بررسی مجدد عملکرد سیستم های اطلاعات کاربر: استفاده از داده کاوی برای شناسایی ویژگی هایی که منجر به بالاترین سطح از عملکرد کاربرست

As competitive pressures increase, managers try to realize every bit of productivity from people, business processes and new information technologies (IT). This leads one to ask, how can managers configure information systems to achieve higher levels of performance from end users? In this regard, managers continually seek advice on how to meet the promises and expectations of continued increases in productivity through the use of IT. However, results from research on how to achieve higher performance through the use of IT in organizations has been mixed. Consequently, it has been difficult for IS researchers to give managers any advice on investing in specific aspects of IS that would lead to the highest performance possible. We focus on this question in this research. We use a data mining approach to tease out information about specific characteristics of IS that managers can manipulate to achieve desired outcomes with regards to individual performance. Our findings offer both researchers and managers significant new knowledge that can make a difference to IT user performance research theory and the practice of user performance management. Further, our research method offers a novel approach to linking theory and practice in IS research, a problem that is of great concern to many IS researchers. The approach is generalized and can be implemented by academic or industry researchers who are interested in generating hypotheses from data for the purpose of theoretical or applied research.

Achieving higher performance in the use of IT in organizations is a continuing problem within information systems research. While firms have continued to invest in information technology (IT), realizing the promises and expectations that IT would provide increasing productivity gains has been difficult (Ward, 2002). As competitive pressures increase, managers of all types are looking to wring every possible bit of productivity out of their investments in IT. The fundamental question for every manager is how to get better performance out of end-users of IT applications. While the question of end-user performance has been much researched, and some answers have been provided, the situation is still unclear. Information systems researchers continue to have difficulty telling managers what they need to do to achieve the highest level of performance from end users of IT applications. As a practical matter, managers want to be able to identify the characteristics of an information system that can be managed to obtain the highest end-user performance. Although this question is implied in user performance studies, it has not be been investigated directly. The answer requires identification and analysis of relationships that may exist between systems characteristics and individual performance. Previous studies of user performance have not systematically examined this issue. In this paper, we pursue the question by trying to identify those properties of IS which tend to lead to the highest levels of individual end-user performance. We apply a data mining-based approach in this investigation that involves the use of decision trees (e.g. Samoilenko & Osei-Bryson, 2008). Our reason for using this method is that we wanted a formal approach for reasoning from the data to derive both hypotheses for future testing and actionable rules that managers can use. In this paper, we use decision tree (DT) analysis of questionnaire data to explore the impact of certain properties of IS on individual performance.

We have presented a DT-based approach to generating and doing preliminary verification of a model that describes the relationship between IS properties and end-user performance. This approach can also be used in conjunction with confirmatory analysis (e.g., regression) in a multi-stage data analysis process. It is general and can be implemented by any academic or industry researcher who is interested in generating hypotheses from data for theoretical or applied research. Many DM software packages (e.g., C5.0, SAS Enterprise Miner, IBM Intelligent Miner) provide facilities that make the generation of DTs a relatively easy task. Given this fact, the major decision to be made by the researcher is the determination of the target events (e.g., PERF01 is in the [5–7] interval) that are of interest. Once this decision has been made, many DM software applications provide convenient facilities for discretizing the target variable (e.g., PERF01) into two or more distinct events (e.g., PERF01 is in the [5–7] interval, PERF01 is in the [1–4] interval). Alternately, the discretization could be done outside the DM tool in other widely available software such as EXCEL. Once this has been done, the research can conveniently generate multiple DTs by varying the choice of splitting method (e.g., Gini, Entropy) and other parameters. It should be noted that our data mining-based approach could have used constructs instead of individual items. A construct’s score is based on a weighted linear combination of the scores of the relevant individual items, where the weights are the relevant factor loadings. Some decision tree construction algorithms allow splits to be based on weighted linear combinations of the individual input variables, and so could accommodate the use of constructs. We chose to use individual items rather than constructs primarily for the following reasons: (1) The individual items take their values from an ordinal Likert scale. Doing factor analysis involves the implicit (but not automatically valid) assumption that the ordinal Likert scale is an interval scale. Our data mining based approach does not require this assumption. (2) A construct’s score is not easily interpretable (because it is based on a weighted linear combination of the scores of the relevant individual items). Thus, rules that are based on constructs would not be easily interpretable, particularly for decision makers. One goal of this research project is to identify characteristics of information systems that could be used by managers to configure the system in a manner that would lead to high end-user performance. Therefore, rules that are based on constructs rather than individual variables would likely be less useful to such managers.