Skip to content

Latest commit

 

History

History
64 lines (50 loc) · 2.29 KB

PLANNING.md

File metadata and controls

64 lines (50 loc) · 2.29 KB

Design + Discussion for PetML

Source: https://huyenchip.com/machine-learning-systems-design/design-a-machine-learning-system.html#design-a-machine-learning-system-dwGQI5R

Project Setup

Goal

  • Allow aspiring pet ownners to enter their info
  • Allow pet info to be obtained/requested from various shelter listing sources
  • Give users maximally appropriate pet recommendations based on their information and our modeling prediction/recommendation
  • Maximize pets adopted from kill shelters
  • Understand the nuance behind pet searching
  • Show users pets they may not know are perfect for them
  • provide a means to contact pet's owner/shelter

User Experience

  • A "Tinder-like" user experience
  • Pets in a balanced number considering a limit of matches per day (?)
  • Rate the quality of matches to supply retraining features/variables for user specific training

Performance Constraints

  • Compute limitations of on device training
  • Limited # of total pets
  • No need to be perfect predictions in this case
  • Ask more behavioral questions

Evaluation

  • Metric showing accepted matches
  • Metrics from users rating their opinion of the matches

Personalization

  • Ideally one lite model layer can be used for specific user(s) recommendations but is this technically feasible, given performance constraints?
  • Could be imeplemented as uers specific feature store that supplies some inputs to a base recommendation model
  • Adaptation to specific users situations is a good metric imho

Project Constraints

  • Limited time to implement
  • No budget for spending
  • 3 person 5 animal team
  • Full stack deliverable

Data Pipeline

Inputs

user profile environmental profile pet profile organization/shelter profile

Outputs

pet recommendation (binary classification)

Bias Considerations

  • How do we avoid reinforcement of societal biases, such as who can adopt a pet, which pets are worth adoption
  • Some have a bias against rescues, mixed breed pets, and certain breeds
  • We can actually highlight genetic disease resistance, life span, temperment metrics to try and "train out" biases when they arise

Model Selection

supervised or unsupervised? Generation or prediction - prediction Regression or classication?

Baseline

Simple heuristic baseline 60% of adopted pets are within 20mi radius (example only)