ClayRS 0.5.1 - ClayRS can see!

Release which includes image support for the Content Analyzer and RecSys modules!

• This release was co-developed with @m-elio

NOTE: The minimum Python version has been bumped up from Python 3.7 to Python 3.8 in order to use @functools.cached_property decorator

---

Added

Content Analyzer

• Implemented visual preprocessors thanks to torchvision library
  • Also torch augmenters were implemented
  • All of them can be checked in the docs
• Implemented postprocessors techniques which also work for textual techniques
  • Visual bag of words (with count and tfidf weighting schema)
  • Scipy vector quantization
  • Dimensionality reduction techniques from sklearn (PCA, Gaussian random projections, Feature agglomeration)
• Images path to process specified in the raw source could be an absolute_path, relative_path, online url!
• Implemented several content techniques which extract embedding features from images
  • Pre-trained models from timm
  • Pre-trained caffe models using opencv.dnn
  • Hog descriptor, Canny edge detector, LBP, SIFT from skimage
  • Color histogram
  • Custom filter convolution
• Implemented FromNPY technique, which imports features from a numpy serialized matrix

RecSys

• Implemented VBPR technique following the corresponding paper
  • The implementation has been tested thoroughly by experimental comparison with cornac (experiment repository can be found here)

Changed

Content Analyzer

• Changed Ratings class to use numpy arrays and integer mappings instead of relying on python dictionaries and strings
• Adapted FieldContentProductionTechnique to consider the distinction between textual and visual techniques
• Added possibility to serialize contents produced with multi threading

RecSys

• Vectorized computation of CentroidVector algorithm
• Adapted content based algorithm abstraction to make room for neural algorithms
• Fixed missing Bootstrap partitioning technique from online documentation
• AllItemsMethodology by default now considers as items catalog the union between train and test set
• HoldOutPartitioningTechnique can now accept an integer value representing the n° of instances to hold rather than a percentage
• Changed log of users skipped in partitioning/algorithm fitting: a single print with total number of skipped users is fired instead of a single one for each skipped user

EvalModel

• Changed NDCG implementation to allow the choice of the gain weights (linear or exponential) and the definition of a discount function
• Improved visualization of statistical tests results