Can We Really Differentiate Between Generations? A Data Mining Approach

Generations

As early as 18th century, philosophers, and scientists, e.g. Comte and Quetelet, have started to discuss the linkage between the date of birth and development in the society. In the 1920s, the sociologist Karl Mannheim claimed that sharing same experiences contributes a frame of reference, a distinct consciousness that can be influential in people’s lives (Alwin & Mc Cammon, 2003; Joshi et al., 2011).

Based on this theory, generations are cohort-groups who have similar values, ideas, and attitudes with common experiences of living in the same timeframe. However, the idea of distinctive generations is, quite complex and even though there is generally a consensus on approximate time lines and generation names, their existence and impacts are not easily documented. Although sources disagree on the specific dates separating the generations from each other, the most frequently used time lines and generation names are as the Greatest Generation (born between 1901-1924), who experienced profound economic and social turmoil, World War I and eventually World War II; the Silent Generation (born between 1925-1945) who experienced the Great Depression and lived through the world wars; the Baby Boomers (born between 1946-1964) who lived the cold war and the human rights movement; the Generation X or Gen X (born between 1965-1979) who lived oil crisis in the world, the continuation of the Cold War; the Generation Y, Gen Y or sometimes called the millennials (born between 1980-2000) who lived during the Gulf War in Iraq, September 11, and globalization; Generation Z, or Gen Z (born between 2001-2010) who lived after terrorism, global recession, climate change; Generation alpha born after 2010 and expected to born until 2025 (Kelan, 2014; Lyons & Kuron, 2014).

Apart from intercultural differences, countries have roughly similar generations since majority have experienced the same historical events and still experiencing with the world as the World Wars, oil crises, global recessions, climate change and likewise. In Turkey, we can also include the country specific events and separate generations as generations who lived through last days of Ottoman Empire and World War II; generations who lived through the foundation and early days of The Turkish Republic; generation who lived through the multiparty system and later the military takeover; the generation who lived through the left-right conflict; the generation who lived during terrorism, and military coup period, lived both welfare and crisis; and last the generation who experienced better economic conditions and increase in conservatism and all these fall approximately the same time line as the above named generations (Adıgüzel et al., 2014). In today’s technology driven life, technological developments also started to separate cohorts from each other and we can easily name the baby boomers as radio, Gen X as television, Gen Y as computer and Internet, and Gen Z as smartphone generations. This categorization is also in line with the technological development in Turkey, hence these generation cohorts are relevant for Turkey as well.

Generations in the organizational context

When the generations are considered from organizational context, we consider generations as cohorts who share similar life styles, work values, and work life expectations. For example, the baby boomers are designated by their high sense of loyalty and commitment to their workplace for long years. The Gen Xs are labelled with high level of motivation and respect for authority. There was also a speedy increase in women’s labor force rates among Gen Xs (Dixon et al., 2013; Sparks, 2012). The Gen Ys have quite different values and expectations compared to generations preceding them. They are said to be more confident and autonomous, have greater mobility and less commitment, prone to technology and social networks (Alexander & Sysko, 2013; Dixon et al., 2013; Holt et al., 2012). They want to shape their lives; work-life balance is more important than ever, and they desire flexibility (Yüksekbilgili, 2013). Compared to Baby boomers, and Gen Xs, Gen Ys are more inclined towards entrepreneurship (Holt et al., 2012; Keleş, 2013; Yurtkoru & Elber Börü, 2019). It is known that since Gen Ys did not witness the times of economic crisis, because they have grown up with the Internet, interactive video games, and TV game shows as entertainment, they like to enjoy fun and win, they consume fast, they have high expectations (but they do not want to pay the price) this generation does not like to work (Alexander & Sysko, 2013; Kuyucu, 2014; Yüksekbilgili, 2013). The Gen Z’s are very new to the work life and there is not enough evidence to profile, however the early signs indicate some characteristics that are emerging as tech savvy, pampered, protected, risk averse and empowered (Queen, 2015).

Data mining classification models

Data mining applications are popular tools in understanding and solving business problems. However, they are not frequently used to analyze survey data in business research although they can be very helpful in revealing useful hidden information. Data mining can be best described as a process for exploring large amounts of data to uncover meaningful patterns and rules and making predictions for behaviors or outcomes (Linoff & Berry, 2011; Luan, 2002). It is the process of extracting valuable information and knowledge including trends, associations, changes, anomalies, and meaningful structures that are hidden in complex or large datasets (Han et al., 2012). Classification models are one of the most widespread techniques used in data mining.

Assigning a class label from a set of possible values to a set of variables using a classifier model is the purpose of classification. A learning algorithm is applied on a training dataset in order to build the classifier model. Before training phase, the class belonging of each record is evident for the training dataset. The classifier model built based on the training set is evaluated on an independent test dataset after this learning phase. If the classifier’s performance is high enough, it will be used for prediction purposes.

There exist numerous algorithms that can be used to build classifier models such as artificial neural networks (ANN), decision tree algorithms, support vector machines (SVM), logistic regression, and discriminant analysis (DA). As mentioned above ANN, SVM, and three different decision tree algorithms are used in this study.

Inspired from human brain, ANN is a system where information is exchanged between interconnected neurons. Using the given information, the connections between the neurons can be improved by modifying the varying weights in between, thus makes ANN capable of learning. Since ANN can be used even in the case with no relationship among classes and variables, it can be used in any complex pattern modelling. ANN can work for any classification problem and tolerate noisy data. However, it requires a long time to build an appropriate model with training and because of this it is often hard to interpret due to its hidden layer structures and nodes (Agaoglu, 2016; Han et al., 2012).

SVM is a regression and classification and technique which tries to locate a hyperplane in order to differentiate the classes by maximizing the margins and minimizing the classification error (Cortes & Vapnik, 1995).

Decision tree algorithms split the instances recursively into subgroups which are mutually exclusive. This process stops when it is impossible to get a further split with an improvement in statistical or impurity measures. Gini Index, Gain Ratio, and Information Gain (used in ID3, C5.0, CART respectively) are the most well-known impurity measures in decision trees (Breiman et al., 1984; Quinlan, 2004). Additionally; in order to improve the performance boosting, an ensemble learning method taking C5.0 as base classifiers is used (Breiman, 1998; Schapire, 1990). In boosting, a strong classifier can be built from weak classifiers that learn iteratively. On the other hand, the CHAID algorithm uses chi-square test for multiway splitting.

In order to evaluate classification models or classifiers, performance measures regarding the decision correctness of the classifiers are used. These performance measures are precision, specificity, recall, and accuracy, and so forth. If a binary classification task is assumed, values of the class can be either Positive (P) or Negative (N). There are four terms to be mentioned to understand these performance measures and confusion matrix. True Positive (TP) values are the actual positive (P) ones correctly identified by the classifier as positives, whereas False Positive (FP) values are the actual negative (N) ones incorrectly identified by the classifier as positives. In a similar fashion; True Negative (TN) values are the actual negative (N) ones correctly identified by the classifier as negatives, whereas False Negative (FN) values are the actual positive (P) ones incorrectly identified by the classifier as negatives (Agaoglu, 2016). These terminologies are also given in the confusion matrix of Table 01 .

The calculations of performance measures are given in Equation (1), (2), (3), and (4). Accuracy is calculated by dividing the number of correct predictions to the number of all predictions. Similarly, precision is calculated by dividing the number of correct predictions for positives to the number of predicted positives. Likewise, recall is calculated by dividing the number of correct predictions for positives to the number of actual positives. Lastly, specificity is the rate of correctly predicted negatives over actual negatives (Agaoglu, 2016).

$Accuracy=\frac{TP+TN}{P+N}$ (1)

$Precision=\frac{TP}{TP+FP}$ (2)

$Recall=\frac{TP}{P}$ (3)

$Specificity=\frac{TN}{N}$ (4)

Instrument

As briefly discussed above, literature on generations imply that attitudes toward work and work environment expectations differ among different generation employees. Yet, especially when we consider Gen Y employees their technology proneness and entrepreneurial intention should be taken into consideration. The attitudes toward work and work environment expectations are measured by Yurtkoru & Elber Börü’s Work Life Expectations & Attitudes scale (WLEA) (Elber Börü & Yurtkoru, 2016; Yurtkoru & Elber Börü, 2015). Since WLEA has considerable number of items measuring technology proneness, no additional scale is used to measure technology. But, to measure entrepreneurial intention Liñán and Chen’s (2009) 10 items scale is also included in the questionnaire. The items are measured on a five-point interval scale where “totally disagree” equals 1 and “totally agree” equals 5. All together the instrument consists of 110 items including entrepreneurial intention, which is used as the average of ten items.

Sample

Data for our study are collected from 1633 employees from Istanbul, Turkey. The sample consists of 737 (45.1%) employees born between 1965 to 1980, and belong to Gen X; and 896 (54.9%) employees born between 1980 to 1995 and belong to Gen Y. The demographics of the sample are given in Table 02 .

Prior to the data mining process, the dataset is randomly divided into train and test data. 80% (1297) of the dataset is used for training and the rest 20% (336) is used to test the classifiers developed. Generations of the respondents are taken as the class variable (target value). Class variables are like the dependent variables in the classical statistical methods. All the other variables are input variables, which will be used to predict the class variable.

ANN classifier

A feed-forward backpropagation ANN method was implemented as the build setting. Two hidden layers, the first with 30 and the second with 10 nodes, were used in the classifier topology. The final model, which is the neural network that has the maximum accuracy, is reached after the stopping criterion has been satisfied for all neural networks. As a result, classifier achieved 103 True Positive (TP) and 127 True Negative (TN) instances, which indicates that 230 records of all test dataset are correctly labeled. On the other hand; there are 50 False Negative (FN) and 56 False Positive (FP) instances, totally 106 records are incorrectly labeled (See Table 03 ).

SVM classifier

We used an RBF function as the kernel function in SVM classifier build settings. The confusion matrix, given in Table 04 , shows the distribution of predictions. In the result of SVM classifier, there are 93 TP and 128 TN instances, which indicates that 221 records of all test dataset are correctly classified. On the other hand; there are 60 FN and 55 FP instances, totally 115 records are misclassified.

C5.0 classifier

We implemented multiway splits using gain ratio as impurity measure in the build settings. Minimum instances per child was taken as two for the stopping rule. Additionally, boosting with 50 trials was applied. Results given in Table 05 , shows the distribution of predictions. As a result, there are 92 TP and 146 TN instances, totally 238 of all instances are correctly labeled. On the other hand, there are 61 FN and 37 FP instances, totally 98 records are incorrectly labeled.

CART classifier

In this classifier, Gini index impurity measure with binary splits was used. Minimum change for impurity was taken as 0.0001 for the stopping rule. Table 06 shows the result of predictions. In the result of CART classifier, there are 78 TP and 134 TN instances, totally 212 of all records are correctly classified. On the other hand, there are 75 FN and 49 FP instances, totally 124 records are misclassified.

CHAID classifier

The CHAID classifier uses chi square test for multiway splitting. Epsilon for convergence was set as 0.0001 for the stopping rule. As shown in Table 07 , there are 81 TP and 126 TN, totally 207 instances are correctly labeled whereas 72 FN and 57 FP, totally 129 instances are incorrectly labeled in the classification matrix.

Confusion matrices of all five classifiers are combined in Table 08 , which shows the differences between classifiers in terms TP, TN, FN, and FP. According to TP values, ANN is the best and CART is the worst; whereas C5.0 is the best and CHAID is the worst in terms of TN values. In a similar fashion, ANN is the best and CART is the worst in terms of FN values; whereas C5.0 is the best and CHAID is the worst in terms of FP values.

In summary, when we compare the performance of the different classifiers, we see that C5.0 outperformed other classifiers according to specificity, precision, and accuracy. Only in recall measure, ANN performed better than the other techniques (See Table 09 ).