Algorithmically Mediated User Relations: Exploring Data's Relationality in Recommender Systems

Repository for the paper "Algorithmically Mediated User Relations: Exploring Data’s Relationality in Recommender Systems" presented in the Regulatable ML Workshop @ NeurIPS 2023.

About the Project

Personalization services, such as recommender systems, operate on vast amounts of user-item interactions to provide personalized content. To do so, they identify patterns in the available interactions and group users based on pre-existing offline or online social relations, or algorithmically determined similarities and differences. We refer to the relations created between users based on algorithmically determined constructs as algorithmically mediated user relations. Taking as a case study collaborative filtering recommendation algorithms where users are interrelated by design, we empirically examine whether algorithmically mediated user relations should be taken into account in practice when quantifying the influence of users’ data on the recommendations of others.

Installation

To install the needed dependencies run:

• conda env create -f requirements.yml
• conda activate user_intercon_env

To use matrix factorization algorithms:

• conda install -c anaconda cython
• Compile Cython: cd RecSys2019_DeepLearning_Evaluation, /opt/miniconda3/envs/user_intercon_env/bin/python run_compile_all_cython.py

Execution

To execute use the command /opt/miniconda3/envs/user_intercon_env/bin/python main.py --flagfile=flagfile.cfg.

Specify all relevant parameters on flagfile.cfg. The main flags to be specified are:

Flag Name	Description	Possible Values
dataset	the dataset to be used	'sample-test', 'MovieLens_100k'
algorithm	the algorithm to be used for tuning, training, etc.	['UserKNNCFRecommender', 'MatrixFactorization_FunkSVD_Cython']
operation	the operation to execute	['get_dataset_statistics','generate_trainset_testset', 'algorithm_hyperparameter_tuning', 'compute_individual_influences', 'compute_group_influences', 'check_independence_assumption']

Data Preprocessing

All relevant code is in data_preprocessing.py.

To get common statistics of a given dataset, set the flag operation to get_dataset_statistics on flagfile.cfg.

To generate the training, validation, and test sets of a given dataset and compute statistics of the resulting sets, set the flag operation to generate_trainset_testset on flagfile.cfg and specify the flags:

Flag Name	Description	Possible Values
ratio_split_train	the train-test splitting ratio	float in [0,1]
ratio_split_validation	the train-validation splitting ratio	float in [0,1]

The splitted train, test, and validation sets for the specified random seed are in ./Data/.

Algorithms & Hyperparameter Tuning

The implemented algorithms and tuning source code of Dacrema et. al are used.

For hyperparameter tuning, set the flag operation to algorithm_hyperparameter_tuning on flagfile.cfg.

Computation of Influence

Individual Influence

Set the flag operation to compute_individual_influences on flagfile.cfg.

Relevant code:

• remove_n_train.py: remove each user's data from the training set and re-train the model with the reduced training set
• compute_ratings.py: compute the ratings of all users for each re-trained model
• compute_individual_influence.py: compute the influence of each user

Group Influence

As discussed in the paper, we define 2 types of influnce of a group of users. To compute:

• $I_{independence}$ set the flag operation to compute_group_influences_independence on flagfile.cfg. Individual influences need to be computed first. Relevant code:
  • find_influential_users.py:
• $I_{relations}$ set the flag operation to compute_group_influences_relations on flagfile.cfg. Individual influences need to be compute first. Relevant code:
  • find_influential_users.py:
  • remove_n_train_group.py

Ploting

Set the flag operation to check_independence_assumption on flagfile.cfg.