I think it makes sense to incorporate lots of kinds of data into a search based rec engine and then pick which kinds you use by adjusting queries rather than changing data.
On Mon, Aug 18, 2014 at 2:25 PM, Fernando Fernández < [email protected]> wrote: > I agree, l also found that situation where I was requested to somehow > include views in one recommender (Following the intuitive idea that if we > put more data into the algorithm it should work better) but it was > counterproductive. If you are trying to generate purchases, many times it > makes no sense to add or use views data when you already have purchase > data. > > > 2014-08-18 20:33 GMT+02:00 Ted Dunning <[email protected]>: > > > Can you say more about the data you have? How are you processing that > > data? > > > > I also had a situation where "also viewed" performed poorly. The problem > > was actually that viewing is a very poor indication of engagement. > Getting > > a better indication (viewed for 30 seconds) made a world of difference in > > the results. No amount of fiddling with the raw view data itself made > any > > difference. > > > > > > > > > > > > > > > > On Mon, Aug 18, 2014 at 3:15 AM, Sigmund Lee <[email protected]> wrote: > > > > > I used to using Mahout's Log-likelihood and Tanimoto coefficient as > > > similarity algo for this scenario, but the results was not so good. So > I > > > wondering if there are another algos that can be used to fit this > > scenario > > > better? For example, co-occurrences matrix that introduced in Mahout In > > > Action? > > > > > > > > > Thanks in advance. > > > > > > Bests. > > > > > >
