AN INTELLIGENT RECOMMENDATION SYSTEM BASED ON

FUZZY LOGIC

Shi Xiaowei

Philips (China) Investment Co.,Ltd. Shanghai R&D Centre

38F,Tower1 Kerry Everbright City 218 Tian Mu Xi Road Shanghai, P.R.C.200070

Keywords: Recommendation, User profile, Fuzzy-logic, Multi-agent, Metadata

Abstract: An intelligent recommendation system for a plurality

of users based on fuzzy logic is presented. The

architecture of a multi-agent recommendation system is described. How the system simulates human

intelligence to provide recommendation to users is explained. The recommendation system is based on the

fuzzy user profile, fuzzy filtering and recommendation agents. The user profile is updated dynamically

based on the feedback information. Fuzzy logic is used in fuzzy filtering to integrate different types of

features together for a better simulation of human intelligence. Ambiguity problems can be solved

successfully in this system, e.g., deducing whether a programme with both interesting features and

uninteresting features is worth recommending or not. The application scenario shows that it is more

convenient for users to find programmes of their interest with the proposed recommendation system.

1 INTRODUCTION

Due to the digitalisation of television, in the near

future we will be able to receive hundreds of

channels via satellites, terrestrial antenna, cable and

even phone lines. At the same time, it is becoming

increasingly challenging for television viewers to

identify television programmes of their interest.

Recommendation systems, such as TV-Advisor,

Personal TV, etc, have been studied to help viewers

to find, personalize, and organize the contents

[1,2,3,4,5].

The following recommendation process is

considere

d: matching Electronic Programme Guide

(EPG) metadata to the user preference knowledge,

filtering out the tedious programmes and

recommending interesting programmes to users, and

updating the user profiles based on feedback

information.

The traditional methods to evaluate if a

rogramme is good enough to be recommended are

based on the explicit (i.e. non-fuzzy) inference. In

other words, the programme evaluation result can

either be “interesting” or “non-interesting”. As we

know, explicit mathematics cannot intelligently

simulate human’s flexible inference. Especially, it is

difficult to decide whether a programme with both

interesting features and uninteresting features should

be recommended by the existing filtering and

recommendation methods.

This paper presents an al

gorithm about the

integration of fuzzy logic into a multi-agent

recommendation system to better recommend

interesting programmes to users. The architecture of

the fuzzy recommendation system is described.

Moreover, the reason why fuzzy logic is adopted in

the recommendation system is explained. The fuzzy

recommendation algorithm is also developed.

Finally, the recommendation process is illustrated

with an example.

2 RECOMMENDATION SYSTEM

2.1 System Architecture

The recommendation system is provided to generate

programme recommendations for multiple users

based on programme metadata and user profiles. The

architecture of this system is illustrated in Figure 1.

The system uses a central recommendation server

unit that includes a Fuzzy User Profile Database,

Fuzzy Filtering Agent, Fuzzy Recommendation

Agent, Profiling Agent, and Interface Agent.

180

Xiaowei S. (2004).

AN INTELLIGENT RECOMMENDATION SYSTEM BASED ON FUZZY LOGIC.

In Proceedings of the First International Conference on Informatics in Control, Automation and Robotics, pages 180-184

DOI: 10.5220/0001136301800184

 SciTePress

Figure 1: System architecture.

• Fuzzy Filtering means to gather and make the

incoming live programme broadcasted based on

programme metadata match to fuzzy user

profiles, which is based on fuzzy logic, and then

filter the programmes based on fuzzy threshold;

• Fuzzy Recommendation generates an optimal

recommendation list to every user according to

the learned user preference knowledge, and

transmits it to user terminals;

• Profiling updates the user profile based on both

the explicit and implicit feedback information

from the Interface Agent;

• The Interface Agent handles the interactions

with the user.

2.2 Information Description

In the context of TV-Anytime [7], metadata consists

of two kinds of information: (a) content description

metadata; (b) consumer metadata. Programme

content description metadata includes attributes of a

television programme, such as the title, genre, list of

actors/actresses, language, etc. These data are used

to make a search.

The user profile metadata in this system defines

the details of a user’s preferences and aversion.

These descriptions are closely correlated with media

descriptions, and thus enable the user to search,

filter, select and consume the desired content

efficiently.

2.3 Application of Fuzzy Logic

Control Theory in a

Recommendation System

There are many factors that influence a user if he/she

wants to view a programme or not. The user’s

attitude to a programme is the result of some

complicated reaction. In other words, it is difficult

for a user to describe their emotion about a

programme in quantity. Fuzzy theory can simulate

human intelligence. It owns the advantage of

describing this kind of indefinite object, providing a

possible way to solve the problem. Hence, fuzzy

Recommendat i on Pr ocess

User

Programs+Met adat a

Fuzzy User Profile Dat abas e

Fuzzy Fi l t er i ng Agent

Fuzzy Recommendat i on Agent

Rec ommendat i on

Li s t

Feedback

Informat i on

Profi l ing Agent

Interface Agent

Knowl edge base

Def uzzi f i cat i on

Deci si on

maki ng

Fuzzification

Pr o c e s s

i nput out put

Fuz z y

Figure 2: General structure of fuzzy inference system .

theory is used in programme recommendation in this

paper.

Fuzzy theory includes a series of procedures for

representing set membership, attributes, and

relationships that cannot be described by single point

numeric estimates. The structure of a fuzzy

inference system is shown in Figure 2.

Where,

Knowledge base: parameters of membership

functions and definitions of rules;

Fuzzification: transformation of crisp inputs

into membership values;

Decision-making: fuzzy inference operations on

the rules;

Defuzzification: transformation of the fuzzy

result of the inference into a crisp output

3 FUZZY RECOMMENDATION

3.1 Fuzzy Information Database

3.1.1 The Fuzzy User Profile

The user profile,

, can be represented by a vector

of these 3-tuples. If there are distinct terms in the

profile, it will be represented by:

)),,().....,,,),....(,,((

Where: is a term;

ld is the “Like_degree” of

term

;

is the Weight of term ; i is the order of

in the profile.

(3-1)

wldtwldtwldtUP =

t w

The fuzzy user profile transforms the crisp

parameters (“Like_degree”, Weight) into

membership values. “Like_degree” means the

degree the user likes a feature. The shape and

location may be different for different problems. If

and

represent “Like_degree” and “Weight”

respectively, the fuzzy memberships can be

described as Figure 3. It is known from Figure 3 that

e e

Figure 3: The fuzzy membership function of user profile.

AN INTELLIGENT RECOMMENDATION SYSTEM BASED ON FUZZY LOGIC

181

is always greater than 0. A larger

indicates

that this feature is more important. When

, it

means “like” and a larger

means the user likes the

feature more; if

, it indicates that the user

“dislike” this feature and the smaller , the more

the user dislikes it.

e e

0≥e

ttttC =

0≤

3.1.2 EPG Metadata

A programme can be represented by a vector of size

),...,...,,(

21 ni

Where is the i

(3-2)

feature of the programme.

3.2 Similarity Matching

The Fuzzy Filtering Agent calculates the similarity

between the programme and the user profile. If the

calculated similarity is above the preset threshold,

the user is considered to be interested in the

programme; then the programme metadata is

transferred to the Recommendation Agent.

In this system, the fuzzy similarity matching

process can be divided into two steps:

• Feature matching;

• Programme matching.

3.2.1 Feature Matching

Feature matching decides how a feature of a

programme is related to a user’s preference. In order

to get the feature “interest_degree”， the procedures

is executed as shown in Figure 2.

Here, Like_degree and Weight are set as input

and the feature of “interest_degree”, which means

how much a user likes the programme, is set as

output.

The step of decision making is based on the

fuzzy inference rules such as following,

I.If Like_degree is “dislike” And Weight is

“secondary“ Then

is “disgusted”;

II.If Like_degree is “dislike” And Weight is

“important” Then

is “very disgusted”;

III.If Like_degree is “neutral” And Weight is

“important” Then

is “neutral”;

…

The method of “centre of gravity” takes more

useful factors into consideration. It is adopted in this

system to defuzzicate the feature’s interest_degree.

secondar y

neut r al i mpor t ant

)(

we)(

lde

dislike

neut r al

-0.5

0. 5

- 0. 25 0 0. 25 10. 50. 2500.75

3.2.2 Programme Matching

Programme matching is to evaluate the programme

interest_degree. It can be calculated by the average

interest_degree of the features related with the

programme.

Figure 3: The fuzzy membership function of user profile

3.3 Filtering & Ranking

Next step is to set a threshold to filter the coming

programme metadata, select the interesting

programme, and then rank and recommend them

based on the programme interest_degree. In this

system, ranking and recommendation processes are

performed by the Fuzzy Recommendation Agent.

A threshold is set in the Fuzzy Filtering Agent.

The threshold can be a crisp value, or a fuzzy value

such as “how much does the user like”. If the

programme interest_degree is greater than the

threshold, which means the programme is what the

user wants to watch, then the Filtering Agent will

transfer the programme metadata to the Fuzzy

Recommendation Agent.

Based on the learned user preference

knowledge, the Fuzzy Recommendation Agent

generates an optimal recommendation list and sends

it to the user according to interest_degree.

3.4 Profiling Agent

In the system, the feedback can be explicitly given

by the user or implicitly derived from observations

of the users’ reaction to a recommended programme.

So, the Profiling Agent revises the user profile based

on both the explicit and implicit feedback

information. In this section, how to update the user

profile by the implicit feedback information is

mainly explained.

For a recommended programme, the user

always has two attitudes to it: skipping over, or

watching. In other words, the user will skip (delete)

the programme he/she dislikes, watch the

programme he/she likes or he/she is not sure. If the

user has watched the programme for a period of

time, the user’s profile will be refined and revised

based on the viewing behaviour.

In this system, for programme

, the algorithm

for updating user profile is depicted as follows:

WeighttWeigh

)-(

•+=

′

(3-3)

ICINCO 2004 - INTELLIGENT CONTROL SYSTEMS AND OPTIMIZATION

182

eLike_degreeLike_degre

)-(

•+=

′

(3-4)

Where,

: The time duration watched;

: The real time duration of the

programme;

: The threshold of the time duration. If

is less than WD

, that means the user is

not interested in that programme;

and

: are less than 1. They are used to

slow down the change of Weight and

Like_degree. Because Weight is more

stable than Like_degree,

≤

is larger than its higher-boundary, let

;

tWeigh

′

ndaryhigher_bouWeight'

is less than its lower-boundary, let

;

tWeigh

′

darylower_bounWeight'

is larger than its higher-boundary,

let

;

eLike_degre

′

ndaryhigher_boue'Like_degre

is less than its lower-boundary,

let

eLike_degre

′

darylower_boune'Like_degre

4 EXAMPLE

4.1 Similarity Matching and

Preference Learning Example

Table 1 shows an assumed user profile A and

upcoming programme. The problem to be solved is

to determine whether the user likes the programme

and how much he/she does.

Based on the above described procedures, the

programme fuzzy filtering inference can be

illustrated as Figure 4.

For the user profile, the actor LiQinqin’s

Like_degree is -0.125, which indicates the user’s

emotion about him is between “dislike” and

“neutral”. In this case, both values of

neutral=ld

and

dislike=ld

are 0.5. In addition, the feature of “Actor” 's

Weight is 0.8, so this feature is “important” and

important

Matching by fuzzy logic rules, the actor

LiQinqin meets both rule II and III. For rule II，

is 0.5, which means the user is “very disgust” at this

feature; for rule III,

is 0.5 and the user feels

“neutral” about this feature.

Through defuzzification, interest_degree

for

the actor LiQinqin is about -0.4. It shows that the

user’s emotion about this feature is mainly “much

disgusted”;

Considering other features, the calculated value

of the programme interest_degree P is 0.45. From

Figure 4, when P (0.45) is mapped into its fuzzy

Table 1: Initial conditions.

User profile A Programme”Cala

is a dog”

Genre Weight=0.9

(Movie Like_degree=0.5

…);

Actor Weight=0.8

(Ge you Like_degree=0.5

…..

LiQinqin

Like_degree=-0.125);

Genre is movie

Actors are

GeYou, LiQinqin

Duration=2hour

membership，the emotion of the user can be

obtained. In the case discussed, it is between “much

interested” and “interested” (

2.0 ≈

interested

8.0 ≈

interestedh

muc

According to the user’s behaviour, the user

profile is updated. For the discussed programme,

duration

is 2 hours. Assumed, the threshold of

the watching time duration

20=

minutes, the time

duration watched

=2 hours,

=0.01，

=0.1,

then:

83.0

)-(

For the fuzzy model of user profile A assumed,

the user profile is updated as following:

Genre Weight=0.9083

(Movie Like_degree=0.5083

Comedy Like_degree=0.3

News Like_degree=-0.2);

Actor Weight=0.8083

(XuJing Like_degree=0.1

GeYou Like_degree=0.583=0.5

LiQinqin Like_degree=-0.125+0.083=-0.042);

4.2 Application Scenario

An application scenario is provided. Figure 5 shows

the main user interface of the fuzzy recommendation

system. Functions are listed in the left column. A

programme recommendation list is on the topside of

middle column. The description of a selected

programme is presented in the bottom of middle

column. For a selected programme, three choices are

provided to the user, which are “Display”, “Delete”,

and “Skip”.

5 CONCLUSION

An intelligent Multi-agent recommendation system

is developed to provide programme

AN INTELLIGENT RECOMMENDATION SYSTEM BASED ON FUZZY LOGIC

183

recommendations for multiple users based on

programme metadata and fuzzy user profiles.

Figure 4: The programme filtering inference.

Different from traditional methods, the user profile

is not based on the viewing history but on a compact

and low-cost structure, including all terms that user

“likes” and “dislikes”. The user profile is updated

dynamically, e.g. “increase” or “decrease” the

corresponding preference parameters according to

the feedback information. The filtering agent uses

fuzzy logic to integrate different types of features

together for a better simulation of human

intelligence. This system shows a better capability in

solving problems of ambiguities in programme

recommendation, e.g., deducing whether a

programme with both the interesting features and

uninteresting features is worth recommending or not.

The application scenario shows that this

recommendation system can help users enjoy life

more freely.

Figure 5: Main interface.

REFERENCES

Kaushal Kurapati, Srinivas Gutta, 2001. A Multi-Agent

TV Recommender. http://www.di.unito.it/~liliana/

UM01/kurapati.pdf

)(

lde

di sl i ke

neut r al

-0.25

-0.125

-0.25

-1

-0.5- 0. 75 0 0. 25 1

ver y

di s gust ed

much

di s gust ed

neut r al

interested

ver y

interested

di s gust ed

much

interested

0. 750. 5

0. 5

0. 6

-0.4

0. 875

-1

ver y

di s gust ed

much

di s gust ed

neut r al

interested

ver y

interested

di s gust ed

much

interested

0. 45

0. 2

0. 375

0. 5

Fuzzi f i cat i on of

f eat ur e" Li Qi nqi n"

Feat ur e " Li

Qi nqi n"

si milarity

inference

Pr ogr am

si milarity

i nf er ence

Belkin, N.J. & Croft, W.B., 1994. Information filtering

and information retrieval: two sides of the same coin.

Communications of the ACM,1994

Foltz, P. W. and Dumais, S. T., 1992. Personalized

information delivery: An analysis of information

filtering methods. Communications of the ACM, 1992

Ardissono, L. and Buczak, 2002. A. Personalisation in

future TV”. Proceedings of TV’02, the 2nd workshop

on personalization in future TV.

Masthoff, J., 2002. Modeling a group of television

viewers, Proceedings of the workshop Future TV,

2002

Babuska, R., 1998. Fuzzy Modeling for Control”.

International Series in Intelligent Technologies,

Kluwer Academic Publishers

TV-Anytime Forum, 2003. Metadata Specification S-3

Part A: Metadata Schemas, Version 1.3,

http://www,tv-anytime.org

ICINCO 2004 - INTELLIGENT CONTROL SYSTEMS AND OPTIMIZATION

184