Sunday, February 21, 2010

The importance of business analyst interfaces

Carole-Ann wrote a blog about what she considers to be the #1 pitfall in the implementation of proper Decision Management applications. Her observations are based on a vast experience of multiple implementations of such applications in multiple vertical domains.

I fully agree with Carole-Ann’s position highlighting the importance of getting the user interface for business analysts well defined and implemented, with their workflows in mind, and with special attention paid to their concerns.

I have been in a position to review multiple implementations, made with both the products I have been responsible for as well as their competitors, and, in a fairly significant number of cases, what I have reviewed has amounted to little more than direct translations of implementation concepts, disconnected from the concepts and workflows of the business analysts. Some of these implementations are a consequence that the tools used insist in presenting very low level interfaces – typically, single ‘if-then-else” rules built by point and click – or a single representation – typically, a complex table-based interface, or a tree representation. But others leverage tools that can do much more, yet they fall down to low level implementations that end up leading to significant frustrations by the business users, and, worst, a breakdown of their workflow, and their ability to actually manage the decisions, contrary to the promise of the Decision Management approach.

This is serious, and I will go as far as Carole-Ann on this – the ultimate success of the discipline hinges on getting this right.

What are the two key issues I have seen in these failed implementations?

Lack of consideration for the concepts and representations used by the business analysts themselves

Business analysts do think in terms of policies, constraints, pricing structures, etc… They do not think in terms of “if-then-else” rules, or “event-condition-action”. Business analysts need to be involved in the decision on which representation to use from the onset: they should go to the whiteboard, and, with no prompting from the actual business rules or decision management implementation specialists, specify how they see their concepts, policies and workflow. More frequently than not, they will use a combination of textual form and graphical representations, and fairly well defined workflows. The role of the implementation specialist will have to be then to select the right representations in the tools to reflect as well as possible the representations and flows used most frequently.

I really cannot understate the importance of this approach. One of the key tenants of the Decision Management approach put forward in the early days and that I still believe in after all these years is that the business analyst must be able to understand the decisions implemented, and the implementation specialist must understand what the changes to the decisions need to be. Ideally, the business analyst can implement the changes directly – but even that requires the implementation specialists and the business analysts to have a proper common ground to effectively defined and enforce the boundaries of what can be done that way.

There is one key consequence of this approach, and that is that there is no way a single representation approach will be sufficient to elicit, implement and manage decisions through their lifecycle. I know that there are whole companies out there pushing to the front their graphical or textual representation as the solution to all decision logic, but on this problem, they fail. Yes, I do not doubt that they can express all possible decision logic – but that’s not the question: the question is whether they can do it in a way that is efficient for the business analyst and that will scale through their workflow and the evolution of decisions.

Let’s take a few examples:
  • I am very familiar with a tool that mostly uses decision trees as the way to express all decision logic, including all potential initial states and all potential outcomes. No surprise there – the decision trees in question grow to the thousands, and tens of thousands of nodes, becoming unusable. Patch solutions, like tree simplification algorithms, are not an answer: the issue is that the tree metaphor is not adapted for all types of decision logic. The internet is full of references to studies that demonstrate that.
  • Another tool uses a table like metaphor as its only representation for decision logic. Again, we hit a similar scalability problem, and for different technical reasons as why the decision trees don’t scale. Tables are awfully poor at representing disjointed logic, and even worse at handling exceptions. But on the other hand, there are many, many, cases in which a single decision is in fact a multi-step decision with a number of exceptions. The implementations end up with a large number of tables that are difficult to navigate through, with close-to-empty tables representing exceptions parallel to huge tables mixing multiple steps in the decision, with tables with extreme density variability (sections with a lot of content, and then sections mostly empty where the user needs to be very good at finding the relevant information), etc…
  • Others used exclusively semi-formal languages. These present the advantage that they are closer to the way the policies would be documented in textual form, but lose the significant synthesizing power of metaphors. Explaining a scorecard in text is incredibly cumbersome compared to what can be done in a real scorecard representation.

Lack of consideration to the availability of context in the business analysts interfaces

This is a constant concern of mine, and something not that many business interfaces do provide: availability of context directly where the business analyst is working. Carole-Ann refers to a huge aspect of this when she focuses on the navigation issues in the user interfaces.

What the business analyst is doing when she/he is manipulating a particular representation is contextual – and to do it right, she/he needs access, directly there, and in the most up to the point form, to the overall context in order to do her/his work right.

Take the example of working on a decision step in a credit card originations decision that refers to the fact that a given customer is a high value customer. Well, it is likely that high value is an attribute of a customer that is defined by business logic somewhere else in the application, and the business analyst may have the need to understand as she/he is implementing the decision step. Similarly, she/he may need to understand where else that attribute is used and how, etc…

This is essential – and forgotten very frequently.

It actually reaches extremes not just with the single-metaphor approaches as described above which again have difficulties coping with varying context, but with single “if-then-else” or “when-then-else” rule representations. It is not true that single rules act alone – that is far more the exception than it is the rule (sorry). Those environment do not scale beyond a few tens of rules, and that, however powerful said rules are, is nowhere close what is needed in real applications.

There is much more to be said about this subject, but Carole-Ann is absolutely right on this issue being the biggest one that needs to be addressed to achieve successful decision management applications.

It is actually telling to see the difference in the implementation steps between how Carole-Ann would do it and how it is done frequently elsewhere. Carole-Ann will focus first and mostly on eliciting the proper concepts, representations and flow, then design the interface for business analysts to a point where it is functional for them and they can start using them, and then worry about the actual implementation behind. This has led to large scale implementations with complex but appropriate life-cycle, very dynamic evolution and still excellent run time performance.

Let’s avoid this pitfall going forward.

Tuesday, January 26, 2010

Decision Management as an Academic Discipline

Those of you who follow this blog (or know me) know that I am quite passionate about decision management, and have essentially articulated the past decade of my professional career around the many problems it requires us software and analytic people to solve.

One key issue that keeps bugging me is the lack of support in the academic world for decision management as a full blown discipline available to both business types (your MBA suits) and technical types (your typical CS or analytics specialist).

I wrote a post on this in Carole-Ann Matignon's TechDec - would love to read your comments.

Monday, January 25, 2010

Books: Adaptive systems - Dynamic Networks - Evolutionary Dynamics

I just finished reading "Complex Adaptive Systems: An Introduction to Computational Models of Social Life", by John H Miller and Scott E Page. Interesting book, dealing with problems that are dear to me, although from a fairly different perspective from the one I am used to.

I come to this from the angle of decision management. As Carole-Ann Matignon writes in her blog, a lot of developments are expected to make 2010 an interesting year for the discipline. Dealing with uncertainty is one of them, and one of the key sources of uncertainty is related to complexity.

One aspect that deserves attention is how we can manage decisions in light of the effects of the very high level of interconnectedness between multiple entities, with relations that are complex, non linear, etc. In the business world I have up to very recently been active in, this translates into the effects of customer psychology, mass effects, etc. All problems that frequently vex traditional approaches, and tend to create challenges to standard modeling approaches.

The book properly highlights the key difference between a complex system and a complicated system: the complex system's behavior cannot be explained by reducing it to sub-components and explaining the components and their interactions. Which, in a certain sense, creates a significant challenge to some of the traditional scientific or engineering approaches. That is what makes these complex systems so fascinating: they are composed of a multitude of agents inter-related and inter-acting in massive terms, yet they exhibit emerging behavior similar, a posteriori , to that of a single agent but that cannot be explained by decomposing it.

John H Holland, from the Santa Institute, has written a lot about this subject. His analysis of the characteristics of ComplexAdaptive Systems is often reduced to this: order is emergent (or behavior is emergent), history is irreversible, future behavior is often unpredictable. I am not sure about the last point - it's not as much a question of whether the future behavior is predictable, it's much more a question of whether traditional approaches allow the prediction of future behavior.
The key challenge is how to model that behavior. The typical techniques used in most decision management approaches today do not deal with complex systems - they focus only on simple agent behavior, modeling away through extreme simplification the network effects. When those intervene in real life, the models become totally irrelevant - and I would venture to say that a lot of the events we've seen in the past decade and in particular the last couple of years have shown the limits of the simplification. After all, it could be claimed that network effects are largely responsible for the propagation of the bad loan practice, as well as the reactions when the bubble collapsed.

It reminds me of the classical work "The Structure and Dynamics of Networks", another nice - though less accessible, collection of papers (Jean-Marie Chauvet refers to it in another post - in French but I am sure he will gladly translate should you ask for it). Read it, challenging by parts (some I ended up skipping) but good. And let me know what you got from it.

Not in the same space, but connected, I got "Evolutionary Dynamics: Exploring the Equations of Life" by Martin Nowak for my 14 yo old. Probably a little bit too ambitious for him (I read it first and am still waiting for him to pick it up), but the book is beautiful, and the theme is connected to Complex Adaptive Systems. Nowak is an expert, and his book covers a wide range of analytic approaches that should become part of the arsenal of those studying complex systems.

This whole thing may seem like a scientist's dream (remember, I am not a scientist), but the reality is that the large consumer-oriented companies are already dealing with this kind of problem. Just think for a minute about what happens in an EBay auction, or an Amazon recommendation.

I am looking forward to significant synthesis of the approaches, and, as a result, techniques that will enrich in the future the way we approach decision management.

What's your take?

Wednesday, January 6, 2010

A discovery: PLINQO

I have recently spent some time looking at LINQ-to-SQL to see whether there would be an opportunity to benefit from it and replace things like NHibernate which I am used to. The best result of that effort was the discovery of PLINQO provided by the CodeSmith team you may be familiar with.

Before reading the rest of this short post, check them out: (and take also a look at

Initially, I was just looking for simple time saving template-based code generation to facilitate the usual fastidious work. I had the surprise to discover that PLINQO also removes a number of the most annoying limitations of LINQ-to-SQL, and that it does that in a safe (no loss of customization, yeah), fairly nice (a lot of developer friendliness in what gets generated) and extensible way.

Three key distinctive features of PLINQO:
- Many-to-Many relationships
- Entity detach
- Complete serialization and cloning
The availability of these three features essentially removes a significant amount of tedious, error prone code with problematic maintenance that LINQ-to-SQL required.

One key area I need to look into is support for the LINQ Dynamic Query Library.

There is much more to PLINQO. More to come.