Skip to content

RecBole v0.2.0
Compare
Choose a tag to compare
@2017pxy 2017pxy released this 14 Jan 02:46
· 1739 commits to master since this release
v0.2.0
3c2b691
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.

RecBole v0.2.0 Release Notes

  • Highlights
  • New Features
  • Improvements
  • Bug Fixes

Highlights

The RecBole v0.2.0 release includes a number of new features, model efficiency improvements and bug fixes. A few of the highlights include:

  1. We add 12 new models into RecBole, including several non-sampling models and an external algorithm lib model: XGBoost.
  2. Case study is added to RecBole, which is helpful for users to analyze the model result (e.g: give an item ID and a user ID and get the score and ranking position of the item).
  3. We improve the efficiency of data loading and negative sampling.
  4. We now support the full ranking evaluation for context-aware recommendation models.

New Features

  • Add 12 new models:
    • Sequential recommendation(6) : RepeatNet(#584), Fossil(#580), HGN(#575), SHAN(#573), NPE(#573)，HRM(#573)
    • General recommendation(5) : CDAE(#642), MacridVAE(#637), MultiVAE(#626), MultiDAE(#626), LINE(#591)
    • External algorithm lib model(1) : XGBoost(#557)
  • training_neg_sample_num of pairwise_loss model now can be greater than 1 (#533).
  • We add training_neg_sample_distribution in config setting to choose the negative sampling strategy during training (#534).
  • We add benchmark_filename in config setting to load pre-split dataset (#596).
  • Progress bar is added for training and evaluating (#618).
  • We add loss_decimal_place and metric_decimal_place in config setting to control decimal place of loss and metric results separately (#625).
  • We add GAUC metric into evaluation (reference: Deep Interest Network for Click-Through Rate Prediction, KDD 2018) (#572).
  • We add unused_col in config setting to drop the columns only used in data preparation but not used in model (#559).

Improvements

  • We support the ranking evaluation for context-aware recommendation models (#503).
  • We improve the efficiency of data loading and negative sampling (#559).
  • We remove the pre_neg_sampling in Dataloader, which is helpless to model training (#559).
  • We improve the underlying data structure of RecBole, which can promote efficiency of data processing (#559).
  • We refactor the evaluation code (#572) and reformat the mode code (#647).

Bug Fixes

Model

  • Fix a bug in NeuMF model: this bug may cause dropout_ratio disable (#629).
  • Fix a bug in NGCF and GCMC model: now the sparse dropout is disable during evaluation in NGCF and GCMC (#601).
  • Fix a bug in DCN model: this bug may cause crash when running on CPU (#633).
  • Fix a bug in BERT4Rec model: this bug may cause crash when running on CPU (#556).

Trainer

  • Fix a bug in the Trainer._generate_train_loss_output(): this bug may cause the training log is missing (#559).

Data

  • Fix a bug in the sampler: this bug may cause runtime error (#559)
Assets 2
Loading