Animal-adoption-EDA-ML---Capstone-project

• Thank you! https://www.kaggle.com/jinbonnie/animal-data/code

Table of contents

• Problem Statement
• Goal
• EDA interesting points
• Data pre-processing
  • Data cleaning
    • Missing Values
  • Data engineering
    • Feature Engineering
      • Total time spent in shelter
      • Adoption rate
    • Example of features with data set errors
      • animalage
        • Before & after data-preprocessing
• Label manipulation before training models
• ML models used
• ML EDA

Problem statement

• need to maximise adoption rates to ensure adoption shelters are not completely dependent on government funding as well as to improve the quality of life of animals and pet owners.
• This is based off our label: (Adopted, adopted and Returned & Not adopted)
• 

Goal

• to assist shelters in identifying features to take note of when trying to improve adoption rates.

EDA (exploratory data analysis) interesting points:

Side note: The most notable issue with this dataset is that it does not have any numerical/continuous data for me to manipulate. I must adjust everything to turn them into numbers

• Black cats and dogs have the highest abandon and adoption rates
• 
• cats & dogs are the most popular animals to be abandoned
• 
• Adoption rate is not time sensitive
• 
• The most abandoned breed is the domestic shorthair in both cats and dogs (so much so that we can shorten them to a new column)
• 

Data pre-processing

Data cleaning

Missing values

• Such as deceaseddate, returneddate & identichip indicate alive, not returned and not chipped respectively
• Other items missing will have their entire row dropped
• 

Data engineering

• Data needs to be in numbers for the Machine to read, these numbers shoud also make sense in the real world.

Feature engineering

Total time spent in shelter

• getting Total time spent in shelter with intake and movement date
• 
• 
• Negative timing is due to the same day dates being reversed is terms of newer and older dates. easily solved with abs() function
• 

Adoption rate

• Obtained via dividing current adopted with total number of animals. Caveat is that first few entries will be very skewed and jumpy on the graph
  • 

Example of features with data set errors

animalage

• Dataset column is riddled with:
  • strings where days, weeks & months need to be converted into years
    • Regular expression usage to achieve this
  • outlier errors which need to make sense. Data that is on extreme ends can affect certain ML models
    • better to keep the data's SD relatively low
    • Here i chose to replace the outlier age (100 year old cats/ dogs) with the googled average maximum age of a cat/dog

Before and after data-preprocessing

Label manipulation before training models

• SMOTEtomek was chosen to remove the nearest neighbours smoted to ensure data is not to0 cluttered
• 

ML models used

Utilizing various Machine learning algorithmns and libraries to get the best model

ML EDA

• MIP plot (we can see that the top 4 features that affected the Model the most were:
  • returned reason
  • Time spent in shelter
  • chipped or not
  • animal's age
• 
• Ultimately the Support vector Machine model was the best:
• 
  • 100% recall on adopted & not adopted labels
  • 100% precision on adopted & returned label
  • 
  • ROC OVO (micro on graph) 98%
  • log loss very very low at 0.119 where logloss = 1.1 is when i myself randomly choose and label the data without any model.
  •