toan01-uet / Solution-for-HackerEarth-Machine-Learning-challenge

HackerEarth Machine Learning challenge: Of Genomes And Genetics

Solution-for-HackerEarth-Machine-Learning-challenge (26st place)

HackerEarth Machine Learning challenge: Of Genomes And Genetics link.

Why I took part in this competition?

Because:

• I have wanted to practice some feature engineering techniques about tabular data, ensemble,... that I found on Kaggle link.
• I have wanted to take part in a competition which was ongoing. That has made me feel like I'm taking part in the Olympics in which a lot of competitors have to compete against each other.

Time to complete

That time when I found this competition, It still had 1 week to close. So my work still had concentrated on EDA and Feature engineering. In this post, I will share how I did feature engineering.

Train and Test

When performing Label Encoding below, you must encode train and test together (reference from)

Feature Engineering Techniques

Label Encode ( Categorical features )

• Features have 2 values:
  • Genes in mother's side
  • Birth defects
  • History of anomalies in previous pregnancie
  • Assisted conception IVF/ART
  • H/O serious maternal illness
  • Folic acid details (peri-conceptiona)
  • Place of birth
  • Heart Rate (rates/min
  • Respiratory Rate (breaths/min)
  • Follow-up
  • Inherited from father
  • Maternal gene
  • Status
  • Paternal gene From what values of features are? (Quantifier) then I have chosen values. For example: with Follow-up feature: High --> 2, Low ---> 1
• Features have more than 2 values:
  • It's the same before. But some new values such as: -, Not available, Not applicable,.. so I had to label them.
  • With some text features like : Location of Institute, Institute Name, Family Name, Father's name I have extracted to had new features then encodes them:
    • Location of Institute: for examples: 125 PARKER HILL AV\nJAMAICA PLAIN, MA 02120\n(42.329611374844326, -71.10616871232227). I had created some features before:
      1. 1. JAMAICA PLAIN : district
      2. MA 02120 : POST CODE
      3. 42.329611374844326 : Latitude
      4. -71.10616871232227 : Longtitude
    • Then hash code: district, POST CODE, Family Name, Father's name

Transforming

• Log transform some numerical features: 'Patient Age', 'Blood cell count (mcL)', "Mother's age", "Father's age", 'White Blood cell count (thousand per microliter)'
• Interaction (ratio): create ratio columns like as in df['patient_per_mom'] = df['Patient Age']/df["Mother's age"] df['patient_per_dad'] = df['Patient Age']/df["Father's age"] df['age_per_bcc'] = df['Patient Age']/df['Blood cell count (mcL)'] df['age_per_wbcc'] = df['Patient Age']/df['White Blood cell count (thousand per microliter)'] df['wbcc_per_bcc'] = df['White Blood cell count (thousand per microliter)'] /df['Blood cell count (mcL)']
• Coordinate features: lat = df["latitude"] lon = df["longtitude"] df["x_dimen"] = np.cos(lat) * np.cos(lon) df["y_dimen"] = np.cos(lat) * np.sin(lon) df["z_dimen"] = np.sin(lat)

Create IS_NULL some impact features

Target Encoding

USING SMOTE TO DEAL WITH IMBALANCE DATASET

MODEL

Because time was limit, so I had choose autoML for model. And this is source code about my solution.