Personality-Aware Link Predictor Boosting, a system to be infused with other topological link predictors to optimise their performance; taking advantage of (computationally recognised) personalities.

palp
├── data
│   ├── network (*)
│   ├── normalised_twitter_personality.csv
│   ├── raw_twitter_personality.csv
|   └── stats.csv (*)
├── gae
├── predictors
├── saved
├── analyser.py
├── helper.py
├── manager.py
└── palpboost.py

(*) not included in repository.

• data:
  • network: Dataset containing Twitter network, obtained from [1]. It should contain a raw and a filtered folder, the former is the raw ego-network edgelists obtained from the aformentioned source; while the latter is a filtered version, where a node is kept if it's found to have a recognised personality.
  • normalised_twitter_personality.csv: Personality file for the Twitter network, normalised between [0-1] for each trait.
  • raw_twitter_personality.csv: Personality file for the Twitter network, as obtained from the trained models.
  • stats.csv: Enumerated following edges, enlisted by personality - used for SPSS analysis.
• gae: TensorFlow implementation of the (Variational) Graph Auto-Encoder model, used for pre-processing and splitting. Please refer to [2].
• predictors: As of 09/05, 5 Link Predictors are incoporated (AdamicAcar, JaccardCoefficient, PreferentialAttachment, SpectralClustering, Node2Vec). Please refer to [3].
• saved: Saved graph objects, train-val-test splits, and calculated centroids containing each personality tendency given the train split for some graph. Filenames are split as follows twitter-{N}-{connected_components}-{p} where:
  • N = it is a subset containing 1/N of the entire graph.
  • connected_components = which kind of connected components are considered (this can be 'strong', 'weak' or nothing at all, to which that would be indicate that it does not filter by components).
  • p = it is a split, where (p*100)% is the percentage of the test data.
• analyser.py: Analyses personality dimensions, followee tendencies, degree count and PALPBoost score effectiveness.
• helper.py: Helper functions.
• manager.py: To execute code from.
• palpboost.py: PALPBoost, containing calls to helper.py to construct the needed centroids, reflecting personality tendencies.

[1]: https://snap.stanford.edu/data/ego-Twitter.html
[2]: https://github.com/tkipf/gae
[3]: https://github.com/lucashu1/link-prediction