lucene-java-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Mathieu Lecarme <math...@garambrogne.net>
Subject Re: Several questions about scoring/sorting + random sorting in an image/related application
Date Fri, 15 Jun 2007 21:53:06 GMT
Compass use a trick to manage father-son indexation.

If you index "collection", with a fields Date, wich are the newest  
picture inside, and putting all picture's keyword to it collection?

Then, with a keyword search, you will find the collection with the  
most tag occurence number and date score.
It could be quick and clean, each night, you recompute score with age  
factor, your magic bonus and th random stuff. Iterating over only 300  
things.
The problem is if your photograph knows how it works and cheating  
with uploading every day a new silly photo, just for boosting all  
their collection.

If age of picture in a same collection are not too differents, this  
trick can be magic, if new pictures are uploaded often, it will  
generate silly score. Maybe with small pack of picture with near date  
and collection score indexed and merging the firts same collection  
part after sorting?

M.

Le 15 juin 07 à 22:33, Antoine Baudoux a écrit :

> Well maybe i didnt explain my problem very well. I have a database  
> with over 3 million images, with each image belonging to one out of  
> 300 possible collections. A query could return more than 100.000  
> images (for example if they search for a popular  image keyword).
>
> I want to sort my result with a combination of image date/ 
> collection scoring : Recent images with high-score collections come  
> first, then recent image with lower collection. As the user  
> navigate through the images, less recent images start to appear,  
> also sorted by collection score.
>
> So imagine i make a query for "nature". This query would return  
> 100.000 images. Your sugestion is that i dont make any attempt to  
> sort the images in Lucene.i just make a query with no sort. I would  
> then need to load the 100.000 rows from my database, then sort  
> those 100.000 rows with my custom-defined ordering. Are-you sure  
> this method would be faster than custom Query, or a ValueSource  
> query? Since lucene already indexes and caches field values, I'm  
> not very sure.
>
>
>> Walt explain differently what I said.
>> Lucene can be efficiently use for selecting objects, without  
>> sorting or scoring anything, then, with id stored in Lucene, you  
>> can sort yourself with a simple Sortable implementation.
>> The only limit is that lucene gives you not too much results, with  
>> your 300 maximal responses, you can play with it easily.
>>
>> M.
>> Le 15 juin 07 à 19:07, Walt Stoneburner a écrit :
>>
>>> Antoine Baudoux writes:
>>>> I want to be able to give a score to each collection.
>>>
>>> Keep in mind, Lucene is computing a score based on quite a number of
>>> things from how often a term is used in a document, how often it
>>> appears in the collection of documents, how long the query is, etc.
>>>
>>> If your concept of a document's score changes, then I'd be  
>>> inclined to
>>> think you're possibly using Lucene in a manner it wasn't designed  
>>> for.
>>> That said, I have two thoughts.
>>>
>>> THOUGHT ONE
>>> Use Lucene to locate "records" for you --- what you really are
>>> interested in getting back _from Lucene_ is the primary key.  Then,
>>> use this key to do a lookup in your database of the score of the day
>>> and sort accordingly.  The idea is that Lucene finds, your table
>>> scores, and because of that you won't need to re-index when  
>>> something
>>> changes.
>>>
>>> THOUGHT TWO
>>> Use boosting.  COLLECTION_ONE^5  COLLECTION_THREE^10  etc.  That way
>>> /if/ the Lucene document appears in the collection, it's score is
>>> weighted according to your preferences.  You're free to change the
>>> boosts on a query-by-query basis without having to re-index.
>>>
>>>
>>>> I can use a Very big ... query ... I am afraid that it will be  
>>>> slow.
>>> Try it.  I think you'll find Lucene is _fast_.  We do some pretty  
>>> HUGE
>>> and complicated queries and Lucene just screams.
>>>
>>>
>>>> I can add another field to each document, containing a computed
>>>> custom score, then i could sort on that field. But i want to avoid
>>>> this solution at all costs, since it would mean re-indexing all the
>>>> documents each time the collection scores change.
>>> Or, use indirection - instead of keeping the score, keep the primary
>>> key of a score table.  Then in a database, where speed won't be the
>>> issue, perform the look up.   Honestly, if you're only got 300
>>> categories, you could keep that simple table in memory using less
>>> space than a small text file.
>>>
>>>
>>>
>>>> I would also like to implement random-sorting. ... Is it a good  
>>>> solution?
>>>> Is there another way to do it?
>>>
>>> This really, really, really feels like you're force fitting  
>>> Lucene to
>>> do some business logic piece of a larger application.  May I be so
>>> bold as to ask what's the _actual_ problem you're trying to solve.
>>> ("I'm trying to make a hole in a piece of oak" as opposed to "What's
>>> the best way to sharpen a Phillips screwdriver enough to cut wood?")
>>>
>>> Keep in mind that the forum is for Lucene, so parts of your  
>>> questions
>>> may be answered outside of the forum.
>>>
>>> -wls
>>>
>>> -------------------------------------------------------------------- 
>>> -
>>> To unsubscribe, e-mail: java-user-unsubscribe@lucene.apache.org
>>> For additional commands, e-mail: java-user-help@lucene.apache.org
>>>
>>>
>>
>>
>> ---------------------------------------------------------------------
>> To unsubscribe, e-mail: java-user-unsubscribe@lucene.apache.org
>> For additional commands, e-mail: java-user-help@lucene.apache.org
>>
>>
>
>
> ---------------------------------------------------------------------
> To unsubscribe, e-mail: java-user-unsubscribe@lucene.apache.org
> For additional commands, e-mail: java-user-help@lucene.apache.org
>
>


---------------------------------------------------------------------
To unsubscribe, e-mail: java-user-unsubscribe@lucene.apache.org
For additional commands, e-mail: java-user-help@lucene.apache.org


Mime
View raw message