Vous êtes sur la page 1sur 7

Proceedings of SOM 2014

December 12-14, 2014


Department of Management Studies, IIT Roorkee
pp.635-641

A Neural Feature Extraction Model for Clustering


Supplier Firms for Supply Chain Collaboration
Pankaj Kumar Medhi1, Sandeep Mondal2, Jash Jhaveri3, Nikhil Gupta4
1

JRE Group of Institutions, 2Indian School of Mines, Dhanbad


3
JRE Group of Institutions, 4JRE Group of Institutions

Abstract: Clustering is a family of classification techniques, often preceding further analysis or application in a number of
fields like data analysis, strategy selection, supplier selection, etc to name a few. While principal component analysis (PCA) is
often used for grouping based on covariance matrix, data based neural techniques are gaining popularity due to flexibility
and adaptability. Among neural clustering techniques Self organizing map (SOM) or Kohonens map is often used when the
objects to be clustered has many attributes. In both supervised and un-supervised modes, Kohonens map exhibits good
capability to extract a classification which assigns highest weight to the most important attribute among all the attributes of
an object.
In this paper we have applied SOM or Kohonens map for classification of firms based on their attributes of information
processes used for market information acquisition. Additional data regarding the success of the said firms success in
collaboration with other firms or supplier is added to see if there is a particular association between a set of information
processes used by firms and its ability to collaborate with partner firms successfully.
Using data from World Bank enterprise survey (2005) of German industries, we could produce three distinct clusters of
industries on the basis of information processes used. When collaboration success data was included, it still showed three
clusters. Hits were obtained using specific support vector for identification of clusters. We found evidence of association
between information processes based on relational self and firms ability to collaborate successfully with partners.

1. INTRODUCTION
The process of choosing suppliers for an organizations
resource need or needs is called partner selection (Wu &
Barnes, 2010). Supplier selection is critical for supply chains
(SC) as suppliers performance on cost, quality, delivery and
service plays a key role for achieving SC objectives.
Segmentation or classification of suppliers is needed to
identify and align the suppliers to the needs of the customers
(Barrat, 2004). Choosing right partner has become more
important by day as supply chains rather than individual firms
compete against each-other (Lambert and Cooper, 2000).
Among all supplier relationships, collaborative partnership or
collaboration has emerged as a strategic value creator for
market competitiveness. Collaboration is the highest level of
SC relationships in terms of strategic importance and
complexity which facilitates SC relationships to integration
level (Davis and Spekman,2004). As a result a shift from
compliance to agreed responsibilities to commitment towards

the partner happens (Braziotis and Tannock, 2011). While


partner selection in SC is a complex problem due to its multiobjective nature (Ho et al. 2010), the strategic nature of
collaboration adds new dimensions to the already complex
problem of partner selection.
As organizations can offer substantial business to a few
organizations and hence creates a set of qualification criteria to
reduce the number of potential partners and supplier
development costs (Dowlatshahi, 2000). While relation to
operational performance and competitive priorities of cost,
quality, delivery and flexibility to partner selection is broadly
agreed upon (Lin and Chen, 2004), the emerging scenario of
business demands a wider range of criteria. The supplier firms
are rated on the basis of these parameters and classified
accordingly.
Collaboration is often said to be strategic in nature and two
streams of theory, namely, Resource Based View (RBV) and
Transaction Cost Economy (TCE) is cited to explain the
phenomenon of collaboration. The concept of collaboration

XVIII Annual International Conference of the Society of Operations Management (SOM 2014)

635

A Neural Feature Extraction Model for Clustering Supplier Firms for Supply Chain Collaboration
itself has been comprehensively defined as consisting of seven
sub-components like goal congruence, information sharing,
incentive alignment, etc (Cao et al., 2010). While RBV and
TCE explains the need of collaboration for organizations, the
efficient execution of the seven sub-processes or at least dome
of them will decide whether collaboration between two
organizations will be successful. Partner selection for
successful collaboration by a firm depends on its ability to
predict whether a chosen partner firm will be able to
participate successfully in the execution of the sub-processes
of collaboration process. Often firms look for certain set of
characteristics as predictors of such capability in a firm. These
characteristics or criteria may include location of a firm,
willingness to share information, ability to provide design and
technological capabilities to customers, etc.
While research on collaboration is plenty as mentioned above,
research focused on the possibility or probability of forming
successful collaborative relationships by firms is rare. What
type of firms or which firm or firms have higher probability of
forming successful collaboration with partner firms? What
processes practiced by firms enable them to become successful
in forming collaboration with other firms?
While there have been a number of studies related to partner
selection for collaboration, market information processes or
simply information processes used by firms have never been
the focus of any such study.Firms use information processes to
collect, disseminate and use market information (Parry &
Song, 2010) for assisting in decision making process. These
processes include actions like meeting with customers,
retailers, surveys etc., (Jaworski&Kohli, 1993). Often the firms
use formal as well as informal processes for information
collation. Market orientation research assumes that formal and
informal information processes are highly correlated and have
same impact on firm performance (Moorman, 1995; Parry &
Song, 2010). Information processes may have impact on the
collaborative relations of a firm, because from process point of
view collaboration will need accurate and complete data
regarding a number of aspects of partner firms. Collaboration
also involves a number of decision making at various levels
and hence dependent on accurate data for good decisions.
Hence ability of a firm to collect appropriate data about a
potential partner firm and its ability to form successful
collaboration will depend on a large scale on its information
processes.
Here we propose to fill this gap with a classification procedure
based on Artificial Neural Network (ANN) to classify firms
based on their use of information processes. Our classification
is not about supplier selection based only on efficiency, but
also to locate them suitably on the basis of their suitability for
strategic relationships. A typology or clustering based on
information process will not only correctly identify suitable
group of partner firms, but it will also help a firm to choose
appropriate strategy regarding information collection.

Information collection is important as collaboration is a longterm partnership and demands relationship specific investment,
hence increasing the importance of accurate and complete
information.
Clustering or classification of firms on chosen criteria of an
organization will make supplier selection easier for it. In most
of the situations there may be more than one capable supplier
for the same task. Many models of supplier selection uses a
number of judgment based rating input and hence their results
are not hundred percent accurate. Rather than finding the best
one for supplier selection, clustering of the suppliers on
selected set of criteria will give us a cost effective way of
finding our suitable set of suppliers for a specific task. This
will also help us in the event our first choice fails to complete
the task.
Criteria for supplier selection is often soft and organization
specific (Wu & Barnes, 2010). Previous studies have often
attempted to point out a single best supplier from a set to
optimize the organizational outcome. Optimization techniques
like fuzzy logic, AHP, Neural network, genetic algorithm and
combinations of them were often used to formulate the
requirements, constraints and then to solve them.
We intend to use the existing performance data of various
suppliers to train our network for classification. We intend to
use existing theoretically supported criteria as constraints in
the decision making process for supplier selection. The
problem will be modeled using neural network, and then
classify the firms into a number of clusters. We propose to
verify our model with existing data. While there is a wealth of
literature on partner selection and a number of models
proposed on it, very few have proposed to validate their
models using data. Adding to it, very few models of
classification of supplier firms for suitability for strategic
relationships like collaboration have been proposed.

2. METHODOLOGY
Partner selection is a multi-criteria decision making problem
(MCDM) involving a number of conflicting factors or
dimensions. As the complexity of the situation increases,
number of factors affecting the problem increases. The trade
off among the criteria must be analyzed before any decision is
taken and a high number of factors make it highly complex.
MCDM supporting techniques based on artificial intelligence
(AI) are among the most promising in the domain.
The huge advantage of this type of techniques is derived from
the fact that computer based AI system can be trained by
historic data or by an expert. Following which a non-expert
can consult the system when faced with a new but similar
situation. While supplier selection MCDM methods based on
genetic algorithm, AHP or fuzzy logic requires formulization
of the decision making process to select one best supplier, AI

XVIII Annual International Conference of the Society of Operations Management (SOM 2014)

636

A Neural Feature Extraction Model for Clustering Supplier Firms for Supply Chain Collaboration
based systems do not require it. AI based systems are designed
to function more like human judgment process. As a result AI
systems can cope better with complexity and uncertainty than
other systems (Guo et al., 2009).
AI based systems need to be provided with the characteristics
on which supplier performance is to be assessed. The decision
of such systems is decided by what the system learn from
historical data or use cases.
Extensive research in partner selection and collaboration has
found an exhaustive list of characteristics firms look for in
partner firms. We chose the following set of characteristics for
our problem of classification of supplier firms. The
characteristics are listed in Table I.
TABLE I: Selection criteria

Supplier
Cost
Design Quality
Conformance quality
Delivery dependability
Delivery speed
Customer service
Order size flexibility
Green index
Social commitment

3.

Partner
Ability for organizational
learning
Ability to develop trust
Financial health of supplier
Technology
Innovation capability
Ability for communication
Operational Flexibility
Proactive approach for
relationship
Brand value
Strategic fitness
Criticality of the input
Capability / Market position

CLUSTERING AND FEATURE DETECTING


NEURAL NETWORK

Neural network can be used for clustering of data. Networks


can be tuned to certain similarity aspects in the data being
evaluated. Supplier performance data can be used to train such
networks and can be used for not only classification of the
existing suppliers but also to group new vendors into an
existing group or class. Suitability of suppliers for strategic
relationship like collaboration can also be determined through
membership of vendors to a particular cluster or class.
Cluster detecting networks exhibit remarkable ability of selforganization. This property has close relation with the
formation of knowledge representation in artificial
intelligence. Another important function of neural networks is
feature detection. Feature detection is related to data
dimensionality reduction which can be fruitfully used for preprocessing of data. Self-organizing maps are used for
clustering data with many advantages compared to the classical
hierarchical methods of classification. One such network is

Kohonens (1982) network. Although such networks have


simple architecture, the self-organization process involves
considerable subtleties.

4. FEATURE MAPPING
Feature extraction is used at preprocessing stage of data to
facilitate classification or clustering. At times it may be
difficult to observe the natural structure in data due to a high
number of dimensions or characteristics. Natural structure is
often related to pattern clustering or their mutual location in
the original pattern space (Jurada, J.M., 2006). We develop our
perception of classes aided by this natural structure. Contrast
to harmonious perception, a chaotic mixture presents difficulty
for their clustering. This may be due to a mismatch between
the data space and our perception space. An increase in the
dimensionality of the pattern space obviously makes it difficult
for us to observe the natural pattern in the data space.
Feature extraction or feature mapping helps to reduce the
dimensionality of vectors in pattern space when representing
these vectors in feature space, while providing as natural a
structure of features as possible. Lower dimension in the
feature space makes for easy perception compared to that in
the original pattern space and makes the clustering process
easier.
Feature mapping capabilities of neural network can be used
without supervision if the architecture is identified. In this way
the higher dimensionality of object definition will not be a
constraint for their classification. To achieve this objective,
feature mapping algorithm can be used to convert patterns of
arbitrary dimensionality into the responses of one or two
dimensional arrays of neurons. Our idea is to convert the
original complicated structure of the objects in the pattern
space into clusters in a one, two or three dimensional feature
space.
Feature mapping algorithm converts pattern of arbitrary
dimensionality into the responses of one or two-dimensional
arrays of neurons. We propose to use one-dimensional linear
array system developed by Kohonen (1982) for topology
preserving mapping for a select set of n-dimensions
representing every supplier.
Let us assume a set of input patterns {xi, i = 1, 2, ..}, where
Xirepresent a supplier with n-dimensions, is arranged in an
ordering relation with respect to a single feature within the set.
A linear array of neurons receive the set as parallel input
signals to produce localized responses. Kohonen (1982)
provides for the following set of equations as one-dimensional
topology-preserving mapping for i1> i2> i3when
yi1 = max {yi(x1), i = 1, 2, 3, , n}
yi2= max {yi(x2), i = 1, 2, 3, , n}
yi3 = max {yi(x3), i = 1, 2, 3, , n}

XVIII Annual International Conference of the Society of Operations Management (SOM 2014)

(1)

637

A Neural Feature Extraction Model for Clustering Supplier Firms for Supply Chain Collaboration
The output in the mapping is topological or preserves the
neighbourhood. The advantage of this method is the input
order need not be specific and can follow arbitrary metrics and
order. Additionally, the dimensionality of the input space
vectors need not be restricted. No restriction on the dimensions
creates big flexibility in dimensions considered for defining a
suitable supplier in any situation. A network for feature
mapping from an original input space into a two-dimensional
rectangular neuron array is shown in Figure 1.

As the training progresses the radius of Nm should be


decreasing, implying Nm(t1) >Nm(t2) >Nm(t3) ..where
t1<t2<t3 .
The weight updating rule for self-organizing feature maps is
given belowwi(t) = [x(t) - wi (t)]

for i m(t)

(4a)

where m(t) denotes the current spatial neighborhood.

Fig. 1

Since learning constant depends both on training time and


size of the neighborhood, equation (4a) can be re-written in
detail as
wi(t) = (Ni, t) [x(t) - wi(t)] for i m(t) --

(4b)

whereis a positive-valued learning function, 0 <(Ni, t) < 1.


As needs to decrease as learning progresses, it can be
expressed as a decreasing function of time. Following is an
example using time and neighborhood radius yielding good
results(Ni, t) = (t) exp [- ||ri - rm || / e2(t)]

W1
X1

W2

wherermandri are the position vectors of the winning vectors of


the winning cell and the winning neighborhood nodes,
respectively and (t) and e(t) are suitably decreasing functions
of learning t.

W3

X2

X3

Inputs

5.

The output yi of a neuron within the array is a function of


similarity S between the input vector xandwidefined as
yi = f (S(x, wi )) ---------------------------

(2)

Learning within self-organizing maps is collective and


collaborative. The best matching neuron cells also activate
their spatial neighbors to react to the same input. Such learning
tunes the network in an orderly fashion by defining some
feature coordinates over the trained network. After learning the
network gives a localized response to every input which
reflects the dominant feature characteristics of it.
We propose to apply input x simultaneously to all nodes and
use spatial neighborhood Nm as a more adequate measure S of
similarity between x and wi than scalar product metric of
similarity.In the current learning step the weights affecting the
winning neighborhood, Nm, undergo adaptation, other weights
remain unaffected. Nm is found around the best matchingnode
m using the following equation|| x - wm|| = min{|| x - wi ||}-

(6)

(3)

SIMULATION

AnR program was developed to represent the above feature


extracting neural network using Kohonen package. Selforganizing maps (SOM) can be used in both unsupervised and
supervised mode. Our objective here is to classify or cluster
firms based on the market information processes used by them
and then predict the ability of the firms to form successful
collaboration with other firms based on their use of
information processes.
A set of training data with 15 dimensions as mentioned in
Table 1 was used to train our network. The measure for the use
of information processes by the firms and the response data set
used for training and verification has been adapted from the
World Bank enterprise survey, 2005. Data from Germany and
Ireland has been selected fro this purpose due to their matured
and developed economic system. For simulation we use
Kohonens network, also known as self-organizing maps
(SOM) in R statistical package. The reason for selection of
SOM is its suitability for high dimensional data analysis. SOM
function in R does not handle data with missing values for
training purpose in R, but for verification purpose such data
sets are acceptable.

XVIII Annual International Conference of the Society of Operations Management (SOM 2014)

638

A Neural Feature Extraction Model for Clustering Supplier Firms for Supply Chain Collaboration

TABLE 2: Study Measures


Factors identified for simulation

Informal Information Process 1 (IIP1)


q21b Importance of former employees who now work for a
potential customer or supplier as a potential source for new
customer
q21c Importance of prior employment of managers by a
potential customer or supplier as a potential source for new
customer
q22b Importance of former employees who now work for a
potential customer or supplier as potential sources of
information about new suppliers for your firm
q22c Importance of prior employment of managers by a
potential customer or supplier as potential sources of
information about new suppliers for your firm
Formal Information Process (FIP)
q21e Importance of Government agencies as potential sources
of information about new customers for your firm
q21f
Importance of Business associations/chambers of
commerce as potential sources of information about new
customers for your firm
q22e Importance of Government agencies as potential sources
of information about new suppliers for your firm
q22f
Importance of Business associations/chambers of
commerce as potential sources of information about new
suppliers for your firm
Informal Information Process 2 (IIP2)
q21a Importance of family and friends as potential sources of
information about new customers for your firm
q21d Importance of existing customers or suppliers as
potential sources of information about new customers for
your firm
q22a Importance of family and friends as potential sources of
information about new suppliers for your firm
q22d Importance of existing customers or suppliers as
potential sources of information about new suppliers for your
firm
CollaborationPerformance (COL_PER)
How important in retrospect was this initiative for the survival
and/or growth of your company over the last 36 months?
q60b4 Agreed to a new joint venture with foreign partner
q60b5 Obtained a new product licensing agreement
q60b6 Outsourced a major production activity that was
previously conducted in-house

An R program was developed for the purpose of simulation


and a number of graphs were plotted for visualization of
results.Kohonen package in R provides a number of options
for visualization of data to facilitate the process of
summarizing the results. Among the options codes, counts,
properties and clusters are the prominent ones. A number of
plots were generated for our simulation and important amongst
them are incorporated below for reference purpose.

XVIII Annual International Conference of the Society of Operations Management (SOM 2014)

639

A Neural Feature Extraction Model for Clustering Supplier Firms for Supply Chain Collaboration
and third cluster uses either formal or informal information
processes, while the fourth type uses minimally any of the
information processes. While the first cluster can be called
information rich, the second and third cluster is partially rich in
information. The fourth cluster is the poorest in terms of use of
the information processes for market information acquisition.
Events of successful collaborations show a significant
coincidence with mainly two of the four clusters of the firms.
Firms using both the informal information processes (IIP1 &
IIP2) together or only IIP2 appear to have higher instances of
successful collaboration. But firms using formal as well as
both the informal information processes appear to have very
little instances of successful collaboration. Same is the case for
firms using very little of all the three types of information
processes. Hence we can conclude that informal information
processes used by firms has a significant bearing on the
collaboration success of firms.
However future research needs to identify which type of
informal information processes, whether using relatives or
friends or using relationships developed at workplaces is more
effective in forming successful collaboration with other firms.
In addition, the effects of firm size and how old a firm is on the
information processes used by firms also need to be verified.
There is strong possibility that size and age of a firm may
influence the choice of its information processes.
REFERENCES

6. DISCUSSION AND CONCLUSION


A number of significant observations can be made from the
graphs that were plotted for visualization of the results.
The plots of properties showed that the weights of formal
processes used for information collection decreased along Xaxis.
Informal Information Process 1 (IIP1) or the use of informal
relations developed in work places showed a general decrease
in use along the Y-axis.
Informal Information Process 2 (IIP1) or the use of informal
relations with relatives and friends outside the work places
showed an interesting finding. While they are used by almost
every firm as a preferred sources for new customers, their use
as information sources for new suppliers is very limited.
Overall, four distinct clusters of firms are visible. One cluster
uses all the types of information process available, the second

[1] Barrat, M. (2004), Understanding the meaning of collaboration


in the supply chain, Supply Chain Management: An
International Journal, Vol. 9 No. 1, pp. 30-42.
[2] Braziotis, C. and Tannock, J. (2011). Building the extended
enterprise: key collaboration factors. The International Journal
of Logistics Management, Vol. 22 No. 3, pp. 349-372.
[3] Cao, M., Vonderembse, M.A., Zhang, Q. and Ragu-Nathan, T.S.
(2010), Supply chain collaboration: conceptualisation and
instrument development, International Journal of Production
Research, Vol. 48 No. 22, pp. 6613-6635.
[4] Davis, E.W. and Spekman, R.E.(2004), The Extended
Enterprise: Gaining
Competitive
Advantage
through
Collaborative Supply Chains, New Jersey: Prentice Hall.
[5] De LurdesVeludo, M., Macbeth, DK. And Purchase, S. (2004),
The Extended Enterprise: Gaining Competitive Advantage
through Collaborative Supply Chains, Prentice hall, Upper
Saddle River, NJ.
[6] Dowlatshahi, S. (2000), Developing a Theory of Reverse
Logistics,Interfaces, Vol. 30 No. 3, pp. 143-55.
[7] Guo, L.,Rivero, D. and Pazos, A. (2010), Epileptic seizure
detection using multiwavelet transform based approximate
entropy and artificial neural networks, Journal of Neuroscience
MethodsVolume 193, Issue 1, 30 October 2010, Pages 156163
[8]

Ho, W., Xu, X., and Dey, P.K. (2010), Multi-criteria


decision makingapproach for supplier evaluation and
selection: A literature review, European Journal of Operations,
Res 202(1), pp. 1624.

XVIII Annual International Conference of the Society of Operations Management (SOM 2014)

640

A Neural Feature Extraction Model for Clustering Supplier Firms for Supply Chain Collaboration
[9] Jaworski, B.J., &Kohli, A.K. (1993), Market orientation:
antecedents and consequences, Journal of Marketing, Vol. 57
No. 3, pp. 53-70.
[10] Kohonen, T. (1982), A simple paradigm for the self-organized
formation of structured feature maps, in Competition and
Cooperation in Neural Nets, Lecture Notes in Biomathematics,
ed. S. Amari, M. Arbib. Berlin, Springer-Verlag.
[11] Lambert, D.M. and Cooper, M.C. (2000), Issues in Supply
chain Management,Industrial Marketing Management, Vol. 29,
No.l, pp. 65-83.
[12] Lin C.W.R., Chen H.Y.S. (2004), A Fuzzy Strategic Alliance
Selection Framework for Supply Chain Partnering under Limited
Evaluation Resources, Computers In Industry, Vol. 55, No. 2,
pp. 159-179.
[13] Moorman, C. (1995),Organizational market information
processes: Cultural antecedents andnew product outcomes,
Journal of Marketing Research, Vol. 32 No. 3, pp. 318-335.

[14] Parry, M.E., & Song, M. (2010),Market information


acquisition, use and new venture performance, Journal of
Product Innovation Management, Vol. 27, pp. 1112-1126.
[15] Spekman, R.E. and Davis, E.W. (2004), Risky business:
expanding the discussion on risk and the extended enterprise,
International Journal of Physical Distribution & Logistics
Management, Vol. 35 No. 5, pp. 414-33.
[16] Wu, C. and Barnes, D. (2009) A model for continuous
improvement in supplier selection in agile supply chains,
Knowledge and Process Management, Vol. 16 No. 3, pp. 85-110.
[17] Wu, C. and Barnes, D.(2010), A dynamic feedback model for
partner selection in agile supply chains, International Journal of
Operations & Production Management, Vol. 32 No. 1, pp. 79103.
[18] Zurada, J.M., Introduction to Artificial Neural Systems, Edition
2006, West Publishing Company, St. Paul, Minnesota.

XVIII Annual International Conference of the Society of Operations Management (SOM 2014)

641

Vous aimerez peut-être aussi