|
Current Projects
The research for my second Ph.D. in
Computer Science at Aalborg
University is in progress and will cover the following topics:
Improving Business Intelligence Speed and Quality through
the OODA Concept
This article introduces the Observation-Orientation-Decision-Action (OODA)
concept as a mean to identify three new desired technologies in business
intelligence applications that improve the speed and quality in the decision
making processes. Specifically, the article identifies: artificial
intelligence to reduce human interaction in the OODA loop, “sentinels” that
can give early warnings about a later influence on a business critical
measure, and finally the ability to analyze the speed and quality of an OODA
loop in order to quantify and analyze organizational talent and core
competencies. In this project the technologies for sentinels and improvement
of OODA effectiveness are pursued further. This article was presented on
DOLAP ’07.
[Full-length
article from DOLAP 2007]
Discovering
Sentinel Rules for Business Intelligence
In this paper, we introduce the concept of sentinel rules. Sentinel rules
are schema-level rules that provide the user with an early warning,
typically when data concerning the external environment changes. For
instance if there is a surge in negative blogging about a company’s
products, a sentinel rule can warn that revenue will go down within two
months if no course of action is taken. By doing this, we expand the window
of opportunity for the organization, and therefore render the organization
capable of successful navigation even though the world behaves chaotically.
Since sentinel rules are at the schema level as opposed to the
data level (such as association rules and sequential patterns),
we are able to provide the user with fewer, more general rules, which means
that less time is needed to interpret the output. In addition, our solution
handles the fuzziness of real-world data by applying a weighted elimination
process which eliminates the contradictions that occur in real-world data.
The paper presents a method for sentinel rule discovery and an SQL
implementation of this based on Microsoft SQL Server 2005. The
implementation is assessed in computational complexity and subjected to
experimental evaluation. It is verified that the implementation scales
linearly on large volumes of data. Furthermore, the implementation is tested
on real-world data where it identified useful and relevant rules for
decision making.
[Full-length DB Technical
Report as published on
vbn.aau.dk]
Efficient Discovery of Generalized Sentinel Rules
This article introduces an algorithm for sentinel discovery and uses the
simple solution from previous article as a baseline. In contrast to the
previously proposed SQL implementation, this algorithm is expected to be
superior in performance as well as allow multiple source measures to be
combined into one sentinel rule. In addition, the algorithm will also be
targeting streaming data sources which means that it will also take into
consideration a sliding window for observation and retirement of relevant
data. The goal is to produce an algorithm that scales linearly while at the
same time outperforms the previous SQL implementation. It is expected that
superior performance can be achieved by eliminating some rules that do not
apply during the discovery process as opposed to in the end, in addition,
this algorithm will take greater advantage of main memory and is thus
expected to perform much faster when all data can fit into memory. This is
particularly interesting in cases with streaming data where we intend to fit
the sliding window to match the size of the main memory in order to optimize
performance.
Sentinel Discovery on Dimension
Hierarchies
This article introduces a parameterized implementation of the sentinel
algorithm that exploits the dimensions to discover sentinels that only
appear at some levels of the dimensional hierarchies. The approach will
expand the algorithm from the previous article to exploit the hierarchical
data to identify clusters where certain sentinel rules apply. Most likely
the algorithm will rely on a bottom-up approach where sentinel rules are
grown from the lowest granularity of the data, and as we travel higher up in
the hierarchy some rules are retired if they do not apply on these levels;
effectively we expect to end up with a set of sentinel rules that apply to a
cluster bounded by the dimension levels over which no rules are applicable.
For example one can think a scenario where the rule: ”IF negative blogs go
up THEN revenue goes down within two months AND IF negative blogs go down
THEN revenue goes up within two months” applies to the state of California,
this rule is then tested and found to apply to all states in the United
States, however it is not found to apply to any other countries in the
world. In other words, we grew the rule from the lowest level, in this case
California, and from there we went to the next hierarchical level, United
States, and found the rule to be true as well. At the global level we found
the rule not to be true, and thus the largest cluster for which the rule is
true is United States. Using this bottom-up approach is expected to give a
very fast performing and useful addition to the algorithm for sentinel
discovery.
Sentinel Discovery in the Real-World
This article will be a field study on real-world data to assess the
feasibility of sentinels. We will test the ability to discover meaningful
and useful sentinel rules on a number of real-world datasets, e.g., a
dataset with three years of weather observations and a dataset with at least
three years of financial, project and support data. If possible, it is also
desired to use data such as those from Google Trends to discover sentinel
rules that deal with the relationship between number of searches for a
company or product and the respective financial figures. We expect to find
that the algorithm will find meaningful rules and while doing so give us
interesting insights by identifying certain clusters where specific rules
apply. Moreover, we expect to find that in many cases the entire discovery
process can run in main memory which means that performance on real-world
data is highly feasible and within reach for many organizations.
Is Few Clicks a Valid Measure for
Effectiveness in Business Intelligence?
This research will investigate the feasibility of few clicks being a valid
measure for the effectiveness of the user traveling through the phases of
the OODA loop. In practice we will set up a number of tasks that a number of
test users will have to do in a number of applications. During these tasks
we will observe the number of clicks as well as the success of the tasks,
and in addition we will ask the users own opinion of perceived usability.
Using these data we will investigate if any correlations exist between
number of clicks, number of successes/errors and perceived usability. We
will select a number of tasks and applications that are as closely related
to the field of Business Intelligence as possible. We will also conduct
research in the real-world by observing the number of clicks of users on a
real business intelligence application to assess few clicks as a meaningful
measure. In this context we will look into the number of clicks, the time
between clicks as well as the complexity of the information presented to the
user, where complexity is a measure based on the number of cells, dimensions
and measures that are returned to the user. Through these experiments we
hope to gain important insight into how we can better design business
intelligence applications that move the user through the OODA loop faster,
and we expect to gain the most insight from the discrete observation of
users since this is not biased as opposed to the users’ opinions gathered.
---
In summary the scientific method
applied in this project is a blend of analytical, constructive and
experimental approaches. During the four articles about discovering
sentinels (section 3.2) a series of prototypes will be developed. These
prototypes will be validated on synthetic data to assess functional
correctness and on real-world data to assess them for usefulness. In
addition, both types of validation will assess the usefulness of the
algorithms from a performance point of view. In the final article about few
clicks as a measure for usability (section 3.3) data from real-world usage
of business intelligence will be collected and analyzed in order to test the
hypothesis that few clicks is indeed a meaningful measure to assess
usefulness of a business intelligence application.
We expect that the research and the
six articles compiled during this project will be a valuable contribution of
concrete solutions to problems that organizations face in global
competition. In addition, we hope to provide a good research foundation that
others can use in their research going forward.
[My Computer Science Ph.D. Thesis]
[Short
Article About this Ph.D. Project]
[See
Complete Project Plan]
Other Topics of interest...
Aside from currently researching for
my second Ph.D. in
Computer Science at Aalborg
University, I am privileged to be
working in the
TARGIT organization which recognizes that great thoughts and products
are inspired by the time we play -and in my case this involves technology.
Playing with technology can lead to new products or it can lead to
inspiration of how to make existing products better within the current
portfolio; no matter what the outcome, playing is fun and it feeds the mind
and soul...
Currently, I am playing with three
projects that have all been inspired by the
CALM book:
Autonomous OODA Cycles in Stock
Trading
This project is an AI classic! Nevertheless, it has been my passion since
the age of 15 to develop a system that is able to trade stocks with an
optimal profit within a span of calculated risk. This project is not driven
by a desire to make a lot of money; it is driven by the fact that this
market is the most perfect in terms being almost as visible for computers as
it is for humans. Therefore this is one of the first fields where the
application of autonomous
OODA cycles seems reasonable.
To some this project might be
considered alchemy, but my findings so far seem promising. What seems to be
working is essentially a decision model based on technical analysis across
larger periods and period segments. In other words simple models based on
large amounts of data.
Time will show the effectiveness of
this project:
Genmab is a benchmark and my first investment that was partially based
on this theory; I bought at 133 on 28th December 2005.
Body
Computing
In this project we are a team working on in-flight computing during
skydiving. The vision is to extend the human abilities in skydiving by
integrating a computer with the body of the skydiver. The computer will be
able to improve free-flying skills allowing the skydiver to monitor the
performance on video as well as tracking vertical and horizontal movement.
At a later stage we should be able to work with the entire body position and
posture in the air.
Another
objective on this type of body computing is to improve skydiving safety by
allowing the skydiver to get warnings in a variety of situations that are
unavailable with contemporary equipment today.
Overall
this project has a broader perspective since the idea to extend human
potential in a given situation by the use of computers can of course be
harvested in a number of situations other than skydiving. The idea is to
create something that generates real-time, intelligent and relevant
information in an extreme -or any- situation. The knowledge generated in
this project can of course also be fed back into the traditional discipline
of Business Intelligence.
The
Battle Droid
My dream with this project is to create a supercomputer that is able to
battle with a human. I believe that a lot can still be learned from
freestyle battle rap as a metaphor for modern strategizing. Such a system
would need to be able to compose sentences that rhyme, address the
weaknesses of the opponent, and defend itself from the opponent’s attacks.
This system would need to be able to understand attacks in words and counter
them intelligently, so building this computer would be a more sophisticated
version of chess computing. One could say that suddenly the outcomes would
be limited only by an entire language rather than by the permissible moves
on a chessboard. Once the system is ready for testing, I will have it battle
great rappers, like Deep Blue played world chess champion Gary Gasparov.
In
computing nature, the system I design will perhaps not beat the opponent at
first, but with every loss it will gain in strength. By pure computing
advancement, it will be a matter of time before it can beat a skilled
opponent, and I am confident that valuable lessons in both artificial
intelligence and speech synthesis from this project could then be used in
other areas to assist in computerizing autonomous
OODA cycles.
----
As my work and
experiments proceed, more information will be posted on this website, so
please stay tuned. Also, please feel free to
contact me
for sharing ideas and suggestions with regards to the articles and projects.
|
