Shlomi Abuchatzera Green
Data Science Fellows February 2021 Cohort
Abstract
Using a mobile app, local businesses, also known as Points of Consumption (POC), can order required goods (such as brands beer) and have their spent money (sell-in) tracked. Knowing the location of the POC and their spends in any given period, our goal is to create a new framework currently named PLU (POCs Like You). The framework will be able to use different clustering algorithms adjustable for different datasets (in example different countries or locations) and partitioning criteria.
Challenges
- Handling outliers: outliers should be detected and handled in each of the clustering steps.
Outliers can be geographical or spend-based. The proper way to handle outliers can vary. - Support for different features for clustering: for certain areas, product requests (raised request from product level) can vary and the generated process should be agile for those requests.
- Evaluation of clustering: defining and finding the right target functions for evaluation. These can be specific to the clustering method or to subjectable to product request.
Achievements
- Understand each step of the existing non-generalized pipeline.
- Simplify the process for an easy-to-run configuration.
- Optimize outlier detection and handling as an integrated part of the process.
- Automate hyperparameter tuning for an optimal target function.
Future project development
Generating an automated framework will make it easy to find the best clustering algorithm and parameters for particular use cases according to a given dataset; Given a dataset of POCs’ location and spends, find the most similar POCs under any given conditions. Provide clustered POCs insights to improve based on more successful POCs in their cluster.
