Explainable Recommendation Through Attentive Multi-View Learning - - PowerPoint PPT Presentation

Explainable Recommendation Through Attentive Multi-View Learning

Advisor: Jia-Ling Koh Presenter: You-Xiang Chen Source: AAAI ‘19 Data: 2020/03/02

Content

Introduction Method Experiment 01 Conclusion 02 03 04

Introduction

Introduction

Recommendation System

Introduction

user feature × user latent item latent × item feature Matrix Factorization

Introduction

Deep but unexplainable

Neural Collaborative Filtering

Introduction

We propose a Deep Explicit Attentive Multi-View Learning Model (DEAML) for explainable recommendation:

• 1. improves accuracy from noisy and sparse data
• 2. formulates personalized explanation generation as a

constrained tree node selection problem

Problem Definition

• User set 𝑉
• Item set 𝐽
• Explicit feature hierarchy Υ
• Set the node in Υ as

ℱ = {ℱ1, … , ℱ𝑀}

• Input
• Output
• Predicted rating Ƹ

𝑠

𝑗𝑘

• Feature-level explanation 𝐹 (𝑡𝑣𝑐𝑡𝑓𝑢 𝑝𝑔 𝐺)

Microsoft Concept Graph

e.g. Pork

Microsoft Concept Graph

https://concept.research.microsoft.com/

• New York (is-a) state
• Name (is-a) information
• Facebook (

(is-a) social medium

• 5 million concepts
• 85 million “IsA” relations

Relate work

• Explicit Factor Models

Enrich user & item representation by adding set of latent factors learned from explicit feature.

capture both explicit & implicit factor

Explicit Factor Models for Explainable Recommendation based on Phrase-level Sentiment Analysis

explicit factor explicit factor

Relate work

• User/Item-feature attention matrix 𝒀, 𝒁

ℱ = ℱ1, … , ℱ

𝑞 , set of explicit feature in review

• Integrating Explicit and Implicit Features

𝑊𝑈: projection matrix Explicit Factor Models for Explainable Recommendation based on Phrase-level Sentiment Analysis

Factorization Model over matrix 𝒀,𝒁 Factorization Model over matrix A

X, Y are in the range of [𝟐, 𝐎]

Method

Framework

Deep Explicit Attentive Multi-View Learning Model

Hierarchical propagation

• Personalized User Attention
• Attn. score

𝒚𝒋𝒎 measures how much user 𝒋 cares about feature 𝑮𝒎

Attentive Multi-View Learning

h=1

• Latent factors learning from explicit features

concatenation

Latent factor learn from explicit feature

(EFM model)

Latent factor learn from implicit feature (EFM model)

item representation at view h user representation at view h

rating prediction in h view

Attentive Multi-View Learning

• Loss of each view

projection matrix rating prediction for each view estimating hidden representation of user/item

• Co-regularization loss

enforcing agreement

Attentive Multi-View Learning

• Weighted sum prediction in each view

Calculate attention weight

Objective function

• Jointly learning

loss of each view Co-regularization loss Weighted sum prediction

Personalized Explanation Generation

• Utility function

user interest at level h item interest at level h weight of view h

4 5

• 1

6 2

Personalized Explanation Generation

• Constrained tree node selection
• max. utility of s-th child
• max. utility (s-1)-th node to t’

Experiment

Dataset

Dataset User# Item# Reviews# Toys&Games 19,412 11,924 167,597 Digital Music 5,541 3,568 64,706 Yelp 8,744 14,082 212,922

• Statistics of the evaluation datasets

5-core 5-core 10-core

Baselines

• Observed rating matrix
• NMF
• PMF

PMF

• SVD++
• Knowledge-based method
• Reviews-based method
• HFT

HFT

• EFM

EFM

• DeepCoNN
• NARRE
• CKE

Single layer structure Deep learning base

RMSE comparison

same weight to all views

Effect of number of latent factors

Case Study

Conclusion

1. We build an initial network based on an explainable deep hierarchy (Microsoft Concept Graph) and improve the model accuracy by optimizing key variables in the hierarchy 2. We propose a Deep Explicit Attentive Multi-View Learning Model (DEAML) for explainable recommendation, which combines the advantages of deep learning-based methods and existing explainable methods. 3. Experimental results show that our model performs better than state-of-the- art methods in both accuracy and explainability.