AN INVESTIGATION INTO DYNAMIC CUSTOMER

REQUIREMENT USING COMPUTATIONAL INTELLIGENCE

Yih Tng Chong and Chun-Hsien Chen

Nanyang Technological University,50 Nanyang Avenue, Singapore

Keywords: Dynamic customer requirement, Product design, Artificial immune system, Artificial neural network,

Design space framework.

Abstract: The twenty-first century is marked by fast evolution of customer tastes and needs. Research has shown that

customer requirements could vary in the temporal space between product conceptualisation and market

introduction. In markets characterized by fast changing consumer needs, products generated might often not

fit the consumer needs as the companies have originally expected. This paper advocates the proactive

management and analysis of the dynamic customer requirements in bid to lower the risk inherent in

developing products for fast shifting markets. A customer requirements analysis and forecast (CRAF)

system that can address the issue is introduced in this paper. Computational intelligence methodologies, viz.

artificial immune system and artificial neural network, are found to be potential techniques in handling and

analysing dynamic customer requirements. The investigation aims to support product development

functions in the pursuit of generating products for near future markets.

1 INTRODUCTION

In the twenty-first century, new product

development (NPD) businesses face the

uncertainties of highly turbulent market. Changing

customer needs is known to be a key driver of

uncertainties that are inherent in NPD projects

(Herstatt et al., 2006). While the processes of

requirement management in product development

have been relatively well-investigated, only few

studies considered the temporal dimension of the

critical information. The status is similar in the

industry; studies have shown that companies in

general have unjustly neglected the aspect during

product development (Kärkkäinen et al., 2001). In a

volatile market, it is clearly imprudent for one to

comment on the validity of customer requirement

data without making references to the frames of

time. Requirement is the basis of product

development, and it can vary with time. In other

words, the requirements of the same customer can be

different in different points of time. The variation of

customer satisfaction attribute weights along the

temporal dimension was elucidated in a longitudinal

study (Mittal et al., 1999). Prior to ramping up

production, design specifications are by necessity

frozen. In cases where customer requirements shift

substantially between the points of design freeze and

market introduction, the product may satisfy the

customers to a lesser extent than intended. Product

developers, when unwary of this variable, might end

up generating products not wanted by the customers.

As the issue of dynamic customer requirements is

increasingly valid due to globalisation and stiffening

competition, it is urgent and critical to recognize

customer needs as time-based variables, in both

practice and research.

2 A BRIEF OVERVIEW OF THE

CURRENT SOLUTIONS

Reichwald et al. (2005) commented that traditional

market research methodologies focus only on

current situation and often do not contribute to the

correct assessments of future customer requirements.

With historical data, time series methods can be

employed for forecasting. Xie et al. (2003)

employed the double exponential smoothing

technique in projecting the importance level of the

requirements, i.e. quantitatively. The method is

however limited to forecasting quantitative data, and

only of linear trend. Raharjo et al. (2006) proposed a

113

Chong Y. and Chen C. (2009).

AN INVESTIGATION INTO DYNAMIC CUSTOMER REQUIREMENT USING COMPUTATIONAL INTELLIGENCE.

In Proceedings of the 11th International Conference on Enterprise Information Systems - Artiﬁcial Intelligence and Decision Support Systems, pages

113-117

DOI: 10.5220/0001960801130117

 SciTePress

method that prioritises the quality characteristics in

the context of a dynamic QFD. The method

identifies quality characteristics that have greater

confidence in meeting future customer requirements.

However, the method does not produce future

customer requirement information; on the contrary,

it requires the information as input. Chen and Yan

(2008) analysed customer utility based on radial

basis function neural network. While the method

does not specifically ascertain the future customer

needs, it predicts the future customer preferences

over a range of design options.

Despite the importance of ascertaining future

customer needs, product developers generally placed

little attention to it (Kärkkäinen et al., 2001). Of the

sparse studies in dynamic customer requirements,

the direction is in general to counter the

uncertainties the variable contributes to NPD. Given

the increasingly fast moving product markets,

research and development effort in the area is ever

more valid. This work advocates product

development organizations to proactively look into

the future requirements of customers; an approach

that analyses the variations of the customer

requirements is introduced in this paper.

3 DATA REPRESENTATION

OF DYNAMIC CUSTOMER

REQUIREMENTS

A customer requirements analysis and forecast

(CRAF) system is proposed to address the issue of

dynamic customer requirement. In the CRAF

system, dynamic customer requirement (DCR)

information is modelled based on the Design Space

Framework proposed by Chong et al. (2009). An

adapted structure employed in this study is hereby

referred to as the Requirement Space Framework

(RSF), as shown in Figure 1. For the robustness of

the data, customer requirements are represented in

multiple levels of abstraction, as described along the

Z axis of the RSF. Chong et al (2009) postulated that

alternative values can exist for requirements, and

details of the requirements can be further

represented by co-requirements. Correspondingly (as

shown in Figure 1), requirement options are

represented along the Y axis while co-requirements

are modelled along the X axis of the RSF. Apart

from the three dimensions of the information, a

fourth facet of the user needs data, i.e. time, is

prescribed in this study. Figure 1 depicts a schematic

snap-shot of the user needs data in a time instance.

m m

Bio-molecule Option link ARB instances

Abstraction plane Co-requirement link

Figure 1: The Requirement Space Framework.

In this work, the learning and analysis of the

dynamic customer needs data is primarily based on

techniques of the artificial immune system (AIS). As

such, the data representation scheme is based on the

basic conventions in the AIS research (de Castro and

Timmis, 2002). The following describes the

mapping of the RSF model to the AIS-based data

representation schemes. Customer requirements are

represented by two types of data classes – the

antigens (Ag) and B-cells. The DCR input data is

modelled as Ag that ‘infiltrates’ the CRAF system,

while the system memory is represented by B-cells,

or more specifically, the artificial recognition balls

(ARB) (see Table 1 and Figure 2). ARB represents a

group of identical B-cells in the system (Timmis and

Neal, 2001). States of ARBs are dichotomy, either

active or inactive. Bio-molecules - the building

blocks of the system data - are modelled with

symbolic attribute-value pairs, e.g. ARB

3,2

= <colour,

black>. Undefined (i.e. ‘don’t care’) bio-molecules

of Ag and ARB are represented using the symbol #.

Table 1: The AIS-based system.

Vertebrate immune

system

CRAF System

Antigens

Customer requirements (in

product market)

B-cells/ ARB

Customer requirements (in

system memory)

Biological environment Product market

ICEIS 2009 - International Conference on Enterprise Information Systems

114

Environment

Antigens

CRAF System

B-cells

ARB

Figure 2: A schematic diagram of the CRAF system.

Concentration of B-cells within the system

represents the importance level of attributes and the

preference level of values of the respective ARB.

The antibody concentration fluctuates as the result of

the interacting mechanisms within the biological

immune system, involving biological processes such

as cloning, apoptosis, and cell activation. In this

work, the dynamics of the B-cell population is

metaphorically applied to model the dynamism of

customer requirements in product markets.

4 THE CUSTOMER

REQUIREMENT ANALYSIS

AND FORECAST ALGORITHM

Immune system is described as adaptive, self-

organizing in nature, maintaining a memory of past

encounters, and has the ability to continually learn

via new encounters (Dasgupta, 2006). An intrinsic

and unique ability of AIS is its continuous capability

to adapt to and to co-evolve with the environment.

Based on established machine learning algorithms

(e.g. Timmis and Neal, 2001), this work proposes a

domain specific dynamic population-based clonal

selection algorithm, i.e. the customer requirements

analysis and forecast (CRAF) algorithm.

The concentration levels of the B-cells are of

interest to this work. The temporal characteristics of

the concentration levels represent the time-based

profiles of the customer needs. Customer needs data

can be non-linear, non-stationary, noisy and limited

in quantity, all which poses challenges for the time-

based learning in the CRAF system. Focused time

delay neural network (FTDNN) with

backpropagation training is embedded in the AIS-

based system to learn the temporal system data. The

proposed CRAF algorithm is described below.

1. Initialisation. A set of active ARBs at a

predetermined concentration level of ARBT

(ARB Threshold) at t = 1 is preinstalled in the

system.

2. Introduction of Antigens. Antigenic data, Ag,

is introduced to the system at time t

3. Secondary Immune Response. Each of the

introduced data packet Ag is presented to all

active ARB for stimulation.

4. Primary Immune Response. Stimulation of

inactive ARB is performed; cross-reactivity

does not apply in the primary response, i.e.

stimulation here requires exact match.

5. Apoptosis. The population of the B-cells in the

system is maintained by the process of

programmed cell death.

6. Concentration Update. The concentration

level of each ARB is refreshed.

7. Antigenic Learning. Antigenic data unfamiliar

to the system will be learnt when foreign

antigen encountered is beyond the recognition

of the existing ARB.

8. Activeness Update. If the concentration level

of active ARB falls below the ARBT level for a

specified period of time, it will be deactivated.

This period is termed as persistency period, or

PP. On the other hand, inactive ARB will be

activated when the concentration level is

maintained above the ARBT level over the

specified PP.

9. Neural Network Training. The currently

available set of B-cell concentration data is

utilized to train the embedded focused time

delay neural network.

10. B-cells Concentration Forecast. Trained

neural network is simulated to forecast the

future B-cell concentration of the respective

ARB. The algorithm is looped (t=t+1) by

proceeding to Step 2.

5 A CASE STUDY

The development of personal computer is

challenging due to the fast moving market. The

requirements of the customers evolve rapidly,

relative to the time span of the product development

cycle. Market foresight is in this case critical to the

developers in acting and reacting to the industry. In

highly competitive market, product customisation is

an important strategy to increase market share.

Robust customer requirements information therefore

plays key role in the process of product planning,

design, marketing and research and development,

AN INVESTIGATION INTO DYNAMIC CUSTOMER REQUIREMENT USING COMPUTATIONAL INTELLIGENCE

115

Figure 4: Mean percentage error of concentration forecast.

etc. Customer requirements information can be

derived from market surveys, as the input to the

CRAF system. Customer requirements may also be

inferred from the sales of products, by regarding the

acts of purchase as customer needs and preferences

expressions.

In this study, an appropriately scaled and

disguised proprietary dataset from a personal

computer manufacturer is employed (Kurawarwala

and Matsuo, 1998). The data relates the purchases

of five types of PC over thirty-eight months in a

particular market segment. P1 to P5 are the models

of PC having the assumed features, as shown below.

• P1: Brand X processor 1.6GHz, Brand A

operating system (OS), tower chassis, 1GB

memory (RAM) (667MHz), 80 GB HDD (5400

RPM), Integrated graphic processor, CD-RW

drive and diskette drive.

• P2: Brand Y processor 1.6GHz, Brand A OS,

tower chassis, 1GB RAM (667MHz), 160 GB

HDD (5400 RPM), Integrated graphic

processor, DVD-ROM and diskette drive.

• P3: Brand X processor 2.0GHz, Brand B OS,

desktop chassis, 2 GB RAM (800 MHz), 320

GB HDD (5400 RPM), discrete graphic

processor Type 1, DVD-ROM drive.

• P4: Brand Y processor 2.0GHz, Brand A OS,

desktop chassis, 2GB RAM (800 MHz), 320

GB HDD (7200 RPM), discrete graphic

processor Type 1, DVD recordable drive.

• P5: Brand X processor 2.4GHz, Brand B OS,

compact chassis, 4 GB RAM (1067MHz), 500

GB HDD (7200 RPM), discrete graphic

processor Type 2, DVD recordable drive and

media card reader.

The CRAF algorithm has been implemented in

the Matlab environment for the purpose of the study.

The simulated sequential process resulted in a set of

ARB objects that reflects the patterns in the data, i.e.

the qualitative and quantitative representations of the

customer requirements at various levels of

abstraction. Post-analysis of the data produced

useful information including the relative preferences

amongst the customer requirements. For instance, in

this case study, the system reported the discrete type

graphic card being relevant around the 25

month,

and thereafter started to gain popularity against the

integrated type (see Figure 3).

The forecast of the B-cells repertoire is made

possible having modelled the dynamics of the

evolution. Dynamic forecast of the B-cells

concentration levels in the ARB was found to be

improving over the months. The mean percentage

error of forecast, in the current study, drops below

5% for the final sampled 6 months, with each lower

than the previous (see Figure 4). The mean

percentage error of forecast during the 38

month

was noted to be 0.79%.

This study demonstrated the functions of the

CRAF system as an indicator of the future customer

requirements. The generated sets of customer needs

data are cross-referenced to the temporal space,

making them exceptionally valuable in highly time-

sensitive markets.

Figure 3: B-cells concentration of graphic card types.

ICEIS 2009 - International Conference on Enterprise Information Systems

116

6 CONCLUSIONS

The CRAF system introduced in this paper is aimed

at addressing the problem of fast shifting customer

needs. In general, the system studies the dynamics of

pattern evolution to further perform data forecast.

The algorithm is specified to operate continuously so

as to track and to analyse the customer needs data

dynamically. Such characteristic is in contrast with

traditional methods that treat singular temporal

customer data in discrete approaches. It is

envisioned that the intelligence that could be derived

from the proposed system may serve to reduce the

uncertainties inherently found in product

development projects. In view of the increasingly

fast changing market, dynamic customer

requirement analysis and forecast (as well as the

applications of the generated intelligence on

downstream activities) are vital areas for future

research.

REFERENCES

Chen, C.-H., Yan, W., 2008. An in-process customer

utility prediction system for product conceptualisation.

Expert Systems with Applications. 34(4): 2555-2567.

Chong, Y.T., Chen, C.-H., Leong, K.F., 2009. Human-

centric product conceptualization using a design space

framework. Advanced Engineering Informatics,

Special Issue on Human Centered Product Design and

Development. 23(2), in press.

Dasgupta, D., 2006. Advances in artificial immune

systems. IEEE Computational Intelligence Magazine.

Nov 2006:40-48.

de Castro, L.N., Timmis, J., 2002. Artificial immune

systems: a new computational intelligence approach.

Heidelberg Germany : Springer-Verlag.

Herstatt, C., Stockstrom, C., Verworn, B., et al., 2006.

“Fuzzy front end” practices in innovating Japanese

companies in new product development. Journal of

Product Innovation Management. 3(1):43-60.

Kärkkäinen, H., Piippo, P., Puumalainen, K., Tuominen,

M., 2001. Assessment of hidden and future customer

needs in Finnish business-to-business companies.

R&D Management. 31(4): 391-407

Kurawarwala, A.A., Matsuo, H., 1998. Product growth

models for medium-term forecasting of short life cycle

products. Technological Forecasting and Social

Change. 57: 169-196.

Mittal, V., Kumar, P., Tsiros, M., 1999. Attribute-level

performance, satisfaction, and behavioral intentions

over time: A consumption-system approach. Journal

of Marketing. 63: 88-101.

Raharjo, H., Xie, M., Brombacher, A.C., 2006. Prioritizing

quality characteristics in dynamic quality function

deployment. International Journal of Production

Research. 44 (23): 5005-5018.

Reichwald, R., Stotko, C.M., Pillar, F.T., 2005.

Distributed mini-factory networks as a form of real-

time enterprise: concept, flexibility potential and case

studies. In: Kuhlin, B., Thielmann, H., (editors), The

practical real-time enterprise – facts and perspectives.

Springer Berlin Heidelberg. 403-434.

Timmis, J., Neal, M., 2001. A resource limited artificial

immune system for data analysis. Knowledge-based

Systems. 14: 121-130.

Xie, M., Tan, K.C., Goh, T.N., 2003. Advanced QFD

applications. ASQ Quality Press. Wisconsin.

AN INVESTIGATION INTO DYNAMIC CUSTOMER REQUIREMENT USING COMPUTATIONAL INTELLIGENCE

117