Glassbox Dashboards
- Jana Volencová (Deactivated)
The Glassbox charts target audience is Configuration Engineers, Business Analysts, and team members working on improving a new optimization model. Glassbox charts are shown on a tab in accelerators such as Negotiation Guidance, Price Waterfall Optimization, Shelf Prices Optimization or Markdown Optimization. The tab provides insights into how the Optimization Engine reached its final state. One needs to understand the main concepts of the Optimization Engine to benefit from this dashboard. Two interesting pages to help users are Glossary (Optimization Engine) and Explainability (Glassbox).
The data for some of these charts are grouped by “agent key”. An agent key depicts the nature of the agent within the waterfall. For instance, UnitGlobalListPrice can be a key, and the agent representing the unit global list price for any specific product a value finder, having this key, so we call UnitGlobalListPrice a “value finder key”. Criteria also have keys, like RevenueTarget for instance.
Criteria Satisfaction
Satisfaction Pie Chart
The whole goal of the Optimization Engine is to satisfy criteria. These charts enable the user to assess at a glance whether or not the engine was successful at it.
The Satisfaction pie chart displays the total number of criteria by satisfaction status (satisfied, acceptable, unacceptable). The meaning of this status is explained in Criteria Description.
Criteria Compare Sankey Chart
This count itself is not sufficient, the two next charts provide more clarity about the criteria satisfaction.
Because some criteria could be satisfied in the current state end less acceptable after the optimization, to compensate for other criteria that go from an unacceptable state to an acceptable or satisfied one, the Sankey chart shows the global journey of the satisfaction state of all the criteria.
Satisfaction Bar Chart
Because some criteria may be much more important than others, the satisfaction bar chart criteria displays the satisfaction status by criteria type. For instance, in the example below, the global RevenueMarginMixMaximization criterion is only a single criterion but it impacts the whole scope of the optimization, whereas there is the MaxUnitInvoicePrice criterion for each product x customer pair but each one is individually of lesser importance. This is why a bar chart provides more details, by displaying the satisfaction status by criteria key and space, i.e. by criterion type and its level of granularity. The tooltips also indicate how many criteria are instantiated.
Value Finders and Criteria Charts
Value finders and criteria are both ends of a feedback loop, their interactions are fundamental to the convergence of the system towards a solution. These charts provide insights into these interactions.
Impact and Satisfaction Bar Chart
This chart displays data per criteria keys x value finders keys. Each criterion has an impact on some value finders and is more or less satisfied at the end of the run. The satisfaction is not related to any specific value finders, it is only related to the criteria. The Impact vs. Satisfaction bar chart is sorted by total impact amount. The more a criterion is impactful on a value finder, the higher it is in the chart. The expected output is to have more satisfied criteria at the bottom of the chart: these criteria do not impact the value finders because they are satisfied.
Questions one may want to ask:
Why does a criterion have not much impact if it is not totally satisfied? It could be because the criterion has an impact mostly on another value finder key. Does it make sense for this customer or this model?
Why does a criterion have much impact on a value finder that is already mostly satisfied or acceptable? It may be because of its importance or because of interactions.
Impact and Influence Bar Chart
While the criteria have an impact on the value finders, the value finders also influence the criteria. This bar chart, which is also at a granularity level of criteria key x value finders key, compares the criteria’s impact (on value finders) to the value finders' influence (on criteria). The Y axis indicates the pairs criteria keys - value finders key. It is also sorted by impact in descending order, hence it is in the same order as the previous chart. You can check that high impact is generally correlated with high influence, although that may be scrambled by the various priorities and levels of granularities handled by the model.
Questions one may want to ask:
Which criteria have more impact on which values finders?
Which value finders have more influence on which criteria?
It can help tune the priorities of the criteria.
Value Finder Influence Bubble Chart
This bubble chart offers a quick view of the scale of value finders/criteria, both in raw quantity and in terms of influence. The X-axis represents the number of value finders, and the Y axis the number of criteria. It is aggregated by the value finders keys.
Each bubble represents a value finder's key, whose size represents the overall influence of the value finders with this key on the criteria of the model. The coordinates represent how many value finders provide this influence and how many criteria they act on. For instance, in the example above, few value finders have much influence on many criteria and many value finders (blue bubble) have not much influence on few criteria.
Criteria Impact Bubble Chart
This bubble chart offers a quick view of the scale of criteria/value finders, both in raw quantity and in terms of impact. The X-axis represents the number of value finders, and the Y axis the number of criteria. It is aggregated by the criteria keys.
Each bubble represents a criteria key, whose size represents the overall impact of the criteria with this key on the value finders of the model. The coordinates represent how many value finders are impacted by this criterion and how many criteria act.
Value Finders Variations Chart
The boxplot chart is represented by the value finder key. A single box is a rectangle starting from the first quartile and ending in the third one, with a horizontal line representing the median. The whiskers have a maximum length of 1.5 times the difference between the median and the quartile it starts from. Moreover, each point of data is represented by a gray dot, slightly jittered horizontally. Hence, if some dots are beyond the whiskers, it means that the data has outliers. For all the boxplot charts, some useful information:
There is a legend on the right of each chart, that separates boxplots from the scatter plot. Some value finder keys represent many points of data. Clicking on the legend “Scatter” will remove the dots and let the user see the box better.
There is a tooltip for each box on mouse-over. It provides the minimum value, the three quartiles, including the median, and the maximum value. The minimum and maximum values may not be the extremities of the whiskers (if there are outliers).
Sometimes the boxes have very different heights, for instance when a value finder key represents some prices or even revenues, and another one represents some percentages like discounts or margins. It is always possible to zoom in on the charts. The user has to simply grab a rectangular with their mouse.
The boxplot chart shows the raw variation of the value finders. It tells the user which value finders move more and if there is a pattern in their variations. Each box represents a value finders' key and the scatter plot represents the density of the values. The tooltip provides the quartiles of the data, from minimum to maximum, and the space where the value finders' key is defined. The color code is shared by all the boxplot charts (but not with the previous bubble charts). I.e. a given value finder's key is represented by the same color in all the boxplot charts but may be represented by another bubble color in the value finders' bubble chart.
Value Finders Relative Variations Chart
The boxplot chart shows the same variations but in percentage. It is useful mainly to focus on small values, like the discount rates, that could have moved a lot relative to their initial value.
Value Finder - Criterion Influence
This network chart offers a quick view of the influence that each criterion had on a particular value finder. In this graph, the value finder is represented by a square, and the criteria linked with this value finder are represented by circles.
At first sight, the graph displays a lot of information. Without exploring each element, we can see:
Which criterion has pressured the value finder by looking at the size of the circles. In the first example, “RevenueMarginMixMaximization” and “MinSpotDiscountRate“ seems to be the most (and only) pressuring criteria.
How the criterion influenced the value finder. In this example, “RevenueMarginMixMaximization” influenced the value finder to increase its value, which is indicated by the arrow sign
▲displayed after the name. Nothing is displayed next to “MinSpotDiscountRate“ because, although the criterion put a lot of pressure on the value finder, it had a small effect on it. The effect is displayed over the link if it is important enough. In this example, “RevenuMarginMixMaximization” had an effect of 0.08 on the value finder, and “MinSpotDiscountRate“ effect of the value finder is 0.01.Whether the criterion is satisfied, we can tell by looking at the color of the link: Green link means the criterion is satisfied, yellow-green means the criterion is Acceptable, and red means the criterion is not satisfied. Here, “RevenueMarginMixMaximization” is Acceptable, “MinSpotDiscountRate” and all other criteria are Satisfied.
When hovering over a criterion, various information are displayed:
Its name.
Its coordinates.
The effect it had on the value finder, i.e., how much the value finder changed its value for it.
The pressure it put on the value finder.
The Input Value, the value it perceives and uses to take decision. Depending on the problem description, this value may not be the value finder’s value. This is the value the criterion used to verify its constraints.
The parameter of the criterion. In this example, the criterion is a minimal threshold, with a threshold value of 0.02. Since the input value is of 0.10, it is satisfied following the description of the threshold criteria.
When hovering over the value finder, additional information are displayed:
The name.
The value “From”, the value before optimization.
The value “To”, the value after optimization.
Information about the stability of the value finder, either “Still moving” or “Stabilized”.
Debug Charts
Initial Movements Boxplot Chart
The Initial Movements boxplot chart represents the variation caused by the “probe” of the value finders, i.e. the initial action the value finder does on its own, without being pushed by any neighbor. This chart allows engineers to check if these probes are small enough not to disrupt the convergence.
Initial Movements Relative Weights Boxplot Chart
The Initial Movements Relative Weights boxplot chart represents how prevalent the probe-induced movement is, compared to the overall variation of the value finder. Hence the users can assess whether the value finder has been mainly pushed by its criteria or disrupted by its probe.
Criticality Profile
This chart displays the evolution of max(dissatisfactions_in_the_sample) during the run, but not all criteria are present in each sample. Hence, each value may correspond to a different criterion. This curve lets the user know whether the convergence is finished or if it still needs some more steps. It is also a good tool to tune the various value finders parameters (for details see Parameters (Pricefx staff only)).
It is not a really good tool if maximization criteria are present in the problem. These criteria work in a special way and often have a negative dissatisfaction, making the profile curve hard to read. For instance, in the image below, the chart alternates between a highly dissatisfied criterion (in the 500M-1000M range), a moderately dissatisfied one (near zero), and a maximization one with negative dissatisfaction.
The user can filter this chart, for instance, by groups of criteria. In the chart below, we can see the evolution of the criterion which corresponds to the “near-zero values” on the overall scope chart above. We see that this criterion’s dissatisfaction increased (probably because the system was helping other criteria at that time), but it decreased by the end of the run and probably needs more steps to settle.
Warning Matrix
This matrix contains the warnings emitted during the Glassbox creation, if some charts cannot be built because of missing information.