Sequential User History Features Not Utilized in Model Input Pipeline in train_amazon_inttower.py

[CHETAN1KUKREJA]

Brief

@archersama
In the file IntTower/train_amazon_inttower.py in line 141 the user history is calculated but the feature is never used for making prediction why?

Current Behavior

The implementation processes user historical interaction data through get_user_feature() and get_var_feature(), generating train_user_hist, but this processed sequential data is not included in the final model input. Currently, only sparse and dense features are passed to the model:

train_model_input = {name: train[name] for name in sparse_features + dense_features}

Expected Behavior

Sequential user history should be incorporated into the model input to leverage historical user-item interactions for better recommendations. The train_user_hist generated from get_var_feature() should be included in the model's input features.

Impact

This oversight means the model is currently making predictions based only on:

• Static user features (reviewerID, user_mean_rating)
• Static item features (asin, categories, item_mean_rating, price)

It's missing the valuable sequential patterns in user behavior that have already been processed but aren't being utilized.

bug enhancement feature-engineering

[archersama]

We found that this feature could cause data leakage, so we removed this. As for the results below autoint, can you report both autoint and inttower results?

[CHETAN1KUKREJA]

a data leak? how exactly?
and without that wouldn't the recommendations be based on just the average rating of the user and the categories, item_mean_rating, price of the item?
so that would be not that personalized,right?

[archersama]

"User history is important. However, constructing an appropriate user history is crucial. You might try creating an example of user history yourself."

[CHETAN1KUKREJA]

Ok, I will try. Can you guide me on this topic, It will be much easier for me if you guide me on what to do, as I don't have to go through the entire code then.

Meanwhile, can you please tell me what the data leak was that you were talking about?