Re: Malicious users on recommender system

[Zia mel <zi...@gmail.com>]

@Steven Bourke	

I am using collaborative filtering and thinking to use item-based .

On Wed, Aug 29, 2012 at 10:54 AM, Steven Bourke <sb...@gmail.com> wrote:
> Can you tell us which type of algorithm you are using? Depending on what
> you are using that will affect the answer.
>
> On Wed, Aug 29, 2012 at 7:16 AM, Ted Dunning <te...@gmail.com> wrote:
>
>> First off, it looks like Amazon is not filtering for engagement here.
>>
>> Second, you have to have Amazon's prominence before attacks by large groups
>> of people are worth it.
>>
>> Third, to quote Amazon "these happen once in a blue moon".  That means you
>> can correct for them manually.
>>
>> So pragmatically speaking, this isn't a big deal if you do the basics
>> right.
>>
>> On Tue, Aug 28, 2012 at 11:23 PM, Zia mel <zi...@gmail.com> wrote:
>>
>> > Thanks Ted. If you can please elaborate on this , Let's say for
>> > example I am recommending online books and 1000 users joined and added
>> > most of the popular books to their list and rate them high to be
>> > similar to other users , then they start adding books they want to
>> > advertise , how can I detect this attitude ? and how can I know if
>> > these are malicious users or true users that just have common
>> > interests ? Is there a way that I can solve this case that happened to
>> > Amazon
>> >   http://news.cnet.com/2100-1023-976435.html
>> >
>> > Thanks
>> >
>> >
>> >
>> >
>> > On Tue, Aug 28, 2012 at 8:23 PM, Ted Dunning <te...@gmail.com>
>> > wrote:
>> > > The single most effective thing you can do with malicious users like
>> this
>> > > is to let them think that they have won.  In the ideal case, you can
>> > detect
>> > > simple click frauds and maintain a per user play adjustment so that
>> they
>> > > see the fraudulent stats and everybody else sees the corrected stats.
>>  If
>> > > you can, this should even extend to your leader board pages.  Once you
>> > have
>> > > this, the fraudsters will generally not increase the sophistication of
>> > > their attacks and you have a fairly simple situation.
>> > >
>> > > You also will have a bit of an advantage if you pick a metric that
>> > > indicates fairly serious engagement.  With videos, for instance, I have
>> > > used plays > 30 seconds as the metric and this was handled by a beacon
>> on
>> > > the page while the 30 second delay measurement was on the server side.
>> > >  This requires a browser to be live and in focus for 30 seconds in
>> order
>> > to
>> > > get a play event which substantially increases the cost of committing
>> the
>> > > click fraud on the fraudsters side.
>> > >
>> > > With the recommendation analysis itself, the key is to flatten all
>> > > frequency metrics per user.  With unsophisticated click fraud, the
>> abuse
>> > > will center on creating high play frequencies for a few users which
>> will
>> > > then be counted as a very small input signal since so few users are
>> doing
>> > > it and their high play rates won't matter.  Also, the major effect if
>> any
>> > > will be to simply give the fraudsters recommendations for their own
>> items
>> > > which will make them happy and won't matter to anyone else.
>> > >
>> > > On Tue, Aug 28, 2012 at 6:29 PM, Zia mel <zi...@gmail.com>
>> wrote:
>> > >
>> > >> Hi ,
>> > >>
>> > >> Is there any way to check for malicious users in mahout so I can
>> > >> remove them from the recommendations or reduce their effect ?
>> > >> Malicious users are the ones that want to play with the ratings and
>> > >> increase or downgrade it.
>> > >>
>> > >> Thanks,
>> > >>
>> >
>>