SachaIZADI/predict-drug-price

Drug Price Prediction

Project Description

The objective is to predict the price for each drug in the test data set (drugs_test.csv). Please refer to the sample_submission.csv file for the correct format for submissions.

Installation

The project was developed with Python 3.9.5, but should be compatible with older version of Python 3 (although this wasn't tested)

1. Create a virtual environment with venv (or any method of your choice) and activate it:

python3 -m venv venv
source venv/bin/activate

2. Install requirements:

pip install -r requirements.txt

3. Unzip the data in data.zip and put all .csv files under the data folder. You can run the following shell script

. data/unzip_data.sh

Run the code

The package can be called via CLI - 2 pipelines (train and predict) are implemented

• Train pipeline

python -m src --do-train

• Predict pipeline

python -m src --do-predict

Advanced parametrization of the train pipeline can be done via src/config.py, where you can:

1. Update the hyperparameters of the Regressor model
2. Precise if you want to:

use_grid_search: bool = False # <-- Set to True if you want to tune the hyperparameters of the model
use_cross_validation: bool = True # <-- Set to True if you want to compute the performance of the model using cross-validation
visualize_results: bool = True # <-- Set to True to get some plots to visualize the output of the model
save_model: bool = True # <-- Set to True to save the model in a pickle file

Modeling aspects & discussion

Evaluation metric

The price variable we are modelling has a pretty wide distribution

- min: 0.6
- max: 990.4
- mean: 28.5
- std: 81.4

Depending on the business question we want to solve, we might prefer

• to be more accurate on higher-priced products or on lower-priced products
• to predict a price range for products (e.g. [0-1], [1-5], ... [100-500], [500+])
• etc.

This utimately changes the evaluation metric we optimize for, and the way we model price.

Missing a bit of business contest I chose to optimize for the mean squared error of the log-price, such that a 50% error on a cheap drug has the same "weight" as a 50% error on an expensive drug.

Feature engineering

The feature engineering code is available at /src/feature_engineering. Most features provided were plugged as is into the model, except for :

• categorical variables, that were one-hot encoded
• active ingredients, where we added a "number of active ingredients per drug" feature, as well as dimensionality reduction because the number of active ingredients was pretty high and plugging them directly would have been prone to overfitting.

Next steps:

• I did not use the pharmaceutical manufacturer information. I expect it to have quite some impact by integrating a notion of generic drugs vs. branded-drugs (which are usually 1.x to 2 times more expensive).

Model choice

I chose to go for a XGBoost regressor as core model, that aims at predicting log(price) to be aligned with the final metric we optimize for. I also added a grid search method for fine-tuning the hyperparameters.

Overall performance of the algorithm

The model performs significantly better than a "baseline" model (that simply predicts the mean):

- Baseline 'dummy' model `mean_squared_log_error`: 2.2
- Model ‘mean_squared_log_error’ on TEST set: 0.50

The performance should still be improved:

• the model still has a high loss (the model has the capacity to overfit the dataset - in some experiments I reduced the error to ~0 on the training set with high tree_depth)
• the model is overfitting on the training set:

- Mean of ‘neg_mean_squared_log_error’ on TRAIN set: -0.27
 Mean of ‘neg_mean_squared_log_error’ on TEST set: -0.50

Feature importance:

We see that active ingredients and count of gellules/plaquettes/etc. have the highest feature importance. This makes sense since

• molecules can have very different production & R&D costs, yielding to very different price of the final drug
• a pack of 20 tabs is likely to be x2 the price of a pack of 10 tabs

Distribution of the errors/predictions:

We see that the model tends to over-predict on low prices, and under-predict on high prices. This should be investigated in a second iteration.

Personal comment

This was a fun and interesting project. I'm particularly happy & proud to have had the chance to hack around with the sklearn library :

• designing an end-to-end sklearn pipeline
• adding a TransformedTargetRegressor layer on top of the XGBoost
• play with the feature importance API of XGBoost.

Files & Field Descriptions

You'll find five CSV files:

• drugs_train.csv: training data set,
• drugs_test.csv: test data set,
• active_ingredients.csv: active ingredients in the drugs.
• drug_label_feature_eng.csv: feature engineering on the text description,
• sample_submission.csv: the expected output for the predictions.

Drugs

Filenames: drugs_train.csv and drugs_test.csv

Field	Description
drug_id	Unique identifier for the drug.
description	Drug label.
administrative_status	Administrative status of the drug.
marketing_status	Marketing status of the drug.
approved_for_hospital_use	Whether the drug is approved for hospital use (oui, non or inconnu).
reimbursement_rate	Reimbursement rate of the drug.
dosage_form	See dosage form.
route_of_administration	Path by which the drug is taken into the body. Comma-separated when a drug has several routes of administration. See route of administration.
marketing_authorization_status	Marketing authorization status.
marketing_declaration_date	Marketing declaration date.
marketing_authorization_date	Marketing authorization date.
marketing_authorization_process	Marketing authorization process.
pharmaceutical_companies	Companies owning a license to sell the drug. Comma-separated when several companies sell the same drug.
price	Price of the drug (i.e. the output variable to predict).

Note: the price column only exists for the train data set.

Active Ingredients

Filename: active_ingredients.csv

Field	Description
drug_id	Unique identifier for the drug.
active_ingredient	Active ingredient in the drug.

Note: some drugs are composed of several active ingredients.

Text Description Feature Engineering

Filename: drug_label_feature_eng.csv

This file is here to help you and provide some feature engineering on the drug labels.

Field	Description
description	Drug label.
label_XXXX	Dummy coding using the words in the drug label (e.g. label_ampoule = 1 if the drug label contains the word ampoule - vial in French).
count_XXXX	Extract the quantity from the description (e.g. count_ampoule = 32 if the drug label the sequence 32 ampoules).

Note: This data has duplicate records and some descriptions in drugs_train.csv or drugs_test.csv might not be present in this file.