| jupyter | ||||||||||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
|
The aim of this report is to summarise, describe, and assess the process of solving two questions using a Data Science approach: Using information from job postings/listings, what factors impact the listed salary of data-related jobs the most (1), and what factors correlate with the industry/job type of the listed data job(2).
The first section of the report outlines the methodology and computer code that was designed and used to collect enough data to answer the project questions. Code was written to rapidly extract important information from over 18 000 job listings from SEEK.com, and more was written to 'clean up' the retrieved information for efficent Data Science analysis. Technically, these two processes are referred to as 'Web Scraping' and 'Data Cleaning'.
The second section describes the data science approach to answering question 1. We classified job listings as either high salary or low salary jobs based on whether the salary was above the aggregate median. This step allows us to use the data to train two classification Machine Learning models (BaggingClassifier & LogisticRegression) to their optimal accuracy levels. We can then perform analysis on our well-trained models to understand the key features/factors that the models used to accurately classify high and low salary jobs, and use these insights to futher understand the current data job market. Technically, this process is referred to as 'Inferencing'.
Section three describes the same inferencing approach to answering question 2. First, data jobs were classified as either being from Science-related industries or not. We also use difference Machine Learning models to perform inferencing to answer this question (SupportVectorClassification and AdaBoost ExtraTreeClassifier).
KeyFindings
Higher Paying Jobs are usually:
- Senior, more leadership type roles
- Experience with cloud computing and data infrastructure/architecture
- Involved with stakeholders
- Experience with SQL
Lower Paying Jobs are usually:
- Analyst Roles
- Have some type of training
Science Related Data Jobs are usually:
- Have something to do with childcare
- Deal with equity
Non-Science Related Data Jobs:
- Deal with acquisitions
- Risk
- Transactions
This section contains the code to scrape around 9000 job postings. Data Munging and Data Cleaning will be in a separate notebook.
The impetus of this section was to build a holistic bundle of code. It even sends text messages when the web scraping is completed!
# List of data job and page amounts to scrape
job_search_links = [['data-analyst', 200],
['data-engineer', 80],
['database-administrator', 40],
['database-admin', 30],
['machine-learning-engineer', 7],
['machine-learning-developer', 10],
['data-scientist', 15],
['data-architect', 30],
['big-data-architect', 5],
['business-analyst', 200],
['machine-learning-scientist', 5],
['data-warehouse', 35],
['business-intelligence-analyst', 40],
['data-science-manager', 35],
['data-science-consultant', 8],
['data', 200]]# Gather ONLY the Links of Data Job Postings
data = []
for title, pages in job_search_links:
for i in range(1,(pages + 1)):
# Getting URL INTO BTS OBJECT
url = f"https://www.seek.com.au/{title}-jobs?page={i}"
# Set the url
url_data = requests.get(url)
# Reqeuest url data
url_bts = BeautifulSoup(url_data.content, 'lxml')
# Turn url data into BS object
# GETTING JOB LINKS FROM BTS OBJECT
page_links = url_bts.find_all("a", {'class': "_2iNL7wI"})
hiring_company = url_bts.find_all('a', {'class':'_3AMdmRg'})
# Get all links
for i in range(len(page_links)):
row = {}
row['Title'] = page_links[i].text
row['Link'] = page_links[i].attrs['href'].split('?')[0]
data.append(row)
if len(data)%100 == 0:
print(f"{len(data)} links scraped")
data_df = pd.DataFrame(data=data)# Drop Duplicate Links
data = data_df.drop_duplicates()# IMPORT Twilio Rest API. Will send a text message to my phone during the scraping process
from twilio.rest import Client
def send_sms(message="Default SMS"):
account_sid = 'ACXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX'
auth_token = '017dXXXXXXXXXXXXXXXXXXXXXXXXXXXX'
client = Client(account_sid, auth_token)
message = client.messages \
.create(
body=f'{message} - Cong',
from_='+1XXXXXXXXXXX',
to='+61XXXXXXXXXX'
)
print(message.sid)
# MAIN BLOCK OF WEB SCRAPING CODE. PERFORMED ALL WEBSCRAPING OF RELEVANT CODE.
# Create empty lists
data_details = []
data_questions = []
data_core = []
data_hirer = []
list_job_types = ['Full Time','Contract/Temp','Part Time','Casual/Vacation','Dail Rate Contract']
for link in data['Link']:
# GRAB THE WEBPAGE
url = f"https://www.seek.com.au{link}"
url_data = requests.get(url)
bts = BeautifulSoup(url_data.content, 'lxml')
# ESSENTIAL JOB INFO SECTION
pandas_row = {}
pandas_row['Link'] = link
try:
core_job_details = bts.find('div', {'class': 'K1Fdmkw JyFVSRZ'}).find('div', {'class': "Pdwn1mb"})
core_job_details = core_job_details.find_all('dd')
except:
pass
# Seek will always includes Date Posted and Job Type info into each Job Posting
try:
if core_job_details[3].text in list_job_types: # If a reward is present in the description, Job Type will ALWAYS the 3rd iterable
pandas_row['Date_Posted'] = core_job_details[0].text
pandas_row['Location_City'] = core_job_details[1].find_all('span', {'class':""})[0].find('strong').text
try:
pandas_row['Location_Region'] = core_job_details[1].find_all('span', {'class':""}
)[0].find('span').text.replace(",","",1).strip()
except:
pass
pandas_row['Reward'] = core_job_details[2].text
pandas_row['Job_Type'] = core_job_details[3].text
pandas_row['Industry'] = core_job_details[4].find_all('span', {'class':""})[0].find('strong').text
pandas_row['Specialisation'] = core_job_details[4].find_all('span', {'class':""}
)[0].find('span').text.replace(",","",1).strip()
else:
pandas_row['Date_Posted'] = core_job_details[0].text
pandas_row['Location_City'] = core_job_details[1].find_all('span', {'class':""})[0].find('strong').text
try:
pandas_row['Location_Region'] = core_job_details[1].find_all('span', {'class':""}
)[0].find('span').text.replace(",","",1).strip()
except:
pass
pandas_row['Job_Type'] = core_job_details[2].text
pandas_row['Industry'] = core_job_details[3].find_all('span', {'class':""})[0].find('strong').text
pandas_row['Specialisation'] = core_job_details[3].find_all('span', {'class':""}
)[0].find('span').text.replace(",","",1).strip()
except:
pass
data_details.append(pandas_row)
# JOB QUESTION INFO
new_row = {}
new_row['Link'] = link
try:
questions = bts.find("ul",{'class': "_34zKk91"}).find_all('span', {'class':''})
for i in range(len(questions)):
new_row[f"Q{i+1}"] = questions[i].text
data_questions.append(new_row)
except:
pass
# JOB DESCRIPTION TEXT INFO
new_row = {}
new_row['Link'] = link
try:
new_row['Text'] = bts.find('div', {'data-automation': "jobDescription"}).text
new_row['Bullet_Text'] = [x.text for x in bts.find('div', {'data-automation': "jobDescription"}).find_all('li')]
new_row['Strong_Text'] = [x.text for x in bts.find('div', {'data-automation': "jobDescription"}).find_all('strong')]
new_row['Par_Text'] = [x.text for x in bts.find('div', {'data-automation': "jobDescription"}).find_all('p')]
except:
pass
data_core.append(new_row)
# JOB HIRER INFO
new_row = {}
new_row['Link'] = link
try:
new_row['Hirer'] = bts.find('span' ,{'class': "_3FrNV7v _2QG7TNq E6m4BZb"}).text
except:
pass
data_hirer.append(new_row)
# PRINT STATEMENTS AND SMS TESTING
if len(data_details)%50 == 0:
print(f"{len(data_details)} links scraped")
if len(data_details)%4300 == 0:
send_sms(f"{len(data_details)} links scraped")
# FINAL PRINT STATEMENT AND SMS SENDING
print("ALL LINKS SCRAPED")
send_sms('ALL LINKS SCRAPED')
# Send scraped data into DataFrames and then into csv's
data_hirier = pd.DataFrame(data=data_hirer)
data_details = pd.DataFrame(data=data_details)
data_core = pd.DataFrame(data=data_core)
data_questions = pd.DataFrame(data=data_questions)
data.to_csv('data_.csv')
data_hirier.to_csv('data_hirer_.csv')
data_details.to_csv('data_details_.csv')
data_core.to_csv('data_core_.csv')
data_questions.to_csv('data_questions_.csv')In this section I:
- Performed a brute force method to create the 'best' possible decision tree classifier model.
- Created a Logistical Regression Classifier model that is easily intrepretable using ELI5 to gain key word insights
- Describe insights gained from inferencing our Logistical Regression Model
The impetus of this section was to understand and constrast computationally construsted model and relatively simple to build (and understand) models.
VIEWING DATASET
import pandas as pd
import numpy as np
import seaborn as sns
import matplotlib.pyplot as pltdf = pd.read_csv("./SubmissionProject4/data_science.csv")
df.head()I performed two GridSearchCV's on two pipelines.
- To find the optimal CountVectorizer, TfidTransformer, SelectFdr, and DecisionTreeClassifier hyperparameters
- To find the optimal BaggingClassifier hyperparameters that used the above Pipeline as it's base estimator
from sklearn.model_selection import train_test_split, cross_val_score, GridSearchCV
from sklearn.feature_extraction.text import CountVectorizer, TfidfTransformer
from sklearn.feature_selection import SelectFdr
from sklearn.metrics import plot_precision_recall_curve, plot_roc_curve, classification_report
import sklearn.metrics as metrics
from sklearn.pipeline import Pipeline
from sklearn.tree import DecisionTreeClassifier
from sklearn.ensemble import BaggingClassifier# Pipeline and GridSearchCV for best DecisionTreeClassifier
# Features and Target
X = df.job_description
y = df.above_ave_salary
print("===={:=<60}".format('Initialising GridSearch: DecisionTree Classifier'))
print("===={:=<60}".format(""))
# Train Test Split
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.25, random_state=101)
print("===={:=<60}".format('Splitting Data into Training and Validation Models'))
print("===={:=<60}".format(""))
# Making Pipeline
print("===={:=<60}".format('Making Pipeline'))
cv = CountVectorizer()
tf = TfidfTransformer()
rdf = SelectFdr()
tree_c = DecisionTreeClassifier()
pipe = Pipeline([('cv', cv),
('tf', None),
('rdf', None),
('tree', tree_c)])
print("===={:=<60}".format(""))
# Defining GridSearch Parameters
print("===={:=<60}".format('Performing GridSearch'))
param = {'cv__stop_words': [None, 'english'],
'cv__max_df': [1.0, 0.5],
'cv__min_df': [1, 0.1],
'cv__ngram_range': [(1,1), (1,3)],
'tf': [TfidfTransformer()],
'tf__norm': ['l1', 'l2'],
'rdf': [SelectFdr()],
'rdf__alpha': [0.05, 0.1],
'tree__criterion': ['gini', 'entropy']}
# Performing GridSearch
grid_search = GridSearchCV(estimator=pipe, param_grid=param, cv=5,
n_jobs=-1, verbose=1)
grid_search.fit(X_train, y_train)
print(grid_search.best_params_)grid_search.best_score_# Pipeline and GridSearchCV for BaggingClassifier
# Features and Target
X = df.job_description
y = df.above_ave_salary
print("===={:=<60}".format('Initialising GridSearch: BaggingClassifier'))
print("===={:=<60}".format(""))
# Train Test Split
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.25, random_state=101)
print("===={:=<60}".format('Splitting Data into Training and Validation Models'))
print("===={:=<60}".format(""))
# Making Pipeline
print("===={:=<60}".format('Making Pipeline'))
cv = CountVectorizer(max_df=0.5, min_df=1, ngram_range=(1,1), stop_words=None)
tf = TfidfTransformer(norm= 'l2')
rdf = SelectFdr(alpha=0.1)
tree_c = DecisionTreeClassifier(criterion='entropy')
bag_c = BaggingClassifier(tree_c)
pipe = Pipeline([('cv', cv),
('tf', tf),
('rdf', rdf),
('bag', bag_c)])
print("===={:=<60}".format(""))
# Defining GridSearch Parameters
print("===={:=<60}".format('Performing GridSearch'))
param = {'bag__n_estimators': [10, 50, 100, 400],
'bag__max_samples': [1.0, 0.5, 0.2],
'bag__max_features': [1.0, 0.5, 0.2]}
# Performing GridSearch
grid_search_2 = GridSearchCV(estimator=pipe, param_grid=param, cv=5,
n_jobs=-1, verbose=1)
grid_search_2.fit(X_train, y_train)grid_search_2.best_params_grid_search_2.best_score_# 'Best' Bagging Classifier on whether job is above salary
# Features and Target
X = df.job_description
y = df.above_ave_salary
print("===={:=<60}".format('Initialising Model: Bagging Classifier'))
print("===={:=<60}".format(""))
# Train Test Split
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.25, random_state=101)
print("===={:=<60}".format('Splitting Data into Training and Validation Models'))
print("===={:=<60}".format(""))
# Count Vectorising and Tfid Vectorising
cv = CountVectorizer(ngram_range=(1,1), stop_words=None, max_df=0.5, min_df=1)
X_train = cv.fit_transform(X_train)
X_test = cv.transform(X_test)
print("===={:=<60}".format('Count Vectorising Text Data'))
print("===={:=<60}".format(f"No. of Features: {X_train.shape[1]}"))
print("===={:=<60}".format(""))
tf = TfidfTransformer(norm='l2')
X_train = tf.fit_transform(X_train)
X_test = tf.transform(X_test)
print("===={:=<60}".format('Tfidf Vectorising Text Data'))
print("===={:=<60}".format(f"No. of Features: {X_train.shape[1]}"))
print("===={:=<60}".format(""))
# Reducing Demensionality based on Fase Discovery Rate
fdr = SelectFdr(alpha=0.1)
X_train = fdr.fit_transform(X_train, y_train)
X_test = fdr.transform(X_test)
print("===={:=<60}".format('Reducing Feature Dimensions using FDR'))
print("===={:=<60}".format(f"No. of Features: {X_train.shape[1]}"))
print("===={:=<60}".format(""))
# Bagging Tree Classifier
print("===={:=<60}".format('Fitting Bagging Decision Tree Classifier'))
tree_c = DecisionTreeClassifier(criterion='entropy')
bag_c = BaggingClassifier(base_estimator=tree_c,
n_estimators=100,
max_features=1.0,
max_samples=0.5,
n_jobs=2)
bag_c.fit(X_train, y_train)
# Create Model Predictions
y_pred = bag_c.predict(X_test)
print("===={:=<60}".format('Creating Model Predictions'))
print("===={:=<60}".format(""))
# Validation of Our Model
print("===={:=<60}".format('Validating Training Model'))
print("===={:=<60}".format('Validation Scores'))
scores = cross_val_score(bag_c, X_train, y_train, cv=5, n_jobs=2)
print(scores)
print("===={:=<60}".format('Mean'))
print(scores.mean())# Bagging Decision Tree Classifier Metrics
print("===={:=<60}".format('Classification Report: Bagging Decision Tree Classifier'))
print(metrics.classification_report(y_test, y_pred))
print("===={:=<60}".format('ROC Curve & Precision Recall Curve'))
metrics.plot_roc_curve(bag_c, X_test, y_test)
metrics.plot_precision_recall_curve(bag_c, X_test, y_test)
plt.show()I created two Logistical Regression Models. Only a CountVectorizer was used for feature extraction and preprocessing to maintain intrepretability of the model using eli5. The two models differ based on the ngram_range parameter in the CountVectorizer: The length, in words, that a feature extracted from the text can be. This was done to corroborate the insights gained from both models.
# Logistic Regression Classifier on whether job is above salary
# Features and Target
X = df.job_description
y = df.above_ave_salary
print("===={:=<60}".format('Initialising Model: Logistic Regression'))
print("===={:=<60}".format(""))
# Train Test Split
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.25, random_state=101)
print("===={:=<60}".format('Splitting Data into Training and Validation Models'))
print("===={:=<60}".format(""))
# Count Vectorising
cv = CountVectorizer(ngram_range=(1,1))
X_train = cv.fit_transform(X_train)
X_test = cv.transform(X_test)
print("===={:=<60}".format('Count Vectorising Text Data'))
print("===={:=<60}".format(f"No. of Features: {X_train.shape[1]}"))
print("===={:=<60}".format(""))
# Logistic Regression Classifier
print("===={:=<60}".format('Fitting Logistic Regression'))
lr = LogisticRegression(max_iter=500)
lr.fit(X_train, y_train)
# Create Model Predictions
y_pred = lr.predict(X_test)
print("===={:=<60}".format('Creating Model Predictions'))
print("===={:=<60}".format(""))
# Validation of Our Model
print("===={:=<60}".format('Validating Training Model'))
print("===={:=<60}".format('Validation Scores'))
print(cross_val_score(lr, X_train, y_train, cv=5, n_jobs=2))
print("===={:=<60}".format('Mean'))
print(cross_val_score(lr, X_train, y_train, cv=5, n_jobs=2).mean())from sklearn.linear_model import LogisticRegression# Logistic Regression Classifier on whether job is above salary
# Features and Target
X = df.job_description
y = df.above_ave_salary
print("===={:=<60}".format('Initialising Model: Logistic Regression'))
print("===={:=<60}".format(""))
# Train Test Split
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.25, random_state=101)
print("===={:=<60}".format('Splitting Data into Training and Validation Models'))
print("===={:=<60}".format(""))
# Count Vectorising
cv = CountVectorizer(ngram_range=(1,1))
X_train = cv.fit_transform(X_train)
X_test = cv.transform(X_test)
print("===={:=<60}".format('Count Vectorising Text Data'))
print("===={:=<60}".format(f"No. of Features: {X_train.shape[1]}"))
print("===={:=<60}".format(""))
# Logistic Regression Classifier
print("===={:=<60}".format('Fitting Logistic Regression'))
lr = LogisticRegression(max_iter=500)
lr.fit(X_train, y_train)
# Create Model Predictions
y_pred = lr.predict(X_test)
print("===={:=<60}".format('Creating Model Predictions'))
print("===={:=<60}".format(""))
# Validation of Our Model
print("===={:=<60}".format('Validating Training Model'))
print("===={:=<60}".format('Validation Scores'))
print(cross_val_score(lr, X_train, y_train, cv=5, n_jobs=2))
print("===={:=<60}".format('Mean'))
print(cross_val_score(lr, X_train, y_train, cv=5, n_jobs=2).mean())# Logistic Regression Metrics
print("===={:=<60}".format('Classification Report: Logistic Regression'))
print(metrics.classification_report(y_test, y_pred))
print("===={:=<60}".format('ROC Curve & Precision Recall Curve'))
metrics.plot_roc_curve(lr, X_test, y_test)
metrics.plot_precision_recall_curve(lr, X_test, y_test)
plt.show()# Logistic Regression Classifier on whether job is above salary
# Features and Target
X = df.job_description
y = df.above_ave_salary
print("===={:=<60}".format('Initialising Model: Logistic Regression, ngram_range=(1,3)'))
print("===={:=<60}".format(""))
# Train Test Split
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.25, random_state=101)
print("===={:=<60}".format('Splitting Data into Training and Validation Models'))
print("===={:=<60}".format(""))
# Count Vectorising
cv_2 = CountVectorizer(ngram_range=(1,3))
X_train = cv_2.fit_transform(X_train)
X_test = cv_2.transform(X_test)
print("===={:=<60}".format('Count Vectorising Text Data'))
print("===={:=<60}".format(f"No. of Features: {X_train.shape[1]}"))
print("===={:=<60}".format(""))
# Logistic Regression Classifier
print("===={:=<60}".format('Fitting Logistic Regression, ngram_range=(1,3)'))
lr_2 = LogisticRegression(max_iter=500)
lr_2.fit(X_train, y_train)
# Create Model Predictions
y_pred = lr_2.predict(X_test)
print("===={:=<60}".format('Creating Model Predictions'))
print("===={:=<60}".format(""))
# Validation of Our Model
print("===={:=<60}".format('Validating Training Model'))
print("===={:=<60}".format('Validation Scores'))
print(cross_val_score(lr_2, X_train, y_train, cv=5, n_jobs=2))
print("===={:=<60}".format('Mean'))
print(cross_val_score(lr_2, X_train, y_train, cv=5, n_jobs=2).mean())# Logistic Regression Metrics
print("===={:=<60}".format('Classification Report: Logistic Regression ngram_range=(1,3)'))
print(metrics.classification_report(y_test, y_pred))
print("===={:=<60}".format('ROC Curve & Precision Recall Curve'))
metrics.plot_roc_curve(lr_2, X_test, y_test)
metrics.plot_precision_recall_curve(lr_2, X_test, y_test)
plt.show()ELI5 is a simple library that lets you quickly view the heaviest weights of a model. For NLP, the weights reveal insights on word association to the target variable.
import eli5 as eli5# Showing key features from LogisticRegression
eli5.show_weights(estimator=lr, top=30, feature_names=cv.get_feature_names(),
target_names=['Low','High Salary Data Job'])# Showing key feautres from LogisticRegression ngram_range=(1,3)
eli5.show_weights(estimator=lr_2, top=30, feature_names=cv_2.get_feature_names(),
target_names=['Low','High Salary Data Job'])In this section I attempted to gain insights into the textual differences between science relate and non-science related data jobs.
I used two models with emphasis on intrepretability using ELI5:
- Support Vector Classifier
- AdaBoost ExtraTreeClassifier
The impetus of this section was to understand and create models I typically wouldn't use for this approach.
VIEWING DATASET
df.head()df.classification.value_counts()# Why not classify jobs based on whether they are from a science related industry
science = ["Information & Communication Technology", "Science & Technology",
"Healthcare & Medical", "Mining, Resources & Energy",
"Farming, Animals & Conservation"]
y = [1 if industry in science else 0 for industry in df['classification']]
print("===={:=<60}".format('Baseline'))
np.mean(y)from sklearn.svm import SVC
from sklearn.preprocessing import StandardScaler# SVM Classifier on whether job is science related
# Features and Target
science = ["Information & Communication Technology", "Science & Technology", "Healthcare & Medical", "Mining, Resources & Energy", "Farming, Animals & Conservation"]
X = df.job_description
y = [1 if industry in science else 0 for industry in df['classification']]
print("===={:=<60}".format('Initialising Model: SVC'))
print("===={:=<60}".format(""))
# Train Test Split
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.25, random_state=101)
print("===={:=<60}".format('Splitting Data into Training and Validation Models'))
print("===={:=<60}".format(""))
# Count Vectorising
cv = CountVectorizer(ngram_range=(1,1))
X_train = cv.fit_transform(X_train)
X_test = cv.transform(X_test)
print("===={:=<60}".format('Count Vectorising Text Data'))
print("===={:=<60}".format(f"No. of Features: {X_train.shape[1]}"))
print("===={:=<60}".format(""))
# Scaling our Data
print("===={:=<60}".format('Scaling Data'))
scaler = StandardScaler()
X_train = scaler.fit_transform(X_train.todense())
X_test = scaler.transform(X_test.todense())
print("===={:=<60}".format(""))
# SVM Classifier
print("===={:=<60}".format('Fitting SVC'))
svc = SVC(kernel='linear')
svc.fit(X_train, y_train)
# Create Model Predictions
y_pred = svc.predict(X_test)
print("===={:=<60}".format('Creating Model Predictions'))
print("===={:=<60}".format(""))
# Validation of Our Model
print("===={:=<60}".format('Validating Training Model'))
print("===={:=<60}".format('Validation Scores'))
scores = cross_val_score(svc, X_train, y_train, cv=5, n_jobs=2)
print(scores)
print("===={:=<60}".format('Mean'))
print(scores.mean())# SVC Metrics
print("===={:=<60}".format('Classification Report: SVC'))
print(metrics.classification_report(y_test, y_pred))
print("===={:=<60}".format('ROC Curve & Precision Recall Curve'))
metrics.plot_roc_curve(svc, X_test, y_test)
metrics.plot_precision_recall_curve(svc, X_test, y_test)
plt.show()eli5.show_weights(estimator=svc, top=(15, 15), feature_names=cv.get_feature_names(),
target_names=['Non Science Data Job','Science Data Job'])from sklearn.ensemble import AdaBoostClassifier
from sklearn.tree import ExtraTreeClassifier# AdaBoost Classifier on whether job is science related
# Features and Target
science = ["Information & Communication Technology", "Science & Technology", "Healthcare & Medical", "Mining, Resources & Energy", "Farming, Animals & Conservation"]
X = df.job_description
y = [1 if industry in science else 0 for industry in df['classification']]
print("===={:=<60}".format('Initialising Model: AdaBoost Classifier (ExtraTreeClassifier)'))
print("===={:=<60}".format(""))
# Train Test Split
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.25, random_state=101)
print("===={:=<60}".format('Splitting Data into Training and Validation Models'))
print("===={:=<60}".format(""))
# Count Vectorising
cv = CountVectorizer(ngram_range=(1,1))
X_train = cv.fit_transform(X_train)
X_test = cv.transform(X_test)
print("===={:=<60}".format('Count Vectorising Text Data'))
print("===={:=<60}".format(f"No. of Features: {X_train.shape[1]}"))
print("===={:=<60}".format(""))
# # Scaling our Data
# print("===={:=<60}".format('Scaling Data'))
# scaler = StandardScaler()
# X_train = scaler.fit_transform(X_train.todense())
# X_test = scaler.transform(X_test.todense())
# print("===={:=<60}".format(""))
# AdaBoost Classifier (ExtraTreeClassifier)
print("===={:=<60}".format('Fitting AdaBoost Clasifier (ExtraTreeClassifier)'))
e_tree = ExtraTreeClassifier()
ada = AdaBoostClassifier(e_tree, n_estimators=200)
ada.fit(X_train, y_train)
# Create Model Predictions
y_pred = ada.predict(X_test)
print("===={:=<60}".format('Creating Model Predictions'))
print("===={:=<60}".format(""))
# Validation of Our Model
print("===={:=<60}".format('Validating Training Model'))
print("===={:=<60}".format('Validation Scores'))
scores = cross_val_score(ada, X_train, y_train, cv=5, n_jobs=2)
print(scores)
print("===={:=<60}".format('Mean'))
print(scores.mean())# AdaBoost (ExtraTreeClassifier) Metrics
print("===={:=<60}".format('Classification Report: AdaBoost Classifier (ExtraTreeClassifier)'))
print(metrics.classification_report(y_test, y_pred))
print("===={:=<60}".format('ROC Curve & Precision Recall Curve'))
metrics.plot_roc_curve(ada, X_test, y_test)
metrics.plot_precision_recall_curve(ada, X_test, y_test)
plt.show()eli5.show_weights(estimator=ada, top=30, feature_names=cv.get_feature_names(),
target_names=['Non Science Data Job','Science Data Job'])