2013/10/9 Eustache DIEMERT <[email protected]>:
> Another issue that comes from times to times is the fitness of the sklearn
> API wrt to recommendation tasks.
>
> I believe it's pretty good if one has to manipulate - e.g.factorize - (item,
> user) matrices, but it falls short when dealing with explore/exploit
> scenarios.
>
> An example of that is the bandit [1] family of algorithms, where one knows
> the payoff of an action iff the action is chosen by the algorithm as the
> next step.
>
> [1] http://en.wikipedia.org/wiki/Multi-armed_bandit

I agree, and evolving the API to address the multi armed bandit /
reinforcement learning kind of task is probably out of the scope of
the scikit-learn project (at least for in the short and medium terms).

Something that would help for the recsys / personalization kind of
applications though would be to make it easier to address
learning-to-rank problems for instance using pairwise reductions as
blogged by Fabian some time ago:

http://fa.bianp.net/blog/2012/learning-to-rank-with-scikit-learn-the-pairwise-transform/

Note that pointwise regression models such as
GradientBoostedRegressor, ExtraTreesRegressor or even penalized linear
regression models with suitable features would already work as good
baselines to predict click through rates or relevance scores to rank
query (recommendations) provided that you can collect to this kind of
supervised / feedback signal from a production recsys.

-- 
Olivier
http://twitter.com/ogrisel - http://github.com/ogrisel

------------------------------------------------------------------------------
October Webinars: Code for Performance
Free Intel webinars can help you accelerate application performance.
Explore tips for MPI, OpenMP, advanced profiling, and more. Get the most from 
the latest Intel processors and coprocessors. See abstracts and register >
http://pubads.g.doubleclick.net/gampad/clk?id=60134071&iu=/4140/ostg.clktrk
_______________________________________________
Scikit-learn-general mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/scikit-learn-general

Reply via email to