A Multi-Data Mining Approach for Shelf Space Optimization

Considering Customer Behaviour

Sheng-Hsiang Huang

, Chieh-Yuan Tsai

and Chih-Chung Lo

Department of Industrial Engineering and Management, Yuan Ze University, Taoyuan, Taiwan

Department of Applied Informatics, Fo Guang University, Yilan County, Taiwan

Keywords: Data Mining, Shelf Space Assignment, Moving and Purchase Patterns, Customer Behaviour.

Abstract: A well product-to-shelf assignment strategy can help customers easily find product items and dramatically

increase the retailing store profit. Previous studies in this area usually applied the space elasticity to

optimize product assortment and space allocation models. However, a well product-to-shelf assignment

strategy should not only consider product assortment and space elasticity. Thus, this study develops a

product-to-shelf assignment approach by considering both product association rules and traveling behaviour

of consumer. Specifically, the first task of this research is to develop a method to discover traveling

behaviour of consumer, which includes both product association rules and traveling behaviour of consumer,

in the store. The second task is to construct and solve a product-to-shelf assignment model, based on the

information provided in the first task. In this research, products are classified as major item, minor item and

the others. Only minor will be reassigned. Experimental result shows our proposed method can reassign

minor items to suitable shelves and increase cross-selling opportunity of major and minor items.

1 INTRODUCTION

To implement a retail strategy, store managers

should develop a retail mix that satisfies the need of

its target market. The elements in retail mix include

store location, product assortment, pricing,

advertising and promotion, store design and display,

services and personal selling (Levy and Weitz,

1995). Among that, product assortment and shelf

space allocation are two important issues which

dramatically affect customers’ purchasing decisions

(Yang, 1999). However, using space elasticity for

shelf space allocation needs to estimate a great

quantity of parameters which results in high cost and

errors in the mathematical model.

Recently, the progress of information technology

makes retailers easily collect daily transaction data.

With the novel information technology, retailers can

solidify ephemeral relationships with customers into

long-term and fruitful relationships if they can

discover customer behavior from collected data.

Data mining is one of the most popular technologies

that discover potential customer knowledge from

business databases to assist a policy decision. Chen

and Lin (2007) applied the multi-level association

rule mining to explore the relationships between

products as well as between product categories for

resolving the product assortment and allocation

problems in retailing. Although the association rules

to assist managers in developing better layout for

stores, their method is more suitable for the case of

new stores or joint sales. For an existing store, the

frequent purchase pattern may not maintain if the

customer’s interesting products are not at the

original locations or shelves anymore.

Except purchasing association between products,

customer traveling behaviors/patterns should be

considered in solving product-to-shelf assignment

problem. When shopping in a store, a customer

travels around the aisles of a store, stops at certain

locations, deliberates about his/her consideration,

and chooses the best options. This process is

repeated until the whole shopping trip completes.

Recently, some studies tried to tackle the shopping

path problem. Larson et al. (2005) presented

exploratory analyses of an extraordinary new dataset

that reveals the path taken by individual shoppers

around an actual grocery store, as provided by RFID

tags located on their shopping carts.

As mentioned above, some researchers employed

product association rules mining to improve shelf

space allocation, while other researchers focused on

Huang S., Tsai C. and Lo C..

A Multi-Data Mining Approach for Shelf Space Optimization - Considering Customer Behaviour.

DOI: 10.5220/0005037700890095

In Proceedings of the 11th International Conference on e-Business (ICE-B-2014), pages 89-95

ISBN: 978-989-758-043-7

 2014 SCITEPRESS (Science and Technology Publications, Lda.)

how to derive shopping paths of consumers.

However, to maximize the cross-selling possibility

for retail stores, product association and traveling

patterns should be integrated and considered at the

same time when dealing with shelf space

management. This paper, therefore, solves the

product-to-shelf assignment problem by taking both

product association rules and traveling patterns of

consumers into consideration. By combining moving

logs and payment records, customer mobile

transaction sequences (i.e., sequences of moving

path with purchased transactions) can be used to

represent the customer purchasing behavior in detail

(Yun and Chen, 2007). Furthermore, valuable

behavior patterns should be discovered to reflect the

actual profit of product items, utility mining can find

the patterns not only with high appearing frequency

but with high utility values (Shie et al., 2012).

Specifically, the first task of this research is to

develop a method to discover traveling behavior of

consumer, which includes both high-utility mobile

sequential patterns and product association rules, in

the store. The second one is to construct a complete

model, based on the information of the first task, for

solving product-to-shelf assignment problem.

The remainder of this paper is organized as

follows. Section 2 formally defines the research

problem and introduces important components of the

proposed method such as store layout, association

rule, high-utility mobile sequential pattern mining,

and product assignment procedure. In Section 3, an

empirical performance evaluation is conducted.

Finally, conclusion is summarized in Section 4.

2 THE PROPOSED METHOD

The framework of the proposed method consists of

three main stages as shown in Figure 1. The main

task in stage 1 is to collect required data. When the

customer completes the shopping, purchasing

transaction recording customer id, purchased items

and quantities are stored into the transaction

database. Next, the traversal path represented by

readers is retrieved from the traversal-temp database

and transformed into a traversal sequence

represented by sections. Finally, the system will

combine purchasing transactions and traversal

sequence of the customer as a mobile transaction

sequence and store it into the mobile transaction

sequence database. The system architecture in the

first stage of the proposed method is shown in

Figure 2.

Figure 1: The framework of the proposed product

assignment method.

Traversal-temp database

(CID, RID, time)

Shopping cart

( RFID tag is attached

on the bottom)

RFID reader

store

detect

transform

Check out

Mobile transaction

sequence database

(CID, mobile transaction

sequence)

Integrate

Transaction database

(CID, item, quantity)

Figure 2: Information architecture for collecting behaviour

data.

Typically, customers tend to follow certain

sequential patterns to pick up their required items.

For example, a customer might go to section A and

take nail polish first, then move to section C and

take pants, and move to section F and take diamond

rings before check out. Although previous studies

did apply different sequential pattern mining

approaches to obtain frequent sequential patterns,

these patterns tends to prefer frequent patterns

instead of valuable patterns. Thus, the major task in

the second stage is to explore customers’ high-utility

mobile sequential patterns based on the mobile

ICE-B2014-InternationalConferenceone-Business

transaction sequence database and the utility of

items. The high-utility mobile sequential pattern is

the sequential pattern containing a list of frequent

visiting sections and frequent purchased items at the

corresponding paths. Meanwhile, another major task

in the second stage is to derive the product

association rules from the transaction database.

In the third stage, based on the product

association rules and high-utility mobile sequential

patterns generated in stage 2, a three-step product

assignment method is proposed to rearrange items

into suitable shelves. The first step is to find all

items ever appeared in the high-utility mobile

sequential patterns. These items are defined as

“major item”. The second step is to find minor items

related with major items from the product

association rules. The final step is to rearrange minor

items to suitable shelves based on the information

derived from high-utility mobile sequential patterns

and product association rules.

2.1 Store Layout

The gray oval in Figure 3 indicates the coverage area

of an RFID reader. Let I ={i

, i

,…,i

} be the set of

all product items sold in the store, Z = {z

, z

, …, z

}

be the set of all shelves in the store, S = {S

, S

, …,

} be the set of sections in the store, and R = {R

, …, R

} be the set of RFID readers.

Figure 3: The layout of an example store.

With the purchasing transaction and traversal

sequence, a mobile transaction sequence of a

customer can be derived and defined as TS = <T

,…,T

> where transaction T

= (Se

; {[it

, q

], [it

],…,[it

, q

]}) represents that a customer

purchases {[it

, q

], [it

, q

],…,[it

, q

]} in section

where q

is the purchased quantity of pth item



I in transaction T

. A path is denoted as

…Se

, where

SSe



and

1 jr

. All the

mobile transaction sequence (TS) will be stored in

the mobile transaction sequence database (TSD).

2.2 Association Rules and Mining

The association rule mining is to discover the rules

of the presence of one set of items implies the

presence of another set of items from a given

transaction database. The form of rule can be

represented as X → Y where X and Y are the

antecedent and consequent of the rules respectively.

The Apriori algorithm, one of the most popular

methods for frequent pattern mining introduced by

(Agawal and Srikant, 1994), is adopted in this

research to get item associations from transaction

database. In the algorithm, L

denotes the set of all

frequent k-itemset and C

denotes the set of

candidate k-itemset. In this research, only “single

item to singe item” rules are needed. Therefore, this

research will stop at L

in the Apriori algorithm and

then generate rules based on provided minimum

confidence. All association rules then are stored in

product association rule database (ARD).

2.3 High-Utility Mobile Sequential

Patterns and Mining

The UMSP

(high Utility Mobile Sequential Pattern

by a Level-wised method) algorithm proposed by

Shie et al. (2012) is adopted in this paper to obtain

high-utility mobile sequential patterns. The UMSP

algorithm consists of four steps. The inputs of the

UMSP

algorithm include a mobile transaction

sequence database (TSD), a pre-defined utility table,

a minimum support threshold (δ), and a minimum

utility threshold (ε). The first three steps are to find

WUMSPs based on the sequence weighted

downward closure (SWDC) property (Liu et al.,

2005), while the forth step is to find high-utility

mobile sequential patterns (UMSPs). In step 1, the

mobile transaction sequence database (TSD) is

scanned several times to generating all WULIs (high

sequence weighted utilization section-itemset) and

each WULI is mapped to a specific identity in a

mapping table. Note that the mapped WULIs are 1-

WULPs (high sequence weighted utilization section-

pattern). In step 2, the mobile transaction sequence

database (TSD) is transformed into a trimmed

database (TD) by mapping the WULIs to their new

identities. The section-items which are impossible to

be the elements of high-utility mobile sequential

AMulti-DataMiningApproachforShelfSpaceOptimization-ConsideringCustomerBehaviour

patterns are removed from the database. In step 3,

the trimmed database (TD) is utilized to find the

WUMSPs (high sequence weighted utilization

mobile sequential pattern) by the proposed level-

wised method. This step is the key to mining

performance and its procedure is shown in Figure 4.

In step 4, the WUMSPs are checked to find UMSPs

(high-utility mobile sequential patterns) by an

additional scan of the mobile transaction sequence

database (TSD). The WUMSPs whose utilities are

larger than or equal to ε are regarded as high-utility

mobile sequential patterns. All UMSPs are then

stored in high-utility mobile sequential pattern

database (SPD).

Input: All 1-WULPs, a trimmed database TD, a minimum

support threshold δ, and a minimum utility threshold ε

Output: WUMSPs

Join the 1-WULPs to form candidate 2-WULPs and then

store them into 2-candidate trees;

For each candidate 2-WULP X

Perform an additional scan of TD;

If sup(X

) ≥δ and SWU(X

) ≥ε

is a 2-WULP;

End If

Generate 2-WUMSPs by joining the 2-WULPs with their

corresponding paths in the 2-candidate trees;

k = 3;

While (candidate WULP is generated)

Generate candidate k-WULPs by combining the (k-

1)-WULPs of the two (k-1)-WUMSPs whose paths

are equal to each other;

Store the generated candidate k-WULPs into k-

candidate trees;

For each candidate k-WULP X

Perform an additional scan of TD;

If sup(X

) ≥δ and SWU(X

) ≥ε

is a k-WULP;

End If

Generate k-WUMSPs by joining the k-WULPs with

their corresponding paths in the k-candidate trees;

k = k + 1;

End While

Figure 4: Third step of the UMSP

algorithm.

Let’s take the following simple example to

explain the computation process of the UMSP

algorithm. Assume the minimum support threshold δ

is 2 and the minimum utility threshold ε is 100. In

addition, the utility table and trimmed database TD

is shown in Table 1 and Table 2, respectively. In TD,

the original mobile transaction sequences are parsed

into the sequences of section-itemsets and paths. For

instance, < S

; t

; 3> in customer CID 1’ means that

occurred in S

, and S

is in the third position of the

path. Note that if there is no item in the start or end

location of a path, the location in a path will be

trimmed.

Table 1: Utility table.

Item Profit ($ per unit) Item Profit ($ per unit)

20 i

5 i

6 i

15 i

Table 2: Transformed mobile transaction sequence

database TD.

CID Sequence of WULI

Path SU

1’

; t

; 1>, <S

; t

;

3>, <S

; t

; 7>

2’

; t

, t

; 1>,

; t

; 3>, <S

; t

;

5>, <S

; t

; 7>

153

3’

; t

; 1>, <S

; t

;

3>, <S

; t

; 5>,

; t

; 7>, <S

; t

;

160

4’

; t

; 1>, <S

; t

;

3>, <S

; t

; 7>,

; t

; 9>

5’

; t

, t

; 1>,

; t

; 3>, <S

; t

;

5>, <S

; t

; 7>

134

In the third step, the candidate 2-WULPs are

generated by joining the 1-WULPs in the mapping

table, and the result is stored into k-candidate trees

(k is the length of the patterns). Each k-candidate

tree stores the candidate k-WULPs whose last

section-itemsets are the same. After constructing 2-

candidate trees, an additional scan of TD is

performed to check the path support and SWU of

each candidate 2-WULP and to form the paths in the

moving patterns. After generating 2-WUMSPs,

candidate 3-WULPs are generated by combining the

2-WULPs of two 2-WUMSPs if the path of one 2-

WUMSP contains the path of another 2-WUMSP.

The processes will be recursively executed until no

further candidate moving pattern is generated. In this

example, 2-candidate tree and 4-candidate tree with

root of <S

; t

> are respectively shown in Figures

5(a) and 5(b). Figure 5(a) indicates five 2-WUMSPs

marked with solid lines are generated, while Figure

5(b) shows three 4-WUMSPs are obtained. In the

fourth step, after all WUMSPs are generated, an

additional scan of the database is performed to check

ICE-B2014-InternationalConferenceone-Business

for real high utility mobile sequential patterns. The

WUMSPs whose utilities are greater than or equal to

the minimum utility threshold are regarded as high

utility mobile sequential patterns. For example, five

2-UMSPs, seven 3-UMSPs, and three 4-UMSPs of

; t

> are found.

Figure 5: (a) 2-candidate tree of <S

; t

>; (b) 4-candidate

tree of <S

; t

2.4 Item Classification

In this study, an item in the store will be classified as

major item or minor item based on the following

definitions.

Definition 1: a major item is the item ever

appeared in high-utility mobile sequential patterns.

Major items are considered as important

commodities attracting customers to purchase. If

major items are reassigned to other

section(s)/shelves(s), the high-utility mobile

sequential patterns might be invalid since the major

attractions are not at the original place anymore.

Therefore, major items are considered as the items

not been rearranged. In the following discussion, the

set of major items is denoted as MA.

Definition 2: a minor item is the item allowed to

be reassigned. Minor items are considered as

affiliated commodities related to major items. Minor

items can be found according to the following rules.

First, all association rules in product association rule

database (ARD) are checked. If the item on the

consequent of an association rule is a major item, the

item on the antecedent of the association rule is a

candidate minor item. Next, if the candidate minor

item is not a major item, the item will be a minor

item and added into the set of minor items MI.

2.5 Product Reassignment

According to definitions 1 and 2, major items should

not be rearranged because frequent customer visiting

behaviours will not maintain if major items are not

displayed at their original positions. Therefore, only

minor items will be rearranged.

To increase the cross sale possibility of minor

items, minor items should be rearranged to the

location as close as possible to its related major

items according to previous customer visiting and

purchasing behaviors. Therefore, based on the

product association rule database and high-utility

mobile sequential pattern database, this study

develops an Item Location Preference Evaluation

(ILPE) procedure to calculate location preference if

a minor item is placed at a section in the store.

First, for each minor item mi

in MI, the

procedure scans product association rule database

(ARD) and retrieve all major items in the consequent

of a rule while the antecedent of the rule is mi

. The

set of major items related to mi

is denoted as GM

Then, for each major item ma

in GM

, the procedure

will scan high-utility mobile sequential pattern

database (SPD) and find out the set of high-utility

mobile sequential patterns containing ma

, which is

denoted as GP

. For each high-utility mobile

sequential pattern UMSP

in GP

, the procedure

will evaluate the movement distance that minor item

is assigned to section s

. Let

be the

movement distance in UMSP

if mi

is moved from

the section that major item ma

is located at to

section s

AMulti-DataMiningApproachforShelfSpaceOptimization-ConsideringCustomerBehaviour

If no relationship among minor item mi

, major

item ma

, section s

, and high-utility mobile

sequential pattern UMSP

can be found,

D is set

as β. Notes that β is the threshold of maximum

section movement and is provided by users.

D is close to 0, minor item mi

should have

high possibility to be rearranged to section s

Therefore, the standardization of assigning mi

to s

under the condition of high-utility mobile sequential

pattern UMSP

and major item ma

is defined as:

jn jn

km km



 (1)

where β is the threshold of maximum section

movement and

W.

The input of the procedure is high-utility mobile

sequential patterns database (SPD), products

association rule database (ARD), major item set

(MA) and minor item set (MI), while the output is

the item location preference matrix [f

j,y

Note that it is assumed that every product item in

this research has the same size so that two minor

items in different shelves can be exchanged directly.

After that, this paper will try to reassign products to

most suitable shelves based on information of matrix

j,y

]. The objective of product rearrangement is to

rearrange minor items and keeps the numbers of

section movement as less as possible. Hungarian

method (Kuhn, 1955) is adopted in this study. The

Hungarian method is a combinatorial optimization

algorithm that can solve the assignment problem.

3 IMPLEMENTATION AND

EXPERIMENTAL RESULTS

3.1 Data Description

A simplified supermarket as illustrated in Figure 6 is

used to demonstrate the feasibility of the proposed

shelf space allocation method. The supermarket is

divided into 37 sections (s

to s

) and 52 shelves (z

to z

) according the instruction mentioned in

Section 2.1. Customers enter the supermarket from

entrance s

and check out their purchase from

section s

or section s

. There are 119 product

items in this store in which an item belongs to one of

the 16 product classes.

However, the RFID system is not deployed in

this example store right now. Thus, a mobile

transaction sequence generator is developed to

simulate the shopping behaviors in the supermarket.

In this study, the total number of mobile transaction

sequences in the generator is set as 1,000. With the

mobile transaction sequences, the transaction of each

customer can be obtained.

Figure 6: Physical location of all products.

3.2 Experimental Results

Based on the transactions from the generator, 24

product association rules are generated using Apriori

algorithm when minimum support = 10% and

minimum confidence = 60%. Part of product

association rules is illustrated in Table 3. Next,

UMSP

algorithm is applied to generate high-utility

mobile sequential patterns based on the utility data

in Table 4. There are 16 high-utility mobile

sequential patterns generated when minimum

support count is 6 and the minimum utility is 150.

Part of the high-utility mobile sequential patterns is

shown in Table 5.

Table 3: Association rule (10%, 60%).

ID Rule ID Rule ID Rule

1 i

→i

… … 24 i

105

→i

Table 4: The utility table for items.

Item Profit Item Profit Item Profit

100 … … i

119

Table 5: High-utility mobile sequential pattern (minimum

support count=6).

PID Pattern

1 <{<S

; i

><S

; i

><S

; i

><S

; i

105

>}; S

, S

…

16 <{<S

; i

><S

; i

>}; S

, S

ICE-B2014-InternationalConferenceone-Business

After association rules and high-utility mobile

sequential patterns are generated, major items and

minor items can be found. In this simulation, 9

major items including i

, i

and

105

are found. In addition, 7 minor items including

, i

and i

107

are identified. With

major and minor items,



is set as 3, the location

preference weight matrix [f

j,y

] can be obtained

according to Equation (1).

The final stage of the proposed shelf space

allocation method is to reassign minor items to their

best shelf location using Hungarian method. Table 6

shows the reassignment result after taking

Hungarian method. We find that i

is strongly related

to major item i

, i

is strongly related to major item

, i

is strongly related to major item i

, and i

107

strongly related to major item i

. Thus, the four

minor items (i

, i

107

) are re-organized to the

location close to their major items (i

, i

Minor item i

does not change shelf location since

already located on the shelf very close to its

major item i

at the original layout. Minor item i

not assigned to the best shelf z

since the location

preference weights are calculated based on average

concept.

Table 6: Result of assignment.

Minor Item Original Shelf New Shelf

107

4 CONCLUSIONS

In retailing business, a well product-to-shelf

assignment strategy will affect customers’

purchasing decision and increase profit for a

retailing store. Thus, this research proposes a novel

method for product-to-shelf assignment taking both

frequent purchased product relationship and

shopping path knowledge into considerations. With

the proposed method, market managers can generate

a better products’ layout. Our method determines

major items and minor items before product

reassignment. Instead of reassigning all of items in

the store, this research reassigns minor items only.

As mentioned, this research trends to rearrange

products depend on information of product’s

relationship and utility, and customer’s shopping

path.

ACKNOWLEDGEMENTS

This work was partially supported by the National

Science Council, Taiwan, R.O.C. under No. NSC

102-2221-E-431-002

REFERENCES

Agrawal, R., Srikant R., 1994. Fast algorithms for mining

association rules’,

Proceedings of the 20th

International Conference on Very Large Data Bases

VLDB, 487-499, Santiago, Chile.

Chen, M.C., C.P. Lin, 2007. A data mining approach to

product assortment and shelf space allocation.

Expert

Systems with Applications

, 32 (4), 976-986.

Kuhn, H.W., 1955. The Hungarian Method for the

assignment problem.

Naval Research Logistics

Quarterly

, 2, 83-97.

Larson, J.S., Bradlow, E.T., Fader, P.S., 2005. An

exploratory look at supermarket shopping paths.

International Journal of Research in Marketing, 22

(4), 395-414.

Levy, M. and Weitz, B.A.,

1995. Retailing management,

Irwin, Chicago.

Liu, Y., Liao, W.K., Choudhary, A., 2005. A fast high

utility itemsets mining algorithm.

Proceedings of the

1st International Workshop on Utility-Based Data

Mining

, 90-99.

Shie, B.E., Hsiao, H.F., Tseng, V.S., 2012. Efficient

algorithms for discovering high utility user behavior

patterns in mobile commerce environments.

Knowledge Information Systems, 37 (2), 363-387.

Yang, M.H., 1999. An efficient algorithm to allocate shelf

space.

European Journal of Operational Research,

131(1), 107-118.

Yun, C.H. Chen, M.S., 2007. Mining mobile sequential

patterns in a mobile commerce environment.

IEEE

Transactions on Systems, Man and Cybernetics-Part

, 30 (2), 278-295.

AMulti-DataMiningApproachforShelfSpaceOptimization-ConsideringCustomerBehaviour