1.test on four square data
a) Binary matrix
b) Matrix counting in the times
c) Hybrid Model
d) consider the context (for example:time, location)
2. Similarity metrics:
since we are going to use "Friends' checkins" to predict, it is important to choose friends. And we wnat to see whether there are some biases in exsting similarity metrics.
We plan to use artifiicial matrix to evaluate the metrics.
3.Comparision between old-fashioned CF and MF:
ajust the parameters and see the gap between the best of old-fashioned CF and MF.
Updated Aug,8th
After reading some papers, I began to believe that there is little possibility that CF will outranged MF, so the point why we are experimenting on this may be that we want to know the probabilities that a user goes to a place he has gone to before. How much is the influence from his similarity friends in this decision?
But still what is the point in doing this? Is their any possibility we could dig something here?
And of course, if we are going to test the possibility that a person go to some place he has gone to before, then the construction of the train and test data matrix may be a little different from what we did before. We will not use the leave-out-k method, insted we will seperate the data by a time line.
Here is something on the data split by time:
http://scikit-learn.org/stable/modules/cross_validation.html
But I do not see much worth looking into in this direction. So I will put it in on hold and go to the next step to see the domain specific biases.
4.Domain Specific Biases:
We preprocessing the train data considering some geographical factors. There are already some works in it, some build a muticenter joint model which consider the user interest and the geographical influence seperately and multily the two probablities int the final step.
What we trying to do here: filter some negtive examples, and one step further strengthen some positive examples
for example:
A user in Enschede does not go to a chinese restaurant in Amsterdam does not mean that he does not like chinese restaurant, that is to say, it should not be cosidered as a negative example.
Besides, ther should be some positive example which should be strengthened. If a user go to a place far from his activitiy center(say Rotterdam) and he goes to a chinese restaurant , it probably means that this user does have a strong favor for chinese restaurant.
So it became some kind of using data far from activity center here?
Note: the blue color are the content written latest.