Really good stuff thank you for all responses. More below: > You are talking about feedback on recommendations, right? When a user > says "that recommendation you presented to me is not a good > recommendation"? I think this is fairly different information from a > user volunteering that "I don't like that item."
This is a good point, I might be naive to think users will actually consider the item being recommended and rate the recommendation on the merits of the item. I might be able to help this by changing the user interface of how we would allow the recommendations to be rated. Rescoring on negative feedback is one way of going by but I really think there should be a way to properly use the data (blending might be an option, I haven't though about). We might be able todo some A-B study once I have enough negative feedback. Is there way today on the map/reduce implementation to do re-scoring? (Chen) If interested, I could probably provide our dataset (anonymized for users and items). Our system is providing recommendations for multiple silos (different set of unrelated items), specific silo we will be getting negative feedback have 120K+ user ratings for 1200+ items. Different silos have different characteristics for example recommendation for urls has alot more items then this specific silo. Ted, I am not sure if I follow what you mean here: > This is also very dangerous because negative ratings often correlated much > more tightly with what people like than with what they don't like. 2010/4/28 补丁象夫 <mahout.c...@gmail.com> > By the way, it's quite a pity that no such dataset is publicly > available with feedbacks on recommendations. As a result, researches > on this topic is quite limited. > > 2010/4/28, 补丁象夫 <mahout.c...@gmail.com>: > > You have two sources of information, feedbacks on recommendations and > > ratings volunteerly provided by the users. Therefore, the model should > > balance those two types of information. > > > > As to how to utilize the feedbacks on recommendations, I think it's a > > learning-to-rank task. Recommended items with negative feedbacks > > should not be taken as irrelevant items, they are only less relevant > > compared with recommendations with positive feedbacks or no feedbacks. > > There are many works on learning to rank, and CofiRank by Yahoo is > > probably the first work to adopt such method for collaborative > > filtering. > > > > > > > > > > > > > > 2010/4/28, Sean Owen <sro...@gmail.com>: > >> You are talking about feedback on recommendations, right? When a user > >> says "that recommendation you presented to me is not a good > >> recommendation"? I think this is fairly different information from a > >> user volunteering that "I don't like that item." An item that a user > >> rates poorly is still, strangely, an item that is relevant to that > >> user. > >> > >> I kind of wrote about this in the book -- let's say you are a > >> hard-core classical music fan. You love Brahms (5 stars) but can't > >> stand the showy Vivaldi (1 star). If I recommended Led Zeppelin to > >> you, you would undoubtedly say it's a bad recommendation. But it is > >> 'bad' in a different way that Vivaldi is bad. Vivaldi is something you > >> knew about already because it was similar to things you do like -- Led > >> Zeppelin was completely foreign and unrelated. > >> > >> Negative ratings like the 1 star that the user volunteered for > >> Vivaldi, therefore, mean some different and should be treated > >> differently than the user's response to the Led Zeppelin > >> recommendation that was pushed at him or her. > >> > >> > >> I would agree with Ted, that negative feedback on recommendations > >> doesn't somehow belong in the data model. You can use it to filter > >> recommendations -- you would want to purge any Led Zeppelin records > >> from recommendations going forward. In the (non-distributed) > >> framework, this is done with a Rescorer. > >> > >> That is, you don't need to or want to add this negative feedback into > >> the model. There's no problem of boolean ratings and such since the > >> association would not exist in the model -- just in recommendation > >> filters. > >> > >> > >> > >> On Tue, Apr 27, 2010 at 6:27 PM, Tolga Oral <tolga.o...@gmail.com> > wrote: > >>> We are building a system on top of provided recommendations allowing > >>> user > >>> to > >>> provide feedback. I have a question on how to apply the negative > >>> feedback > >>> for both boolean and none boolean scores. > >>> > >>> Assuming its a boolean item (item with no score), I cant see how we can > >>> provide negative feedback for a given recommendation. Does this mean we > >>> should convert our boolean items to a score of 1 and give score of -1 > >>> (or > >>> 0) > >>> for recommendations that user didnt like? > >>> > >>> This applies to items with score too I am not sure if the correct way > of > >>> going about this is to provide 0 or a negative value. > >>> > >>> Thank you. > >>> > >> > > > > > > -- > > http://blog.sina.com.cn/apachemahout > > > > > -- > http://blog.sina.com.cn/apachemahout >
