My name is Tirtha Changmai, Machine Learning Data Associate and Alexa Data Services – Amazon. I have pursued my Undergraduate in Mechanical Engineering. I have over 5 years of industry experience ranging from technical customer support, content writing, and data pre-processing over the years. Pursuing PGP data science is undoubtedly the best passion that I developed during my tenure working in the data industry. It is fascinating to see how much knowledge is available in the world that we can consume to solve an existing problem within our professional domain.
My Machine Learning Data Associate duties primarily involve Goal Evaluation to generate high-quality labeled data. There are various challenges associated with various domains of data pre-processing. As working from home has become the new normal, we have resorted to various tactics to keep the quality of our work afloat. However, one major problem we faced within our team was failing to generate consistent labeled data with every new joiner.
To facilitate their learning and reduce contact with other associates, I have come up with an idea to create a decision tree structure that would be applied to the existing knowledge forum. This would focus on optimizing the mentioned KPIs:
- Reducing errors in pre-processing
- Minimizing AHT (Average Handling Time)
- Reducing dependency on other Associates
The first step was to understand the challenges faced by the new joiners. I have facilitated team connections to collect their needs. As I understood from our connections, the major challenge was navigating through the knowledge forum, contributing to an increase in AHT and lack of visual interpretation leading to confusion and controllable human errors.
As decision trees help us visually understand the flow of decisions, I have built a structure considering the various steps we take while evaluating a user goal.
This has helped everyone immensely as evaluations were streamlined to be followed across the team while improving quality.
Key takeaways from this project involve:
- Reduction in AHT by 20% as referencing is limited to a single source in comparison to the multi-level structure of the knowledge forum.
- Quality improvement from 87% to 90% and above.
- Visualization of flow is easier to read without spending an enormous amount of time that would affect the overall AHT.
