How to better debug advanced recommender results

How Advanced Recommender works

1. We first ask customers for information like age, weight, height, and bra size (for female tops). Then we use that information to estimate the person's body measurement. There is a linear regression model involved. When bra size is provided, the chest size is directly calculated from the bra size.
2. Given the estimated body measurement, we then compare it with the size chart table and use the additional recommender field configuration provided by the merchant to compute a "fit score" for each size. And then recommend the size with the highest score.

Common recommender issues

• Including the wrong recommender fields. For example, only the actual measurements should be used for the recommender. Other information like "UK sizes" should not be included. Generally, just 1 or 2 recommenders need to be used.

• Incorrect canonical measurement linking. It is important to confirm whether the measurements in the size charts are the intended body measurement OR the actual garment measurement. If it is garment measurement, a common mistake is choosing between "Circumference" and "Length only

How to debug recommender issues

• The best way is to provide different recommender inputs using the user match feature. If none of the inputs recommend any sizes, this is usually because of an incorrect recommender field or canonical field setting (see above). The user match can give a quick way to see how our recommended size match up with the expected size.

• If there are recommendations, but they are off, then we can check if they are consistently larger or smaller. This may be an indicator of how much to tweak the easing. Lowering the easing will nudge the recommendation to smaller sizes, whereas bigger easing indicates looser fit and larger sizes.

• When working with our customer support team, it is also helpful to provide the actual body measurements for the users you are testing with. This way we can compare with our estimated body measurement. 

How to tweak the recommender setup

• if there is "length" recommender field involved, check if the "ideal length reference point" is reasonable based on the type of the product. For example, a shirt usually falls around the crotch area, but a jacket usually fits 10-15 cm under the crotch. The"Acceptable range from reference point"can also be increased to make the length a less sensitive factor

• For chest, waist, and hip, tweak the easing. Use the user match statistics as a guide and tweak easing manually to find the ones with the highest match rate. Note that easing can only be tweaked when the canonical measurement in product measurement

How to use User match

Guide: Adding Advanced Recommender User Match

User match is a feature where merchants can provide different potential input to the recommender and see how our recommendation matches up against the expected size. There is also an experimental feature for "automatically" tweaking the configuration. This may not always work, however. But merchants can also tweak the settings manually and directly see how this affects the matching rate.