But employing a usability expert when designing the tools and observing how users interact with them would go a long way toward improving their usefulness.

My mind utterly boggles each time I discover that a large software vendor still doesn’t seem to have realized this. Or maybe Fair Isaac did do usability engineering, but entrusted it to a blithering incompetent. That frankly would be more reassuring than them not having tried at all.

• Access to information
• Application software

More recently – especially the past five years – there’s been a third way:

• Infrastructure upgrade

as early-generation implementations get replaced by newer ones.

At the 50,000 foot level, here’s some of what I see going on:

• Classical BI marketing is floundering. BI vendors don’t know whether they’re in the business of quick/easy information access, analytic apps, or better-enterprise-system-software.
• A few areas of analytic application are being packaged and marketed well, with solid business-process stories and good customer acceptance of same. The biggies are budgeting/planning and CRM analytics. On the whole, however, analytic apps are floundering, or else are little more than reporting front-ends on operational systems (e.g., in network management).
• Data warehouse software startups are on a roll. Especially at the high end, this is a pure infrastructure-upgrade business. There’s plenty of room still for improvement, but multiple vendors each are doing good jobs of marketing on the basis of:
  • Speeds and feeds
  • Ease of deployment
  • Ease of administration
  • Price
  • Credibility
• Data integration is mainly an infrastructure improvement play. After all, that integration COULD be hand-coded. Automating the process is usually a better-infrastructure story.
• Text search is still an information-access story. There are multiple niches where search is booming. But in all cases the story is information access. Evidently the technology and/or market aren’t mature enough yet for strong infrastructure stories. And in the limited cases where text search gets integrated into general application software packages, it’s usually just for information access rather than a real business process.
• Data mining and predictive analytics are mainly information access plays. Yes, the information being accessed is calculated rather than raw. Yes, I believe that the heart of the data mining market is continuous process improvement. Even so, what users buy from the vendors is usually little more than information toolkits.
• Text analytics is mainly an information access play. Text mining and information extraction have two main uses right now. Either they resemble – and indeed often feed into — data mining, or else they are used to enhance search and search-like document access.
• Information services have always been an information access play. When you think about it, the financial-quote-machine business is a huge part of the whole decision support market. Lexis/Nexis is no slouch either. And they’re all about providing information access.

Related links

• This three-headed taxonomy of strategies is similar to one I previously postulated for Microsoft, SAP, and IBMOracle.
• I covered analytic business processes at length in a November, 2004 white paper. Unfortunately, industry progress since then has been relatively slow.
• I’ve written voluminously about data warehouse software startups on DBMS2.
• One example of infrastructure focus is the ease-of-deployment trend.
• Web search and generic enterprise search aren’t the only search areas to focus on information access. (And yes, they’re most definitely separate areas.) Even customer-facing structured search does; the information is just tailored according to different criteria.

• Business Objects and Netezza have announced a mid-range BI appliance.
• Ingres is headed in the same direction.
• QlikTech is enjoying great growth for its fast-deploying BI technology.
• KXEN and Verix offer “easy” data mining technology.
• Search-based BI is trying to circumvent the data warehouse deployment process.

It’s about time.

I think there are lots of problems like that in BI, and what we need to do is step back and consider all the different kinds of BI that enterprises value and need. More precisely, let’s consider the major kinds of use of BI, because it seems that each calls for different kinds of technological support. Here’s one possible list:

• Early warning of situations that require action.
• Communication of company results.
• Deep analysis and decision support.
• Operational analytics.

Here’s what I mean by each category.

Early warning of situations that require action. This is the classic image of BI. People get reports or graphs on paper or on screen, see that some numbers are out of whack, and react accordingly. Dashboards can be a prettier version of the same thing, or they can be more focused on alarms and alerts. Nowadays, alarms and alerts can also arrive by IM, email, text message, pager, and so on.

On the one hand, a large fraction of the economic value in the history of BI has been generated in this area. On the other, technology for doing so is continually perceived as inadequate. I continue to think that KPI management, alerting technology, and so on are advancing much more slowly than they should be. Even the buzz around Business Activity Monitoring doesn’t seem to have accelerated things much.

Nor does this change when the warnings are the product of text or data mining. For example, despite a very interesting approach to generating alerts, at this point in its development Verix delivers them in uninspired ways.

Mainly, the supporting technology here is the standard query/reporting stack, including new dashboard/scorecard/alerting tools. It’s also the main place where data mining and text mining should be more integrated into standard BI than they are – i.e., to define and populate metrics and KPIs.

Communication of company results. Back in the 1980s, the conventional wisdom was that half the benefit of reporting tools was for actual analytics, while half was just for communicating among enterprise employees. That’s probably still valid.

BI vendors had the good idea a few years ago to build out their collaboration capabilities, but generally didn’t follow through on it, SAP somewhat excepted. Bad choice, in my opinion. The use of text search to get at BI results is something of a plausibility argument for my views in this area.

Again, the supporting technology here is largely the standard reporting stack. Portal/collaboration tools should be more involved than they are.

Deep analysis and decision support. Routine, scheduling reporting was covered in my first two categories. But this third one is where the bulk of ad hoc query and data mining fall. Generally, it’s where lots of specialized and/or calculation-intensive analytic technology comes into play. It’s also where the drilldown aspect of standard reporting shows up. Also, this is the area that is driving much of the recent transformation and disruption in the data warehouse market, because different kinds of BI need different kinds of data warehousing technology.

Operational analytics. In operational analytics, a small amount of analysis is done real-time or near-real-time, in connection with an operational business process. The most technically demanding examples are probably the customer-facing ones, such as call centers or personalized e-commerce sites. This is where the buzz around “active/enterprise data warehousing” is concentrated. There may also be interesting messaging aspects. And data mining scoring may be a consideration.

1. Data is assembled in a data warehouse from transactional information, with all the effort and expense that requires. Maybe more data is even deliberately gathered. Or maybe the data is in large part acquired, at moderate cost, from third-party providers like credit bureaus.
2. The database experts fire up long-running, expensive data extraction processes to select data for analysis. Often, special data warehousing technology is used just for that purpose.
3. The statistical experts pound away at the data in their dungeons, torturing it until it reveals its secrets.
4. The results are made available to business operating units, both as reports and in the form of executable models.

Each in its own way, KXEN and Verix (the imminent new name of the company now called Business Events) want to change all that.

KXEN believes they have found a one-size-fits-all set of data mining models and algorithms. This is not an SVM (Support Vector Machine), which they actually don’t offer any more, but rather something else from the fertile brain of SVM co-inventor Vladimir Vapnik, called Structured Risk Minimization. While the details have been published, they asked me not to write about them anyway for some kind of security-through-obscurity competitive reason. So let’s just say that these are not just the linear models they previously were or seemed to be stuck with. (For a small company with limited footprint, there sure is a lot of false information out there about how the whole thing works.)

A limited set of models lets one design a fairly simple user interface, especially when the models are good at helping one zoom through what otherwise can be annoying steps (like variable reduction, in which you choose which 80-90%+ of the data columns to disregard). Based on that relative simplicity, KXEN wants to let business users data mine directly, without being dependent on statistical specialists and their machinery. They position this as providing better results, because it allows rapid-cycle-time data exploration.

They also have a pickier statistical point to make, which is that their model-building process is streamlined and automated enough that it’s realistic to build lots of parallel “local” models, e.g. for each store or region in a retail chain. By way of contrast, in traditional data mining one would normally have one model used for all localities, but perhaps with additional variables indicating which locality the model was currently referring to. KXEN confidently believes that its way is superior, but in a recent discussion didn’t actually provide me with much beyond hand-waving to back that claim up.

I don’t actually have a good feel for how well these pitches are being received by the market. KXEN’s biggest sales successes seem to be via partnerships with various other analytics players, and it’s tough to judge whether that’s due more to price or to embeddability or to the fundamental merits of their overall case.

Business Events, imminently to be renamed Verix, is a raw start-up with a story even more extreme than KXEN’s: Sophisticated analytic results just delivered on a SaaS basis, with no thinking required by the customer at all. Obviously, this can only make sense if the universe of possible results is rather limited, and indeed it is.

Verix’s approach assumes a classical star set-up: A single measure/fact table and a complexly hierarchical set of dimensions. Verix looks exhaustively at time series on the facts, pulls out all series that are showing anomalies in two or more dimensions at once, and isolates exactly the point in the dimension network where the anomaly is occurring. If sales of frobalizing widgets in Houston are off plan, it identifies whether this is really a Houston issue or a Texas one, and whether it’s a problem just for frobalizers or – gulp – for the entire widget category.

The company claims that some insights you get this way just wouldn’t have been found by conventional BI. E.g., if frobalizers are down in Houston but up in Dallas, and the analysis stopped at Texas, nobody (not even the Houston district manager?) would ever know of the great Houstonian frobalizer downturn.

The company sounds like they’re working on all the right things to generalize this model. Initial interest in what they have seems to be concentrated in the pharmaceutical and CPG (Consumer Packaged Goods) industry, although there are a couple of paying telecom customers as well. One thing pharma and CPG have in common is that a lot of your raw data comes from third parties, such as IMS, and so your sales data are visible to your competitors anyway. Given that, it’s easy to believe that the SaaS nature of the service isn’t causing a lot of customer discomfort.

And by the way, IT departments aren’t involved in the Verix buying process whatsoever.

Analytic business processes — or the areas of overlap between analytics and business process — are poorly understood. Business Activity Monitoring and Operational BI? Great buzzwords, but there’s way too little thought put into figuring out exactly which metrics are most useful for making which kinds of business decisions. Continuous planning/budgeting? The surface has only been scratched. A numerate, “one-truth” enterprise culture? Hah. When we identify an enterprise that truly has a pervasive numbers-oriented culture, it usually is one that winds up pathologically managing to a purely short-term set of goals. (But some exceptions to that rule are among the great corporations of the world.)

One area that really needs more consideration is data capture. You can’t analyze data you don’t have. Certain industries have indeed recognized this. E.g., travel and gaming have been hugely successful with loyalty cards; indeed, casino giant Harrah’s probably gets over 100% of its profits via targeted marketing based on the mining of its loyalty card data. Credit transaction data and the like is of course also heavily exploited. I made this whole case in a Computerworld column a year ago, and if you missed it I suggest still checking that column out today.

But that’s all transactional data. The story for text data is much worse. Indeed, survey forms typically try to force people away from just saying what they think, instead giving them endless checklists that bring back unhappy memories of SATs and #2 pencils. Yet text mining technology now exists that makes it possible to glean crucial information from free-form text. If you haven’t already checked it out, you should.

Particularly interesting, I think, are some examples in the area of text data and analytics.

• There’s a large set of “early warning” apps where text mining is being deployed. Many of those same apps are addressed by data mining of tabular data too – antifraud, to start with, and also warranty tracking and indeed most of the rest.
• Data mining has been huge in CRM.
• The use of data mining in manufacturing to do failure analysis, improve quality, etc. is really on the rise. This goes at least somewhat beyond what one could reasonably pigeonhole as “early warning.”
• Data mining plays a big role in the life sciences, and is being applied to a broad range of other sciences as well.
• Data mining is a huge part of R&D at search engine and antispam vendors.

By the time I submitted the column, the list had morphed into:

• Customer offer targeting.
• Other CRM applications, often of text mining, such as reputation management or just sentiment tracking.
• National security, antifraud, and crime prevention.
• Purer portfolio/risk management applications.
• Defect tracking.
• Health care and scientific research.

For lots of examples and explanation of the categories, please see the column when available. (Theoretically that should be on the inauspicious date of September 11. In practice, it could be any time next week. I’ll post a link here when I know of one that works.)

While the latter version of the list may be slicker and more precise, which is why I went with it in the column, I think the former is more useful for a discussion of why those particular apps are the ones that get adopted. Simply put, data mining apps are concentrated at two extremes:

1. Seeking “gold nuggets” of insight.
2. Continuous process improvement.

What’s more, if I had to pick just one of those categories, I’d pick #2. The annals of BI are replete with examples of insights that just leapt out of reports and danced straight to the bottom line. But those stories are generally about reports and OLAP analyses, not full-blown statistical workups. Don’t get me wrong — I’m sure there are plenty of cases of data mining producing hugely valuable sudden insights. But, uh, I can’t think of any right now, at least not in the mainstream statistical analyses we usually think of when we hear “data mining.” (Perhaps some kindly product vendors will help me out with examples. If nothing else, there should be examples in the life sciences, forensics, product quality, etc. – i.e., in applications where there only ever was one single answer to discover in the first place. )

Where data mining does succeed all the time is in areas such as marketing efficiency improvement – mailing smarter, better targeting customer offers, and of course avoiding “bad guy” customers such as fraud or default risks in the first place. Text mining is something of an exception to that rule – but then, despite its name, it’s not clear that all of text mining should be classified as data mining anyway. Some of it is just “knowledge/fact/information extraction”, which generally is used to inform analytic technologies of some sort or other. But those can be regular BI or text search or whatever, with data mining just being one of the candidates on the fact-consumer-technology candidate list.

• I have posted extensively on text mining over on the Text Technologies blog.
• In particular, much of the column was based on a post in which I discussed “early warning” applications of text mining.
• The research was informed by a trip to the KDD 2006 conference, about which I’ve blogged separately.
• SAS is world’s biggest vendor of this stuff, so if you want to know what the applications are, you might want to start with their website.

At the academic part, I didn’t actually experience all that much. In part, this was because of the accents. I don’t think I’ve ever been around as many people with Indian surnames whose English was so incomprehensible. And while one would typically be prepared for thick accents from Chinese-named folks, usually there are a LOT more exceptions to that stereotype than there were in this particular case. Even keynote addresses were not immune. So I didn’t actually attend very many sessions.

That said, I attended a few talks, browsed a lot of papers, chatted with a number of attendees, and came away with a few observations, including:

• Text mining is a big area of research.
• So is network/link/graph analysis.
• Knowledge is terribly stovepiped. Graduate students working on fast algorithms didn’t think about parallel processing. A Yahoo presenter with a small state-transition matrix didn’t think about Markov chains.
• Notwithstanding the stovepiping, there’s great interest in applying data mining to other disciplines, especially academic ones.
• The handwringing about our educational system has merit. Very few of the young people there were actually Americans, and the young Americans I did talk to seemed to have more personality than actual grasp of their material.

The vendor part consisted mainly of a few usual suspects, led by SAS, SPSS, and Oracle, and a few outfits there to hire researchers, led by Google, Microsoft, and Yahoo. I had the chance to talk for hours each w/ Oracle and SAS, which has heavily informed a column I just submitted to Computerworld. (Watch this blog for a link on or soon after September 11).