CODEWITHSUNDEEP
tensorflowtensorflow-transform
Utkarsh D
Utkarsh D

Reputation: 11

Creating Tensors from features that are linked together

I have a set of multi valued features which are linked together. As an example,

ItemCodes Scores
AK, NA, UY 0.6, 0.2, 0.2
KG, AK 0.5, 0.5

Each Item has a corresponding score associated with it. Some rows might not have any items/scores. I want to convert this dataset (there are other numerical features) into a form that can be fed to an Neural Net. The data comes from a different part of the system with its own API.

I was trying to create a vector of item codes (binary, if the item is present or not) and a second vector with score values at the corresponding indices. If I only had Item codes, I could do a multi-hot encoding and get a feature vector of items. So far I am using

values = tft.compute_and_apply_vocabulary(itemcodes)

to get the indices, which I can then set to 1 in the output Tensor. But, if item AK is allotted index j in the multi-hot Tensor, how do I ensure 0.6 is also set at index j in the second Tensor? Because there can be missing values, the ItemCodes are Scores are available as a SparseTensor and so I am unable to iterate directly. How can I achieve what I want? Or is there a better way to represent such features?

Upvotes: 1

Views: 44

Answers (1)

Pritam Dodeja
Pritam Dodeja

Reputation: 326

I think the fundamental question to answer is what does each row represent? If ItemCodes are static, you can create a lookup table, and place the associated score at the right index. If you provide more code/details, I can likely help you with this.

Upvotes: 0

Related Questions