Predicting customer purchase behavior within Telecom: How Artificial Intelligence can be collaborated into marketing efforts

IN THE FIELD OF TECHNOLOGY DEGREE PROJECT

INDUSTRIAL ENGINEERING AND MANAGEMENT AND THE MAIN FIELD OF STUDY

INDUSTRIAL MANAGEMENT, SECOND CYCLE, 30 CREDITS STOCKHOLM SWEDEN 2020,

Predicting customer

purchase behavior within Telecom

How Artificial Intelligence can be collaborated into marketing efforts

JOHN FORSLUND JESPER FAHLÉN

KTH ROYAL INSTITUTE OF TECHNOLOGY

SCHOOL OF INDUSTRIAL ENGINEERING AND MANAGEMENT

Predicting customer purchase behavior within Telecom

How Artificial Intelligence can be collaborated into marketing efforts

By

John Forslund Jesper Fahlén

Master of Science Thesis TRITA-ITM-EX 2020:356 KTH Industrial Engineering and Management

Industrial Management

SE-100 44 STOCKHOLM   

Förutspå köpbeteenden inom telekom

Hur Artificiell Intelligens kan användas i marknadsföringsaktiviteter

Av

John Forslund Jesper Fahlén

Examensarbete TRITA-ITM-EX 2020:356 KTH Industriell teknik och management

Industriell ekonomi och organisation

SE-100 44 STOCKHOLM   

ABSTRACT

This study aims to investigate the implementation of an AI model that predicts customer purchases, in the telecom industry. The thesis also outlines how such an AI model can assist decision-making in marketing strategies. It is concluded that designing the AI model by following a Recurrent Neural Network (RNN) architecture with a Long Short-Term Memory (LSTM) layer, allow for a successful implementation with satisfactory model performances. Stepwise instructions to construct such model is presented in the methodology section of the study. The RNN-LSTM model further serves as an assisting tool for marketers to assess how a consumer’s website behavior affect their purchase behavior over time, in a quantitative way - by observing what the authors refer to as the Customer Purchase Propensity Journey (CPPJ). The firm empirical basis of CPPJ, can help organizations improve their allocation of marketing resources, as well as benefit the organization’s online presence by allowing for personalization of the customer experience.

KEYWORDS

Recurrent Neural Networks, RNN, Long Short-Term Memory, LSTM, Clickstream Data, Telecom, Consumer Behavior, Customer Journey, Purchase Prediction.

SAMMANFATTNING

Denna studie undersöker implementeringen av en AI-modell som förutspår kunders köp, inom telekombranschen. Studien syftar även till att påvisa hur en sådan AI-modell kan understödja beslutsfattande i marknadsföringsstrategier. Genom att designa AI-modellen med en Recurrent Neural Network (RNN) arkitektur med ett Long Short-Term Memory (LSTM) lager, drar studien slutsatsen att en sådan design möjliggör en framgångsrik implementering med tillfredsställande modellprestation.

Instruktioner erhålls stegvis för att konstruera modellen i studiens metodikavsnitt.

RNN-LSTM-modellen kan med fördel användas som ett hjälpande verktyg till marknadsförare för att bedöma hur en kunds beteendemönster på en hemsida påverkar deras köpbeteende över tiden, på ett kvantitativt sätt - genom att observera det ramverk som författarna kallar för Kundköpbenägenhetsresan, på engelska Customer Purchase Propensity Journey (CPPJ). Den empiriska grunden av CPPJ kan hjälpa organisationer att förbättra allokeringen av marknadsföringsresurser, samt gynna deras digitala närvaro genom att möjliggöra mer relevant personalisering i kundupplevelsen.

NYCKELORD

Recurrent Neural Networks, RNN, Long Short-Term Memory, LSTM, Clickstream Data, Telekom,

TABLE OF CONTENTS

1 INTRODUCTION 11

1.1 BACKGROUND 12

1.1.1 THE BENEFITS OF CONSUMER INSIGHTS 12

1.1.2 DIGITAL TRANSFORMATION TREND 13

1.1.3 DIGITAL EFFORTS INITIATED IN TELECOM 14

1.1.4 AI IN MARKETING 15

1.1.5 NEURAL NETWORKS 15

1.1.6 CHARACTERISTIC PRODUCTS WITHIN TELECOM 17

1.2 PURPOSE 17

1.2.1 PROBLEM FORMULATION 17

1.2.2 RESEARCH QUESTION 19

1.2.3 DELIMITATIONS 19

2 THEORY AND LITERATURE REVIEW 20

2.1 CONSUMER BEHAVIOR & DECISION-MAKING 20

2.2 COGNITIVE PROCESSES 22

2.3 LOW AND HIGH INVOLVEMENT PURCHASING 23

2.4 DATA-DRIVEN MARKETING 24

2.4.1 CUSTOMER PROFILING 26

2.4.2 PROMOTION STRATEGIES & AD TARGETING 27

2.4.3 DEMAND FORECASTING 27

2.4.4 PRICING STRATEGY 28

2.5 AI MODELS FOR SEQUENTIAL DATA 28

2.6 STATE-OF-THE-ART AI MODELS FOR CLICKSTREAM DATA 29

3 METHODOLOGY 31

3.1 RESEARCH DESIGN 31

3.2 DATA 31

3.3 SOFTWARE TOOLS 33

3.4 PRE-PROCESSING AND LABELING THE DATA 33

3.5 MODEL DESIGN 35

3.6 TRAINING THE MODEL 37

3.7 VALIDATING THE MODEL 38

4 RESULTS 40

4.1 MODEL PERFORMANCES 40

4.1.1 BASELINE MODELS 41

4.1.2 RNN-LSTM-CURRENT 41

4.1.3 RNN-LSTM-UPCOMING 42

4.1.4 MODEL SUMMARY 43

5 DISCUSSION 47

5.1 MODEL VIABILITY 47

5.2 MODEL NOVELTY 48

5.3 A DATA-DRIVEN CONSUMER BEHAVIOR FRAMEWORK 48

5.3.1 OVERVIEW 48

5.3.2 LEARNINGS FROM MARKETING LITERATURE 49

5.3.3 CUSTOMER PURCHASE PROPENSITY JOURNEY 49

5.4 MANAGERIAL IMPLICATIONS 50

5.4.1 MANAGERIAL AND ORGANIZATIONAL BENEFITS 51

5.4.2 PREREQUISITES FOR IMPLEMENTATION 51

5.4.3 INDUSTRY APPLICABILITY 51

6 CONCLUSIONS 53

7 REFERENCES 54

ACKNOWLEDGEMENTS

The authors of this thesis would like to extend their gratitude towards persons who have contributed significantly towards the completion of this study. In particular:

Yury Kucheev for constantly providing invaluable feedback along the way. A better supervisor for this thesis could not have been found.

David Norell for believing in us and taking us under his wings at Telia. Through challenges David was always there to assist and guide us forward. Despite the troublesome time with the Corona pandemic, we always felt supported by David and the Telia organization.

Lastly​, the authors, as arrogant as it may sound, want to show appreciation to each other. Not only does this thesis mark the end of a comprehensive research project - it also marks the end of an intensive university period filled with highs and lows. One road to the Millennium - there is but one, my brother; tis founded on unselfish love, the joy you give another.

1 INTRODUCTION

The recent decade has seen a radical transition in physical shopping transforming into or expanding towards digital shopping (ITN, 2018), accessible by anyone from anywhere.

Consequently, consumer behavior patterns, which are deemed highly valuable by companies, have changed trajectory. Understanding the customer is vital for companies in that it helps them operate more efficiently in different scenarios, ranging from production to sales.

However, the process of gathering, analyzing and leveraging the information about your customers has not evolved at the same pace as the transition of shopping (Quinton & Simkin, 2016); observing a customer stopping by a shelf of t-shirts can easily be visually observed by a sales representative in order to guide the customer accordingly - in contrast, what measures should be taken, and approaches used, when a customer stops by the same t-shirt online? This lag in adapting to changing consumer behavior creates an opportunity for existing companies (Scott et al., 2016). Recent academic research indicates that Artificial Intelligence (AI) models could help companies achieve competitive advantages by delivering obscure and advanced insights about customers (D’arco et al., 2019; Bekavac & Praničević, 2015), resulting in improved operational performances (McKinsey, 2016).

This research concentrates on the Swedish telecom industry and its pertaining products. One important aspect to acknowledge is that telecom products such as broadband, mobile postpaid plans and phones are by nature more risky for the customer due to its lifecycle and price, and thus require a higher rate of involvement in comparison to purchasing fast-moving consumer goods (FCMG) such as t-shirts, drinks or toothpastes (Rossiter, Percy & Donovan, 1991;

Rossiter & Percy 1997). This nature of products within telecom shapes the industry’s customer journeys and consequently marketing strategies applied by companies.

The commissioner for this thesis project is the telecom organisation Telia - the leading market actor in Sweden (PTS, 2018). Telia has been collecting consumer behavior data on their website, Telia.se, since April 2019. Utilizing this data, the authors aim to investigate the application of an AI model, within the telecom industry, that predicts a visitor’s purchase propensity based on their online session. The thesis will uncover whether such an AI model could be a useful tool for decision makers in marketing strategies. Furthermore, the thesis will provide fruitful insights to the field of state-of-the-art AI models applied within the telecom

industry. Additionally, marketing professionals will gain an increased understanding of how AI techniques in general can be used together with traditional marketing practices.

This study is first of its kind for two reasons:

a) ​A cross-disciplinary approach​. Although our study applies a quantitative AI model, which is data-intensive, we focus our discussion on the business and marketing perspective of this application.

b) ​The telecom industry​. This study is the first in investigating the application of an AI model for predicting customer purchase behavior in the telecom industry and its complex products.

Such studies have mainly focused on the FCMG industry, such as fashion (Lang &

Rettenmeier, 2017).

The thesis aims to address two research questions. Namely, how can an AI model be implemented that successfully predicts customer purchase propensity in telecom; and how can such a model help marketers in their decision-making processes.

1.1 BACKGROUND

This segment initially presents background on consumer insights from the business perspective. This perspective is complemented by academic findings on consumer research and the integrated field of big data, AI and customer journey mapping. Brief information about AI is described, both in general and its marketing applications. Lastly, product characteristics within telecom is introduced.

1.1.1 THE BENEFITS OF CONSUMER INSIGHTS

Every Business-to-Consumer (B2C) company should have a profound interest for understanding what motivates the consumer, as well as how they shop and make decisions.

This is highlighted in an article from McKinsey (2016), where they explain that a successful capability in this regard, can result in the development of new products, services or markets and hence create a competitive edge towards competitors. The strategy has shown to be successful throughout multiple industries, as companies with largely divergent properties have managed to meet their customer needs better than their respective competitors by focusing on

consumer insights . The McKinsey (2016) article further emphasizes that generating¹ consumer insights is an antecedent for organic growth - a growth strategy deemed as most successful for company growth (Bain & Company, 2000). The link between consumer focus and company growth is in other words prominent.

While it is a strategic necessity to develop consumer insights, it remains a constant challenge to develop a better understanding for consumer behavior patterns. This is because market trends are changing fast and technological paradigm shifts are changing the ways consumers shop. The McKinsey (2016) article argues that a combination of insight tools is required to capture consumer insights, but that many companies tend to undermine or leave out some. In response, the article proposes research approaches in order to generate efficient consumer insights. One of these approaches is digitizing the recording of customers’ daily decisions and purchases; activities that in marketing terms are most commonly known as customer journey (Baines et al., 2016). A second approach in the article is for companies to use advanced analytics to capture detail insights. Findings suggest that companies investing in big data and advanced analytics tend to achieve a 10 percent increase in sales, up to 5 percent higher return on sales, and a margin increase of 1-2 percent (McKinsey, 2016). This illustrates the potential competitive advantages that can be achieved.

1.1.2 DIGITAL TRANSFORMATION TREND

Companies’ ignorance of digital investments for consumer insights is quite contradictory considering the customer oriented fashion of B2C companies. Literature on consumer behavior as a phenomenon emerged as early as in the 1960’s and remains to be an active field of research at current date (Barmola et al., 2010). The principle of knowing the customer’s behavior and preferences ought to be a central point in marketing strategies (Barmola et al., 2010). Considering recent trends in AI development, scholars state that companies may obtain new benefits to support marketing decision-making from the use of big data and advanced analytics (D’arco et al., 2019). The authors of this thesis found it interesting as to why businesses have not embraced and implemented the directions that McKinsey (2016) suggests, when the benefits are evident from both academic and non-academic perspectives.

1​The discovery of a fundamental consumer need that companies can use to create value for the customer and the

One reasonable explanation could be due to the fact that the relationship between big data, AI and customer journey mapping is in a primitive, young phase. D’arco et al (2019) present findings, from a literature review on the aforementioned field, which show that research in topics such as the adoption of big data and AI tools in the marketing area did not become prominent until 2015, where it has since increased in popularity. The same authors further suggest that more research on the field of big data, AI and customer journey mapping should be conducted, in order to get a better understanding of how they can collaborate. This, as they agree that its application is dependent on different factors such as the nature of the industry, marketing objective and technological capabilities.

1.1.3 DIGITAL EFFORTS INITIATED IN TELECOM

There are signs of initiatives taken in the telecom industry which are in line with McKinsey’s (2016) suggestions. At present, Telia has operations in place to reach and engage customers.

This is done through marketing efforts such as personalization and retargeting advertisements, i.e deliver individualized messages and product offerings, attempting to convert previous visitors into buyers. There are, however, room for improvements in this regard, especially with website personalization. In order to improve personalization techniques, gaining a better understanding of consumer intentions could enable new developments and ultimately improve sales conversion rates (Shi & Ghedira, 2016).

Another distinct indicator for measures taken within the industry towards a more updated and digitized consumer insight strategy, is the adoption of the new field in the interface between technology and marketing known as MarTech, which stands for marketing technology. One of the key trends of marketing in the recent decade is the growing importance of technology - digital touchpoints with the customers require digital solutions, which may be a reason why the number of active companies in the MarTech industry has grown from roughly 150 companies in 2011 to 7040 in 2019 according to the annually released MarTech Landscape Supergraphic (ChiefMartec, 2019). The emergence of MarTech has led to a natural rise in data-driven decision making as well as a rise in digital tools and platforms used by marketers, which in turn has enabled more advanced analytics such as AI to be integrated with the technology or even offered as an additional tool in the MarTech platforms (Salesforce, 2020;

MarTech Cube, 2020). In fact, the global AI marketing industry is expected to grow at a

compounded annual growth rate (CAGR) of roughly 27% until 2025 and will then be worth USD 37 billion (BMRC, 2018).

1.1.4 AI IN MARKETING

Artificial Intelligence (AI) was broadly defined by John McCarthy, who himself coined the term in 1965, as “The science and engineering of making intelligent machines, especially intelligent computer programs.” (McCarthy, 2007). The different subfields of AI is varying and plenty, however, the subfield which is of interest to this thesis is machine learning. Much as the name suggests, it concerns how to make a machine learn from data. There are different types of learning, but most relevant for this thesis is supervised learning, in which a model is trained on input data ​and the output data in order to learn how they map to each other, so it can be used to produce output data based only on input data. For instance, to predict lung cancer survival (Lynch et al., 2017), input data could be data points with the characteristics gender, age, tumor size and so on, and the second type of input data would be how long that patient survived the cancer. After an adequate amount of training, it could then be used to predict the survival prospects of a patient given only the first type of inputs. This of course assumes that there indeed exists a pattern between the chosen input and output data. Thus, the model learns this seemingly complex pattern through training, and can then apply its newly developed knowledge to predict outcomes. For simplicity, the term AI is used throughout the thesis, even though it often refers to (supervised) machine learning.

Predictive analytics that take advantage of AI techniques can be found in companies’

marketing efforts today for a broad range of use cases. Netflix uses it to power their recommender system, which leads to personalized recommendations (Walch, 2019). Google (2018) uses it to offer their advertising clients the service of responsive search ads that helps deliver personalized ads. Amazon (2020) uses AI to enable their staffless stores to keep track of which customers who are in the store and what they are buying. H&M uses the technique to predict market trends as well as to optimize their logistics of getting the correct amount of clothes to stores (Reuters, 2019).

1.1.5 NEURAL NETWORKS

There are different types of machine learning models, and they all have fundamentally different architectures and logics. In the recent decade, neural networks has become one of the

leading models due to the research advancements made in the subfield called deep learning (Parloff, 2016). Deep learning has the ability to find complex patterns by analyzing vast amounts of data, due to adding several layers on top of eachother, whereas most of the previous model architectures’ (including simple neural networks with one layer) learning capabilities were limited after a certain amount of data. Different types of neural network architectures have been defined, such as CNNs (Convolutional Neural Networks) which have excelled at image recognition, RNNs (Recurrent Neural Networks) which takes a sequence ​as input and therefore has excelled at text and speech recognition (known as Natural Language Processing; NLP) since words in a sentence can be seen as sequential.

Consumer behavior can be seen as a sequential process. Current literature within marketing present variations of behavioral models similar to Baines et al’s (2016) model of the consumer proposition acquisition journey (Scott et al., 2016). The models suggest that there is a linear or step-by-step behavior from the consumer. In online consumer behavior, the pattern is more tangible, as website analytical tools allow for tracking of the user’s interactions and URL history (Bekavac & Praničević, 2015). This information about a website visitor is called clickstream data. Regarding the selection of an AI model which suits the task of predicting consumer behavior based on clickstream data, RNN has been deemed as the most prominent.

This was found in research conducted with the fashion retailer Zalando, as they concluded that the model produces the most accurate predictions of consumer behavior while also requiring the least amount of preparation (Lang & Rettenmeier, 2017). Furthermore, RNNs provide a high level of explainability, which means that at each sequence step in the clickstream data, it is possible to retrieve the purchase propensity. Thus, it is possible to gain insights into where exactly in the customer journey the probability of a purchase is declining or increasing.

Other than using RNNs to predict purchase propensity, it can also be used to predict a visitor’s next URL on the website (Hidasi et al., 2015). This could have potential use-cases like choosing appropriate content dynamically on the website, such as banners or links in order to create a more personalized customer experience.

1.1.6 CHARACTERISTIC PRODUCTS WITHIN TELECOM

Products pertaining to the telecom industry are fairly different from FCMGs. Typical B2C product types for telecom companies involve electronic devices (such as smartphones and laptops) and associating accessories, mobile prepaid and postpaid plans, as well as internet service. Additionally, telecom products are often accompanied by service offerings.

Contrasting the telecom products against the aforementioned FCMGs, one can conclude that they invoke a complex buying behavior. Further background about human thought and decision processes are presented in the literature review segment of this study. The main intuition behind telecom products invoking a complex buying behavior for the customer, is the long lifetime cycle and relatively high prices of the products. Moreover, FCMGs include a one-time payment, whereas telecom products often involve a stream of payments. Using the Rossiter & Percy grid, presented as Figure 3 later in the literature review, one can conclude that the telecom products classify as high involvement (high price) and informational (rationale).

These differences in the nature of products and services entail that the market communication strategy for telecom products should be tailored differently compared to communication from FCMGs. As such, the telecom industry can not fully embrace the literature findings from past AI research on purchase prediction in the FCMG industry; there is a need for similar research to be conducted in settings which are adequately fitted to the telecom environment. Should such research be conducted, it would result in significant explanatory value to consumer behavior within telecom, which marketers can incorporate into marketing activities.

1.2 PURPOSE

1.2.1 PROBLEM FORMULATION

The shopping landscape is undergoing a transformation. Physical shopping is rapidly being overshadowed by the eruption of online shopping and digital stores. There is a race for companies to reach the desired state where marketing insights are in sync with the new market trends, in order to achieve competitive advantages. Business practitioners agree on the fact that firms need to modernize their consumer insight tools, due to the digital transformation era (McKinsey, 2016). More specifically, firms should 1) aim for digitizing their customer journey, as well as 2) utilize advanced analytics to capture detailed insights. Firms that are

successful in this consumer focus aspect, see lucrative returns thanks to a competitive advantage (McKinsey, 2016).

While scholars within the field of innovative marketing agree that a good customer journey mapping is a key source for competitive advantage, they also acknowledge the fact that the relationship between big data, AI and customer journey mapping is a relatively unexplored area, marking a prominent research gap (D’arco et al., 2019). Hence, the directions that McKinsey (2016) proposes has not yet been established on an academic level; applying advanced analytics on consumer big data includes a wide range of quantitative methods and their suitability depends on the nature of the industry, marketing objective and technological capabilities, to name a few. D’arco et al. (2019) suggest that more research should be carried out on the relationship between big data, AI and customer journey mapping, with emphasis on the technologies behind big data and AI.

Recent literature studies indicate that attempts have been made to reduce the information gap on big data, AI and customer journey mapping in the FMCG industry. One study applied a state-of-the-art AI model known as Recurrent Neural Network (RNN) which successfully predicted customer purchase propensity in the fashion industry (Lang & Rettenmeier, 2017).

Hence, their work produced significant value to the literature field D’arco et al. (2019) are referring to as well as function as a tool for decision making processes in marketing strategies.

With background to this, the authors of this thesis want to investigate the application of a similar AI model applied on the telecom industry - an industry that is different from the FMCG industry and where similar research is, to the authors’ knowledge, lacking to this date.

The thesis will contribute to reducing the aforementioned research gap within the relationship of big data, AI and customer journey mapping. Practical implications for the telecom industry will potentially lead to a more competitive customer journey management, with use cases such as dynamic personalization & ad targeting.

This thesis will investigate the application of an AI model, within the telecom industry, that predicts a visitor’s purchase propensity based on their online session and clickstream data.

The AI model is applied in a telecom setting based on data from the leading Swedish telecom

company, Telia. Furthermore, the thesis will uncover whether such an AI model could be a useful tool for decision makers in marketing strategies.

1.2.2 RESEARCH QUESTION

RQ 1 » How can an AI model be implemented that successfully predicts customer purchase propensity in telecom?

RQ 2 » How can such a model help marketers in their decision-making processes?

The thesis will address the questions in separate manners. The first research question is addressed in the methodology section, by applying techniques found in state-of-the-art AI literature. The second question is treated mainly in the discussion section, where the discourse is motivated by literature on marketing, AI and big data, as well as findings from the first research question.

1.2.3 DELIMITATIONS

Although this study focuses on the telecom industry, data is generated by and collected from merely one company. Moreover, the data is based on customer interactions from the company’s unique website(s) which is tailored to the organization’s specific brand and values.

As such, while findings could contribute with meaningful findings for the telecom industry, consideration should be taken for the one-sided data sample collection. However, given that the data comes from the leading Swedish telecom actor, as well as the telecom industry having homogenous products and services in general (Qayyum, 2017), the data collection is likely to represent the market behavior.

2 THEORY AND LITERATURE REVIEW

This segment will treat literature and theory pertaining to the marketing field to illustrate traditional marketing practices and underlying decision models, as well as consumer behavior decision-making processes to illustrate the cognitive activities consumers take to retrieve and act on information. Furthermore, literature on state-of-the-art AI models will be presented to show different applications and business use cases.

2.1 CONSUMER BEHAVIOR & DECISION-MAKING

Hawkins and Mothersbaugh (2010) define consumer behavior as “the study of individuals, groups, or organizations and the processes they use to select, secure, use, and dispose of products, services, experiences, or ideas to satisfy needs and the impacts that these processes have on the consumer and society consumers and the processes they undertake in their purchase of a product or service”. Furthermore, it is argued that all marketing decisions are based on assumptions and knowledge of consumer behavior. Organizations apply theories and information about consumer behavior on a regular basis, as knowledge about the field is essential for influencing consumers’ decisions about which products to purchase.

Barmola et al. (2010) claim that literature on consumer behavior as a phenomenon emerged in the 1960’s and has since significantly progressed through obscure and sporadic research.

Today, literature on the consumer behavior field constitutes a rich cumulative collection of distinct subfields relevant to the knowledge of consumer behavior (Barmola et al., 2010). For example, subfields range from firms’ processes of market segmentation to individual and situational characteristics that influence attitude change (Hawkins & Mothersbaugh, 2010).

Furthermore, consumer behavior is interdisciplinary and has been developed by scientists, philosophers and researchers from fields of psychology, sociology, social psychology, cultural anthropology, and economics (Barmola et al., 2010). Additionally, the broad and varied cumulative collection - that makes up consumer behavior - has been subject to many critical reviews, both from scholar and business perspectives.

Stankevich (2017) summarizes main trends, theories and gaps in the fields of consumer behavior and decision-making. It is stated that marketing has one goal - to reach consumers at

consumers’ decision-making, in turn, originates from a sequential 5-stage model named

‘consumer buying process’. The model, which is presented in Figure 1, explains the phases a consumer experiences when buying a product or service. (Stankevich, 2017; Kotler & Keller, 2012; Barmola et al., 2010).

Figure 1.​ Five-stage model of the consumer buying process. (Kotler & Keller, 2012).

Stankevich (2017) refers to the 5-stage model as the traditional model of consumer decision-making process, and argues that the model has served as basis for new sophisticated concepts to be developed, giving scholars and marketers a more modernized and clear view of understanding how touchpoints and decision processes interact with each other. Scott et al.

(2016) observes that a broadly similar model is variously referred to within the literature of consumer decision-making, of which he mentions the terms ‘customer journey’, ‘consumer buying process’, ‘consumer proposition acquisition process’, among others.

Baines et al. (2016) narrate the ‘consumer proposition acquisition process’, which is the mapping of each phase that a consumer goes through from initial motive development to the actual purchase and re-evaluation. This model is presented in Figure 2. Each of the phases represents a customer touchpoint between the company and the consumer. Flowing top-bottom, the model can be seen as a sequence; through each phase potential customers drop out from the acquisition process, resulting in only a fair share of initial consumers who actually ends up being converted into actual customers. Baines et al. (2016) emphasize that the challenge for companies is to minimize the dropout of potential customers between each phase in the model, thus improving the transition rate of potential customers to actual customers.

Figure 2.​ The consumer proposition acquisition process (Baines et al., 2016).

2.2 COGNITIVE PROCESSES

Theory within consumer behavior helps give an understanding as to why people consume in certain ways. The notion is that buying behavior patterns show a predictable mental design, and buying habits are developed as tendencies and become spontaneous over time. Thus, buying habits can to some extent be manipulated by psychological techniques present within behavior patterns (Hanson & Kysar, 1999; Kahneman, 2011). Barmola et al. (2010) reinforces the notion of cognitive mechanics’ involvement in, and implications on, the buying process.

In their article, authors articulate that consumer behavior entails exchanges between various social and psychological variables at play. Barmola et al. (2010) emphasize that the exchanges compromise dynamic interactions of cognition, behavior and environmental events.

Psychologist Kahneman (2011) demonstrates the two systems inside the human brain which control the form of thoughts. System I is fast and automatic, and does not require any effort when used. This system is governed by habits and experiences, resulting in immediate decision-making. System I is always active and governs most of our daily life decisions.

System II on the other hand is slower than I, and is often triggered when system I does not manage to comprehend complicated information. System II entails reasoning and analysis, requiring effort from the individual to perform deliberate type of thinking. System I can be resembled to intuition, and that decision-making purely based on intuition can be a good tool, but at the same time it can be a risky one. The reasoning behind this is that the intuitive

Furthermore, Kahneman (2011) emphasizes that the two systems compensate one another as they function in parallel to each other; we can not function without any of the systems. In everyday life, system I is most prominent as the majority of decisions taken are trivial.

However, individual preferences come into play as well, which in turn determines to what frequency individuals exercise each system; some groups prefer to make decisions based on system I as the primary tool, while others prefer to use system II.

2.3 LOW AND HIGH INVOLVEMENT PURCHASING

Further research that is likened to Kahneman’s (2011) systems is found in the notion of purchase involvement. Products purchased for the first time generally require more involvement than repeatedly purchased products (Boyd et al., 2002). Similar information about consumers’ different levels of involvement in purchase scenarios, stems from research looking at the decision-making process from a cognitive orientation. With the motivation that additional perspectives should be taken into account regarding how consumers acquire knowledge, and what experiences that they use, in their purchase decision-making - Belch G.

and Belch M. (2009) conducted a discussion about the differences between low- and high-involvement in decision-making. Belch G. and Belch M. (2009) examined different approaches to learning and their implications for advertising and promotion. They argue that the five-stage model of consumer decision-making process, views the consumer as a problem solver and information processor. The consumer therefore engages in a number of mental processes to evaluate various alternatives in order to determine the degree to which they might satisfy needs or purchase motives. Thus, in reality the authors argue that the consumer buying process can deviate from the theoretical model, depending on the amount of information the consumer is faced to process and evaluate.

The Rossiter & Percy grid is used as a framework to distinguish different types of products.

The model, which is presented in Figure 3, highlights levels of risk and motivation triggered by certain product types, and what associating cognitive systems are active in the consumer’s decision-making process. The grid has the shape of 2x2 and its vertical axis covers the consumer’s rate of involvement/risk; the determinant factor here is what perceived level of risk the individual consumers experiences, a higher risk (price) indicates a higher involvement. The horizontal axis in the model covers underlying motivation for the purchase;

is the purchase part of solving a problem (and thereby informational motivation) or is it

addressing an emotion (transformative motivation). Products can be categorized to fit into cells within the grid. The model can be a useful tool in helping marketers understand what market communication strategy to use, depending on their product(s) (Rossiter, Percy &

Donovan, 1991; Rossiter & Percy, 1997).

Figure 3. The Rossiter & Percy grid (Rossiter, Percy & Donovan, 1991; Rossiter & Percy, 1997).

2.4 DATA-DRIVEN MARKETING

Modern technology has turned consumers into constant generators of behavioral data that organizations can utilize by exercising data-driven marketing to achieve competitive advantages (Camilleri & Miah, 2017). Data-driven marketing allows organizations to provide customers with product offerings that they want before they even know it, based on predicted needs (Camilleri & Miah, 2017). Such marketing innovation captures market opportunities by generating customer demand and mapping it to the company’s customer value proposition.

However, in the midst of fast changing customer trends and companies’ eagerness and ability to match these, there is a risk of companies diffusing their core propositions or being driven into non-profitable or over-competitive markets. To mitigate these risks, the framework presented in Figure 4 was introduced by Camilleri and Miah (2017) which can be used to

objectives. Figure 4 represents a conceptual framework for consumer analytics strategy, illustrating how organizations should utilize customer data in order to evolve marketing activities and innovations. Thus, the study merely treats the strategical process of data-driven marketing and does not mention which practical tools should be present.

Figure 4. ​Proposed consumer analytics framework (Camilleri & Miah, 2017).

D’arco et al. (2019) carried out a literature review on the impact of big data and AI on marketing. Based on the fact that applications of AI and big data have increased in popularity recent years in several business environments - including marketing - the authors set out to systematize how these technologies should be leveraged strategically to plan the customer journey. The customer journey used in the literature review follows Lemon and Verhoef’s (2016) idea, which is a simplified version of Kotler and Keller’s (2012) consumer buying process model. Lemon and Verhoef’s (2016) model divides the five-stage model into three distinct phases: the prepurchase, purchase and postpurchase phase. Based on this model, authors mapped out strategic applications of AI and big data for each phase, which culminated in a theoretical framework for marketing managers to understand how such analytical tools

can increase the marketing performance and reduce the complexity of the purchase patterns and consumer activities (D’arco et al., 2019). In Figure 5, D’arco et al. (2019) present ten areas within their framework where application of big data and AI technologies can be used to support marketers in decision-making. These areas, marked along the customer journey, pertain to (1) customer profiling; (2) ad targeting; (3) promotion strategy; (4) client acquisition; (5) pricing strategy; (6) purchase history; (7) demand forecasting; (8) predictive analytics; (9) monitor consumer sentiments; (10) CRM activities. A selection of these areas are explored further in the next sections to highlight useful applications of AI in marketing.

Figure 5. ​Big Data and AI framework for the customer journey mapping (D’arco et al., 2019).

2.4.1 CUSTOMER PROFILING

Trusov et al. (2016) suggest that consumer behavioral profiles can be identified by processing a user’s online activity coming from web visits and display advertising. The data retrieval - and customer profiling - is enabled thanks to advancements in information and communication technology. Practitioners and researchers can aggregate and synthesize the vast amount of information that users, voluntarily and involuntarily, leave in most of their

approaches of customer profiling has been implemented by several companies through their specific apps, in which they track user data to help them create customer profiles and utilize data-driven decisions in individualized marketing efforts (Pousttchi & Hufenbach, 2014).

According to D’arco et al. (2019), employing customer profiling will help marketers understand their customers’ characteristics (demographics) and behavior (psychographics), by answering questions about who they are, what their interests are, and what they want.

2.4.2 PROMOTION STRATEGIES & AD TARGETING

Possessing information about customers’ characteristics and behavior is useful for promotion strategies such as sales promotion (Buhalis & Foerste, 2015). Miralles-Pechuan et al. (2018) highlight that machine learning techniques can be used to micro-target potential customers and expose them to advertising campaigns. In their suggested machine learning model, data from customer profiles makes for highly useful input in terms of configuring the target segments of a campaign, such as time, age, browser, operating system and device type. By selecting a specific audience for their campaign, marketers will boost the performance of their online publicity. This ultimately leads to an improved allocation of their resources, for instance in terms of advertising spend and ad targeting (Edelman, 2010; Miralles-Pechuan et al., 2018).

2.4.3 DEMAND FORECASTING

Monitoring and analyzing web traffic volume is a useful metric in predicting product and service demand, as well as future revenue and performance (Yang et al., 2014). A major contributing element to this fact, is explained by the introduction of mobile apps and the enormous amount of data from consumers that can be collected. As Trabucchi and colleagues (2017) suggest, the phenomenon of mobile apps represents a powerful new source of value.

They argue that companies using the information from these apps can utilize the information to improve their understanding of customers’ needs and wants, but also how they use products and services associated to the apps. The strengthened understanding enables companies to make advancements in product and service development, as well as predicting demand (Trabucchi et al., 2017). Furthermore, researchers propose that investigating online comments and questions from customers as well as online variables pertaining to products - such as free delivery and discount offerings, are additional information that is useful as input metrics for

developing models that are capable of predicting sales patterns of products online (Chong et al., 2016; Chong et al., 2017; Park et al., 2019).

2.4.4 PRICING STRATEGY

D’arco et al. (2019) claim that the retail industry is changing due to available automatic algorithms which are able to carry out dynamic pricing. This pricing strategy suggests that companies adjust the prices of products and services in real-time to match the current market demand. D’arco et al. (2019) go on to state that similar price calculations for human beings are highly resource-intensive due to complex computations. In contrast, AI models enable the same computation to be automated and performed quickly. As highlighted by Wirth (2018), AI is a powerful tool in determining whether a product is attractive to a person and also to determine whether a customer will buy a product at a certain price point. His research was inspired by investigating the potential of AI in marketing decisions regarding to the prominent marketing mix; product, price, place, promotion (Wirth, 2018). On the same theme, scholars suggest that big data analytics techniques are suitable for optimizing pricing strategies when factors such as seasonalities and trends affect product demand. In particular, big data analytics have been a useful technique in determining intra- and inter-price elasticities (Danaher et al., 2014).

2.5 AI MODELS FOR SEQUENTIAL DATA

Prediction based on sequential data was formerly deemed as a key problem in AI, especially in fields such as natural language processing (NLP) where sentences can be seen as sequential words with a wide range of different sizes (Mahoney, 1999).

Bengio et al. (2003) proposed that neural networks were a better alternative to solve statistical language modelling than the current NLP-specific models and methods at that time. This idea was applied and proven by several researchers around that time (Goodman, 2007; Schwenk et al., 2005). However, regular neural networks (known as feedforward neural networks) has the disadvantage of requiring a fixed size of the sequence, and furthermore could only successfully work on sequences of a length between 5 and 10 (Mikolov et al., 2010).

Recurrent neural networks (RNN) solves the aforementioned problem of requiring the input

input. Mikolov et al. (2010) concluded that an RNN language model performed better than the former state-of-the-art language models. A simplified RNN model is illustrated in Figure 6 below. One of the success factors of RNNs is attributed to the advanced technique known as backpropagation through time (Bodén, 2001). However, Sundermeyer et al. (2012), highlighted a potential drawback with the technique, referring to vanishing gradient descent, which in essence means that the longer the sequence, the more difficult it is to train properly.

Figure 6. ​A simplified overview of an RNN.

Long short-term memory (LSTM) neural network architecture was introduced by Hochreiter and Schmidhuber in 1997, which differed from regular RNNs by being able to not only learn short-term dependencies, but also long-term (hence the name). Hochreiter and Schmidhuber (1997) concluded that LSTM caused less vanishing gradient descent problems and was successful on very long sequences. Furthermore, Sundermeyer et al. (2012) implemented an LSTM language model which out-performed the previous state-of-the-art RNN language model.

2.6 STATE-OF-THE-ART AI MODELS FOR CLICKSTREAM DATA

Given the success of applying RNNs (with LSTM architecture) on sequential language data, it has subsequently been tested in other fields with similar type of sequential data, such as in e-commerce where consumer behavior data is represented in the form of clickstream data (Bekavac & Praničević, 2015). Bekavac and Praničević (2015) imply that clickstream data corresponds to the question ‘what happens on websites’ or ‘visitor behavior while browsing

the website’ and ‘how many conversions have been achieved on the website’. Wu et al.

(2015) showed that it was possible to predict consumer purchase behavior with RNNs based only on clickstream data, and that this was comparable to previous state-of-the-art models which required much heavier feature engineering than RNNs. Simultaneously, Hidasi et al.² (2015) showed that it was possible to build a product recommender system with RNNs based on clickstream data. Researchers from Zalando, the leading fashion e-commerce website in Europe, concluded that RNNs could predict a visitor’s purchase propensity within the upcoming 7 days better than their previous model could, which was based on a logistic regression model (Lang & Rettenmeier, 2017). Similar to Wu et al. (2015), the researchers at Zalando concluded that RNNs required significantly less feature engineering (Lang &

Rettenmeier, 2017). As a future area of research, Lang & Rettenmeier (2017) suggested that similar investigation should be done on product level to enable sophisticated recommendation products.

Zalando’s conclusions regarding the capabilities and suitability of RNNs applied to clickstream data were further strengthened by additional RNN articles reaching similar findings (Toth et al., 2017; Sheil et al., 2018; Santolaya, 2018; Tan et al., 2016), although these articles implemented different technical approaches. Most notably, Tan et al. (2016) implemented a technique widely used in NLP called word embeddings in order to represent different click events in the clickstream data. Instead of representing them as a one-hot encoding vector , which is the norm for neural networks, each click event becomes³ represented as a vector of floats. Word embeddings are beneficial for two reasons; 1) they may reduce the dimensionality of the input vector, and 2) they can help the RNN to identify patterns in the input data, thus increasing the learning capabilities overall (Tan et al., 2016;

Mikolov et al., 2013).

2Feature engineering refers to the process of creating features from raw data, often based on extensive domain knowledge.

3 METHODOLOGY

The methodology carried out in this report is mainly focused on answering the first research question, namely; how can an AI model be implemented that successfully predicts customer purchase propensity in telecom? First, there will be a presentation of the research design and data collection. This will be followed by a description of measures taken to prepare the data before it is ready to be used as input to the model. Software environments and tools used throughout the thesis will be introduced. Choices pertaining to model design, training and validation will further be listed with background to the literature review.

3.1 RESEARCH DESIGN

This thesis is conducted as cross-disciplinary research (CDR). Regarding the established research methods of CDR, they have been characterized as “fragmented” and highly context-sensitive (O’Rourke, 2017). Given this context-sensitivity, the research design is based on aspects from both AI research as well as marketing research. Methodology from the AI field is used to answer the first research question - namely, how can an AI model be implemented that successfully predicts purchase propensity in telecom - which is more technical in nature. In order to address the second research question - how can an AI model predicting customer purchase propensity in telecom help marketers in their decision-making processes - marketing literature is researched, where current decision frameworks are reviewed and considered together with the produced AI model. The contributions that the AI model can generate from a marketing perspective are evaluated and concluded.

3.2 DATA

The data collection of this thesis is two-fold. The data collection for the AI model, which is used to train and test the model on, is collected from Telia with the authors working as insiders. This data is discussed further in the following paragraphs. Secondly, literature review conducted on AI and marketing fields independently, as well as their integrated field, constitutes data collection in terms of theory.

The dataset in which the research aims to perform analysis on is already collected and brought to the authors of this study through the thesis commissioner, Telia. The company has been tracking visitors’ website activities since April 2019 and saved the information to an internal

database. Approximately 1.7 million data points are collected daily. The clickstream data is formatted so that each data point represents a customer’s click interaction with the website (and its subsequent domains). If, for example, one customer enters Telia.se and then clicks into the support section (Telia.se/support), two data points will appear in the dataset, along with a significant number of attributes - including the user’s cookie and domain visited. A small sample of the dataset is presented in Figure 7 below. The URL column has been highlighted in Figure 7, as it is planned to function as primary input to the AI model. The selection of URL as primary input variable will be discussed in more detail in the data preprocessing section of the thesis.

A low share of the total data points - and customer interactions - actually results in a purchase being made. This is evident from initial analyses of the dataset. Thus, the ratio between purchase and non-purchase interactions constitutes an imbalanced dataset, similar to the dataset used by Zalando in their research (Lang & Rettenmeier, 2017).

Additionally, it is important to point out that the dataset does not contain any data that can be directly identified to a customer. There is no record on customers’ names, personal security codes nor email addresses in the dataset - same goes for any other sensitive information.

Instead, the dataset is generated based on anonymized cookies. However, for ease of readability, further thesis sections will use the costumer term, but cookies are what we are really referring to.

Figure 7. ​Sample of the dataset.

3.3 SOFTWARE TOOLS

The software environment used throughout the analysis of the thesis, ranging from data preprocessing to the implementation of the AI model, is the programming language Python.

Leading libraries for handling data and performing numerical and statistical operations are used, such as NumPy and Pandas. Regarding AI model implementation, Google’s open AI library Tensorflow is used with its built-in API Keras, which allows for a more high-level and easy-to-use implementation of AI models.

3.4 PRE-PROCESSING AND LABELING THE DATA

The clickstream data from Telia contains information about which URL a user visits, as well as what the user clicks on and when the user makes a transaction. Most previous research has taken all of these different behavior information as inputs when performing similar purchase predictions (for instance Lang & Rettenmeier (2017)). However, the data from Telia contains several behavioral events that is generated automatically from their own systems, which has nothing to do with the actual user behavior. It is therefore decided that only the URLs shall function as input to the model.

Diverse pre-processing and filtering techniques are applied on the URLs, such as truncating them when deemed appropriate. The process of truncating URLs implies that their tails are stripped off in order to exclude irrelevant information. To illustrate, subsequent text following

“?” (i.e. GET-request parameters) and “/community/” (i.e. specific forum threads) in the URLs are removed.

A visitor’s sequence of URLs make up their online session. As a visitor enters Telia’s website, a unique session cookie is created. The session cookie expires after 30 minutes of inactivity from the visitor. Should the visitor come back after this inactive period, a new unique cookie is issued - adding a new session to the database. An analysis of the length of sequences is performed, in order to see if a maximum limit can be placed without skewing the sessions (consumers’ behavior) too much. Less than one percent of sessions that result in successful purchases, have a sequence length of over 50 URLs. This small group will likely cause dimensionality problems if input to the model, as a consequence of too large input vectors. Thus, all sequences are limited to 50 URLs.

Zalando researchers suggest that future research should look into purchase prediction on a product level (Lang & Rettenmeier, 2017). The authors of this thesis draw inspiration from Zalando researchers’ suggestions and will implement purchase prediction both on total transaction level and on a product category level. In the postpaid product category, we will additionally perform predictions on specific bundle offerings. Although some bundle offerings entail prominent brands, such as Apple and Samsung, the brand perspective is disregarded. Instead, focus is on the bundle offering as a product. Hereafter, they will all be referred to as product categories in the thesis.

Labeling the data decides what the model will try to learn, and ultimately predict. In order for the model to learn how to predict purchase propensity, the data that is used to train it needs to be labeled, which means that each URL is marked with either a purchase or non-purchase label. The specific label types (and product categories) are:

List of​ ​product categories.

1) All purchases

2) Broadband purchases 3) TV services purchases ⁴

4) Postpaid purchases (​mobilabonnemang​ in Swedish)

5) Postpaid purchases specifically from non-logged in customers 6) Postpaid with an Apple phone purchases

7) Postpaid with a Samsung phone purchases 8) Postpaid with a Huawei phone purchases 9) Postpaid with a Sony phone purchases

This thesis will test two types of predictions. Contrary to Lang & Rettenmeier’s (2017) approach, which is to predict purchases occurring in the upcoming 7 days (including in the current session), this thesis will use current session purchases and upcoming sessions purchases as two independent prediction types. In sum, the following types of predictions are applied:

List of prediction models.

1) Purchases in current session

2) Upcoming purchases within 14 days, excluding current session

The rationale behind using the variants of predictions as shown above, is that additional explainability is added by dividing Zalando’s “combined” model into two independent models. This provides the advantage of locating where the purchase propensity number is originating from, whether it is the current session or the upcoming sessions.

3.5 MODEL DESIGN

In order to implement an AI model that predicts purchase propensity, a number of decisions have to be made in terms of its architecture. This thesis will apply an RNN with LSTM architecture, which has been regarded as an appropriate model for clickstream data (Lang et al., 2017; Köhn & Lessmann, 2020). In addition, Telia’s dataset collection follows a highly imbalanced distribution, similar to the clickstream data Zalando researchers used for their study (Lang & Rettenmeier, 2017). This means that there are very few purchases in each web session compared to non-purchases . Additional key design choices for the model includes ⁵ selecting what layers to include in the model, and selecting the model hyperparameters . ⁶

Essentially, AI models perform mathematical operations and therefore require inputs to be in the form of numbers. In our case, the URLs need to be translated into numbers. Traditionally, the one-hot encoding technique is an approach to carry out the translation. However, given the large quantity of different URLs on the Telia website, this approach would require enormous computer power to perform. Instead, other sophisticated approaches are available to perform the translation. Tan et al. (2016) and Santolaya (2018) suggest to add a so-called embedding layer to the AI model. This layer is placed before the primary LSTM layer, and converts URLs to uniquely representative vectors containing decimal numbers. An added benefit of the embedding layer is the possibility of making the model learn which URLs are similar to each other in regards to their context of how visitors browse the website.

5 Several techniques were tested in order to mitigate the imbalance of the dataset, but ultimately provided no improvements to the results.

6Model hyperparameters refer to the parameters that are set before the model starts training on data, such as the size and learning rate of the layers. On the other hand, model parameters refers to the trainable weights inside the

As Lang & Rettenmeier (2017) pointed out, RNNs allow for a naturally higher degree of model explainability and interpretability than other AI models, which the field has demonstrated an urgent need for (Vellido et al., 2011; Ribeiro et al., 2016; Goodman et al., 2016). In order to see how the predicted purchase propensity changes over the course of the visitor session journey, the LSTM layer is set to make a prediction at every URL, instead of only the last one which is the default. The hyperparameter “return_sequences” in Keras ⁷ allows for this. From a business perspective, the authors of this thesis perceive this technical attribute of the LSTM layer as an opportunity to access and analyze touchpoints in the customer journey - as the visitor session and customer journey both entail company communication and associating customer reactions along their trajectories. At each URL the visitors enters, she is exposed to various company information (branding, product offerings, customer service, etc), which influences her decision-making process.

The rest of the model hyperparameters are set as follows; the embedding layer contains 64 units, the LSTM layer contains 256 units, the optimizer is the learning method Adam (Kingma

& Ba., 2015) and remaining values are set as the Keras default settings. These hyperparameters are chosen by looking at best practices in the field, previous research articles as well as testing different variations in small-scale tests. Due to the excessive amount of time it takes to perform a full hyperparameter optimization search (Claesen & De Moor, 2015), this operation will not be included in the scope of this thesis. Finally, the last layer is a dense layer, which converts the outputs from the LSTM layer into probabilities ranging from 0 to 1 for each product category prediction. As there are 9 product categories, this results in 9 units for the dense layer. The dense layer is, more specifically, a so-called time distributed dense layer, which is a requirement when setting the LSTM layer to predict at every URL. See Figure 8 for an overview of the model design.

Figure 8.​ A simplified overview of the RNN model for this thesis.

3.6 TRAINING THE MODEL

Whilst there exists clickstream data dating back to April 2019, the model is trained on data between January 1st 2020 - January 31st 2020. The data is divided into two parts that are referred to as training data and validation data, in order to validate the model with different samples than it is trained on. A uniform random subset of 90 percent is kept as training data, the other 10 percent as validation data. The relatively short period of time (January 1st 2020 - January 31st 2020) is mainly due to the expirable nature of URLs, as new ones are developed over time. Furthermore, monthly data may capture factors that affect product demand such as current market trends and seasonality fluctuations (Danaher et al., 2014). January is chosen due to the Swedish consumer habits at that time were deemed to be most up-to-date whilst still remaining rather unaffected by the COVID-19 pandemic - which is taking place as this thesis is being conducted.

Two models are trained, each with different labels (see Section 3.4 for more specifics). For the model that is predicting purchases within 14 days in the upcoming sessions, data between February 1st 2020 - February 14th 2020 are used to set the labels (i.e. for the look-ahead