How to handle multi label features with the dataset construction?

Nobody has claimed this yet.

Hello! Thanks a lot for this amazing recommender system.

I have a question regarding the procedure of creating the one-hot encoding feature matrices.

Example: I have scenario where items have a feature called "tags", which means that an item may be associated with several tags. I really want to use the provided dataset module to create the necessary mappings and matrices, but I do not understand the best way to go about it. I see two options, but would love to know if there is another, more convenient way, and I imagine that more people have this problem.

| Item |provider| Tags |
| ----- |---| ----- |
| itemA | providerX | Tag1 |
| itemB | providerZ | [Tag1, Tag2, Tag3] |

1. Format the input to the build_item_features() function as following:

| Item |provider| Tags |
| ----- |---| ----- |
| itemA | providerX | Tag1 |
| itemB | providerZ | Tag1 |
| itemB | providerZ | Tag2 |
| itemB | providerZ | Tag3 |

This would however increase the one-hot encoded weight on the "provider" feature to 3. Which means that in the end I set all feature values >1 equal to 1. But is this the way to go?

2. The second option is to do it all manually, which would be a shame since you got this powerful functionality for it.

Thanks in advance for any advice.

best,

Morgan

The issue discusses multi-label item features and the dataset construction API, but names no implementation files or tests. Start by locating the dataset feature-building entry point and reviewing how repeated item-feature rows are encoded. A useful outcome would require a decided supported approach for multi-label features, with corresponding documentation and tests.

Written by the indexing model from the issue text.