Updating an unsupervised model but retaining similarity In my example I am using topic modelling (specifically a version of LDA) although I think avenues for exploring this could relate to other unsupervised techniques like clustering.
I train a model and get a good set of topics (or clusters) relating to the data. Data continues to be produced, and after a while I want to update the model so it's reflecting changes in the data. If I train a new model with no reference to the existing model, there's some chance it will pick up on similar topics/clusters to my existing model, but it isn't totally guaranteed this will be neatly done, particularly when I need to define the number of topics/clusters upfront. So if the existing model had 10 topics (and did a good job of reflecting the actual topics in the data), and in the intervening time an 11th topic has started coming up, if I direct the new model to use 10 topics, it will perform less well than if I fed it 11 topics - but I won't know that before training them. After training them then I need a way of spotting that the 11 topic model has better continuation of existing topics and the additional topic is a distinct and useful addition.
I know about topic models that include time as a variable (Dynamic Topic Models, Topics Over Time), but they are still run at a fixed point in time. What I essentially want is to be able to add new data to train several models (mixed with more recent old data) and identify the model that is the best continuation of the previous model. I've tried comparing old and new topics using topic diff and while this is helpful for getting a sense that topics are overlapping, I'm not seeing how to use this to differentiate between model options without having this pretty directed by the number of topics.
I can't find anything about this kind of problem, but I feel sure it must be relevant to a lot of work so I might be searching for the wrong thing or thinking about it the wrong way round. Like I said it could be that the answer isn't specific to topic modelling. It's really about having an approach that will accommodate/recognise changes in the data without totally disrupting continuing patterns.
Thanks for your help.
 A: I have the same question and can't find any answers. All related questions on Cross Validated stay unanswered (here or here).
My idea for keeping topics from previous years while being able to capture new topics of the current year is as follows:

*

*Fit a separate model for each time slice (e.g., years)

*Find matching topics between years using topic similarity measures

*Regard topics with high similarity as "the same topic" and topics
with medium similarity as "subtopics" or "related topics".

The result would be groups of topics from different years. These groups might be connected (e.g., different topics merge over the years or one topic splits into subtopics) or may be incomplete fore some years years, i.e, no matching topic in a specific year as this topic simply was not addressed in this year.
Of course, fitting a separate model is time intensive, but it guarantees that the topics by year remain during every "update". Different numbers of topics in each year might better reflect topic dynamics than a fixed number over all years.
Regarding classification of topics as being "the same" or "similar", thresholds analogous to He et al. (2009) need to be defined. To this end, human coders might rate similarity and their ratings could be compared to similarity measures in order to determine thresholds.
This is just an idea, but in principle, this approach could be applied to different topic modeling variants and update periods.
