neural graph collaborative filtering184pc128.csie.ntnu.edu.tw/presentation/19-12-20... ·...
TRANSCRIPT
Neural Graph Collaborative Filtering
Advisor: Jia-Ling Koh
Presenter: You-Xiang Chen
Source: SIGIR ‘19
Data: 2019/12/20
1
Content
Introduction
Method
Experiment
01
Conclusion
02
03
042
Introduction
3
Recommended System
4
Collaborative Filtering
Matrix Factorization
Neural Collaborative Filtering
5
Introduction
1. Embedding , which transforms users and items to
vectorized representations
2. Interaction modeling , which reconstructs historical
interactions based on the embeddings.
Components of learnable CF models
6
Motivation
Inherent drawback of existing methods
• Most existing methods build the embedding
function with the descriptive features only.
• The embedding function lacks an explicit
encoding of the crucial collaborative signal .
7
Goal
• Proposing a new recommendation framework
based on graph neural network .
• The work explicitly encodes the collaborative
signal in the form of high-order connectivities .
8
Method
9
Neural graph collaborative filtering
Embedding
Embedding Propagation Layer
Prediction Layer
10
Embedding Layer
11
Embedding Propagation LayerFirst-order Propagation
• Message Construction
• Message Aggregation
For a connected user-item pair (u,i) , we define
the interaction message as:
We aggregate the messages propagated
from 𝒖′𝒔 neighborhood as 𝑢′𝑠representation :
𝒆𝒖(𝒍)
𝒆𝒖(𝟏)
𝒆𝒖(𝟐)
12
Embedding Propagation LayerFirst-order Propagation
• Message Construction𝒇 ∙ : 𝒆𝒏𝒄𝒐𝒅𝒊𝒏𝒈 𝒇𝒖𝒏𝒄𝒕𝒊𝒐𝒏
𝒑𝒖𝒊: 𝒄𝒐𝒆𝒇𝒇𝒊𝒄𝒊𝒆𝒏𝒕 𝒕𝒐 𝒄𝒐𝒏𝒕𝒓𝒐𝒍 𝒅𝒆𝒄𝒂𝒚 𝒇𝒂𝒄𝒕𝒐𝒓
𝐖𝟏,𝐖𝟐 ∈ ℝ𝒅′×𝒅
• Message Aggregation
𝑮𝑪𝑵 𝒄𝒐𝒏𝒔𝒊𝒅𝒆𝒓 𝒆𝒊 𝒐𝒏𝒍𝒚
13
Embedding Propagation LayerHigh-order Propagation
• Message Aggregation
𝐖𝟏𝒍 ,𝐖𝟐
𝒍 ∈ ℝ𝒅𝒍×𝒅𝒍−𝟏
14
Embedding Propagation LayerHigh-order Propagation
• Propagation Rule in Matrix Form
It allows us to discard the node sampling procedure
Simultaneously updating the representations for all
users and items in a rather efficient way
(facilitate batch implementation)
15
Embedding Propagation LayerHigh-order Propagation
• Matrix Form of layer-wise propagation rule
𝑹 ∈ ℝ𝑵×𝑴 :user-item interaction matrix
A :adjacency matrix
D :Diagonal degree matrix
𝑬(𝒍) ∈ ℝ 𝑵+𝑴 ×𝒅𝒍
Node degree
16
Laplacian Matrix
1
3
4
2
5
ℒ = 𝐷 − 𝐴 =
−1 −1 0 0 0−1 2 −1 −1 00 −1 3 −1 −10 −1 −1 3 −10 0 −1 −1 2
𝐴 =
0 1 0 0 01 0 1 1 00 1 0 1 10 1 1 0 10 0 1 1 0
𝐷 =
1 0 0 0 00 2 0 0 00 0 3 0 00 0 0 3 00 0 0 0 2
• Adjacency Matrix & Diagonal Degree Matrix
• Laplacian Matrix
17
Laplacian Matrix
ℒ 𝑠𝑦𝑚 = 𝐷−12 𝐿 𝐷−
12 = 𝐼 − 𝐷−
12 𝐴 𝐷−
12
ℒ𝑖,𝑗𝑠𝑦𝑚
• Symmetric normalized Laplacian
18
Model Prediction
19
OptimizationBayesian Personalized Ranking
Dislike? Missing value?
20
OptimizationBayesian Personalized Ranking
𝑦𝑢𝑖𝑗 = 𝑦𝑢𝑖 − 𝑦𝑢𝑗
• Learning models with BPR
• Objective function
𝓡+: observed interactions𝓡−: unobserved interactions
𝐿2 𝑟𝑒𝑔𝑢𝑙𝑎𝑟𝑖𝑧𝑎𝑡𝑖𝑜𝑛
𝑓𝑜𝑟 𝑒𝑣𝑒𝑟𝑦 𝑟𝑎𝑛𝑑𝑜𝑚𝑙𝑦 𝑠𝑎𝑚𝑝𝑙𝑒𝑑 𝒕𝒓𝒊𝒑𝒍𝒆 𝒖, 𝒊, 𝒋 ∈ 𝓞
BPR
21
Experiment
22
Dataset
Dataset Interaction# Item# User# Density
Gowalla 29,858 40,981 1,027,370 0.00084
Yelp2018 31,831 40,841 1,666,869 0.00128
Amazon-Book
52,643 91,599 2,984,108 0.00062
• Statistics of the evaluation datasets
23
Evaluation metrics
𝐷𝐶𝐺𝑘 =
𝑖=1
𝑘2𝑟𝑒𝑙𝑖 − 1
log2(𝑖 + 1)
𝑁𝐷𝐶𝐺@𝐾 =𝐷𝐶𝐺@𝐾
𝐼𝐷𝐶𝐺
• Normalized Discounted cumulative gain
24
Comparison Baselines
• Model-based CF methods
MF
NeuMF
CMN
• Graph-based CF methods
HOP-Rec
• Graph Convolutional Network
PinSage
GC-MC
25
Performance Comparison
26
Comparison w.r.t. Sparsity Level
27
Comparison w.r.t. Sparsity Level
28
Study of NGCF
29
Study of NGCF
30
Conclusion
31
Conclusion
• Extensive results demonstrate the state-of-the-art performance
of NGCF and its effectiveness in improving the embedding
quality with neural embedding propagation.
• The research present the NGCF model, a new recommendation
framework based on graph neural network, which explicitly
encodes the collaborative signal in the form of high-order
connectivities by performing embedding propagation.
32