程序代写代做代考 data mining database information retrieval recommender

recommender

15s1: COMP9417 Machine Learning and Data Mining

Recommender Systems

May 27, 2015

Acknowledgement: Material derived from Adomavicius & Tuzhulin
(2005) IEEE Trans. on Knowledge and Data Engineering, 17(6),

734–749.

Aims

This lecture will enable you to describe and reproduce machine learning
approaches within the framework of Recommender Systems. Following it
you should be able to:

• define the problem of recommender systems

• describe content-based, collaborative and hybrid recommender systems

• reproduce key similarity-based approaches to recommender systems

COMP9417: May 27, 2015 Recommender Systems: Slide 1

Introduction

• Recommender systems – a form of personalization

– “person who liked x may also like y”

• related to instance-based learning

– similarity function

• other forms of learning may be used to model user choices

COMP9417: May 27, 2015 Recommender Systems: Slide 2

A Framework for Recommendation

Example movie rating matrix, where each entry has user c rating item s.

Given: utility u : c⇥ s 7! R

Problem: 8c 2 C, choose s0c = argmaxs2S u(c, s)

This is learning in the sense of requiring extrapolation to predict the
unknown values of the utility funcion.

COMP9417: May 27, 2015 Recommender Systems: Slide 3

Content-based Recommendation

User c is recommended items s that are similar to past choices.

• idea comes from information retrieval

• requires a profile of the content or description of items

u(c, s) = score(ContentBasedProfile(c),Content(s))

E.g.,

u(c, s) = cos(~wc, ~ws) =
~wc · ~ws

k~wck ⇥ k~wsk
where

~wc is a vector of summarising terms of c’s past choices, and

~ws is a vector of most relevant terms describing s

COMP9417: May 27, 2015 Recommender Systems: Slide 4

Content-based Recommendation

Advantages

• well-understood techniques from Information Retrieval

• can extract latent features from text analysis

Disadvantages

• may not have content, or may be limited or sparse

• over-specialisation: recommendations given for known types only

• new user problem: must do some rating to get recommendations

COMP9417: May 27, 2015 Recommender Systems: Slide 5

Collaborative-based Recommendation

User c is recommended items that users with similar taste have chosen.

• a.k.a. collaborative filtering (CF)

• Amazon-style recommender systems

Two main methods: memory-based, and model-based CF.

COMP9417: May 27, 2015 Recommender Systems: Slide 6

Memory-based CF

Predict unknown rating rc,s of user c for item s by aggregating the ratings
of N users c0 most similar to c who have rated s:

rc,s = aggrc02C rc0,s

What aggregation to use ? One commonly used is weighted sum

rc,s = k

X

c02C

sim(c, c

0
)⇥ rc0,s

where

k is just a normalising factor, and the similarity function can be correlation,
cosine distance, etc. on the vector of items rated (e.g., bought) by users.

Alternatively, can use item-based similarity (Amazon).

COMP9417: May 27, 2015 Recommender Systems: Slide 7

Model-based CF

Memory-based CF is like a nearest-neighbour method.

A big problem is sparsity — to address this, often try to find a low-
rank approximation to the matrix (i.e., finding smaller “user-feature” and
“movie-feature” matrices) using a form of stochastic gradient descent.

However, can use other machine learning methods to build a model to
predict directly the unknown rating rc,s from examples in the database.

E.g., Naive Bayes-type approaches.

This is called model-based CF.

COMP9417: May 27, 2015 Recommender Systems: Slide 8

Collaborative-based Recommendation

Advantages

• works well in practice

• does not require content (descriptions)

Disadvantages

• new user problem: must do some rating to get recommendations

• new item problem: must be rated to be used in recommendations

• “grey sheep”: insu�ciently individual !

• “black sheep”: too individual !!

COMP9417: May 27, 2015 Recommender Systems: Slide 9

Hybrid Recommender Systems

Key idea: combine model-based and memory-based approaches

• “cold-start” problem: use model to predict before user activity

• “sparsity” problem: use model to predict missing values

But: learning models may be di�cult or expensive

COMP9417: May 27, 2015 Recommender Systems: Slide 10

Summary

• based on techniques from information retrieval and machine learning

• an application area growing rapidly

• simple systems can do surprisingly well

• many possible extensions, e.g., recommendation in social networks

COMP9417: May 27, 2015 Recommender Systems: Slide 11