UNOBTRUSIVE ACQUSITION OF USER INFORMATION FOR

E – COMMERCE APPLICATIONS

Oshadi Alahakoon, Seng Loke, Arkady Zaslavsky

School of Computer Science and Software Engineering, Monash University,900 Dandenong Rd, Caulfield,3145, Australia

Keywords: Unobtrusive Interfaces, user modelling, e-commerce

Abstract: E-commerce has become a common activity among many people. Although widely used, the interfaces

which users communicate with e-commerce systems are still at an early stage of development in terms of

intelligence and user-friendliness. Unobtrusiveness is recognized as one of the most important desired

attributes of an intelligent and friendly interface. In this paper we describe our work on an information

architecture to minimize obtrusiveness. A layered information architecture supported by a structured user

profile model is described in the paper. As example scenario is presented to clarify the new architecture and

the development of a cost model for measuring the level of obtrusiveness is discussed.

1 INTRODUCTION

Recommender systems are being used by increasing

number of commercial websites to guide consumers

locating the products they will like. To provide such

personalized results, recommender systems require

information about user preferences and needs.

Although the obvious method is asking from the

user, this will result in user filling out lengthy forms

or conducting long dialogs with the system. Such

time consuming activities may make the system

unpopular with the user. Therefore these interfaces

providing personalized services need to be

unobtrusive. Such unobtrusiveness may be achieved

by minimizing direct user inquiry and providing

users with an easy to use interface.

Although the most reliable first hand information

is obtained from direct user inputs, it compromises

the unobtrusiveness. As a solution we can use other

less reliable information sources as web contents,

stereotypes or prediction techniques based upon

available information. Therefore it is apparent that

with unobtrusiveness the precision of the user

information is reduced. In our model we attempt to

achieve a balance between unobtrusiveness and

information precision.

eHermes is a web based multi-agent system

(Jayaputera et al., 2002, Alahakoon et al., 2003),

which is currently being developed at Monash

University. It has a flexible and extensible open

architecture, which can adapt to changing

environments. eHermes helps users with their

information needs such as financial services and

online shopping for goods and services.

Our work is related to the on going work of the

personalization component (front-end) of the

eHermes system. The main challenges of this

component are of two fold.

1. Unobtrusive acquisition of user information

2. Building and maintenance of adaptive user

models

The information (described in 1) is required for

construction of long term, reusable user profiles and

also to identify current user need within a given

application domain. Then storing and management

of such acquired information as user profiles is

required. Such profiles are organized (structurally)

to be used in recommending online goods and

services in different domains. (eg. Real estate,

Insurance polices and Purchasing various

commodities). In that profile structure, some

components are reusable irrespective of the

application domain.

In this paper we focus on the unobtrusive

acquisition of information from the users. We

identify ‘levels of obtrusiveness’ in information

acquisition from a user, and relate these levels to a

user profile with a layered architecture. Therefore

the layers of the user profile represent different

levels (or degrees) of obtrusiveness that will be

‘forced’ on a user during information acquisition.

Using this definition, we present a model to

Alahakoon O., Loke S. and Zaslavsky A. (2004).

UNOBTRUSIVE ACQUSITION OF USER INFORMATION FOR E–COMMERCE APPLICATIONS.

In Proceedings of the Sixth International Conference on Enterprise Information Systems, pages 3-8

DOI: 10.5220/0002633400030008

 SciTePress

minimize the obtrusiveness by acquiring certain

information from other sources and maintaining user

information in structured profiles.

We structure our user profiles to hold user

information belonging to different application

domains in separate modules. Application domain

independent information is to be stored in a common

module. For example an individual using eHermes

to seek for a recipe need not to re-enter some

information (eg. highly obtrusive user demographic

data), when he/she needs a recommendation for a car

insurance policy. In that context, our system

contributes to unobtrusiveness by not collecting the

same information again.To demonstrate our model,

we use a recipe recommender as an example

scenario.

The rest of the paper is organized as follows.

Section 2 describes past work, which are related to

our research. Section 3 introduces our information

architecture and the layered user profile model.

Section 4 describes the user interface, which makes

use of the new architecture described in section 3. A

cost model for measuring the level of obtrusiveness

is also discussed in section 4. Section 5 provides the

concluding remarks with a discussion of the future

work.

2 RELATED WORK

Personalized user adaptive systems are used in many

application areas, as information filtering and

retrieval, email filtering, recommendations in e-

commerce and intelligent user interfaces. These

systems obtain user preferences through interaction

with the users, build user models and utilize these

models to provide users with customized results. In

long term they learn about the individual user and

adapt themselves to give more personalised results.

The work we describe here lines up with the e-

commerce recommendation systems. In addition to

precise recommendations, our intension is to provide

users with an unobtrusive interface, which makes

system-user interaction an enjoyable one. Recent

work on recommender systems includes

personalized systems recommending music

(Shardanand and Maes, 1995), electronic TV

programme guides (Ardissono et al., 2001),

restaurant recommendations (Tewari et al., 2000,

Burke, 1999, Thompson et al., 2002), information

retrieval (Middleton et al., 2001, Balabanovic and

Shoham, 1995, Marko Balabanovic and Shoham,

1995), real estate (Shearin and Lieberman, 2001)

and many other application areas.

The user profiles created in above systems are

only to be used in particular applications. In that

context, users have to employ different systems for

their different information needs. Users have to

disclose their information to each of these

applications. Again these users need to get familiar

with various user interfaces. To avoid such efforts,

we propose a single system, with the ability to create

and maintain an adaptive and application

independent user profile. The profiles created in our

model hold some common information for number

of application domains.

In addition to above application dependent user

modelling systems, there are user modelling shell

systems (Orwant, 1991, Kobsa and Pohl, 1995, Kay,

1995). Shell systems maintain knowledge about

users and assist interactive software systems in

adapting to their current users by providing

assumptions about user requirements.

Generally shell systems too maintain different

user profiles for different applications.

DOPPELGANGER (Orwant, 1991) is a server

based shell system which has centralised user

profiles for providing applications with assumptions

about user behaviour. User information acquisition

in DOPPELGANGER is done from various

resources using number of techniques. Some of this

information is gathered using sensors. This

information is application independent. As most of

the user modeling data that is useful for an

application, remains application specific, this may

not acquire useful information (Pohl, 1996). In our

model, although there is a single user profile for all

the domains, it is a collection of application specific

user profiles.

The uniqueness in our model is in the use of 3-

layered information architecture, which maintains

user information according to obtrusive levels. Such

information structuring gains more control over the

user information. Using the cost model (section 4)

together with structured profiles, it is possible to

control the obtrusiveness of a particular user session.

Again such structured profiles could be used to

impose different security levels over user

information. For example user information in the

first level (more personal and private) of the profile

structure can be regarded as confidential. Access of

such information could be controlled using different

authority levels.

3 INFORMATION

ARCHITECTURE

In this section we discuss the categorization of

information in to different levels and the structured

user profile architecture.

ICEIS 2004 - HUMAN-COMPUTER INTERACTION

3.1 Categorization of User

Information

As our model is used in many domains for different

user needs (using possibly different user interfaces)

the profile generated should have both general and

domain dependent information. We categorized the

information requirement into 3 separate levels.

1. User Classification information – domain/

request independent(Eg. Demographics).

2. Domain dependent information – long-term

user needs in the problem domain.

3. Request dependent temporary information

– spontaneous/temporary user needs

In our proposed information acquisition model,

all required information is categorized into the above

3 levels. Level 1 contains more personal & private

information. The 2

level of information is less

private as it only describes the user with respect to a

particular information domain. Level 3 is very

general temporary information. We propose that a

user would consider an attempt to acquire

information at level 1 (classification information) as

highly obtrusive while level 2 (domain) and 3

(request dependant) information as of lesser

obtrusive nature. Therefore we allocate the degrees

of obtrusiveness as high, medium and low to the

information category levels 1, 2 and 3 respectively.

Since level 1 and 2 information can apply to a

number of transactions we relate this information to

a user’s profile. This information represents a

particular user or user group. On the other hand,

level 3 information (although may be influenced by

level 1 & 2) are only directly related to a particular

transaction or event. To include the concept of this

layered information architecture, we have developed

a two-layered profile architecture of users (described

below).

3.2 Structured User Profiles

As introduced above we create profiles that may be

used in multiple application areas. Therefore we

organize the structure of user profiles in to a two

layered architecture.

The information required for these long term

adaptive profiles are mainly of two types.

1. Information regarding general characteristics

of the user – do not change values depending on the

application area. Example demographic information.

2. User’s likes and dislikes towards item

attributes he/she is expecting to purchase-

application dependent.

We call former attributes as Classification details

and latter as Prediction details. Classification details

are kept as domain independent. They are obtained

from user inputs, past behaviour and using

stereotypes. Some of the attributes never change as

name, date-of-birth or gender; these are called static

attributes. Attributes that can change over time such

as address, occupation or income are called default

attributes.

As an existing user make requests, for items

from different application domains, system updates

the default user classifications (if changed) and re-

generates appropriate prediction details.

Classification part

Default

Suburb Caulfield

Marital status Single

Occupation Clark

Monthly Income $3000

Weight 80kg

Height 170cm

Smoker Mild

Health Issues Diabetes

Profile Id 34

Static

Given Name John

Surname Smith

DOB 30/11/70

Gender Male

Predictive part -

Car Insurance

Car_Used_Locally H

Freq_visited_SC 1,2

LongDis_Travel L

Predictive part - Recipe

Keen_to_Cook H

TSpent_Cooking H

Prefered_food_Cul C

Preferred_Main_Ingredient 1,4,2

BS_Concern M

Figure 1: Layered user profile architecture

Each time a new request is received from a new

application domain the profile grows by adding a

new prediction layer instance. As classification

details are domain independent, they are stored in

the main module to be used commonly within

different application domains.

Figure1 shows how two layers of user

information in two different application domains

(cooking recipe recommendation and car insurance

policy recommendation) are connected to the main

module. There is a main module consists of

application domain independent user information.

Rests of the modules are for different application

domains holding user information corresponding to

that particular domain.

UNOBTRUSIVE ACQUSITION OF USER INFORMATION FOR E – COMMERCE APPLICATIONS

4 A USER INTERFACE FOR

OPTIMISING

UNOBTRUSIVENESS

In user modelling literature, unobtrusive information

(about the user) gathering is considered as one of the

main challenges. In our work we take several

measures to minimize system intrusiveness.

1. Ask for user inputs only when unable to use

an alternative information source.

2. Ask clear and short questions that are quickly

readable.

3. Provide options/possible answers, so that the

user can reply with just a mouse click.

4. Finally, include user selected and system

selected preferences with the results.

To obtain optimal unobtrusiveness, the following

indirect sources are used in our system.

1. A well organized domain database

2. Stereotypes

3. Related information repositories - Eg. online

recipes, food nutrition information, supermarket

websites

4. Past user behaviour – Accepting or rejecting

a system recommendation

4.1 The Unobtrusive user interface

The goal of our user interface is to acquire as much

information as possible regarding the user’s

requirement(s) and preferences. The challenge is to

acquire this information whilst minimising

obtrusiveness.

The system maintains a static question graph

with all the questions it needs in order to determine

the most suitable recipe for a particular user.

Questions in the graph may belong to any of three

information layers described in section 3.1. The

answers to the questions are used to filter out the

large number of recipes. The most suitable once are

presented to the user, with the option of browsing

through a list of close matches. The question graph

consists of

(a) A standard set of questions representing the

domain of interest

(b) A set of links providing directions over

possible alternative question sequences.

When a user logs into the system the questions

will be activated. Depending on the level of

obtrusiveness (calculated using a cost model -

section 4.2), system decides either direct acquisition

(i.e. ask the user) or indirect acquisition of answers

(using one or more of the sources listed earlier).

Progression of questioning is described with an

example in section 4.3.

4.2 The cost model

A cost model has been developed to measure the

unobtrusiveness during a user interaction session.

We are hoping to use this as an optimizer to provide

system users with unobtrusive interactions. The cost

model is described below.

Each question is allocated a level of

obtrusiveness (L), according to the information

categorization level. This value is ‘high’ for level 1,

‘medium’ for level 2 and ‘low’ for level 3. Each

question can obtain its answers from different

resources, namely user input (UI), stereotypes (St),

past user behavior (PUB), Information Repositories

(IR) and available profile information (PI).

Depending on the information source used, the

degree of obtrusiveness (Obs), the certainty of

values obtained (UN), and the time needed for

acquisition (T) will vary. We calculate the cost

incurred for each information acquisition act, by

identifying values for these three parameters. The

information source dependant values for the

parameters are given in table 1.

Since the cost of question q

will depend on the

three values and the level of the question (within the

profile),

Cost(q

) = L (a

Obs + a

UN + a

× T)

where ai ∈ Ν, i∈{1,2,3}, 0

a≤

≤ 1 and a

1_+

+ a

3 = 1; L is

level of obtrusiveness with in the

profile.

Table 1

Info-

Source

Obtrusive

ness (Obs)

Uncertainty

(UN)

Time (T)

High Low Low

Low High Medium

PUB

Low Medium High

Low Medium Low

Low Low High

Costs Obs, UN and T can take values low,

medium or high. For example values 0.3, 0.6 and 0.9

can be used for low, medium and high respectively.

The total cost of a query would be,

∑

Cost(qi)

where N is the number of questions required to

satisfy the query.

Values in table 1 demonstrate that we can build

reliable profiles within a short time span by being

ICEIS 2004 - HUMAN-COMPUTER INTERACTION

highly obtrusive (or vice versa). The cost model can

be used to adapt to the context (whether priority

should be unobtrusiveness or reliability and/or

speed). ie. using a maximum allowable total cost (as

a threshold) the system can select the information

sources for satisfying the questions, maintaining the

cost within the threshold value.

4.3 An example scenario

In our example scenario, system interacts with the

user in an unobtrusive manner and recommends

recipes from a large recipe collection. We developed

a database consisting of 50 recipes for testing

purposes. These recipes are obtained from the

AllRecipes (AllRecipes) website. The recipes are

described in terms of user needs. Some of the

attributes are Dish Type (eg. Main, Dessert,

Appetizer etc), Meal Type (Lunch, Dinner, etc),

Tradition (eg. Chinese, Indian, Malaysian, etc), Food

Type (e.g. Beef, Fish, Poultry, etc).

4.3.1User Interface

In our implementation, the questions are displayed

one after the other in dialog boxes. The user can

either answer or ignore the questions. Most of the

questions come with options (possible answers), for

the user to select with a mouse click. A subset of

questions and answers in a session is shown in the

figure2.

4.3.2 System Questions

The questions ask may belong to different ‘levels’ of

information categories as described in section 3. For

example questions and their possible information

level (given within brackets) are as follows.

Q1: What meal Time? (3)

Q2: What is the occasion? (3)

Q3: What season?

(3)

Q4: What dish?

(3)

Q5: What traditional food preferred?

(2, 3)

Q6: Any health issues?

(1)

Q7: What cooking times preferred?

(2)

Q8: Any preferred main ingredient?

(3, 2)

It is possible for certain questions (Q5 and Q8)

to belong to two or more levels. In such a situation,

the Cost calculation will require a pre-defined

fraction of membership to each level.

Figure 2 demonstrates the question graph with

two alternative routes. The path may divert to

another node depending on the answer/information

for a particular question. For example (figure 2) for

different answers to Q1 (meal time) the path will be

directed to another node as some questions (eg: Q4)

becomes irrelevant.

Enough

in fo rm a tio n

to find

recipes

UI - Dinner

UI - Tea

UI - General

IR -S um m er

UI - Soup

UI - No

St - High

Need m ore

in fo rm a tio n

PUB - B'day party

IR - Summer

S t - Ita lian

St - Stereotypes

PUB - Past User Behavior

IR - Information Repositories

UI - User Input

PI - P rofile Info

PI - Beef

Figure 2: The question graph

5 CONCLUSIONS AND FUTURE

WORK

The paper described some on going work on

building an unobtrusive interface to an e-commerce

system. The main contributions of this paper is the

concept of identifying levels (or layered)

information architecture and the categorization of

the user requirements into this architecture. The

structured user profiles (Alahakoon et al., 2003)

described, supports this model and thus provides the

UNOBTRUSIVE ACQUSITION OF USER INFORMATION FOR E – COMMERCE APPLICATIONS

foundation for our cost model. The structured user

profile provides the advantage of separating the

domain independent information, thereby making it

possible to use the top layer of the user profile for

many domains.

ACKNOWLEDGEMENTS

Support for this project from the Australian

Research Council (ARC) Linkage Grant LPO211384

and Microsoft is thankfully acknowledged.

REFERENCES

Alahakoon, O., Loke, S. and Zaslavsky, A., 2003.

Modular and hierarchical user profile instantiation for

e-commerce applications using mobile software

agents. The 14th Australasian Conference on

Information Systems (ACIS).

AllRecipes,

http://allrecipes.com/.

Ardissono, L., Portis, F., Torasso, P., Bellifemine, F.,

Chiarotto, A. and Difino, A., 2001. Architecture of a

system for the generation of personalized Electronic

Program Guides. Proc. UM2001 Workshop on

Personalization in Future TV (TV'01).

Balabanovic, M. and Shoham, Y., 1995. Learning

Information Retrieval Agents: Experiments with

Automated Web Browsing. Proceedings of the AAAI

Spring Symposium on Information Gathering from

Heterogeneous,Distributed Resources. Stanford,CA.

Burke, R., 1999. The Wasabi Personal Shopper: A

Case_Based Recommender System. In Proceedings of

the 11th National Conference on Innovative

Applications of Artificial Intelligence. 844-849,

AAAI,.

Jayaputera, G., Alahakoon, O., Cruz, L., Loke, S. and

Zaslavsky, A., 2002. Assembling Agents On-Demand

for Pervasive Wireless Services. Proceedings of the

Second International Workshop on Wireless

Information Systems.

Kay, J., 1995. The UM toolkit for reusable, long term user

models. User Modeling and UserAdapted Interaction.

4(3), 149--196.

Kobsa, A. and Pohl, W., 1995. The user modeling shell

system bgp-ms. User Modeling and User-Adapted

Interaction. 4(2), 59--106.

Marko Balabanovic and Shoham, Y., 1995. Learning

Information Retrieval Agents: Experiments with

Automated Web Browsing. Proceedings of the AAAI

Spring Symposium on Information Gathering from

Heterogeneous,Distributed Resources. Stanford,CA.

Middleton, S. E., Roure, D. C. D. and Shadbolt, N. R.,

2001. Capturing Knowledge of User Preferences:

Ontologies on recommender systems. In Proceedings

of the First International Conference on Knowledge

Capture (K-CAP2001), Victoria, B.C. Canada.

Orwant, J., 1991. M.S. Thesis, Doppelgnger goes to

School: Machine Learning for User Modeling.

Program in Media Arts and Sciences. MIT.

Pohl, W., 1996. Learning about the user -- user modeling

and machine learning. Proc. ICML'96 Workshop

Machine Learning meets human-Computer

Interaction. 29--40.

Shardanand, U. and Maes, P., 1995. Social Information

Filtering: Algorithms for Automating ``Word of

Mouth''. Proceedings of (ACM) (HI) 95 Conference on

Human Factors in Computing Systems. 1, 210--217.

Shearin, S. and Lieberman, H., 2001. Intelligent Profiling

by Example. Proceedings of the International

Conference on Intelligent User Interfaces (IUI 2001)

,Santa Fe, NM,. 145-152.

Tewari, G., Youll, J. and Maes, P., 2000. Personalized

Location-Based Brokering Using an Agent-Based

Intermediary Architecture. In Proc. of the Int. Conf. on

E-Commerce, Seoul, Korea,.

Thompson, C. A., Göker, M. and Langley, P., 2002. A

personalized system for conversational

recommendations. Journal of Artificial Intelligence

Research.

ICEIS 2004 - HUMAN-COMPUTER INTERACTION