Image Construction Rules for Visualizing the Non-Visual Managerial Data Dr. Ping Zhang School of Information Studies Syracuse University, Syracuse, NY 13244 pzhang@syr.edu Data representations and their impact on decision making performance have been a subject of Information Systems research for a long time [Benbasat et al. 86, DeSanctis 84, DeSanctis & Jarvenpaa 89, Vessey 91]. The graphical representations in most studies in this area are common business charts, such as pie, line, bar, among others. These charts are not necessarily designed to support any specific decision tasks or processes. Thus the focus of these studies has been limited to simple decision tasks with small data volume and simple data relationships. To date, visualization of large-amount, multi-dimensional managerial data for decision-making support is nearly nonexistent. Most existing visualization tools, such as IBM Data Explorer, AVS, and IRIS Explorer, are designed for scientific visualization purposes and are very difficult to use for visualizing managerial data. Considering the manufacturing production planning problems at the Electronic Card Assembly and Test plant (ECAT) at IBM Austin, Texas, the presented research studied the characteristics of the managerial data, then developed special visualization techniques for constructing visual representations to support planners to develop superior production plans. A visualization prototype VIZ_planner was designed and implemented, and empirically evaluated [Zhang 95]. Hundreds of products, thousands of components, and many other factors can be visualized to provide planners with production planning insight. Manufacturing Production Planning Data In manufacturing production planning, a planner s goal is to maximize overall revenue from the production, subject to resource constraints such as tools and component availability. A typical manufacturer can have hundreds of different products and thousands of different components. Some of the components are used by different products and thus named common components.
Different production assembly lines (also called Production Pull Lines, or PPLs) share tool capacity and components during the production. A production plan crosses several weeks and is done every month. The decision-making environment is usually very dynamic. For instance, there often is a severe component shortage problem. Among all the short components, however, only a very small portion of components is critical or bottleneck components. During the planning period, the planner has many possible actions to take in order to solve some of the problems or sub-problems. For instance, s/he can move products to different assembly lines, move products to different time periods, adjust an assembly line s capacity, change the quantity of products to be made in each week (demand), change the quantity of components available (scheduled receipts), and change the distribution of components over products (production mix). In such a dynamic environment, the planner s understanding of the planning problem situation is crucial. Due to complicated relationships among data and the dynamic decision-making environment, existing production planning systems, such as MRP II [Vollmann et al. 88] and simulation systems, cannot provide solutions of production plans that reflect the changing constrains and environment. Although some of these systems have friendly user interfaces and simple one- or two-dimensional graphs (line charts, bar charts, pie charts, etc.), the data they can provide are basically in tabular format: they are either limited to the size that a computer screen can handle, or on a printed report that may be several hundred pages long. These data could not give planners a clear vision of what is going on from an overall perspective. In other words, they show trees but not the forest. The planners need to use their cognitive powers to figure out the stories behind the data, including identifying the critical components from all shortfall components. This presents an obvious opportunity for enhancement of the planning process through computer generated visualizations [DeFanti et al. 89, McCormick et al. 87] that shift some of the planners mental load to their perceptual system. On the other hand, the production planning data are large in volume, multidimensional, and have complicated relationships among them (not linear, hierarchical, network, nor geographical oriented). These data are not geometric by nature. There is no obvious geometry that can be assigned to each data object involved, nor obvious geometry indicating the relationships among the data objects. This provides a great challenge for the information visualization.
In our approach of visualizing the non-visual, we designed visual abstracts for each data object involved in each of the decision-making processes. Then we link the data objects as images using the rules we developed for this specific problem domain. Visual Abstracts A visual abstract basically provides a geometry for a data object for being presented in a 2-D Euclid plane. In our study, visual abstracts were intended to provide a foundation for (1) supporting visual perceptions, and (2) constructing efficient visual representations that facilitate all levels of readings. According to Bertin [1983], there are four human visual perceptions that correspond to three levels of information organization. The qualitative level of information organization includes all the concepts of simple differentiation. It involves two human visual perceptions or perceptual approaches: association and differentiation. The next level is ordered level that involves all the concepts that permit a ranking of the elements in a universally acknowledged manner. This level corresponds to the ordered perception. The third level is the quantitative level where one makes use of a countable unit. It relates to the quantitative perception. Bertin s three visual perceptions can be considered as three types of comparisons: compare to associate or differentiate, compare for more than or less than, and compare quantitatively. Another consideration in this research was that the final visual representations should not be too complicated for business managers to understand and use. Thus, we extended the traditional bar charts and used them as our theme for the entire visualization system. Most of the values were represented by bars in a broader sense (such as areas or lines) because bars are good for comparisons [Croxton 32] and are well understood by people, including business managers. Image Construction
The geometric structures indicating relationships among data objects need to be carefully designed so that they are meaningful to the planners and representable on a computer screen. An image is composed of elemental symbols (visual abstracts in our study) and their relations. The layout of multiple data objects in one image is determined by the dependency relationships among the data objects. We say data object A is dependent on B, or B determines A (fully or partially), if A is a function of B: A = F(B). In this case we say there is a dependency relationship between A and B. In a virtual multi-dimensional space, each data object has its own axis just as in an ordinary one-dimensional space. The following rules apply to data objects to make them geometrically connected. Figure 1 partially depicts these rules. Rule 1. (Dependency Dimensions) If A is determined or partially determined by B, then A and B construct a 2D plane by sharing the same origin. For example, demand satisfaction (A) is partially determined by component availability (B). We name A and B dependency dimensions. Rule 2. (Time-Space Dimensions) If A and B are time series data and one dimensional location data, and they both determine other data objects at the same time, then A and B construct a 2D plane by sharing the same origin. For example, it is meaningful to refer to the capacity availability (C) for assembly line 2 (B) at planning week 3 (A). In this construction rule, we say A and B construct time-space dimensions. Rule 3. (Parallel Dimensions) If A and B have no dependency relationship, but they both partially determine C, then A and B could be in a parallel position sharing the same origin. For instance, component availability (A) and capacity availability (B) have no dependency relationship between them, but they both determine the demand satisfaction (C). We name A and B parallel dimensions. Parallel dimensions can be more than two, as long as the involved variables are independent of each other and they partially determine other data objects. Rule 4. (Overlap Dimensions) If A and B have a dependency relationship and they are both determined by either time-space dimensions or another data object C, then A and B can share the same axis (dimension) by overlapping each other. In this way, we say that A and B are overlap dimensions. This is a special case of dependency dimensions.
Rule 5. All elementary graphing techniques and rules, when not conflicting with the above rules, apply to up to 3 data objects. Rule 6. Combinations of the above rules may be used to geometrically connect all the data objects involved in one image. Figure 1. Image Construction Rules A C B A (Time) B (Location) Rule 1: Dependency Dimensions Rule 2: Time-Space Dimensions C A B A B C (Time) D (Location) Rule 3: Parallel Dimensions Rule 4: Overlap Dimensions Images for Production Planning 1 Figures 2, 3, 4, and 5 depict a decision-making process that involves 110 products, 1861 total components with 145 common components, 12 planning weeks, six assembly lines (production pull lines PPLs), and two production constraints: tool capacity and components. At a global level, there are two images that complement each other: Global Satisfaction & Potential, which shows the satisfactory side of the planning problem, and Global Shortfall, which indicates the shortfall side. Figure 2 shows demand shortfall based on capacity and component shortfalls. It involves five data objects. Time and PPLs determine other three data objects and 1 Interested readers can contact the author for color images.
thus construct time-space dimensions according to Rule 2. Demand Shortfall is partially determined by Capacity Shortfall, as well as by Component Shortfall (Rule 1). Meanwhile, Capacity Shortfall and Component Shortfall do not depend on each other but they both partially determine Demand Shortfall (Rule 3). The images are resizable in both horizontal and vertical directions and the angle for the entire image changes correspondingly to ensure clear readings. The longer the bar, the higher the value, which is converted as percentage information to ensure that all data values can be represented, as well as to allow planners to make relative comparisons. Figure 3 lists all the products in terms of their production satisfaction (orange lines) in the context of demand satisfaction (green bars). Demand Satisfaction is dependent on Product Satisfaction for each PPL at each planning week. This is the situation of Rule 4 where Overlap Dimensions can be applied. In Figure 4, a detailed image of Product Satisfaction for a specific PPL is zoomed in from Figure 3. Each product is identified by their identification numbers and can be examined clearly. The purple color means there is no demand for that product in that week. All the required components have to be in sets in order to produce one product. If a specific product is of the planner s interest, Figure 5 can provide a detailed view of which component of this product is most shortfall and thus affects the production of this product. The underlined components at right of the image are common components and are used by multiple products. This image indicates to the planner that they should resolve the components with the shortest bars before they put any effort on any other shortfall components. Preliminary Empirical Evaluation After two real world planners from two different manufacturers tested the VIZ_planner, a formal lab experimental study was conducted to preliminarily evaluate the effectiveness of visualization on production planning performance. A total of 13 motivated graduate students with real world or
classroom production planning experience participated in the study. In order to pursue the experiment within affordable mental effort by the subjects, the production planning tasks were simplified considerably. Two randomly assigned groups worked on the same decision problems using two different computer systems respectively: one was a traditional MRP II type system (Norm_planner) with tabular format for data displays, and the other was the simplified version of VIZ_planner called Viz_planner. Within a limited time period, the subjects did what-if analysis by manipulating some of the planning raw data. The results of the experiment indicated that: (1) the Viz group generated more alternatives for solutions than the Norm group did; (2) the Viz group had a higher increased amount of revenue per change in raw data than the Norm group did; (3) the Viz group was more satisfied with the outcomes than the Norm group was. The Viz group had 15 minutes watching a demonstration of Viz_planner, and 40 minutes for practice in getting familiar with it after they knew how to use Norm_planner. This implies that the images for manufacturing production planning are very easy to understand and use. Although they were told before the experiment that they did not have to use the images if they did not like to, 83% of Viz members used the images intensively during their problem-solving processes. Conclusion The current research tested the feasibility of visualizing large amount of managerial data for decision-making support. The current result is very promising. Although a specific management domain is selected for the study, the goal of the research is to discover visualization rules that can be applied in many management domains. Future research includes refining and applying the visualization techniques to other managerial problems and conducting more extensive empirical evaluations. References
Benbasat, Izak, Albert S. Dexter, and Peter Todd, An Experimental Program Investigating Color- Enhanced and Graphics Information Presentation: An Integration of the Findings, CACM, Vol. 29, No. 11, 1986, 1094-1105 Bertin, Jacques, Semiology of Graphics, Translated by William J. Berg, The University of Wisconsin Press, 1983. Croxton, Frederick E., and Harold Stein, Graphic Comparisons by Bars, Squares, Circles, and Cubes, American Statistical Association, Vol. 17, No. 177, 1932, 54-60. DeFanti, T.A., M.D. Brown, B.H. McCormick, Visualization -- expending scientific and engineering research opportunities," Computer, August 1989, 12-25 DeSanctis, Gerardine, Computer Graphics as Decision Aids: Directions for Research, Decision Sciences, Vol. 15, 1984, 463-487 DeSanctis, Gerardine, and Sirkaa L. Jarvenpaa, Graphical Presentation of Accounting Data for Financial Forecasting: An Experimental Investigation, Accounting, Organizations, and Society, 14, 5/6, 1989, 509-525 McCormick, Bruce H., et al. (eds.) Visualization in Scientific Computing, Computer Graphics, Vol.22, No.6, Nov. 1987 Tufte, E.R., Envisioning Information, Cheshire, Connecticut: Graphics Press, 1990. Vessey, Iris, Cognitive Fit: A Theory-Based Analysis of the Graph Versus Tables Literature, Decision Sciences, Vol. 22, 1991, 219-239 Vollmann, Thomas E., William L. Berry, and D. Clay Whybark, Manufacturing Planning and Control Systems, Second Edition, Homewood, IL: Dow Jones-Irwin, 1988 Zhang, Ping, Visualization for Decision Making Support, Ph.D. Dissertation, The University of Texas at Austin, 1995.