Project Overview
Our research focus is on data-analytic methods for quantitatively measuring risks and chances,
and for getting the most out of this data for making business decisions.
Our research includes (but is not limited to) maximization of lifetime value beyond myopic views on near-future gain or loss,
but considering the long time perspective, and risk-sensitive decision making methods
that avoid the risks of big losses, or in contrast, justifies risks to seize opportunities.
We are especially focused on marketing optimization as an application area,
and developing new methods for capturing, analyzing,
and making marketing decisions with uncertain data in collaboration with other IBM Research labs,
such as the T.J. Watson Research Center.
We are also deploying our technologies to leading companies in cooperation with IBM Business Consulting Services.
Research Issues
- Cost-Sensitive and Risk-Sensitive Data Mining
- Customer Lifetime Value Maximization
Cost-Sensitive and Risk-Sensitive Data Mining
Taking appropriate marketing actions for appropriate targets or in given business environments
by utilizing data stored in databases is modeled as a cost-sensitive classification problem in data mining and machine learning.
In cost-sensitive classification problems,
evaluation of the results is not the ordinary binary evaluation of success and failure,
but real-valued evaluations representing economic, psychological, or social gains and losses.
Cycle of (marketing) decision making
A commonly used strategy for cost-sensitive learning is to find decision making rules
that minimize the mean loss (or equivalently, maximize the expected profit).
This strategy generally seems to work well "on average", but does not reflect users' attitudes to risks.
For example, a user might avoid the risks of the occurrences of very rare,
but huge, losses. On the other hand,
another user might be willing to take such risks hoping to seize the opportunities for huge profits.
Alternately, in the context of making very important decisions on cooperate strategies,
where several consecutive misjudges might lead to unacceptably huge losses,
the mean-cost approach is too optimistic.
Risk-sensitive learning is the approach that we propose for such risk-sensitive decision making,
making it possible to derive decision rules that reflect users' risk preferences.
We are tackling the cost-sensitive decision-making problems from the view of financial-risk management,
and developing effective and efficient methods that support risk-sensitive decision making.
The idea of risk-sensitive learning is applicable not only to marketing, but also to other contexts,
such as medical diagnosis where a serious medical error has huge economic and social impacts.
Recently, we developed a novel method to find risk-sensitive decision rules
by optimizing the risk metric called expected shortfall (Kashima: SDM 06).
Expected shortfall is receiving considerable attention in the field of financial engineering,
since it captures risks more accurately than the commonly used measure called value-at-risk,
and it also has several desirable computational properties.
Our approach can be called a breakthrough step
in the sense that we bring the idea of aggressively exploiting information about risk to the field of data mining.
Risk-sensitive data mining via minimization of expected shortfall (See
Kashima: SDM 06）
Customer Lifetime Value Maximization
Traditional optimization approaches focused on return and risk use calibrated and fixed parameters.
In an economically dynamic environment,
these parameters change and are not stationary.
Such a parameter change in a harmful direction can be disastrous for a company,
and the company needs to take optimal actions in consideration of the change.
Such a strategy corresponds to a strategy for maximizing customer lifetime value (LTV).
Proposing optimal products or services for each customer will be beneficial for both customers and the company.
The company which tries to maximize customer lifetime values focuses on
both the currently loyal customers and the potentially loyal customers.
The potentially loyal customers are the ones who currently belong to non-high-profit segments
but have the potential to be high-profit customers in the future.
The tight economic environments of recent years force companies to focus on these potential customers.
Maximization of customer value
These customer dynamics are generally modeled by state transition models.
Markov Decision Process (MDP) is such a model that considers
changes of customer psychological state caused by marketing actions.
IBM Research already has solutions for these marketing optimization scenarios using MDPs.
CELM (Customer Equity and Lifetime Management) developed in Zurich Research
CCOM (Cross Channel Optimized Marketing) developed in T.J. Watson Research Center.
Tokyo Research Laboratory is developing this approach to strictly evaluate the risks in decision making,
and to find optimal policies considering the return-risk tradeoff.