"Tag" API about myrrix-recommender HOT 22 CLOSED

A first pass at this change has been committed. It is fairly complete, but 
still in progress. Testing welcome. Look at the new "setUserTag" and 
"setItemTag" methods.

What would be the effect of using such methods? Say, for example if I tag the 
items with the category to which they belong. 

Also, can one item/user have more than one tag? (I assume they can)

These behave just like users and items, really. For example, tagging a bunch of 
items as "shoes" has an effect as if some new users (named "shoes") interacted 
with all those items. Similarly for user tags. The difference is that these 
tags are not returned in results, clusters, etc. Yes anything can have many 
tags.

It is really just putting a nicer API around the same mechanism that already 
exists, for this kind of use case.

Could this tagging system be potentially used to establish a set of candidate 
items to recommend from? For example, recommend items to user X from items 
tagged as "shoes"?

Not yet. I'm trying to keep the model from becoming a partial database. At the 
same time this is a natural type of feature. It will depend somewhat on how use 
cases turn out, what the real commonly needed functionality is.

Hey Sean, are multiple tags per user or item allowed?

Yes, think of tags as just like items or users. You can interact with many, and 
several times.

Hi Sean,
I updated the PHP client two days ago. But I am not still exactly sure about 
how this tag API is used into the recommendation process? Does it allow to do 
item recommendation based on similar tags as complimentary to based on 
interaction with users ?

This is great -- I've also linked to the PHP client from the web site.

If you think of a user tag as like a "fake item", and then think of this as 
just a superficially different version of setPreference that takes a String, 
then that's 90% of it. Tags never come back out of the system, but they do 
inform it about similarities.

Thanks for the explication. I think I understand better know.
I ran some test with the Tag API. And I told myself, if the model considers 
item's tags as "fake users", I should be able to build a model without users, 
but just with items and tags. So I fed myrrix with the Movielens dataset just 
using movies and genres (each genre converted as a unique number). 
Unfortunately the model wasn't able to compute.
Can you tel me what is the reason behind that ?
Again, thanks for your amazing work.

What do you mean by it wasn't able to compute, didn't finish a model?

Without the StackTrace or the log file, and by looking at the u.genre file for 
the 100K Movielens data set, I can see that there are only 19 unique genres. My 
best guess is that is related to the default number of features Myrrix uses, 
which is 30. Setting the model.features system variable to 19 or below should 
fix the issue.

In this page there is some information on how to do that: 
http://myrrix.com/documentation-serving-layer/

Yes it depends on what the problem was and what the input was. If it low 
dimensionality, yes, you set -Dmodel.features= something lower. the error 
message even shows the flag and the suggested value. Which is why I suspect 
it's something else.

I recreated the data as it was described and I get this:

Fri May 03 13:57:20 CEST 2013 WARNING 30 x 30 matrix is near-singular 
(threshold 0.001); add more data or decrease the value of model.features
Fri May 03 13:57:20 CEST 2013 INFO QR decomposition diagonal: 
[-0.801454214821591, -0.8244992489662669, -0.8127366641701587, 
-0.9010884706991794, -0.7539391390844163, -0.8165742209074729, -0.694458913
9108499, -0.8730410465268026, -0.6132681585565684, -0.5577616105081729, 
-0.6337528980108209, -0.7292905497474778, -0.9539175256073287, 
-0.29446127227258395, -0.730798190947223, -0.6293325894976258, -0
.4385896630760355, -0.39040701839108044, -0.25396291917620395, 
-4.5640997547171797E-8, -5.926015839003651E-8, 1.478182807027981E-8, 
2.476861001015895E-8, -1.5676323782489018E-8, 1.3736249407832068E-8,
 1.6790686826589384E-8, -1.613689043258607E-8, 1.206001330790155E-8, -5.305474385237009E-9, -4.695006275691013E-9]
Fri May 03 13:57:20 CEST 2013 WARNING Suggested value of -Dmodel.features is 
less than 19
Fri May 03 13:57:20 CEST 2013 WARNING Unable to compute a valid generation yet; 
waiting for more data

Ok, let me retry it, I'll give you the stack trace if it fails. Thanks for your 
quick answers.

Thanks, I works like a charm :)
Sorry, I should have read more carefully the logs suggesting to add 
-Dmodel.features=18 to my command line. Now I have the movielens dataset 
working quite great with just genres.
@Sean, do you share any insight about your roadmap ? I would be very happy to 
anticipate your releases and keep the php lib in sync.

Makes sense, yeah. 

There are no API changes in the works at this point; it's all pretty frozen for 
the RC1 and 1.0 release.
The roadmap changes are more related to the infrastructure and deployment, and 
business model.

Later there will be things like a classifier platform, which will be a separate 
product.

Thanks for the heads up!
I can't wait to see your new classifier platform. Do you have a schedule in 
mind?
Also, I know this is not the place, and I have been asked quite a few time, but 
do you plan to include ways to enforce recommendation/similarity based on a tag 
value. For instance, to ask a recommendation for a user only in movies tagged 
with the genre 'comedy'.

Yes I suppose that if there is any API change planned, it would be to make the 
filtering capability more of a first-class API call, rather than something 
coded up in a rescorer. That would be somewhat significant, and wouldn't happen 
until a version 1.1 or something. 

The classifier is still just in planning. It would be a distributed decision 
forest implementation. Input would again be something like CSV: 
"feature1,feature2,...,target" and queries would entail specifying some of the 
features. Again pretty simple API-wise, probably complicated to implement.

Ok, thanks for all. I'll keep in touch to see those new features :)