Market Segmentation With Cluster Analysis Based On Video Streaming Data

1 Introduction The use of machine learning has for a long time been an important part in business areas like finance, telecom, marketing and retail and is continuously becoming more important in almost any business area [5]. One of the tasks that machine learning can be used to address is market segmentation, which has been one of the most researched marketing areas since its birth in the 50’s[35, Chapter 1]. Companies that have managed to adapt their offering to one or more segments of the markets have gained an advantage over the alternatives of mass marketing or mass producing [35]. As companies collect more data and more detailed data about customers, the advantages of targeting specific groups of homogeneous users compared to using a mass marketing approach will continue to grow. Market segmentation is the process of dividing customers into coherent subgroups. Segmenting customers can, for example, be used in product differentiation [9], segmentation, targeting, positioning (STP) approach [22] or most importantly in customer relationship management (CRM) [16]. CRM is an essential part of any customer centric business and can be defined as ”the strategy for building, managing and strengthening loyal and long-lasting customer relationships” [34]. Within CRM specifically, market segmentation plays an important role the customer identification step, where it can be used to divide the customer basis into smaller groups of similar customers [23]. Market segmentation is most commonly performed with cluster analysis [10] [26]. The most commonly used cluster analysis methods for markets segmentation have been K-means and Ward’s minimum variance method [10]. As historically K-means is such a common method with proven usability in almost any field, it is considered the standard approach when it comes to clustering. However, in the last couple of decades numerous methods for clustering have been attempted, such as support vector clustering [15], self organizing feature maps [[16] [19] [18] [1]], genetic clustering algorithms [33] and artificial neural networks (ANN) [14]. Market segmentation can be performed with different segmentation bases, which are a set of input variables that the segmentation is based on. These variables can be categorized as general or product specific [35]. General variables are variables that can be applied to any domain, e.g. demographic or lifestyle. Product specific variables on the other hand are more directed towards behavioral and domain specific variables such as purchase behavior or product usage. The choice of segmentation basis will have a large effect on which market segments are found. When relying on a particular set of variables as input for the market segmentation, the assumption is made that similarities in these variables will correspond to similarities in the response to marketing efforts or product usage. General variables are most commonly used as they are intuitive and easy to implement [33]. However, Tsai et al [33] point out the risks of relying on such variables. Firstly, the assumption that customers with similar demographics and lifestyles will have similar purchasing behaviour may not be true. Secondly, these variables may vary over time and thus be less useful. In their research the purchase behavior is used as input variables. The authors demonstrate that it gives more homogeneous purchasing patterns in the clusters than using demographic variables. In this thesis the segmentation is based solely on behavioural variables; the segmentation was performed on information about video streams started by users. To the best of my knowledge, market segmentation based on video streams has not been attempted before. One of the benefits by using only variables on customer’s behavior as a segmentation basis, is that resulting cluster can be found that are homogeneous in how the product is used, and thus easy to target. If, for example, a cluster that corresponds to watching a specific genre, the content of a marketing campaign could be adapted after this information. This is different from clustering based on demographic variables, where clusters could be found that are homogeneous with respect to 1

demographics but not homogeneous with respect to watched videos. To be able to distinct cluster specific preferences from interests in overall very popular videos, a normalizing metric called video importance score (VIS) was introduced. VIS for a video and cluster indicate the importance of the video for the cluster. The score is calculated by normalizing the popularity of a video within a cluster by the overall video popularity in the service. By sorting the videos after the VIS, the most important videos for a cluster can be found. 1.1 Research question The key research question to be addressed in this thesis is whether a useful market segmentation can be effectively performed on data of video streams started by users. In order to address it, usefulness criteria have to be first adopted. Wedel and Kamakura [35] summarized six criteria, that frequently have been put forward as indicators of effectiveness and profitability of marketing strategies. These are: identifiability, substantiality, accessibility, stability, responsiveness and actionability. The thesis proposes to reformulate and adopt these criteria to the young domain of video streaming. They can be interpreted as follows: understandability - it can be understood what kind of users that are in each cluster, so that the users can be targeted properly. homogeneity - the users of the same cluster should be similar. independence - the users from different clusters behave differently. stability - the clustering is stable over time. actionability - the ability to provide guidance for decisions. To capture the relevance of a market segmentation to the company, the usefulness criteria were created in the discussion with the company. Structured data of started video streams for over 38 000 users and over 2000 different videos were made available. This enables investigation if based exclusively on the information about what users have watched is enough to capture clusters that fulfill the above mentioned criteria and thus can also be used in marketing efforts. The use of such criteria also comes with the task of quantifying the results, which is challenging when there is no ground truth solution to the problem. The combined qualitative and quantitative approach was adopted to determine the level of criteria fulfillment. This approach helps in ensuring that the clusters can be interpreted and understood by any relevant stakeholder of a company. 1.2 Scope and objectives The research in this thesis is not aimed at method comparison so just a few selected unsupervised learning approaches are validated without particular emphasis on parameter tuning. The thesis is intended to investigate the possibility of clustering based on the customer behaviour of watching movies and series and identify a suitable evaluation framework. The main objectives of this work are: To use unsupervised learning in order to create a market segmentation based on solely started streams by users. To find a way to adopt the success criteria for market segmentation for the video streaming service domain, and evaluate the segmentation both quantitatively and qualitatively. 2

To introduce a metric to normalize overall video popularity, in order to be able to find specific video preferences within the clusters. To provide recommendation to the collaborating company on how to use cluster analysis for market segmentation. 3

2 Background 2.1 Market segmentation The concept of market segmentation was formally introduced by Smith [30]. He states that market segmentation ”consist of viewing a heterogeneous market (one characterized by divergent demand) as a number of smaller homogeneous markets in response to differing product preferences among important market segments”. With this way of looking at the video streaming market, the streaming users can be seen as a heterogeneous market, that can be separated into smaller homogeneous groups depending on their video preferences (demand preferences). There are no true labels for the segmentation of a set of users, which makes it difficult to evaluate the success of a segmentation. Wedel and Kamakura [35] have summarized 6 common criteria to determine the effectiveness and profitability of a segmentation: identifiability, meaning the ability to identify customers from each segment based on easily measurable variables. substantiality, which means that the segments should be large enough so that it is profitable for the company to target. How small groups can be and still be profitable to target depends highly on what kind of company is considered. accessibility is the ability for the company to reach the targeted segments. responsiveness means that each segment should respond uniquely to marketing efforts. stability - the segments should be stable and not change in behaviour or composition over time, at least for a time period long enough to be able to identify the segments and implement a segmented marketing strategy. actionability, which is the ability to provide guidance for decisions. Although these criteria generally are effective, not all of them are relevant for all segmentation problems. In the scope of this thesis, it is redundant to analyze identifiability and accessibility, since a modern company with a digital service like Viaplay, fulfils these criteria via database use. The responsiveness criterion is replaced with homogeneity and independence, since this can more precisely describe the segmentation success when using cluster analysis. The substantiality is important, but due to the fact that Viaplay has a large amount of users, the clusters will end up being substantial in all implementations. Segmentation basis is a set of variables used to perform market segmentation. Wedel and Kamakura [35] used the categorization of bases into general segmentation bases, that are independent of the product, service or circumstance and product-specific bases, which are related to the product, service and/or particular circumstances. They also used the categorization of variables as observable (measured directly) and unobservable (inferred). Figure 1, shows a matrix with examples of categorized types of variables categorized by Wedel and Kamakura [35]. The variables used in this thesis is observable and product-specific as we observe customers usage of videos in the streaming website. 4

Product-specific Observable Cultural, geographic, User status, usage demographic and frequency, store socioeconomic loyalty and patronage, variables situations Unobservable General Psychographics, Psycographics, values, benefits, perceptions, personality and elasticities, attributes, life-style preferences, intention Figure 1: Examples of variables for segmentation bases categorised as general/product-specific and observable/unobservable. 2.2 Cluster analysis In this thesis, the market segmentation is performed with cluster analysis. Cluster analysis is the grouping of a collection of objects into smaller subsets based on some similarity criteria [13]. The objects in the subsets, or clusters, should be more similar to each other than to objects in other clusters, with the ultimate goal of finding natural groupings in the data. The absence of known labels for the objects makes cluster analysis unsupervised learning, as opposed to supervised learning/classification where labels of each object are known. There exist a lot of different methods, and there are no general purpose clustering algorithms [17]. This thesis will mainly describe the centroid based method K-means and the density based method Density-based spatial clustering of applications with noise (DBSCAN). Other common types of clustering methods are hierarchical clustering (clustering into hierarchies of clusters), and distribution based clustering (that assumes Gaussian distributed data). Hierarchical clustering was attempted without any successful result and distribution based clustering was not tested within the scope of this thesis, since the data is not Gaussian distributed. 2.2.1 K-means A clustering method that has been around since the 50s and still is the most commonly used algorithm is the K-means algorithm [17]. K-means belong to the family of centroid based methods. Such methods use a centroid to define the clusters, and objects are placed in the cluster with the centroid closest to the objects. Figure 2 shows a generated example of resulting clusters as well as clusters centers when using K-means. The goal of K-means is to minimize the sum of squared distances between all points and their assigned cluster center. Let Xi , i 1, ., n be a set of d-dimensional data points, let Ck , k 1, ., K be a set of K clusters and let µk be cluster centers, or centroids, which is the mean of the points assigned to the cluster. Then the cost function is defined as: 5

K X X Xi µk 2 k 1 xi Ck The K-means algorithm to partition the users is: 1. Randomly assign all points Xi to clusters Ck , and compute cluster centers µk as the mean the assigned points. 2. Reassign points to the cluster with the closest cluster center. 3. Compute new cluster centers. 4. Repeat step 2 and 3 until convergence i.e. no points change cluster in step 2. Minimizing the cost function is an NP-hard problem [11], so only a local minimum is guaranteed. For this reason, the best solution is often selected by taking the partition with the lowest cost function value over several runs. Another problem with K-means is the selection of the number of clusters, K. One way of selecting K is to use some heuristic function like Davies-Bouldin index [8], Silhouette coefficient [29] or Rand measure [27]. However, the selection of K with the best heuristic values does not guarantee to give the most useful solution to a given problem. A common way is to instead let a domain expert select the number of K that gives the most meaningful solution [17]. The standard version of K-means if often modified or extended to fit a certain type of problem. Some common variations are K-medoids [25], where the center is an actual data point (median) instead of the mean, Fuzzy C-means [4] where points has a certain degree of belonging to several clusters and Spherical K-means [37], where the distance metric used is cosine distance and the centers are normalized. Since the data representation is high dimensional and sparse, it is appropriate to attempt using cosine distance (Spherical K-means) and not only the most commonly used Euclidean distance as dissimilarity metric [31]. When using cosine distance, the distance between two vectors Xi and Xj is then measured as 1 cos θ where θ is the angle between the two vectors Xi and Xj . This distance is maximized when the vectors are orthogonal, which means the vectors are completely independent. The main concept behind the usage of cosine metric is that the direction of the vector is more important than the magnitude. With Euclidean distance, the representations of most users may end up lying close to the origin and be considered similar although they might not have similar preferences in videos. 6

Figure 2: Randomly generated data clustered into 2 clusters with the K-means algorithm. Example taken from 1 . 2.2.2 DBSCAN Another common clustering method is DBSCAN [12]. DBSCAN is density based as opposed to K-means and K-medoids that are centroid based. It creates clusters by looking at dense areas where many points are close. If a point is closer (with some distance measure) to a point than a predefined threshold, the points are considered neighbors. In this way, larger groups of neighbors form, and if the number of neighbors passes a predefined minimum threshold, these neighbors define a cluster. A strength of DBSCAN compared to for example K-means is that the number of clusters do not have to be set. The only parameters that DBSCAN needs is the minimum neighbor threshold and the threshold for minimum points needed to form a cluster. Figure 3 shows an example of the result of a DBSCAN clustering. Cluster assignments and noise Noise Cluster #1 Cluster #2 3 2 1 0 -1 -2 -3 -4 -2 0 2 4 Figure 3: Generated data clustered with the DBSCAN algorithm. 1 http://se.mathworks.com/help/stats/kmeans.html, visited 2016-08-25. 7

2.3 Collaborating company Viaplay is a streaming service of Swedish media company MTG, with customers mainly in the Nordic countries such as Sweden, Norway, Finland and Denmark. The service provides streaming of movies, shows, series and various sport events. Viaplay offers two different main types of services: Viaplay and Viaplay Sport. The ordinary Viaplay service provides streaming of a broad range of series, shows and movies and it is from this part of the service data has been made available. Both the service and the customer base of Viaplay have changed over the last years, which means older data may not be relevant. For this reason, only data from the last year is used in the market segmentation. For a streaming service provider, the customers are quite active, so one year worth of started streams is expected to be sufficient to identify individual preferences. 2.4 Related work Cluster analysis for market segmentation has been researched and widely used in countless applications within many differente business areas and with many different clustering methods. Punj and Stewart [26] compiled a list of 20 examples of when cluster analysis had been used for market segmentation. Dolnicar [10] more recently did a study of how market segmentation with cluster analysis is generally conducted and what weaknesses earlier attempts commonly have. Of the 243 articles included in that study, the average sample size is 698 with a maximum of 20 000. The average number of variables is 11.5 with a maximum of 66. In this thesis, a data set with sample size of over 38 000 with over 1700 variables was used, which in comparison to earlier attempts is extremely high both in sample size and number of variables. Dolnicar citeresearchstats found that there is no correlation between the sample size and the number of variables used. This is problematic as much of the research has a small sample size with a high number of variables. The clustering algorithm will always give a result regardless of sample size, but the more variables that is used, the higher sample size is needed to find natural groupings. Furthermore, Dolnicar [10] found Ward’s minimum variance method and K-means to be the most commonly used clustering methods reported in the literature. There is no single algorithm that is fit for any problem, so it is up to the researcher to make sure that the chosen methods is suitable for the segmentation task. As measure of similarity, Euclidean distance was used in 96% of the cases. Euclidean distance is often adequate but has weaknesses when it comes to ordinal data. If Euclidean distance is used with categorical data, it is assumed equal distance for the intervals between categories, which is a doubtful assumption [10]. Strehl et al. [31] pointed out that cosine distance is more appropriate than Euclidean distance when the data is sparse. 2.4.1 Similar Market segmentation attempts Ozer [24] perfomed

Market segmentation can be performed with different segmentation bases, which are a set of input variables that the segmentation is based on. These variables can be categorized as general or product specific [35]. General variables are variables that can be applied to any domain, e.g. demographic or lifestyle.