A Holistic Approach to Proximity Marketing

Dimitris Poulopoulos and Athina Kalampogia

QIVOS, Athens, Greece

Keywords: Proximity Marketing, Recommendation Systems.

Abstract: Proximity marketing, as a promotional technique, can benefit shopping centres and malls in terms of revenue,

as well as customer loyalty, by analysing the customers’ data and using their profiles to better address their

needs and target advertising and promotional campaigns. To this end, retailers exploit cellular technology to

send marketing messages to users’ mobile devices, that are near a specific area of one of its stores. With more

than six billion mobile phones in the hands of consumers today, every consumer with a smartphone is

potentially susceptible to a proximity marketing campaign.

1 INTRODUCTION

IKEA is the world's largest furniture retailer, with a

presence in over 45 countries. Every day, thousands

of customers roam the stores' exhibitions considering

new purchases for their homes or workplaces. In this

work, we present the description and design of the

proximity marketing solution implemented for the

Greek branch of the organisation.

IKEA has several ways of gathering and storing

data derived from its customers. Those data are later

analysed to extract knowledge, that will assist the

brand address the customers’ expectations. QIVOS

has developed a loyalty scheme for IKEA, where

many of its customers are registered. Through this

program, consumers are equipped with a loyalty

member plastic card and collect points upon their

purchases. By the time a customer collects a specific

amount of points, IKEA can offer specific discounts

to the customer’s receipt. However, through this

loyalty scheme, IKEA is not only offering discounts

to its customers but also gathering data about the

customers’ consumer behaviour, that is further

analysed and used for marketing reasons.

Currently, the loyalty program is reactive. In

CloudDBAppliance, the loyalty scheme is

transforming into a proactive, mobile-first loyalty

program. Moreover, QIVOS, in collaboration with

IKEA, installed beacon devices in different areas in

one of the brand’s stores to collect information in

real-time about the customers’ specific location.

To undertake this transformation, offer innovative

services to the customers, facilitate their purchasing

decisions and exploit the new sources of data in real-

time, QIVOS has invested in implementing a novel

incremental recommendation engine.

Chris Anderson, in his 2004 article entitled “The

Long Tail”, said that we are leaving the age of

information and entering the age of recommendation

(Anderson, 2006). Unless we have a way to filter the

information overload that we absorb every day and

retain only what is important to us, data reduce to

noise.

Moreover, we seem to diverge from the notion of

search and embrace that of discovery. The difference

is, that when searching, a user is actively looking for

something. Discovery suggests that something the

user did not know existed or even did not know how

to ask for it, finds him (O’Brien, 2006). A

recommender or recommendation system aims to

predict the behavioural patterns of the user, his

preferences or dislikes, and utterly provide

personalised recommendations on items the user

would be likely to interact.

When explicit feedback from the users is

available, the system tries to solve a surrogate

problem, where we view ratings as a proxy to

preference. The problem then is usually solved as a

regression problem; thus, the algorithm is trying to

predict a real-valued rating to complete the missing

values in the user-item matrix.

However, most of the times - like in our case - we

do not have access to explicit user feedback. In such

scenarios, the system works passively in the

background trying to collect significant features.

Those features can be used to implicitly discover user

Poulopoulos, D. and Kalampogia, A.

A Holistic Approach to Proximity Marketing.

DOI: 10.5220/0007861106290634

In Proceedings of the 9th International Conference on Cloud Computing and Services Science (CLOSER 2019), pages 629-634

ISBN: 978-989-758-365-0

629

habits, preferences and behaviour, with some level of

confidence. In the case of binary implicit feedback,

we simplify the user-item matrix to a Boolean matrix,

where "true" values correspond to positive user-item

interaction, and "false" values indicate no interaction.

In any case, implicit feedback has consequences to

both algorithmic design and evaluation measures.

Our approach covers the following challenging

requirements:

▪ Real-time: Recommendations should arrive in

real-time, in less than 30 seconds, while the

ideal goal is under 10 seconds. This

requirement derives from the fact that

customers that pass a specific location rarely

return to pick up a late recommended product.

Thus, the time requirement drives us to either

scale up complex solutions or turn to more

straightforward techniques.

▪ Incremental Learning: For a real-time

scenario, an ideal recommendation engine

updates its parameters incrementally and

adapts to the ever-changing needs and

behaviours of customers.

▪ Proximity: IKEA exhibitions define a precise

path that every customer follows. For example,

the customer passes through the living room

sector, then enters the bathroom region, and

finally, ends up exploring kitchen products.

Thus, customers should be able to quickly

match what they already have in their basket

with products from the same category.

▪ Up-selling: One of the requirements is to detect

related products to what a customer has already

in the basket, and recommend those that,

although a bit pricier, present an opportunity

due to some in-store special offer, or a stock

policy.

This work is structured as follows: We present a brief

description of the use-case in Section 2. Section 3

presents a bird's eye view of the system. Section 4

provides the requirements, and Section 5 concludes

this paper.

2 USE CASE DESCRIPTION

Shopping centres and malls can benefit (i.e. in terms

of revenue) from analysing its customers’ data, and

using the concept of proximity marketing to send

Proximity Marketing| What Is Proximity Marketing?"

Marketing Schools. N.p., n.d. Web. 20 Mar. 2019

http://www.marketing-schools.org/types-of-marketing/p

roximity-marketing.html.

marketing messages to users’ that are near a specific

area of interest. With more than six billion mobile

phones in the hands of consumers today

every

consumer with a smartphone is potentially

susceptible to a proximity marketing campaign.

2.1 Overview

IKEA is a major retailer, that participates in

CloudDBAppliance project as a partner, with the

interest of trying a cloud-based proximity marketing

solution, developed by QIVOS, a company

specialised in cloud customer solutions for large

enterprises, especially in the retail sector.

To this end, the proposed scenario takes place in

IKEA stores with the aim of analysing customers'

data deriving from multiple sources (e.g. Bluetooth

enabled devices), with an aim to offer personalised

content, encourage specific behaviours, enhance the

shopping experience, facilitate the purchase decision

and predict the needs of its customers. The goal is to

have a platform able to get insights from all IKEA

stores, segment the clients, and produce personalised

offers based on the information extracted globally

from all stores, in real-time.

To achieve this, a recommendation engine will

leverage the vast amounts of data, that will surge from

real-time sensors and historical customer profiles.

Moreover, to adapt to such a dynamic system, it is

essential to refer to the notion of incremental learning

and develop the engine accordingly.

2.2 Problem Statement

In our scenario, user preferences change frequently,

and new data continuously arrive in a real-time

manner. A recommender system should ideally adapt

to these changes as they happen, modifying its model

to always speak for the current status, while requiring

a single pass through the data. This is the idea of

incremental learning.

While most recommender systems utilise some

variation of collaborative filtering, they suffer, most

of the time, from scalability and efficiency problems,

as the computations needed to grow polynomially

with the number of users and items in a database.

To address this problem, researchers have

proposed several approximation methods; Breese et

al. (Breese, 1998) and Ungar et al. (Ungar, 1998)

employ Bayesian network and clustering approaches.

ADITCA 2019 - Special Session on Appliances for Data-Intensive and Time Critical Applications

630

In (Sarwar, 2000, Deerwester, 1990) Sarwar et al. and

Deerwester et al. perform a dimensionality reduction

technique for the user-item interaction matrix, by

applying folding in Singular Value Decomposition

(SVD). Other researchers also turned their attention

to data reduction by removing irrelevant and

redundant elements (Zeng, 2004, Yu, 2002), and

content boosted collaborative filtering methods,

where they score each item relevancy by partitioning

the item space according to categories

\cite{popescul2001probabilistic}. Finally, greedy

algorithms that randomly sample users, or discard

popular and unpopular items have also been

proposed.

Although such approximation methods improve

run-time performance, they do it at the expense of

accuracy. This is, for example, the case of clustering

based methods, and although different optimizations

have been proposed using several fine-grained

segments (Jung, 2001), the cost of computation

approaches this of classic collaborative filtering

approaches. Moreover, there are other disadvantages

that one might consider. For example, Bayesian

networks work fine in environments where user

behaviour changes slowly with respect to the time

required to build the model, but they are not practical

in environments where changes happen rapidly.

Considering all this, there seems to be a trade of

between recommendation quality and performance

efficiency. This is the problem that incremental

learning tries to alleviate, by composing highly

scalable algorithms, that have much faster run-times

with no accuracy degradation.

2.3 Data Acquisition

IKEA has several ways of gathering and storing data

from its customers. To begin with, in that specific

scenario, IKEA customers will be equipped through

their smartphones with a mobile application, in which

customers will take note of the purchases that they are

willing to make. Thus, this “wish-list” will be

digitised, allowing IKEA to know exactly what its

customers will probably buy, and what their itinerary

will be while in store.

What is more, most of IKEA customers are

registered to the IKEA loyalty program, through

which they are equipped with members’ plastic cards

and collect points upon their purchases. By the time a

customer collects a specific amount of points, these

points can be transformed into discount vouchers to

the customer’s receipt. However, through this loyalty

program, IKEA is not only offering discounts to its

customers but also gathering data about the

customers’ consumer behaviour, that is further

analysed and used for marketing reasons. Currently,

the loyalty program is reactive. In

CloudDBAppliance, the loyalty program will be

transformed into a proactive, mobile loyalty program,

and to enrich customers’ data, will leverage beacons

that will be installed in different areas in the IKEA

stores, to collect information in real-time about the

customers’ specific location.

2.4 Methodology

We have developed several ways of collecting and

managing customers’ data, to assist specially

designed mechanisms to process and analyse this

information in real-time, so as to predict customers’

needs and suggest additional purchases, offers and

coupons.

More specifically, by the time a customer with a

balanced consumer behaviour will add to her/his

shopping cart an item, real-time analytics will be

performed on:

▪ Data concerning the customer’s consumer

behaviour according to his/her previous

purchases

▪ Other customers’ consumer behaviour

according to what kind of similar items they

have purchased in combination with that

product

Thus, similar consumer patterns will be identified

and forecasted between customers with similar

behaviours, and suggestions will be provided to the

customers’ device, through predictive real-time

analytical mechanisms, about products that other

customers purchased along with the selected product.

Furthermore, through the installed beacon

devices, IKEA will track, in real-time, the

“geographic location” of all its customers. As a result,

real-time analytical algorithms will be executed

including data deriving from:

▪ A customer’s current location into the store

▪ A customer’s current shopping cart

▪ Past consumer behaviour of customers’ that

purchased same or related products

In that scenario, real-time predictions will be

made utilizing the aforementioned vast amounts of

data, in such a way that customers will be suggested

through their devices about items that

match/correspond to the items that they have already

added to their shopping cart, and are located just a few

steps away from them in the store.

A Holistic Approach to Proximity Marketing

631

3 SYSTEM OVERVIEW

Figure 1 depicts a high-level overview of the IKEA

proximity marketing system.

Figure 1: IKEA proximity marketing conceptual diagram.

Data are retrieved by the operational database and fed to the

incremental recommendation service. In real-time the

customer's location and shopping cart are fed as extra input

to the model and post-filtering service, to create an

incremental context-aware recommendation engine.

Section 3.1 presents the conceptual architecture

depicted in Figure 1 and the initial architecture of the

incremental recommendation service.

3.1 Conceptual Architecture

The process starts with the IKEA mobile app, which

sends back the beacon identifier as well as the

shopping cart for each update (insert, update, delete

item). This data, along with live transactions coming

from QIVOS Cloud (QC), feeds the Streaming

Engine.

The Streaming Engine, in turn, promotes live

transactions to the recommendation engine, which

uses them for incremental training. It also promotes

the user's location and shopping cart in the filtering

service to clean up irrelevant suggestions.

The incremental recommendation service, in

addition to live data, initially receives a batch of

historical training data, for bootstrapping, from the

operational database.

Finally, the request for recommendations comes

from the streaming engine, in the form of an alert,

whenever a user changes position in the store. Thus,

by deploying a continuous query on the beacon

streaming data, we can identify the position of the

customer at any point in time. Whenever we have

something of significance, for example, the customer

changes a showroom in the exhibition (e.g. moves

from kitchens to bedrooms), an alert is triggered, and

new recommendations are produced, concerning the

new context.

3.2 Recommendation Engine

Architecture

Initially, we feed the customer's historical data, stored

in the database, in the pre-processing unit. The unit

must shape them in a form that the recommendation

algorithm accepts. Moreover, it logically transforms

the data so that the algorithm can discover intricate

patterns behind customers' purchases and behaviour.

The training of the algorithm is done

incrementally, in an online manner. In real-time, we

also feed the customer's location and shopping cart

into the model. By the time a customer adds

something to the shopping cart or moves to a new in-

store location, the system produces candidate items.

The primary job of the algorithm is to create a list

of items, that might be of interest to the user and are

located nearby. This list is passed to a ranking

algorithm, that sorts these candidate items to create

personalised recommendations.

The algorithm is inspired by the work done in

language models (Mikolov, 2013, Collobert, 2008,

Bengio, 2003), as well as research done on

incremental collaborative filtering approaches

(Papagelis, 2005, Vinagre, 2014, Miranda, 2008) and

is summarised in Figure 2. Its job is to find the

correlations between the items that customers choose

together, as well as the reasoning behind these

purchases (e.g., same colour, brand, style, etc.). The

research that has explicitly done on the

recommendation algorithm is a subject of a future

publication.

Figure 2: Recommendation engine conceptual diagram.

Data are retrieved by the operational database and fed to the

data pre-processing unit. In real-time the customer's

location and shopping cart are fed as extra input to the

model, to create an incremental context-aware

recommendation engine.

ADITCA 2019 - Special Session on Appliances for Data-Intensive and Time Critical Applications

632

During training, we make use of offline metrics

like precision, recall, and ranking loss, but the real

value of a recommender is not only to predict the

held-out data in a test set but also to discover

additional items that might be of interest to the

customers. This is to balance the exploration-

exploitation trade-off and offer a diverse set of

recommendations, suggestion products that

customers are unaware of their existence.

Consequently, we can only draw a safe conclusion

using specifically designed A/B tests in production.

4 REQUIREMENTS

Most of the requirements of the system have to do

with the real-time nature of the business logic.

Ideally, recommendations should arrive in under

10 seconds, considering the huge traffic that could

occur in a store during rush hour. This requirement

derives from the fact that customers that pass a

specific location rarely return to pick up a late

recommended product. Thus, the time requirement

drives us to either scale up complex solutions or turn

to more straightforward techniques.

Our approach is to build an incremental

recommendation service, that after a bootstrap

training session, adapts dynamically to the ever-

changing customer needs and behaviours. The

training of the system happens in an online manner,

where each new customer-item interaction is

integrated directly into the model.

Moreover, IKEA exhibitions define a precise path

that every customer follows. For example, the

customer passes through the living room sector, then

enters the bathroom region, and finally, ends up

exploring kitchen products. Thus, customers should

be able to quickly match what they already have in

their basket with products from the same category.

This has also some implications in the post-

filtering service. This means that we need to consider

items that the customers have already bought, items

that are in the customers’ current shopping carts and

items that are out of stock.

Finally, there is up-selling. Thus, one of the

requirements is to detect related products to what a

customer has already in the basket, and recommend

those that, although a bit pricier, present an

opportunity due to some in-store special offer, or a

stock policy. This could drive the customer’s average

basket up and have a huge impact on the IKEA’s

revenue.

5 CONCLUSIONS

IKEA is the world's largest furniture retailer, with a

presence in over 45 countries. Every day, thousands

of customers roam the stores' exhibitions, consider

new purchases for their homes or workplaces.

In this work, we presented the description and

design of the proximity marketing solution

implemented for the Greek branch of the

organisation, which is one of the three use-cases of

the CloudDBAppliance project. The primary

objective is the realisation of a cloud-based proximity

marketing solution, developed by QIVOS, tailored to

large enterprises, especially in the retail sector. The

implementation depends on heterogenous data from

various sources, aggregated so as to be fed into an

incremental recommendation engine, that adapts in

real-time to the changing consumer behaviour. Future

work is aiming on parallelizing the algorithm, making

it more scalable, to address the needs of ever growing

datasets.

ACKNOWLEDGEMENTS

This work has received funding from the European

Union’s Horizon 2020 research and innovation

programme under grant agreement number 732051,

CloudDBAppliance project.

REFERENCES

Alexandrin Popescul, David M Pennock, and Steve

Lawrence, 2001. Probabilistic models for unified

collaborative and content-based recommendation in

sparse-data environments. In Proceedings of the

Seventeenth conference on Uncertainty in artificial

intelligence. Morgan Kaufmann Publishers Inc., pp.

437–444.

Badrul Sarwar et al., 2000. Application of dimensionality

reduction in recommender system-a case study. Tech.

rep. Minnesota Univ Minneapolis Dept. of Computer

Science, 2000.

Catarina Miranda and Alípio M Jorge, 2008. Incremental

collaborative filtering for binary ratings. In Proceedings

of the 2008 IEEE/WIC/ACM International Conference

on Web Intelligence and Intelligent Agent Technology-

Volume 01, pp. 389–392.

Chris Anderson, 2006, The long tail: Why the future of

business is selling more for less. Hyperion.

Chun Zeng et al., 2004. Similarity measure and instance

selection for collaborative filtering”. In International

Journal of Electronic Commerce 8.4, pp. 115–129.

Jeffrey M O’Brien, 2006. The race to create a

‘smart’google. In Fortune Magazine.

A Holistic Approach to Proximity Marketing

633

João Vinagre, Alípio Mário Jorge, and João Gama, 2014.

Fast incremental matrix factorization for

recommendation with positive-only feedback. In Inter-

national Conference on User Modeling, Adaptation,

and Personalization, pp. 459–470.

John S Breese, David Heckerman, and Carl Kadie, 1998.

Empirical analysis of predictive algorithms for

collaborative filtering. In Proceedings of the Fourteenth

conference on Uncertainty in artificial intelligence.

Morgan Kaufmann Publishers Inc., pp. 43–52.

Kai Yu et al., 2002. Instance selection techniques for

memory-based collaborative filtering. In Proceedings

of the 2002 SIAM International Conference on Data

Mining. SIAM, pp. 59–74.

Lyle H Ungar and Dean P Foster, 1998. Clustering methods

for collaborative filtering. In AAAI workshop on

recommendation systems. Vol. 1., pp. 114–129.

Manos Papagelis et al., 2005. Incremental collaborative

filtering for highly-scalable recommendation

algorithms. In International Symposium on

Methodologies for Intelligent Systems, pp. 553–561.

Ronan Collobert and Jason Weston, 2008. A unified

architecture for natural language processing: Deep

neural networks with multitask learning”. In

Proceedings of the 25th international conference on

Machine learning, pp. 160–167.

Scott Deerwester et al., 1990. Indexing by latent semantic

analysis. In Journal of the American society for

information science, pp. 391–407.

Sung-young Jung and Taek-Soo Kim, 2001. An

incremental similarity computation method in

agglomerative hierarchical clustering. In Journal of

Korean Institute of Intelligent Systems 11.7, pp. 579–

583.

Tomas Mikolov et al., 2013. Distributed representations of

words and phrases and their compositionality”. In

Advances in neural information processing systems, pp.

3111–3119.

Yoshua Bengio et al., 2003. A neural probabilistic language

model. In Journal of machine learning research, pp.

1137–1155.

ADITCA 2019 - Special Session on Appliances for Data-Intensive and Time Critical Applications

634