Skip to content

EvalNE v0.3.2
Compare
Choose a tag to compare
@Dru-Mara Dru-Mara released this 20 Apr 15:25
· 22 commits to master since this release
v0.3.2
b3c1b0a
This commit was created on GitHub.com and signed with GitHub’s verified signature. The key has expired.
GPG key ID: 4AEE18F83AFDEB23
Expired
Learn about vigilant mode.

EvalNE v0.3.2

Release date: 20 Apr 2020

Main improvements since v0.2.2 release:

  • Old Evaluator class is now LPEvaluator
  • Added Network Reconstruction evaluation (NREvaluator)
  • Added Node Classification evaluation (NCEvaluator)
  • Train/validation splits are now required when initializing Evaluator classes
  • Added 3 new algorithms for computing train/test splits. One extremely scalable up to millions of nodes/edges
  • Improved error management and error logging
  • Edge embedding methods are now always tunned as method parameters. Results for the best are given.
  • For link prediction and network reconstruction, the user can now evaluate the methods exclusively on train data.
  • Added Scoresheet class to simplify output management
  • Export results directly to pandas dataframe and latex tables supported
  • A new heuristic for LP named all_baselines has been added. Generates a 5-dim edge embedding by combining the existing heuristics [CN, JC, AA, PA, RAI].
  • Automated file header detection (in the output of embedding methods) is now a function
  • Functions for reading the embeddings, predictions and node labels have been added
  • The user can now set a timeout for the execution of each method in the conf files. E.g. TIMEOUT = 1800
  • Conf files now support any sklearn binary classifier in the LP_MODEL variable. E.g. LP_MODEL=sklearn.svm.LinearSVC(C=1.0, kernel=’rbf’, degree=3)
  • Conf files also support keyword SVM for the LP_MODEL. This uses the sklearn LinearSVC model and tunes the regularization parameter on a grid [0.1, 1, 10, 100, 1000].
  • Method execution is made safer by using Popen communicate instead of subprocess.run(shell=True)
  • Removed lp_model coefficient output. This could lead to errors and failed evaluations for certain Sklearn binary classifiers
  • Method compute_pred() of LPEvaluator and NREvaluator classes now tries to use lp_model.predict_proba() if the classifier does not have it, the function defaults to lp_model.predict()
  • The scoresheet method get_pandas_df() now includes a repeat parameter which denotes the exact experiment repeat results the user wants in the DF. If repeat=None, the DF returned will contain the average metric over all experiment repeats.

A complete Release Log is available on the EvalNE Read The Docs page here.

Assets 3
Loading