how to use Xt_to_t1 to find the topic weight distribution for higher layers?

[bhomass]

for layer 1 (bottom layer), Xt_to_t1 is

3x1 cell array
{20x9548 double}
{ 0x0 double}
{ 0x0 double}

there is no doubt how to derive the topic distribution of a single document here. But for higher layers there are multiple sets of wights. e.g. for layer 3

3x1 cell array
{20x9548 double}
{20x9548 double}
{19x9548 double}

I can't figure out how to use all three sets of weights to calculate the topic distribution. Especially when the last one is only 19 topics, instead of the 20 I specified.

Anyone figured this out?

[ethanhezhao]

Hi, Phi has three layers because T = 3. Phi{1} can be usually used. T is the number of layers of GBN. T can be set to 1 if multi-layers of topics are not needed. The model will remove inactive topics in higher layers. Thus, it may have fewer topics.

[bhomass]

Thank you for the response. Yes, of course there are 3 Phi if T=3. But I am not looking for word distribution of topic K. I am looking for topic weights for a document j. From the code, I deduced (not 100% sure) that Xt_to_t1 represents this topic weight. Why should there be 3 sets of that, even if its for t=3.

[ethanhezhao]

This is because of the implementation of GBN. The variables in ParaGlobal{T} shall be used and others are just for internal model training. Please refer to the GBN paper and code for details.