0
$\begingroup$

I am doing a project that involves training and testing different algorithms to predict a developer's profile type (Frontend, Fullstack, QA, ML, etc.) using that developer's skills (AWS, Selenium, Hadoop, Java, etc.).

I am looking more in depth at correct and incorrect classifications made by the Logistic Regression (OVR) algorithm to understand whether the presence or absence of particular skills cause this to happen. With my example, I use Fullstack-Fullstack as a correct classification and Fullstack-ML as incorrect.

***# returns skills present in correct classifications which are not present in incorrect classifications***
correct_classification_skills.difference(incorrect_classification_skills)

Returned:
*['ASP.NET', 'AWS', 'Adobe Photoshop', 'Agile', 'Ajax',
       'Amazon Web Services (AWS)', 'Angular', 'AngularJS', 'Apache', 'Azure',
       'Backbone.js', 'Blockchain', 'Bootstrap', 'C++', 'CSS', 'Cassandra',
       'Docker', 'Drupal', 'Eclipse', 'Elixir', 'Express.js', 'Flask',
       'Full-stack', 'Git', 'GitHub', 'Google Analytics', 'HTML', 'HTML5',
       'JIRA', 'JSON', 'Java', 'JavaScript', 'MSSQL', 'MacOS', 'Meteor',
       'Microsoft', 'Mobile', 'MySQL', 'Nginx', 'Node.js',
       'Object-oriented Programming (OOP)', 'PHP', 'PostgreSQL', 'Python',
       'REST APIs', 'React', 'React Native', 'React.js', 'Redux', 'Ruby',
       'SCSS', 'Shopify', 'Swift', 'Ubuntu', 'Visual Studio', 'WordPress']*

***# returns skills present in incorrect classifications which are not present in correct classifications***
incorrect_classification_skills.difference(correct_classification_skills)

Returned:
*['SQL']*

My question is: is one more important than the other? In other words, is it more important to look at the skills absent or present?

Thank you so much for the help!

Improve this question
$\endgroup$
1
  • $\begingroup$ In my opinion it doesn't make a lot of sense to ask which feature causes errors in general: particular features could be "blamed" for an error in a specific instance yes, but across all instances if a feature is a general cause of error then I would assume that its distribution is different in the training and test data (or it's simply not frequent enough so the training set is not a representative sample). What I would suggest is to try to cluster the instances and see which cluster tend to have good/bad performance. $\endgroup$ – Erwan Feb 2 at 12:59

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.