TraceHub

Time-Series and TraceHub

• Time-Series - Series of observations 𝑥_𝑖, where each observation x corresponds to a specific time t
• TraceHub - Platform where separate analytics and dataset owners can plug in time-series analytics and datasets as APIs
• System runs semi-automated transformer pipelines to generate insights
• Reports usefulness of the generated insights

Architecture

• Data Loaders : automatic type detection while loading the data as traces; each trace being sequence of events in time
• Transformer Pipelines : system automates the transformer pipeline by suggesting the applicable ones on the traces; comes with some standard out-of-the-box data transformers like one-hot encoding, missing value, quantiling, binarizing column etc.
• Analytics : after the transformer pipeline is completed, resulting output of traces is fed to analytics which then generates insights along with a usefulness score of it.

Usecase – Simulated Loan Application Dataset

• Loan application types : {medical, vacation, vehicle}
• Events in the application : {receive, evaluate, gather, decision}
  • Receive : new application submitted
  • Evaluate : agent evaluates the application
  • Gather : agent decides to gather additional documents for the application
  • Decision : agent approves or denies the application
• TraceHub generates insights on this data
• Code walk-through on below illustrating the 3 steps :
  • Data Loading
  • Transformer Pipelines
  • Insight generation

loader = TabularData()
data = loader.load(data_path="dataset.csv")
setoftraces = TraceSplit().split(data, "appid")

applicable_tfs = [tf for tf in transformers_list if tf.applicable(setoftraces)]
applicable_tfs[i].transform(setoftraces, *args)

analytics = applicable_tf_analytics(setoftraces)
analytics[i].apply(setoftraces)