Rikiya Takahashi

I am a researcher of data analytics belonging to Mathematical Sciences team at IBM Research - Tokyo. My interest is applying advanced machine learning algorithms for i) modeling complex datasets related to human behaviors, and ii) deriving practical business insights such as marketing expertise.

My research interests are nonparametric Bayesian methods and robust unsupervised structure learning algorithms. Especially, I study several limited classes of statistical models whose parameters can be easily optimized globally, but the optimized models are still sufficient for applying real problems. Basically, many unsupervised learning algorithms involve the local optimality problems as in clustering methods based on k-means, Expectation-Maximization algorithms, and hidden Markov models. We often observe that the poor local optima are certainly possible but unnatural interpretations of the given datasets. Since the local optimality is often caused by too large degrees of freedom in parameter spaces, designing statistical models both whose parameter search spaces are limited and whose expressive powers are still flexible is a key challenge to create reliable and powerful analytics applicable for any kinds of datasets. To put it another way, I always search for the "gaps" between the computational difficulties and practicalities in real applications. While optimization researchers search for the elegant techniques to optimize difficult and intractable objective functions, I am seeking for the simple and tractable objective functions that can be adopted in real applications.

For business perspectives, I aim to apply my research outcomes for marketing research and risk management. In both mass- and direct- marketing problems, structure learning algorithms are effective for modeling and understanding the true behaviors of consumers. One of the crucial issues in marketing is that the behaviors of consumers are fickle and difficult to understand. I believe that firms can acquire high profits by clarifying the underlying mechanisms of customers from responding marketing activities to purchase decisions. Also, quantitative risk assessments in marketing activities are important because several marketing activities involve large risks such as lavish spending of advertisement costs, or long-term damages for the brand reputations. Since the behaviors of consumers are stochastically fluctuated, the predictions of the behaviors inevitably involve errors. Probabilistic prediction models and unsupervised density estimators are ones of the key technologies for such risk-sensitive assessments of impacts of the marketing decisions.

For marketing research with relationship to the risk assessments, I have a personal opinion in which not the firm's but the customer's risk-attitudes are worth considering in designing products and services. Readers who are interested in this issue, please move to this column.

Contact Information

IBM Research - Tokyo
1623-14 Shimotsuruma, Yamato-shi, Kanagawa-ken 242-8502 Japan
e-mail:

Education

• M.S. in Information Science and Technology, 2004, University of Tokyo, Tokyo, Japan
• B.Eng. in Information and Communication Engineering, 2002, University of Tokyo, Tokyo, Japan