Quantified drug consumption

⚠️ This Readme is a short summary of our project and does not reflect the whole process. To get more details about our approach, please refer to the following documents:

• Our complete Notebook: Quantified_drug_consumption_notebook.ipynb
• The html version of the notebook: Quantified_drug_consumption_notebook.html
• Our powerpoint presentation: [PPT]Quantified_drug_consumption_flavien_deseure--charron_benjamin_demouge

Table of Contents

• About
• Getting Started
  • Prerequisites
  • Notebook
  • Website & API
    • Website
    • API
• Data Vizualisation
  • Features
  • Targets
• Preprocessing
  • Features
  • Targets
• Modeling
• Selected models
• Conclusion
• Authors

About

The objective of this project is to carry out a Data Science project from an imposed dataset. We get the following database: Drug Consumption Quantified from UCI Machine Learning repository.

It is derived from an online survey conducted between 2011 and 2012 among 1885 respondents aged 18 years and older from English-speaking countries. It collects demographic informations, three personality tests:

• NEO-FFI-R: The Big Five personality test measures the five personality factors that psychologists have determined are core to our personality makeup.
  • Nscore: Neuroticism - How sensitive a person is to stress and negative emotional triggers.
  • Escore: Extraversion - How much a person is energized by the outside world.
  • Oscore: Openness - How open a person is to new ideas and experiences.
  • Ascore: Agreeableness - How much a person puts others' interests and needs ahead of their own.
  • Cscore: Conscientiousness - How goal-directed, persistent, and organized a person is.
• BIS11: assess the personality/behavioral construct of impulsiveness
• ImpSS: assess various personality characteristics and behaviors related to impulsivity and sensation seeking

and 19 central nervous system psychoactive drugs with the following possibilities:

Never Used	Used over a Decade Ago	Used in Last Decade	Used in Last Year	Used in Last Month	Used in Last Week	Used in Last Day

The authors of the survey showed that there is a relationship between risk of addiction to drugs and personnality attributes.

From this dataset, we choose to address the following problematic:

How can we model the risk of addiction to a drug based on personality and demographic data?

We choose the following drugs with the following classes to answer the problematic:

• For Alcohol, Cannabis, Nicotine, Amphet, Benzos, Coke, Ecstasy, Legalh, LSD, Mushrooms

Not addicted	Addicted

• For Amyl, Crack, Heroin, Ketamine, Meth, VSA

Never Used	Used

• For Caff

Not daily addicted	Daily addicted

This repository contains:

• Drug Consumption Quantified dataset (if the weblink stop working)
• Requirements file
• Python Notebook (ipynb and html)
• Python Web application Django and API which predicts the addiction to the drugs.
• An example of api used
• PowerPoint Presentation of the project
• Final trained models for each drugs

Getting Started

Prerequisites

1. Clone the repository

git clone https://github.com/flaviendeseure/Quantified_drug_consumption.git

2. Install python libraries

pip install -r requirements.txt

Notebook

Simply launch the notebook on jupyter!

Website & API

1. Get to the api directory

cd PATH_TO_API/api 

2. Launch the server

python manage.py runserver

Website

Simply browse on the website

API

Parameters

Feature	Format
age	int >= 18
gender	{"Man", "Woman"}
education	{"Left school before 16 years", "Left school at 16 years", "Left school at 17 years", "Left school at 18 years", "Some college or university, no certificate or degree", "Professional certificate/ diploma", "University degree", "Master's degree","Doctoral degree"}
country	{"Australia", Canada", "New Zealand", "Other", "Republic of Ireland", "UK", "USA"}
ethnicity	{"Asian", Black", "Mixed-Black/Asian", "Mixed-White/Asian", "Mixed-White/Black", "Other", "White"}
nscore	12 <= int <= 60
escore	16 <= int <= 59
oscore	24 <= int <= 60
ascore	12 <= int <= 60
cscore	17 <= int <= 57
impulsivity	1 <= int <= 10
SS	1 <= int <= 11

Results

Result	Description
Addicted	Used last month
Not addicted	Never used in the last month
Used	Used
Never used	Never used
Daily addicted	Used yesterday
Not daily addicted	Not used yesterday

You can use this python file example (example_api.py)

import requests

age = 18
gender = "Man"
education = "Left school before 16 years"
country = "Australia"
ethnicity = "Asian"
nscore = 12
escore = 16
oscore = 24
ascore = 12
cscore = 17
impulsivity = 1
SS = 1

parameters = [age, gender, education, country, ethnicity, nscore, escore, oscore, ascore, cscore, impulsivity, SS]
parameter_names = ["age", "gender", "education", "country", "ethnicity", "nscore", "escore", "oscore", "ascore", "cscore", "impulsivity", "SS"]

url = 'http://127.0.0.1:8000/api/score/?'

for n, p in zip(parameter_names, parameters):
    url += f"{n}={p}"
    if n != SS:
        url += "&"

r = requests.get(url)

print(r.status_code)
if r.status_code == 200:
    for name, result in r.json().items():
        print(f"{name}: {result}")
else:
    print(r.text)

Data Vizualisation

Features

The following graph describes the repartition of the number of drugs used per responders. We can clearly identify two groups of people:

• people who have tried less than 6 drugs (47.53%)
• people who have tried more than 7 drugs (52.47%)

Targets

The following graph describes the repartition for each drug of each classes. As expected, we have an imbalanced repartition of data for each class given that for most drugs, majority of people didn't consume drugs.

Preprocessing

Features

1. Feature selection
   Drop feature that will not be used for modelling: ID

2. Feature encoding
   Encoding have already been performed on the original dataset, so we just changed the encoding for country and ethnicity (one hot) because there are not ordinal variables.

Targets

1. Target selection
   Drop drug that will not be used for modelling: Choc (performances too low due to the popularity of this drug) and Semer(fictif)

2. Target encoding

• First we keep the 7 classes
• Then we group the classes into binary classes

Modeling

Metrics

• Accuracy
• Balanced accuracy
• Confusion matrix

Technics used (imbalanced data)

• Weighting
• Sampling: SMOTE

Phases

• Initial classes
  We keep the original classes and see what we get.

  • Original models and data
  • Weighting features
  • Sampling features

• New classes
  We create our own classes to obtain better results.

  • Original models and data
  • Weighting features
  • Sampling features

• Tuning hyperparameters

Finally we obtained the following evaluation balanced accuracy scores:

Selected models

Drug	Model
Alcohol	Logistic Regression
Amphet	Logistic Regression
Amyl	Logistic Regression
Benzos	Logistic Regression
Caff	Logistic Regression
Cannabis	Logistic Regression
Coke	SVC
Crack	Logistic Regression
Esctasy	SVC
Heroin	SVC
Ketamine	SVC
Legalh	SVC
LSD	SVC
Meth	Logistic Regression
Mushrooms	BernoulliNB
Nicotine	BernoulliNB
VSA	Logistic Regression

Conclusion

Finally the best method to overcome the imbalanced classes in targets was weighting. Thus, we have achieved a mean of 71.24% of balanced accuracy.
To conclude majors difficulties we encounter were:

• Lack of data
• Imbalanced classes for the target
• Preprocessed features
• Biased dataset:
  • young population
  • most people comes from UK

Authors

• Flavien Deseure--Charron
• Benjamin Demouge