jiqicn
commented
11 months ago After checking the paper and the implementation, I found that this is expected by the algorithm itself.
First, we found that the increments happened on those arguments that received no attacks. In this case, their parent vectors should be zero vectors, and thus their aggregated strength in each step should be 0.
Then by looking at the influence function given in the paper:
given aggregated strength equals 0, the output of influence functions will always be something equal to the weight of the input argument. This means that the algorithm forces the updated strength toward the weights given. In case an argument is assigned an initial strength smaller than its weight, an increment can be observed.
I tried to change the initial weights of "uniform" method to 1 instead of 0.5, and observed something that makes more sense, since all argument weights should be smaller than 1:
With an empty support link set given and the strength vector initialized using the "uniform" method, increments were observed for some of the arguments in the stock example. This is anti-intuition as no support should result in no increment on argument strength.