Q3 and Q4 weekly progresses on the Research front for the Image Matching structured task
- Worked on image placeholder identification [T277828]: identified a category-based approach to label images as placeholders, and incorporated it into the algorithm. Trained a simple neural network to distinguish between normal images and placeholder images, with accuracy 88%+
- Worked with Platform Engineering to fix bug T277875 and identified the source of the problem
- Helped Structured Data refining the design of the test for image recommendations POC results T273092
- Ran accuracy diagnositcs on the data from T273057. For the 2 largest sources, it looks like the algorithm has a 64-71% accuracy, meaning the majority of users say@Ai that the algorithm recommends a good match:
It looks like the time it takes a user to give a response is also inversely proportional with goodness of the match:
- Next @AikoChou and I will breakdown these metrics by user expertise and topic.
- Discussed with ML team + Platform + Product about when and how to convert this algo into a model that can be served through LiftWing.
Analyzed with Aiko the results of the Android POC, to understand
(1) The extent to which newcomers behave and annotate data differently
(2) The extent to which non-english users struggle with the POC
(3) The reliability of newcomers annotations (via agreement)
Finalized the analysis of Android data, and helped the Growth team with the decision-making process around whether to deploy the "add an image" task as part of the newcomers structured tasks. They decided to go for it next fiscal year.