Ml application on_student_non_deployment
Download as PPTX, PDF0 likes49 views
This document discusses using machine learning models to predict whether students who complete a technical training program will be deployable or non-deployable. It analyzes demographic, academic, training, and assessment data from over 3,000 students over three years. Feature selection methods identified important variables like education stream, technical skills, training duration, gender, and hiring method. Classification models including logistic regression and two deep learning models were trained on 70% of the data and evaluated on 30%. The deep learning models significantly outperformed logistic regression, with AUCs of 0.86 and 0.73, indicating they can better distinguish deployable from non-deployable students.
Ml application on_student_non_deployment
- 2. What it is all about? 2 Client/Business that provides several technical trainings to fresh graduates are interested in understanding distinguishing factors that differentiated between deployable and non deployable students. Business get badly affected with non deployable ones from both prospect revenues and operational cost perspectives. Hence, interested in leveraging data science process in order to identify certain student attributes which help them in distinguishing in front in order to avoid all the business loss. Current study uses machine learning techniques to provide actionable insights to the businesses. Business has provided past 3 years data that contains students basic demographics, training details and deployment details. Objective: “Model that distinguishes non deployable ones for all future uses.”
- 3. Data Understanding & Preparation 2 Data of 44 columns with 3470 rows have student demographics, academic background, training and assessment details, deployment details, etc. It has around ten to fifteen columns that measured students training and deployment against their difficulty, attendance and repetitiveness. Columns are measured and available in both numeric and character formats. All columns are duly converted to needed formats namely., dates, binary responses, nominal to numeric, etc. Found few columns (5) with missing value percentage more than 50% have been excluded. Columns representing identification ones are ignored for analysis.
- 4. Feature Selection & Training: After the employment of feature selection algorithms infogain and glmnet the following variables have been used for training the model: Education stream belonging to ECE, and IT, Technical skill belonging to Java & Testing, Number of Training Days underwent, gender, quarter of the year, hiring mode method. Then, data has been partitioned into training and testing dataset in the respective ratio of 70:30. Due to nature of binary dependent (target), we employed three classification model namely., Binary Logistic Regression for base model comparison and two deep learning models using two different libraries. Below is glimpse of high p-value independents (features):
- 5. Validation Below are the accuracy with respect to different models on test data set. As, it is evident from above, BLR as base model couldn’t provide appropriate AUC value, where as both deep learning classifier libraries provided very good AUC accuracy measure against BLR. Thus, deep learning classifier provided better classification than traditional binary logistic regression model. Models Accuracy (AUC) Binary Logistic Regression (BLR) 0.54 Deep Learning (using tensor flow) 0.86 Deep Learning (using H2o) 0.73