Tweet classification - use of features I have been given the task to classify some tweets per topic. I have done a classification based only on the per-document-per-topic probability with LDA. I have been suggested to use BTM instead, so I am working on that too.
The dataset I work on is made of 3 columns:

*

*the text,

*the username, and

*the day and hour the post was uploaded.

The problem is that during every classification I made or attempted, I only used the text column. I have been told I should use all three, but I have no idea about how to implement this, and I can't see why usernames should be relevant either for a topic classification. I need to mention that I do not have ground truth, so all techniques I can use are not supervised.
I would be extremely thankful if anyone could share his ideas with me or suggest a path to follow.
Thank you in advance.
 A: Recently I come cross following DataSkeptic podcast
Face Mask Usage during the COVID-19 Pandemic
https://dataskeptic.com/blog/episodes/2020/face-mask-sentiment-analysis
(https://arxiv.org/abs/2011.00336)
In which data harvested from Twitter accounts are used.
As far as I remember,  they used date to analyse the trend over time (temporal aspect) and cross check /validate with other twitters. Username (name, surname, bio) is used to determine demographic .
If you haven't see the paper or listened to podcase, it might give you hints in your problem constraints.
A: Based on your description, the problem is more of a clustering rather than classification problem. The username matters because, well, people care about different things (a reporter is more likely to tweet about a political subject rather than say a chef.)
You can use bi-clustering ideas, simultaneously clustering tweets (represented by words, bigrams, etc.) and usernames. "Bipartite network clustering" or "co-clustering" are other good terms to search. Have a look at this paper

*

*Co-clustering documents and words using Bipartite Spectral Graph Partitioning
Some further thought has to go into it because the resulting say word-username matrix is very sparse. But this should at least give some new ideas.
