hybrid Cross Domain (DHCD)
  • first step towards cross-domain review generation and justification
  • can capture some between-domain relations
  • has better rating prediction than single-domain baselines -> adding cross-

domain information helps

  • has better rating prediction and faster convergence than rating-only baselines
  • > adding review data helps
  • has a good performance in cold-start setting
  • There is a trade-off between review generation and rating prediction

20

slide-21
SLIDE 21

Thank you!

ssahebi@albany.edu code: https://github.com/ssahebi/Neural_Hybrid_Cross_Domain

21