A semi supervised LDA

Why a semi-supervised LDA ?

In many realworld problems, we want to classify document into defined classes. However, building a large corpus can be very costly. LDA is well known algorithm to discover topics in corpus but it. So why we don't learned a LDA that its topics will aligned to our defined topics? My idea is pretty simple. We use pre-defined keywords for each topic to initialize some prior distribution. Better initialization often leads to a better convergence, which is well-known in high dimensional space data.

Requirements

This model is adapted based on Gensim's LDA implementation. I developed this algorithm several years ago, so it might work only with old version of gensim. You need to run following command to install requirements:

$ pip install gensim==0.12.4

Keyword preparation

You need to define some keywords for each topic. See sample for the format of this file.

Training a model

Runing run_lda_v2.py to train a new lda model. The full arguments are as bellowing:

usage: train_lda_v2.py [-h] [-k NUM_TOPICS] [--passes PASSES]
                      [--iterations ITERATIONS] [--eval_every EVAL_EVERY]
                      [--path PATH] [--min_tf MIN_TF] [--max_df MAX_DF]
                      [--vocab_size VOCAB_SIZE] [--kw_file KW_FILE]
                      [--threads THREADS] [--chunksize CHUNKSIZE] [--new]
                      --model_path MODEL_PATH [--build_dict] [--dic DIC]
                      [--tfmod TFMOD]

Training lda model

optional arguments:
 -h, --help            show this help message and exit
 -k NUM_TOPICS, --num_topics NUM_TOPICS
 --passes PASSES
 --iterations ITERATIONS
 --eval_every EVAL_EVERY
 --path PATH
 --min_tf MIN_TF
 --max_df MAX_DF
 --vocab_size VOCAB_SIZE
 --kw_file KW_FILE
 --threads THREADS
 --chunksize CHUNKSIZE
 --new
 --model_path MODEL_PATH
 --build_dict
 --dic DIC
 --tfmod TFMOD

Disclaimer

This algorithm works with my case, but it might not work with your case. And I won't support any technical problems related to this library. So use it at your own risk.