As I understand it, what Mugoma is asking about has nothing to do with
filtering or rescoring candidates (and apparently he is already using
IDRescorer in other settings to do that).
He seems to want to exclude ratings when calculating the user- or
item-similarity (or whatever approach he is using), which almost by
definition does not include the candidates (at least in the case of
user-similarity).
HTH,
Jens
On 05/09/2012 03:18 PM, Sean Owen wrote:
In that case -- the rescoring never operates on "original items". It is
rescoring only estimated ratings.
If you supply no rescorer, no filtering or rescoring happens.
You would never delete data to make an item unrecommendable for one query,
because you would be totally deleting the data!
I hope this finally clarifies -- something like this could happen:
There are 9 items in the world: 1 2 3 4 5 6 7 8 9
User A expresses a rating for 1 2 3, so only 4 5 6 7 8 9 are recommendable
The framework further selects as possible candidates 6 7 8 9
The filter removes 6 7, leaving 8 9 as possibilities<-- THIS IS FILTERING
The recommender algorithm predicts a rating of 3.5 for 8, and 3.2 for 9
The rescorer changes the prediction for 9 from 3.2 to 4.1<-- THIS IS
RESCORING
The final recommendations are 9 (score of 4.1), then 8 (score of 3.5)
To say it a third time: IDRescorer already does *exactly what you are
describing*!!!
On Wed, May 9, 2012 at 11:33 AM, Mugoma Joseph Okomba<[email protected]>wrote:
By 'original items' I mean the items in the database (just the raw table
rows). In the example I gave original items are 10. Actually should be
'original ratings', not 'original items'
If you look at CachingRecommender it has 2 recommend () methods: one with
IDRescorer and other without. My understanding of this is that the one
with IDRescorer has *recommended* items altered by IDRescorer. The one
without doesn't. So, the IDRescorer works on *recommended* items and not
*original* ratings.
The confusion is not about filtering parse but what's being filtered.
For the example I gave I could do an SQL to delete the 'unwanted' ratings
then have a clean set of ratings to feed into the recommender so that the
recommender sees 7 ratings instead of 10. But this doesn't look intuitive
so I thought there's a better way of handling this within mahout.
Probably what I need is a new data model that overwrites
getPreferencesFromUser(long id) and getPreferencesForItem(long itemID)
On Wed, May 9, 2012 12:25 pm, Sean Owen wrote:
What do you mean "original items"? The user's preferred items are already
not candidates for recommendation, but that is nothing to do with the
rescorer. It operates on all *candidate* items, *before* scoring.
What is your distinction between filtering *recommended* items and
*original* items? Either way it is filtered. I don't understand what you
are getting at.
On Wed, May 9, 2012 at 10:09 AM, Mugoma Joseph Okomba
<[email protected]>wrote:
If it's true that IDRescorer works on original items then that's both
bad
and good news for me.
The bad news is is that all the code I had previous written involving
IDRescorer is all bugy since I had assumed that IDRescorer filters
recommendations and not original list
The bit of good news is that I don't have to anything for the new task.
But, if IDRescorer changes original list, what can be used to change
recommendations?
