A data mining approach for desire and intention to participate in virtual communities

ABSTRACT


INTRODUCTION
As virtual community concept emerged during time, new definitions of the term found place in the literature. Porter proposes a virtual community definition that, a virtual community is an aggregation of individuals or business partners who interact around a shared interest, where the interaction is at least partially supported and/or mediated by technology and guided by some protocols or norms [1]. Plant approaches the term from a similar perspective defining a virtual community as a collective group that come together either temporarily or permanently through an electronic medium to enable the interaction of entities, individuals or organizations in a common problem or interest space [2]. In addition to these, Rheingold defines a virtual community as social aggregations that emerge from the Internet when enough people carry on those public discussions long enough, with sufficient human feeling to form webs of personal relationships in cyberspace [3].
The purpose of this study is to investigate performances of some of the data mining approaches while understanding desire and intention to participate in virtual communities and the factors affecting it. For this purpose a model has been developed with the focus on desire and intention to participate to virtual communities and its antecedents. Later following the data gathering phase and pre-processing of the data several data mining approaches have been applied to the data. Some part of the data is used for training purposes whereas remaining is used for testing the model which has been formed following the literature review. Consequently in addition to the studies in the scientific body of knowledge a collaborative and contributive data mining approach is applied to understand desire and intention to participate in virtual communities.

RESEARCH METHOD
Data mining can be defined as the process of extracting hidden patterns from large chunks of data. In doing this knowledge discovery, prediction or forecasting can be in the focus of data mining. While knowledge discovery provides us explicit information about the characteristics of the data set predictive modeling provides predictions of future events. As stated by Simoudis, data mining is the process of extracting valid, previously unknown, comprehensible and actionable information from large databases and using it to make business decisions [4]. Data mining borrows approaches from several disciplines as statistics, mathematics or computer science in order to find useful patterns and knowledge from large data sets. As it is indicated in Shearer's crisp-dm model, a data mining process is composed of business understanding, data understanding, data preparation, model building, testing/evaluation and deployment processes. In the following sections some of the data mining approaches used in analyzing the data set will be introduced [5].

Data gathering and processing
As suggested in literature over 385 observations (425 in our sample in this study) has been found sufficient for the sample size values with an error of 5% and a confidence level of 95% (survey monkey site-sample size calculator). In literature used formula to calculate this has been n= t 2 x (p x q)/e 2 where n refers to sample size, p refers to proportion, percentage or presence of the study characteristics (in literature it is suggested that when we have no prior values for the proportions to be estimated, we can use p-and q-values as 50%.) q=1-p, e refers to margin of error; t = 1.96 (with 95% confidence level). Based on that, n = 1.962 x 0.5 x 0.5 / 0.052 sample size has been found 384. 16 and rounded to 385 [6][7].
Scales used in the study is given in detail. Positive anticipated emotions refer to the pre-factuals hypothesized to influence desires to perform a behavior which can be in the form of positive anticipated emotions or negative anticipated emotions and it's likely to expect its influence on virtual community participation and desire and intention to participate in virtual communities [8]. In the literature, it is pointed out that in general people are in a tendency to expect some return when they share their knowledge. As it is defined by Chiu et al., norm of reciprocity refers to knowledge exchanges that are mutual and perceived by the parties as fair and one of the important factors that leads to knowledge sharing behavior [9].
Perceived usefulness refers to the degree to which a person believes that using a particular system would enhance his or her performance [10,11]. As it is indicated by Porter, in the technology acceptance model, perceived usefulness and perceived ease of use are the beliefs that are presumed to influence attitudes toward new technology [12]. As it is pointed out in Fishbein and Ajzen's theory of reasoned action, attitudes are formed as a result of the beliefs about the outcomes of performing that act and expected outcomes. If the outcome of performing that behavior seems beneficial to the individual, he/she may participate in that particular behavior [13,14].
Early definitions of social comparison theory date back to 1954s that started with Festinger's social comparison theory. As stated in the literature according to social comparison theory, there is a drive within individuals to look to outside images in order to evaluate their own opinions and abilities in the sense that it mainly focuses on explaining and understanding tendencies of individuals in evaluating and comparing their own opinions and desires with others which may lead to an self enhancement in individuals' self images.
As it is pointed out in literature desires provide the motivation to decide in favor of acting as part of a virtual community. Therefore desire construct has been measured with the help of questions adapted from Dholokia's respective scale [15].
As it is defined by Dholokia, We-Intentions construct used in the model refers to the intentions to participate ingether as a group which is to be a function of both individual (i.e., attitudes, perceived behavioral control, positive, and negative anticipated emotions) and social determinants [15]. Desire and intention to participate in virtual communities refers to the merge of we-intention and desires of Dholakia where desires provide the motivation to decide in favor of acting as part of a virtual community and we intentions stand for the intentions to participate together as a group, to be a function of both individual (i.e., attitudes, perceived behavioral control, positive, and negative anticipated emotions) and social determinants (i.e., subjective norms, group norms, and social identity) [15]. Respective scales have been borrowed empirically from the studies as shown in Table 1.

Data mining methods
As part of the research conducted several data mining approaches have been applied to the data set. Data mining methods can be used more accurately with data preprocessing approaches [16]. Such as normalization of the data, discretization the conutinues data and etc. Brief descriptions of the methods that have been used as follow. Then the subsets are expanded in order of their average entropy starting with the smallest. The reason for this is that subsequent subsets will most likely not end up being expanded and the subset with low average entropy is more likely to result in a small sub tree and therefore produce a more general rule [19]. 3) OneR: OneR, generates a one-level decision tree that is expressed in the form of a set of rules that all test one particular attribute. OneR is a method that often comes up with quite good rules for characterizing the structure in data [20]. Pseudo code for 1R is as follow. For each attribute, For each value of that attribute, make a rule as follows: Count how often each class appears Find the most frequent class Make the rule assign that class to this attribute-value. Calculate the error rate of the rules. Choose the rules with the smallest error rates [20].

4) Multilayer Perceptron: A Multilayer Perceptron is a version of the original perceptron model proposed
by Rosenblatt in the 1950s and considered as a type of neural networks (Rosenblatt, 1958). A perceptron (artificial neuron) is a function of several input perceptrons which is formed as a combination of input weights to the hidden layer perceptrons. As stated by Ramchoun in literature multilayer perceptron has one or more hidden layers between its input and output layers, the neurons are organized in layers, the connections are always directed from input layers to output layers and the neurons in the same layer are not interconnected [21]. In this approach hidden layer is a function of the nodes in the previous layer, and the output nodes are a function of the nodes in the hidden layer. 5) Bayesian Network: There are no deterministic rules which allow to identify a subscriber as a risk indicator. Graphical models such as Bayesian networks supply a general framework for dealing with uncertainly in a probabilistic setting and thus are well suited to tackle the problem of prediction. Every graph of a Bayesian network codes a class of probability distributions. The nodes of that graph comply with the variables of the problem domain. Arrows between nodes denote allowed (causal) relations between the variables. These dependencies are quantified by conditional distributions for every node given its parents [22]. A Bayesian network B over a set of variables U is a network structure Bs, which is directed acyclic graph (DAG) over U and set of probability tables Bp={p(u|pa(u))|u Є U} where pa(u) is the set of parents of u in Bs. A Bayesian network represents probability distributions [23,24].

FINDINGS
Reliability of the constructs have been re-assessed and re-evaluated considering suggested lower limit of Cronbach's alpha in literature. As it is shown in Table 2 with the sample size of 425 it has been seen that all Cronbach alpha values for the respective constructs have a value of higher than .70, in other words all the constructs used in the research model are statistically reliable and can be regarded as reliable constructs of the research model [25]. From this reason, In this study, benchmarking of the algorithms of JRip, Part, OneR Method, Multilayer Perceptron, Bayesian Networks have been performed. In testing the research model with each of the data mining approaches 66 percent of the data has been used for the training whereas remaining part of the data set has been used for the testing of the model. . Precision values of JRip and OneR method could not been calculated since proportion of instances truly classified of a class divided by the total instances classified in that class have been calculated undefined in the confusion matrix. Among all the algorithms, multilayer perceptron had the most correct classification rate with 93.007 percent, a good true positive rate of 0.930 and a precision 0.921. Part method had a correct classification rate of 91.60 percent, true positive rate of 0.916 and a precision value of 0.923. Multilayer perceptron had the lowest RMSE with a value of 0.24. Comparison of data mining methods used can be seen in Table 3.

DISCUSSION AND CONCLUSION
In this study, we investigated the factors behind desire and intention to participate in virtual communities following an intensive literature review. This is later followed with the model formation and applying the data mining techniques as suggested in literature. In the analysis part of the study we examined Data mining can be defined as the process of extracting hidden patterns from large chunks of data. In doing this knowledge discovery, prediction or forecasting can be in the focus of data mining. Jrip, part, oner method, Multilayer Perceptron (Neural Networks), and Bayesian Networks have been chosen as the data mining techniques in order to examine desire and intention to participate in virtual communities for this purpose. Among them JRip is a rule learner alike in principle to the rule learner Ripper [17]. The part algorithm combines two common data mining strategies; the divide and conquer strategy for decision tree learning with the separate and conquer strategy for rule learning. Oner generates a one level decision tree that is expressed in the form of a set of rules that all test one particular attribute. A Multilayer Perceptron is a version of the original perceptron model proposed by Rosenblatt in the 1950s and considered as a type of neural networks [26]. A perceptron (artificial neuron) is a function of several input perceptrons which is formed as a combination of input weights to the hidden layer perceptrons which lead them to the output layer. Finaly graphical models such as bayesian networks supply a general framework for dealing with uncertainly in a probabilistic setting and thus are well suited to tackle the problem of prediction.
In this study, we have met our objectives of evaluating and investigating the performances of different data mining techniques for the data set that is being used to understand desire and intention to participate in virtual communities. In addition to the studies in the scientific body of knowledge a collaborative and contributive data mining approach is applied to understand desire and intention to participate in virtual communities. Based on the results, multilayer perceptron had the most correct classification rate with 93.007 percent, a good true positive rate of 0.930 and a precision 0.921. Part method had a correct classification rate of 91.60 percent, true positive rate of 0.916 and a precision value of 0.923. Multilayer perceptron had the lowest RMSE with a value of 0.24. Based on the high correct classification rate and low RMSE measure, multilayer perceptron (neural network) can be considered as an effective method and can be used in understanding desire and intention to participate in virtual communities and its antecedents.