matrix_factorisation_python_sqlite_recommender

Features

• Configurable number of items in Recommendation
• Considers item if skipped
• Favours new items in Recommedation Generation
• Model can be saved to pickle/sqlite
• Option to set a minimum score
• Can recommend both catagory wise and global products
• User can rate the Product on the Scale of 1 to 5

Installation

Data Set

266 Products
7 catagories
14 user
838 reviews [may differ]

Run Command

cd matrix_factorisation_python__sqlite_recommender
python3 main.py

Future Development

• Use of Apiori Algorithm to leverage items that are mostly brought together

Other side of Matrix Factorisation

In Matrix Factorisation the weights are learnt by itself which is great when the data is sparse & most traditional model fails to deal with this scenario but over the time when data gets dense there is no scope of improvement cause:

• We cannot add any extra weight externally
• The vector length is fixed so any feature prior or post User-item interaction cannot be taken into consideration
• All correlation between features are treated naively hence we cannot include any constant or variable human factors into the computation hence it under-fits dense data
• Generalisation is the strength of Matrix Factorisation Models and Also the weakness
• Over the time it’ll start recommending product with very less confidence score
• Appart from Suggestions there are situation like sending Email / Push notifications to the users requires items to have very high confidence score.

Project History

Started as a part of an Job Assignment - "Tinder for Software Stacks" & i got the job

Help and Support

Website: http://qbitdata.in/

Linkedin: https://www.linkedin.com/in/lubdhak/

Special Thanks to www.analyticsvidhya.com