nails

This repository contains the code for the paper: Normative Alignment of Recommender Systems via Internal Label Shift, accepted as an extended abstract at RecSys '25 Late Breaking Results.

---

Model & Implementation Details

We use the NRMSDocVec model from the ebnerd-benchmark repository.

⚠ Important note: Our setup expects that the model outputs logits or prediction scores that sum to 1. We use softmax in the scripts.

Therefore, in our experiments, we commented out the following line in nrms_docvec.py (line 183):

pred_one = tf.keras.layers.Activation(activation="sigmoid")(pred_one)

---

Download Prediction Files

We share the prediction scores obtained from training, which we use to generate the results: Download here

---

Data Format

The shared prediction file is stored in tabular form with the following structure:

• Shape: (13,536,710, 3)
• Columns:
  • impression_id (u32) — Unique identifier for the impression.
  • article_ids_inview (list[i32]) — List of article IDs shown in the impression.
  • scores (list[f32]) — Corresponding model prediction scores for the articles in article_ids_inview.

Example:

impression_id	article_ids_inview	scores
10017530	[9794425, 9794706, ..., 9794673]	[-0.013472, -0.458544, ..., 0.873...]
28473735	[9794845, 9794924, ..., 9794932]	[0.886794, -0.158117, ..., 0.5009...]
32426821	[9797023, 9798775, ..., 9798644]	[0.802967, 0.521927, ..., 0.35544...]
28680972	[9791182, 9789674, ..., 9756075]	[-0.181093, 0.288003, ..., -2.459...]
12308406	[9797733, 9797537, ..., 9798323]	[0.044418, -0.334013, ..., 0.4230...]
...	...	...

---

Setup

conda create -n nails python=3.11
conda activate nails
pip install -r requirements.txt

---

Run Experiments

Editorial distribution:

python exp_nails.py --distribution_type Editorial
python exp_steck.py --distribution_type Editorial
python exp_nails_steck_combine.py --distribution_type Editorial

Uniform distribution:

python exp_nails.py --distribution_type Uniform
python exp_steck.py --distribution_type Uniform
python exp_nails_steck_combine.py --distribution_type Uniform

---

Quick Dummy Run

python exp_nails.py --n_samples 150 --n_samples_test 151
python exp_steck.py --n_samples 150 --n_samples_test 151
python exp_nails_steck_combine.py --n_samples 150