mahout-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Sebastian Schelter <...@apache.org>
Subject Re: [mahout-user] Recommendation engine to support a LAMP application
Date Tue, 07 Dec 2010 20:20:37 GMT
Hello Jose,

I think Mahout has everything you need. All you have to do is to record
user-item interactions after which you can use one of our many recommender
implementations. If you wanna do recommendations based on similar user
tastes you can use a userbased recommender, if you'd like to look at
item-similarities you can use an itembased recommender. It's very likely
that you need to play around a little with both to see which one fits your
needs best.

You can also implement a custom similarity measure that takes content
similarities like matching descriptions into account and rescores the
computed similarities based on that. Last but not least connecting the taste
webservice to PHP shouldn't be a problem either.

I suggest you get a copy of "Mahout in Action" from
http://manning.com/owen/and read the recommender chapter to get a more
detailed impression of what
Mahout currently offers and then talk with your customer to decide whether
it fits your needs.

--sebastian

PS: If you decide to use Mahout, don't forget to put your application on the
Powered By page! :)

2010/12/7 José Moreira <zemanel@zemanel.eu>

> Hello,
>
> i was asked by a friendly company to improve their current product
> recommendation system which is currently a couple of complex SQL queries
> using fulltext search that match similar products based on title,
> description and tags (a single comma separated field) and i don't have
> experience with recommendation engines.
>
>
> The new system must also take into account the click rate of a product
> (in regard of similar ones) when it's being recommended, meaning that
> when products B and C are presented as similar products of A and the
> user (a web visitor, the application has no user accounts) clicks on B,
> it's click rate in regard of A is increased and may influence the next
> recommendation of similar products of A which orders the result by
> similarity and click rate in relation to the displayed product, which i
> think is called item-to-item recommendation.
>
> Eventually, similar personal "tastes" will also need to be implemented.
>
> As a note, my customer is also interested in obtaining for example 8
> products from this set plus 2 more with a low click rate to give them a
> chance of obtaining a higher click rate.
>
> It's also required to be able to configure the weights of the
> "variables", for example, 0.5clickRate+0.3description and so on.
>
> Instead of starting to hack a custom solution i've been looking for
> recommendation engines in PHP (the database is mysql) but haven't found
> much.
>
> I'm also aware of Apache Mahout, i could suggest the use of Mahout or a
> similar java engine as a "web service" on a separate server, although i
> know the customer is very much prone to a "pure" PHP solution, as their
> applications are very basic stack-wise. The end solution will possibly
> be implemented on other similar projects, so perhaps the standalone
> server would be running multiple instances.
>
> Additionally i don't have experience with Mahout or recommendation
> engines in general, so i'm currently not aware how they manage the
> "related click rate" feature. My questions are:
>
> Any suggestions on implementing these features as a stand-alone server
> or otherwise? I apologize if i haven't described the situation using
> more direct technical terms.
>
> Thank you
>
> --
> http://zemanel.eu
> http://github.com/zemanel
> http://pt.linkedin.com/in/josemoreira
> http://djangopeople.net/josemoreira
> irc://zemanel@irc.freenode.net
>
>

Mime
  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message