MALIC

Source code and datasets of CIKM-21 paper, Topic Modeling for Multi-Aspect Listwise Comparisons, by Delvin Ce Zhang and Hady W. Lauw.

MALIC is a topic model that can incorporate document comparisons as meta-data to improve topic modeling quality and produce a list of document comparison outcomes.

Implementation Environment

• numpy == 1.17.4
• scipy == 1.3.1

Run

python main.py

Parameter Setting

• -em: number of EM iterations, default = 20
• -qn: number of iterations of Quasi-Newton method, default = 20
• -dn: dataset name, default = country
• -nt: number of topics, default = 30
• -su: s = supervised setting, u = unsupervised setting, default = s
• -tr2: ratio of training documents, 1 = rank aggregation, 0.8 = partial ranking prediction, aspect assignment, and perplexity, default = 1
• -na: number of aspects, None = use the correct number of aspects in the dataset, other positive integers (1 or 4 or 8 or 16) can be set, default = None
• -p: number of documents added to existing partial rankings, 0 (used in the main paper) = use {3, 4, 5}-way, top-{2, 3, 4}, and choice-{5, 10, 15}, other positive integers (2 or 4 or 6 or 8) can be set, default = 0
• -e: number of partial rankings for each length and each aspect, default = 50, change to 5 or 25 or 100 or 200
• -a: alpha, Dirichlet prior, default = 0.01
• -s: sigma, Gaussian covariance prior, default = 0.01
• -r: lambda, regularizer, default = 0.01
• -rs: random seed, we randomly generate 5 different random seeds to run experiments independently, and report both mean and standard deviation in the main paper

Output

Results will be output to ./results file.

• doc_topic_dist.txt contains #documents row, each row is a topic distribution with #topics diminsions
• topic_word_dist.txt contains #topics row, each row is a distribution over #words words

Reference

If you use our paper, including code and data, please cite

@inproceedings{malic,
  title={Topic Modeling for Multi-Aspect Listwise Comparisons},
  author={Zhang, Delvin Ce and Lauw, Hady W},
  booktitle={Proceedings of the 30th ACM International Conference on Information \& Knowledge Management},
  pages={2507--2516},
  year={2021}
}