Since the middle of the 20th century, various types of automatic music
composition
have been attempted. Usually, these attempts utilized random number or
stochastic information to generate music.
In these trials, the researches focused on the automatic generation of
music.
Indeed, their trials are important for art and science. However,
from the viewpoint of entertainment,
it is more important to generate music by ourselves with the proper
assistance or guidance from a computer.
We are now developing a symbolically rich interactive music composition
system, the Augmented Composer, that uses atomic music patterns
and controls their duration, pitch, and
velocity to generate music. Its interface is quite simple and easy, so even
a non-expert can compose ``music''.
Through analysis of
composed music (MIDI data), we found an explicit
difference between music by expert composers and novices composers.
Therefore, We propose a music composition support
(suggestion) system from the viewpoint of active mining.
Our future goal includes a method to stimulate the users' creativity by
concepts such as chance discovery.
Coming soon.....
- Rule Acquisition from small and heterogeneous data set
If we analyze the data (perform a data mining) that is collected during a medical care, it is slightly difficult to obtain proper results. The reason for the difficulty is followings;
The data is not collected with an aiming to obtain a standard value set for a medical diagnosis but for a medical diagnosis of the special patients. In addition, since the type of medical inspection depends upon patient's health status, items and frequency of the medical inspections are different according to the patients. For example, to not so serious patients, usually, only medical inspections which an insurance can be applied will be performed, and a frequency of medical inspection is not so high. On the contrary, to serious patients, all the necessary medical inspections will be performed even if they are quite expensive, and a frequency of medical inspection is quite high. Thus, the data collected during medical care usually contains various types of data and deviation. In addition, since the amount of data that can be collected in one year is not so large, it will not be sufficient to perform data mining on the data set. Even if the amount of a data set is sufficient, a lot of data is missing and the data is heterogeneous.
When we perform data mining with using such a imperfect data, we cannot obtain any proper result that can be accepted by the doctors. Even if we can obtain a proper result, the generated result tends to have a feature of serious patients. And the result is changeable (less robust). If we add one record to the data set, the result analyzed from the new data set will be different.
In order to obtain a proper result as possible from an imperfect data, we propose a method to increase data and modify the imperfect data to semi-homogeneous data. As a result, the result is as standard as possible, acceptable by doctors and less changeable. In fact, we adopt the method that obtain Gaussian distribution from the original data set, and based on the distribution, the data becomes increased. By the method, we can obtain a data such as 1) increased, 2) as standard as possible, and 3) less effected by eccentric data.
We obtained modified data set such as followings;1) 
Original 
After modification
2) 
Original
After modification
3) 
Original
After modification
1) The results from the origical data is also robust, so it may not be necessary to increase data set.
2) A robustness of generated decision trees is upgraded.
3) No effects on a robustness generated decision trees can be observed.
Thus, especially for the data set such as 2), the proposed method effectively works.
