I want to predict price of products. For each product, I use one-hot encoding to model their features. These features come from a limited set of fields (i.e., product attributes). For example, a field color may have five features: "red", "green", "yellow", "black", and "white". A product can have one of these colors.
I also have a field called "Expert-Suggested Price" (e.g. f51-f54 in the figure). However, for some reasons, the number of feature values a product has is not fixed: Some products have no any suggested price, while some have 3-4 suggested prices.
I know that for categorical field, we can add one more dummy feature such as "unknown". However, the prices here are numerical, instead of categorical. The price is in range [0, +inf).
The problem: If I directly vectorize suggested prices, I am afraid the products which have four suggested prices may have bias compared with those have no suggested price.