Machine Learning (ML) is an extension of
Artificial Intelligence (AI)
where computers are programmed
to make decisions based on events occuring in systems and processes (recognize changes)
with least amont of human intervention. In machine learning, systems
learn from
existing data. This is in essence, enabling computers to create new programs using
statistical models such as MCMC,
and predicting the outcome or making corrections (risk minimization using
ERM
) to existing processes.
|
|
Forecasting is used in many areas such as
weather, short and long term financial performance, stock
performance etc. In forecasting, statistical models are
used forecast future outcomes. Based on the forecasted
output humans make meaningful decisions. There is no
autonomous learning involved. In machine learning, the
forecasted values are used as the basis and for every new
event. New values are predicted in a autonomous manner
based on the newly learnt values. This continously goes
on, and analysis can be made using multiple models.
|
|
In the world of robotics, the
industrial robots perform tasks with greatest precision in a repetitive manner and operate
in fixed set of rules (programs and hardware/software). The ones that are humanoid robots,
have additional intelligence such as pattern recognition, obstacle avoidance, machine
vision, independent decision making and are
programmed to learn as they perform their duties. The modern day automobiles have automated options
that use the concepts of machine learning with embedded hardware/software that assist in
lane drift correction, automated stopping, collision mitigation and so on that are self
managed. As the machine learning technology matures, there will be cars that will
be totally self-driving automobiles with all
mobile/wireless
features embedded in them.
|
|
|
Data Mining, Machine Learning - Venn Diagram, [2] |
|
In current business
operations, a good example is the involvement of humans in decision making based on use of
BI tools to create new sales strategies, incentives,
regional price corrections, inventory adjustments,
SCM,
and so on. By integrating
OLTP
systems, sales transaction data stores (database), BI tools and with statistical models based
prediction, minimize risks using ERM, the various systems can correct themselves and
perform all human tasks in an automated manner without human intervention. The goal of
machine learning is to enable systems to constantly
monitor changes, recognize patterns, learn new trends, use statistical models and make
intelligent decisions.
|
|
Some of the basic steps of machine learning do exist in current systems, where code will
analyze missing data patterns and try to manage data. It stops there in terms of sending
warning emails about new trends or change in patterns. The management has to make
manual corrections by performing statistical analysis, many semi-automated or manual
interventions and corrections. As each step of the process is automated using learning code,
we can achieve total automation and create learning machines that work
in an autonomous manner.
|
|
The social media and ecommerce companies have tried to initiate the machine learning
concepts by providing users with related content and products respectively. There are
still gaps in many cases as the predictive outcomes are biased since they model
trends using previous users patterns to promote sales or get higher "like" counts in
social media content.
|
|
Machine Learning In Financial Institutes
|
Financial institutions/banks use machine learning to
prevent fraud. The machine learning
models analyze card holder's buying pattern
constantly. When there is a extremely large amount
of card purchase transaction, it flags the
transaction to stop the transaction. At a high
level following are the ML steps.
|
|
1. Create a ML Model to analyze card holder's
buying pattern including the stores, resturants,
on-line purchases, locations etc.
|
2. Set base thresholds and let ML Model analyze
the pattern continuously and learn new
patterns as time goes by.
|
3. Analyze each transaction based on existing
ML-Model pattern for the customer and
learn/update pattern.
|
4. Flag the transaction as questionable or
fraudulent if it fails the customer
purchasing pattern (amount, location,
method of purchase etc.).
|
|
Further, if the same card was used to
purchase a flight ticket, say from
Philadelphia to San Francisco, it
should now know (by machine learning)
that after
specific date/time there could be
credit card charges in the destination
city - San Francisco and around. Still
ML-Models have to analyze the
transaction. Now the complexities are,
was the ticket one-way or two-way.
This also creates new ways to analyze
transaction with date/timestamps and
locations.
|
|
Another scenario is, if the ticket was bought by
some other card/mobile transaction, since
there may not be joint data exchange between
credit card systems/financial institutions, the
transaction could be flagged as fraud or
questionable. Some card companies call
the customer to provide verification to
validate the transaction at
POS.
All these actions
will be based ML customer-purchasing pattern
analysis.
|
|
Three Key Machine Learning Methods
|
1. Supervised learning - a planned approach with standard input with several
predicted output (in robots, all possible motions or 3D co-ordinates are
recorded and it can operate autonomously using the learnt models and
alogrithms).
Mathematically the best approach to
supervised machine learning is
SVM.
|
|
2. Unsupervised learning - learning constantly in an autonomous manner
with all types inputs (known and unknown) and system is allowed to predict
output based on trends, statistical models and algorithms. The system has
be configured to all possible statistical models and algorithms. Two well
known methods used in unsupervised learning are
PCA
and clustering.
|
|
3. Semi-supervised learning - a hybrid of supervised and
unsupervised learning. A system is configured to all possible statistical
models and learning algorithms. The system is monitored and supervised
learning is applied when responses are not as expected. This type of
learning is best for situations that are too complex or inputs are unknown,
resulting in unpredictable outcomes.
|
|
Related Information:
|
1. IBM Machine Learning
|
2. Looking backwards, looking forwards: SAS, data mining, and machine learning
|
3. SAS Machine Learning
|
4. Azure - Microsoft machine Learning
|
5. Neural Network Overview
|
6. Medical Conceptual Analysis
|