Course Selector.

An algorithm to predict a recommendation for a learning course.

Manuel Zimmermann | Data Analytics, Campus Ironhack Berlin - 12. March 2021

Content

• Project Description
• Hypotheses / Questions
• Dataset
• Cleaning
• Model Training and Evaluation
• Conclusion
• Workflow
• Links

Project Description

Over 900 learning course needs to be reviewed and migrated to a new software. The workforce is low and the level of effort is high. The purpose of this project is to help the Learning Team to decided wheter an exiting course is worth to relocate in a new Learning Record Store or not.

Hypotheses / Questions

Can a machine learning algorithm decide if learning course is worth to keep and reduce the amount of courses that needs to be relocated?

Dataset

• The dataset is build on multiple reports from the exicting software:
  • Course_details.csv ~1150
  • Completion_count.csv ~163.567
  • Course_categoy.csv ~93
  • User.csv ~6391
  • site_log.csv ~385.682
• To manage the the reports a SQL Database was developed with multiple table
  • 3NF (Third Normal Form)

Cleaning

Before cleaning the data in Jupyter Notebook with pandas, I collected and prepeared the nessceary data from the SQL Database that I created.

connection_string = 'mysql+pymysql://root:' + password + '@localhost/clz'
engine = create_engine(connection_string)
data = pd.read_sql_query('SELECT 
                         count(*) AS completion_count, c.course_category, cc.course_id, date_created, c.course_name, c.location 
                         FROM course_completion cc
                         JOIN course c ON c.course_id = cc.course_id
                         WHERE Completion_Status IN ('Complete', 'In progress')
                         AND c.course_category != 'Archive' AND c.course_category != 'Sandbox'
                         GROUP BY Course_id;', engine
                        )

• Cleaning Porcess (Python (Numpy & Pandas))
  1. Header Standardization:
     • Snake Casing for header standardization due to its simplicity. Snake Casing is a convention which replaces spaces with underscores and converts any upper-case letters to lower-case.
  2. Dtypes
     • Assigning the right data types to the features.
       • date_created.dt.days
  3. Checking for and removing NaN Value
     • At this point I checked for the percentage of null values per column.
       • location ~18%
       • content ~6%
       • course_type ~18%
       • tags ~35%
     • Mostly NaNs are filled with 'Unkown' or 0.
  4. Column Split
     • Feature 'content' & 'tags' contains multiple strings. Both columns are encoded with pd.get_dummies

Model Training and Evaluation

• Encoding Categoricals

  • Categorical features course_type & course_category were encoded with pandas .astype('category')and .cat.codes

• Scaling Numericals

  • I opted to use a sklearn StandardScaler due to the distribution shape being non-normal. It ensures that our variables are scaled to values within -4 and 12.

• Cross Validation

  • I used the Hyperparameter tuning from sklearn.model_selection import GridSearchCV in order to determine the optimal values for a given model.

• Model Training

  • Selecting Classifier:
    • DecisionTreeClassifier
    • LogisticRegression
    • RandomForestClassifier
  • Evaluation Classification Report:

Accuracy of LogisticRegression on test set: 0.84

 [[17  4]
 [ 2 15]]
              precision    recall  f1-score   support

           0       0.89      0.81      0.85        21
           1       0.79      0.88      0.83        17

    accuracy                           0.84        38
   macro avg       0.84      0.85      0.84        38
weighted avg       0.85      0.84      0.84        38


Accuracy DecisionTreeClassifier on test set: 0.79

 [[16  5]
 [ 3 14]]
              precision    recall  f1-score   support

           0       0.84      0.76      0.80        21
           1       0.74      0.82      0.78        17

    accuracy                           0.79        38
   macro avg       0.79      0.79      0.79        38
weighted avg       0.80      0.79      0.79        38


Accuracy RandomForestClassifier on test set: 0.84

 [[18  3]
 [ 3 14]]
              precision    recall  f1-score   support

           0       0.86      0.86      0.86        21
           1       0.82      0.82      0.82        17

    accuracy                           0.84        38
   macro avg       0.84      0.84      0.84        38
weighted avg       0.84      0.84      0.84        38

Conclusion

• With the optimal parameters for the RandomForestClassifier I get an accuracy of 0.84 = 84%. In short, 8 of 10 are correct predicted.
• Best paramters:

    criterion = 'entropy', 
    max_depth = 12, 
    max_leaf_nodes = 4, 
    min_samples_split = 2,
    n_estimators = 69

Workflow

Gathering Data • Building SQL Database • Feature Selection • Cleaning • Scaling • Train&Test • HyperTuning • Applying Model • Visualization

Links

Repository
Presentaion