Social recommender system

Social Recommender System

2015/4/10 1Middleware, CCNT, ZJU

Yueshen XuMiddleware, CCNT, ZJU

[email protected]

[email protected]

Knowledge

Engineering

&

E-Commerce

Outline

2015/4/10 2Middleware, CCNT, ZJU

Where from?

How to recommend?

What to recommend?

What’s the problem?

ML & DM

Related Topics

Trends

What’s your

perspective?

Basic, Generalized, Comprehensible

Introduction

Social Overload

Facebook largest social network site

– 600,000,000 users

YouTube largest video sharing site

– 2,000,000,000

Twitter largest microblogging site

– 65,000,000 tweets per day

Sina microblog largest microblogging site in

China

– 400,000,000 users

2015/4/10 Middleware, CCNT, ZJU 3

Introduction

The Recommender Systems is an augmentation

of the social process

Any CF system has social characteristics

Social Media and Recommender Systems can

mutually benefit each other

2015/4/10 Middleware, CCNT, ZJU 4

Introduction

2015/4/10 Middleware, CCNT, ZJU 5

Real-world examplesWhy?

Different man,

Different news

Pioneer

‘....based on

recommendatio

n algorithms....’

Multi-Media

Fundamental Recommendation Approaches

Collaborative filtering based Recommendation

Aggregate ratings of objects from users and generate

recommendation based on inter-user/inter-item

similarity

Demographic Recommendation

Age,gender,income…

Content-based Recommendation

Music gene

Hybrid Methods

Mixed

2015/4/10 Middleware, CCNT, ZJU 6

Your imagination

Fundamental Recommendation Approaches

In the real world, we seek advices from our

trusted people

CF automate the process of ‘word-of-mouth’

Select a subset of the users(neighbors) to use as

recommenders

2015/4/10 Middleware, CCNT, ZJU 7

Collaborative Filtering

Fundamental Recommendation Approaches

Shall we recommend Superman for John?

Jon’s taste is similar to both Chris and Alice tastes

Do not recommend Superman to him

2015/4/10 Middleware, CCNT, ZJU 8

User based CF algorithm

Fundamental Recommendation Approaches

2015/4/10 Middleware, CCNT, ZJU 9

User based CF algorithm

vi - the mean vote for user i

k - a normalization factor

pij – the predicitive vote

w(i,j ) – the similarity between ui and uk !

Cose based similarity Pearson Based similarity

Fundamental Recommendation Approaches

The transpose of the user-based algorithms

Bob dislike Snow-white(which is similar to Shrek)

Do not recommend

2015/4/10 Middleware, CCNT, ZJU 10

Item based CF algorithm

W(k,j) is a measure of item similarity – usually the cosine measure

Matrix Factorization

Matrix Decomposition

Tri-angle

LU

QR

Spectral

SVD

2015/4/10 Middleware, CCNT, ZJU 11

Matrix Factorization

SVD-like

Non-negative

PMF

BPMF

pLSA, LDA

Matrix

Theory

Machine

Learning

Discriminative Model

Generative Model

Unsupervised

Learning

Matrix Factorization---SVD : the ancestor

Rudiment---Singular Value Decomposition

For an arbitrary matrix A there exists a factorization

named SVD, as follows:

2015/4/10 Middleware, CCNT, ZJU 12

Matrix Factorization---Latent Semantic Analysis PTM LDA

Low-rank matrix factorizationWhy factorizing?

– One is about the interpretation

– You prefer Lost in Thailand ‘cause it’s a drama, and X, and

Y, and Z, and ......

– X, Y & Z are named as latent factors

So matrix factorization can be come across as

another type of LSA(Latent Semantic Analysis)

2015/4/10 Middleware, CCNT, ZJU 13

Share us

sth

corssing

your mind

Probabilistic

Topic Model !

Matrix Factorization---SVD-Like : low-rank matrix factorization

Latent Factor Model Generative Model

Low-rank matrix factorization Latent Factor Space

2015/4/10 Middleware, CCNT, ZJU 14

QPRR T

QPRR T

QPRR T

QPRR T

QPRR T

QPRR T

Rating

Matrix

Approximate

Rating Matrix User Latent

Factor Matrix

Item Latent

Factor Matrix

ff

ifufui fiQfuPqpriuR ),(),(),(

Predicted value ),( jiR

kk

ikuk kiQkuPqpjirjiR ),(),(),(),(

k-rank

factors

Basic Form

Matrix Factorization---SVD-Like : low-rank matrix factorization

Minimize the sum-squared errors

2015/4/10 Middleware, CCNT, ZJU 15

Skip

Details

m

i

n

j

j

T

iijQP

QPR1 1

2

, 2

1min

m

i

n

j

j

T

iijijQP

QPRI1 1

2

,)(

2

1min

Frobenius Form

Just like Quadratic regression

I : the indicator function

Regularization

Avoid overfitting Why? Sparsity/Sample

Shortage

2221

1 1

2

, 22)(

2

1min

FF

m

i

n

j

j

T

iijijQP

QPQPRI

Solution

Stochastic Gradient Descent

Matrix Factorization---PMF : the production of Bayesian Theory

SVD-Like is not perfect Why?

Subject & Object the victim of formalism

Maximum Posterior Probability(MAP)

2015/4/10 Middleware, CCNT, ZJU 16

)()(),|()|,( VpUpVURpRVUp

m

i

n

j

I

Rj

T

iij

Rij

VUrNVURp1 1

2,|),|(

m

i

UiU IUNUp1

22 ),0|()|(

n

j

ViV IVNVp1

22 ),0|()|(

Gaussian

Noise

n

j

Vi

m

i

Ui

m

i

n

j

I

Rj

T

iij IVNIUNVUrNRVUpRij

1

2

1

2

1 1

2 ),0|(),0|(,|)|,(

Zero-mean spherical Gaussian prior

Surroundings

Topics related

Non-negative Matrix Factorization

– Deng Cai etc.

Boltzmann Machines

– Discarded

Heterogeneous networks

– Prof. Han

– Link Prediction & Community Discovery

Transfer Learning & Online Learning

– Qiang Yang etc.

2015/4/10 Middleware, CCNT, ZJU 17

Excavate

Structures

Neural

Network

‘Graph

Regularized

NMF for.....’

Different

Certain

Networks

Online

Algorithms

Others

• Semantic

Web

• Ranking

• Computing

Ads

• Network

Marketing

• Clustering

• NLP

• TM

• Sociology

• Etc.

Trends---Horizontal Expansion

More Relationship More Matrix

Social Network

– Turn to your friends for suggestion

Trust Network

– Turn to who you trust for suggestion

Clarify the connection

– What’s the relationship?

– Why does it work?

2015/4/10 Middleware, CCNT, ZJU 18

Weight &

Relationship

Social/Trust

Network Etc.

Structure of Networks

Trends---Vertical Expansion

3-4-5- Dimensions Tensor

A tensor can be represented as a multi-dimensional

array of numerical values.

– 1-dimensional tensor : Vector

– 2-dimensional tensor : Matrix

Tensor Decomposition & Tensor Factorization

2015/4/10 Middleware, CCNT, ZJU 19

observed

value

3th, Latent factor,

Time or Tag

1th Latent factor

one, User

2thLatent factor ,

Item

2015/4/10 20Middleware, CCNT, ZJU

Social Recommender System