Tracking Dynamics of Topic Trends Using a Finite Mixture Model(KDD 2004)
- Initialize \mu in approprietly.
, and set \mu_i^{(0)} the first x_t s that are different each other.
Fitting Gaussian distribution in streaming documents. Parameters in step t are updated using parameters in step t-1.
make ttfmm
./ttfmm input_file alpha lambda k_max
- alpha[NUM]: Parameter of \alpha( > 0)
- lambda[NUM]: Parameter of \labmda(0 < lambda < 1)
- k_max[NUM]: # of maximum topic( > 0)
Each line is one document.
timestamp \t word_1 \t word_2 \t ...
timestamp \t word_1 \t word_2 \t ...
- Small \lambda causes inf in \lambda^{-(t_{new} - t_{old})}.
- Dont calculate denominator in p(i|x_t) because |\Sigma| sometimes zero.
- Using logsumexp in p(i|x_t) to avoid overflow.