Aside from the NewsAPI API key, all relevant data is included in this repository.
The dataset is frozen. Running make dataset with a live dataset would update the current dataset in datasets/data.json; however, the dataset was frozen to allow for consistent experiments.
All preprocessors are stored in featurizers/ and preprocessed data in featurizers/featurized_data/. To run a given featurizer, run make featurize/<featurizer name>, or make featurize/all to run all preprocessors.
Due to the way our infrastructure was designed, models were not independent of preprocessors. So, a given model file in models/ combines a clustering model, its parameters, and which preprocessed data it ingests. Running any models/<model name>.py trains and saves the model into models/models/. Note that the biterm and tkm models may not be runnable because they require specific installation and setup from https://github.com/markoarnauto/biterm and https://github.com/JohnTailor/tkm, respectively (see the Miscellaneous section for more info).
If using the Makefile, run make model/<model name> to train a given model or make model/all to train all 10-cluster K-means models.
First, for quantitative analysis of the clustering models, run python analyzers/cluster_score.py <model name> (or, to run it on across K-means models using different cluster numbers, run make clusterscore/clusternums). Then, for qualitative analysis, run python analyzers/visualize.py <model name> (or, again, make visualize/all for 10-cluster k-means visualization). Output scatterplots use the first three PCA dimensions and are stored in analyzers/graphs/.
Then, to analyze the results of the keyword extractors, cluster keywords can be examined by running python analyzers/group.py <model name> <extractor name> (or for 10-cluster k-means: make group/all extractor=<extractor name>). The cluster keywords are stored in groupings/. The effectiveness of the cluster keywords can be quantitatively evaluated using python analyzers/regression.py <model name> <extractor name> (or, again, make analyze/exp10 for 10-cluster k-means).
After running make analyze/exp10, the difference values were manually copied into manual/accuracy_diffs.json and manual/graph_acc_diffs.py was run to produce the manual/accuracy_differences.png.
Both repositories mentioned earlier were git cloned into vendor/ and pip install . was ran from within the repository home directory. A Cython environment was used for installing and training both biterm and tkm models.