joseprsm/rexify

🦖 Streamlined Recommender Systems with TensorFlow and KubeFlow

rexify's Issues

Add item-to-item similarity

Create Schema dataclass

Move validations from the HasSchemaInput class to a Schema dataclass.

Tasks

• Schema dataclass
• create dataclass for user and items
• validate inputs (must have id field for users and items, for example)
• validate event types
• validate rating features

Ranking Model

• Update EventEncoder to explode events to columns
• Replace single EventModel and it's Softmax activation by multiple models with Sigmoid activation

Add minimum number of examples for ScaNN

Support text data

• create FeatureTransformer for text data

Kubeflow Pipelines

• kfp components
  • download component
  • transform component
  • train component
  • deploy component

KServe

Serve index models with KServe.

Add tests for callbacks

Add warning for missing optional dependencies

Add `scann` index

Optimize users, items read into FeatureExtractor

Create Input and Output DataFrame objects

Fallback embeddings

User Session Lookup

Based on a window_size, extract the most recent session history for a given user_id.

Checklist

• input is a tf.Tensor of user ids
• output is a tf.Tensor of a list with the last n items user_id interacted with
  • output must be the most recent (aka the last) items, users interacted with
• add attributes to Sequencer to pass data to SessionLookup
• override lookup during inference

Refactor FeatureExtractor

Tasks

• include _EventGenerator in FeatureExtractor
• pass custom transformations
  • create CustomTransformer class
    • feature name must be included in init
    • complements transformations from schema
  • create same base class for current transformers

Create single preprocessing pipeline

Join user, item and event preprocessing pipelines (FeatureExtractors, EventGenerator) in a single one.

Tasks

• Refactor FeatureExtractor
• Pass user and item data to EventGenerator
  • as pandas.DataFrames; or
  • as paths (need to specify something like load_fn=pd.read_csv)
• Should be able to access each (user, item) IdEncoders

Override Index call method