CHANGELOG.md

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog and this project adheres to Semantic Versioning.

[v1.10.0] - 2022-07-31

Added

• Added n_jobs and chunksize arguments to MuminDataset, to enable customisation of these.

### Changed

• Lowered the default value of chunksize from 50 to 10, which also lowers the memory requirements when processing articles and images, as fewer of these are kept in memory at a time.
• Now stores all images as uint8 NumPy arrays rather than int64, reducing memory usage of images significantly.

[v1.9.0] - 2022-07-22

Added

• Added checkpoint after rehydration. This means that if compilation fails for whatever reason after this point, the next compilation will resume after the rehydration process.
• Added some more unit tests.

Fixed

• Fixed bug on Windows where some tweet IDs were negative.
• Fixed another bug on Windows where the timeout decorator did not work, due to the use of signals, which are not available on Windows machines.
• Fixed bug on MacOS causing Python to crash during parallel extraction of articles and images.

Changed

• Refactored repository to use the more modern pyproject.toml with poetry.

[v1.8.0] - 2022-04-14

### Changed

• Now allows instantiation of MuminDataset without having any Twitter bearer token, neither as an explicit argument nor as an environment variable, which is useful for pre-compiled datasets. If the dataset needs to be compiled then a RuntimeError will be raised when calling the compile method.

[v1.7.0] - 2022-03-24

### Added

• Now allows setting twitter_bearer_token=None in the constructor of MuminDataset, which uses the environment variable TWITTER_API_KEY instead, which can be stored in a separate .env file. This is now the default value of twitter_bearer_token.

Changed

• Replaced DataFrame.append calls with pd.concat, as the former is deprecated and will be removed from pandas in the future.

[v1.6.2] - 2022-03-21

Fixed

• Now removes claims that are only connected to deleted tweets when calling to_dgl. This previously caused a bug that was due to a mismatch between nodes in the dataset (which includes deleted ones) and nodes in the DGL graph (which does not contain the deleted ones).

[v1.6.1] - 2022-03-17

Fixed

• Now correctly catches JSONDecodeError during rehydration.

[v1.6.0] - 2022-03-10

Changed

• Changed the download link from Git-LFS to the official data.bris data repository, with URI https://doi.org/10.5523/bris.23yv276we2mll25fjakkfim2ml.

[v1.5.0] - 2022-02-19

### Changed

• Now using dicts rather than Series in to_dgl. This improved the wall time from 1.5 hours to 2 seconds!

Fixed

• There was a bug in the call to dgl.data.utils.load_graphs causing load_dgl_graph to fail. This is fixed now.

## [v1.4.1] - 2022-02-19

Changed

• Now only saves dataset at the end of add_embeddings if any embeddings were added.

[v1.4.0] - 2022-02-19

### Added

• The to_dgl method is now being parallelised, speeding export up significantly.
• Added convenience functions save_dgl_graph and load_dgl_graph, which stores the Boolean train/val/test masks as unsigned 8-bit integers and handles the conversion. Using the dgl-native save_graphs and load_graphs causes an error, as it cannot handle Boolean tensors. These two convenience functions can be loaded simply as from mumin import save_dgl_graph, load_dgl_graph.

[v1.3.0] - 2022-02-18

Added

• Now uses GPU to embed all the text and images, if available.

[v1.2.5] - 2022-02-06

Fixed

• Now does not raise an error if we are not authorised to rehydrate a tweet, and instead merely skips it.

[v1.2.4] - 2022-01-24

Fixed

• Changed the minimum Python version compatible with mumin to 3.7, rather than 3.4.

[v1.2.3] - 2021-12-15

Fixed

• During rehydration, the authors of the source tweets were not included, and the images from tweets were not included either. They are now included.

[v1.2.2] - 2021-12-15

Fixed

• Now replacing NaN values for Numpy features with np.nan instead of an array, as fillna does not accept that. These are then converted in a scalar array with a np.nan value.

[v1.2.1] - 2021-12-15

Fixed

• When running add_embeddings, only embeddings to existing nodes will be added. This caused an error when e.g. images were not included in the dataset.

[v1.2.0] - 2021-12-14

Changed

• If tweets have been deleted (and thus cannot be rehydrated) then we keep them along with their related entities, just without being able to populate their features. When exporting to DGL then neither these tweets nor their replies are included.

Added

• Now includes a check that tweets are actually rehydrated, and raises an error if they are not. Such an error is usually due to the inputted Twitter Bearer Token being invalid.

[v1.1.1] - 2021-12-13

Fixed

• Fixed bug in producing embeddings

[v1.1.0] - 2021-12-12

Fixed

• Updated the dataset with deduplicated entries. The deduplication is done such that the duplicate with the largest relevance parameter is kept.
• Include checks of whether nodes and relations exist, before extracting data from them.

Added

• Added include_timelines option, which allows one to not include all the extra tweets in the timelines if not needed. As this greatly increases the amount of tweets needed to rehydrate, it defaults to False.

[v1.0.2] - 2021-12-09

Fixed

• Removed the relations from the dump which we are getting through compilation anyway.
• Updated the filtering mechanism, so that the relevance parameter is built in to all nodes and relations upon download.
• Deal with the situation where no relations exist of a certain type, above a specified threshold.

[v1.0.1] - 2021-12-05

Fixed

• Added in the POSTED relation, as leaving this out effectively meant that all the new tweets were filtered out during compilation.

[v1.0.0] - 2021-12-03

Changed

• Added new version of the dataset, which now includes a sample of ~100 timeline tweets for every user. This approximately doubles the dataset size, to ~200MB before compilation. This new dataset includes different train/val/test splits as well, which is now 80/10/10 rather than 60/10/30. This means that the training dataset will see a much more varied amount of events (6-7) compared to the previous 2.

[v0.7.0] - 2021-12-02

Changed

• Changed include_images to include_tweet_images, which now only includes the images from the tweets themselves. Further, include_user_images is changed to include_extra_images, which now includes both profile pictures and the top images from articles. The tweet pictures are included by default, and the extras are not. This is to reduce the size of the default dataset, to make it easier to use.

[v0.6.0] - 2021-12-01

Changed

• Split up the include_images into include_images and include_user_images, with the former including images from tweets and articles, and the latter being profile pictures. The former has been set to True by default, and the latter False. This is due to the large amount of profile pictures making the dataset excessively large.

Fixed

• Now catches connection errors when attempting to rehydrate tweets.

[v0.5.3] - 2021-11-26

Fixed

• Masks have been changed to boolean tensors, as otherwise indexing did not work properly.
• In the case where a claim/tweet does not have any label, this produces NaN values in the mask- and label tensors. These are now substituted for zeroes. This means that they will always be masked out, and so the label will not matter anyway.

[v0.5.2] - 2021-11-24

Fixed

• Now converting masks to long tensors, which is required for them to be used as indexing tensors in PyTorch.

### Changed

• Now only dumping dataset once while adding embeddings, where previously it dumped the dataset after adding embeddings to each node type. This is done to add embeddings faster, as the dumping of the dataset can take quite a long time.
• Now blanket catching all errors when processing images and articles, as there were too many edge cases.

[v0.5.1] - 2021-11-09

Fixed

• When encountering HTTP status 401 (unauthorized) during rehydration, we skip that batch of tweets.
• Image relations were extracted incorrectly, due to a wrong treatment of the images coming directly from the tweets via the media_key identifier, and the images coming from URLs present in the tweets themselves. Both are now correctly included in a uniform fashion.
• Datatypes are now only set for a given node if the node is included in the dataset. For instance, datatypes for the article features are only set if include_articles == True.

[v0.5.0] - 2021-11-08

Added

• The Claim nodes now have language, keywords, cluster_keywords and cluster attributes.
• Now sets datatypes for all the dataframes, to reduce memory usage.

### Fixed

• Updated README to a single zip file, rather than stating that the dataset is saved as a bunch of CSV files.
• Fixed image embedding shape from (1, 768) to (768,).
• Article embeddings are now computed correctly.
• Catch IndexError and LocationParseError when processing images.

Changed

• Now dumping files incrementally rather than keeping all of them in memory, to avoid out-of-memory issues when saving the dataset.
• Dataset size argument now defaults to 'small', rather than 'large'.
• Updated the dataset. This is still not the final version: timelines of users are currently missing.
• Now storing the dataset in a zip file of Pickle files instead of HDF. This is because of HDF requiring extra installation, and there being maximal storage requirements in the dataframes when storing as HDF. The resulting zip file of Pickle files is stored with protocol 4, making it compatible with Python 3.4 and newer. Further, the dataset being downloaded has been heavily compressed, taking up a quarter of the disk space compared to the previous CSV approach. When the dataset has been downloaded it will be converted to a less compressed version, taking up more space but making loading and saving much faster.

Removed

• Disabled numexpr, transformers and bs4 logging.

[v0.4.0] - 2021-10-26

### Fixed

• All embeddings are now extracted from the pooler output, corresponding to the [CLS] tag.
• Ensured that train/val/test masks are boolean tensors when exporting to DGL, as opposed to binary integers.
• Content embeddings for articles were not aggregated per chunk, but now a mean is taken across all content chunks.
• Assign zero embeddings to user descriptions if they are not available.

Changed

• The DGL graph returned by the to_dgl method now returns a bidirectional graph.
• The verbose argument of MuminDataset now defaults to True.
• Now storing the dataset as a single HDF file instead of a zipped folder of CSV files, primarily because data types are being preserved in this way, and that HDF is a binary format supported by Pandas which can handle multidimensional ndarrays as entries in a dataframe.
• The default models used to embed texts and images are now xlm-roberta-base and google/vit-base-patch16-224-in21k.

Removed

• Removed the poll and place nodes, as they were too few to matter.
• Removed the (:User)-[:HAS_PINNED]->(:Tweet) relation, as there were too few of them to matter.

[v0.3.1] - 2021-10-19

Fixed

• Fixed the shape of the user description embeddings.

[v0.3.0] - 2021-10-18

Fixed

• Now catches SSLError and OSError when processing images.
• Now catches ReadTimeoutError when processing articles.
• The (:Tweet)-[:MENTIONS]->(:User) was missing in the dataset. It has now been added back in.
• Added tokenizer truncation when adding node embeddings.
• Fixed an issue with embedding user descriptions when the description is not available.

### Changed

• Changed the download link to the dataset, which now fetches the dataset from a specific commit, enabling proper dataset versioning.
• Changed the timeout parameter when downloading images from five seconds to ten seconds.
• Now processing 50 articles and images on each worker, compared to the previous 5.
• When loading in an existing dataset, auxilliaries and islands are removed. This ensures that to_dgl works properly.

### Removed

• Removed the review warning from the README and when initialising the dataset. The dataset is still not complete, in the sense that we will add retweets and timelines, but we will instead just keep versioning the dataset until we have included these extra features.

[v0.2.0] - 2021-10-12

Added

• Added claim embeddings to Claim nodes, being the transformer embeddings of the claims translated to English, as described in the paper.
• Added train/val/test split to claim nodes. When exporting to DGL using the to_dgl method, the Claim and Tweet nodes will have train_mask, val_mask and test_mask attributes that can be used to control loss and metric calculation. These are consistent, meaning that tweets connected to claims will always belong to the same split.
• Added labels to both Tweet and Claim nodes.

Fixed

• Properly embeds reviewers of claims in case a claim has been reviewed by multiple reviewers.
• Load claim embeddings properly.
• Catches TooManyRequests exception when extracting images.
• Load dataset CSVs with Python engine, as the C engine caused errors.
• Disable tokenizer parallelism, which caused warning messages during rehydration of tweets.
• Ensure proper quoting of strings when dumping dataset to CSVs.
• Enable truncation of strings before tokenizing, when embedding texts.
• Convert masks to integers, which caused an issue when exporting to a DGL graph.
• Bug when computing reviewer embeddings for claims.
• Now properly shows compiled=True when printing the dataset, after compilation.

Changed

• Changed disclaimer about review period.

[v0.1.4] - 2021-09-13

Fixed

• Include (:User)-[:POSTED]->(:Reply) in the dataset, extracted from the rehydrated reply and quote tweets.

[v0.1.3] - 2021-09-13

### Fixed

• Compilation error when including images.
• Only include videos if they are present in the dataset.
• Ensure that article embeddings can properly be converted to PyTorch tensors when exporting to DGL.

[v0.1.2] - 2021-09-13

Fixed

• The replies were not reduced correctly when the small or medium variants of the dataset was compiled.
• The reply features were not filtered and renamed properly, to keep them consistent with the tweet nodes.
• Users without any description now gets assigned a zero vector as their description embedding.
• If a relation does not have any node pairs then do not try to create a corresponding DGL relation.
• Reset nodes and rels attributes when loading dataset.
• Add embeddings for Reply nodes.

[v0.1.1] - 2021-09-08

Changed

• Changed installation instructions in readme to pip install mumin.

[v0.1.0] - 2021-09-07

Added

• First release, including a MuminDataset class that can compile the dataset dump the compiled dataset to local csv files, and export it as a dgl graph.