v0.2.2

Improvements

• Expand user path (~) on DB dataset
• Change null interaction pair generator to sample uniformly. This changes the way null interaction pairs are sampled so that in-dataset null pairs (pairs with lower interaction value than the threshold) and out-of-dataset null pairs are sampled uniformly.

Bugfixes

• Fix select_random_generator returning non-int internal ids, that was producing issues with data sampling

Others

• Update urllib3 dev dependency due to vulnerabilities (CVE-2019-11324, CVE-2018-20060)

v0.2.1

Improvements

• Added MAE regression metric
• Add optional ignore_low_predictions_threshold parameter to the recommendation evaluation process, so that items with model predictions bellow this threshold are ignored
• Add Caser sequential recommender (Tang, Jiaxi, and Ke Wang. "Personalized top-n sequential recommendation via convolutional sequence embedding." Proceedings of the Eleventh ACM International Conference on Web Search and Data Mining. 2018.)

Bugfixes

• Fix matrix split to evaluate the number of min user interactions for all users
• Fix potential issue with list sampler not using the provided column to sample negative interaction ids

Others

• Change recommend() method to not filter out item recommendations bellow the model interaction threshold by default

v0.2.0 Alpha Release

Improvements

• Added list-sampler with negative_ratio support, sequence sorting and group records by multiple columns
• Added support for early stopping rules, with the first built-in rule being the MaxValidationRule
• Slightly improve MemoryInteractionDataset performance by removing unneeded type casts and simplifying record conversion
• Improve recommender training workflow, by abstracting gradient computation, weight updates and improve separation of concerns through specific functions that should focus on each subtask (new methods: _sample_batch(), _predict_batch(), _compute_batch_loss(), _compute_reg_loss())
• Improve recommender logging: optional console (defaults to on) and file (defaults to off) logger, with error, warning and info logs
• Allow passing custom parameters to InteractionDataset instances, when instantiating them through the integrated datasets submodule
• Automatically compute regularization for recommenders that use tf.keras.Model or tf.keras.Layer trainable variables with any regularization parameter set
• Change how metrics are passed to evaluation procedures, to improve readability and extensibility. Now instead of receiving tuples with functions and extra params, it receives callable instances of those metrics:
  • From: metrics={'HR': (hit_ratio, {}), 'NDCG': (ndcg, {})}, 'FScore': (f_score, {'beta': 1})
  • To: metrics=[HitRatio(), NDCG(), FScore(beta=1)]

Bugfixes

• Fix bug on integrated datasets get_test_dataset that could end up generating test datasets (via leave_k_out) instead of using the offered test dataset
• Fix temporary DatabaseInteractionDataset files not being deleted at times

Others

• Add error handling to the ranking evaluation process
• Add examples to showcase extending existing recommenders

v0.1.0

Improvements

• Point Sampler now returns triples instead of tuples, now containing the interaction value of the sampled user id and item id as well
• Update recommender workflow to process multiple points in each _do_batch call
• Added regularization to DMF model
• Epoch callbacks are also computed before the first epoch

Bugfixes

• Fix NCE-DMF loss function

Others

• Changed recommender default interaction threshold to 0.001 to avoid assuming interaction values of 0 as positive
• Rename InteractionDataset copy method to copy

v0.0.3

Improvements

• Last_timestamps option for leave_k_out split
• Create matrix_split variant
• Improve leave_k_out split performance by using raw identifiers instead of assigning and using internal identifiers
• Move split methods location from DRecPy.Evaluation to DRecPy.Evaluation.Splits
• Move evaluation processes location from DRecPy.Evaluation to DRecPy.Evaluation.Processes
• Improve select_one() performance for in-memory interaction dataset instances
• Add CDAE support for custom losses and support for sampling data points based on neg_ratio parameter
• Improve CDAE ranking speed
• Add epoch callback results to the training progress bar
• Improve evaluation processes multithreading performance to be 15% faster in the average case
• Improve defaults of evaluation processes
• Add recommendation_evaluation process
• Add novelty argument for the ranking and recommendation evaluation processes

Bugfixes

• Fix predictive evaluation bug when evaluating the performance on the training set
• Fix ranking evaluation n_neg_interactions bounds check
• Fix ranking evaluation repeated generated items on the negative interactions list

Others

• Improve existing examples
• Add consistency evaluation scripts
• Improve README.md and relevant documentation
• Complete tests for the dataset module, evaluation module, and recommender baselines
• Add integration with Travis CI

v0.0.2

• Bug fixes where deterministic evaluations where not always possible (even with a seed parameter passed through)
• Correct requirement versions for critical dependencies
• Bug fixes due to the interaction vectors returned by InteractionDataset instances being of type integer, which caused tensorflow conversion incompatibilities
• Fixed crashes hapenning because DatabaseInteractionDataset instanced would leave open cursors
• Fix wrong retrieved MemoryInteractionDataset column value types
• Improve DatabaseInteractionDataset cleanup method to drop unused tables