home |
copyright ©2016, tim@menzies.us
src |
chat
<td align=center><img src="img/lectures.gif">
</td><td align=center><img src="img/review.gif">
</td><td align=center><img width=64 src="img/books.png">
</td> </tr>
<tr>
<td valign=top xwidth="100px">
</td>
<td valign=top xwidth="100px">
News
|
Tutorials
|
Workshops
|
Cool stuff
|
|
|
Goals: this code is a reference implementation of the ideas in my talk Data Science2 = (Test * DataScience). This is a combined data mining/optimization tool kit that is
- Fast: - All algorithms near linear-time, not RAM hogs;
- Light: - Small memory footprint;
- Explicable: - Can offer a succinct human-understandable presentation of what it has learned;
- Actionable: - Comments not just on what is but also what to do (and when we say "what" and "do", those statements understand local practicalities like what is observable and what is controllable);
- Goal aware: - Mulitple goals = no problem, different goals = different models;
- Stable: - Offers not just point solutions, but decision regions inside of which we can confidently expect some effect
- Humble: - Offers a cerfication envelope where all conclusions come with a note saying "you should (not) trust me since I have (not) seen this kind of thing before";
- Context-aware: - Knows how local parts of the data can lead to different models; knows how to find different contexts;
- Sharable: - Knows how to transfer models/data between contexts;
- Privacy-aware: - Can hid data from individuals while preserving trends in the whole population;
- Self-tuning: - Can do so, very quickly;
- Incremental: - Can update old models with new ideas.
That's "all". Should not take more than a few Ph.D.s to finish.