Learning from Streaming Data when Users Choose

Authors: Jinyan Su, Sarah Dean

ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We also experimentally demonstrate the utility of our algorithm with real world data.
Researcher Affiliation Academia 1Department of Computer Science, Cornell University. Correspondence to: Jinyan Su <js3673@cornell.edu>, Sarah Dean <sdean@cornell.edu>.
Pseudocode Yes Algorithm 1 Multi-learner Streaming Gradient Descent (MSGD)
Open Source Code Yes Code for reproducing these results can be found at https://github.com/ sdean-group/MSGD.
Open Datasets Yes Our first experimental setting is based on a widely used movie recommendation dataset Movielens10M (Harper & Konstan, 2015)... Our second setting is based on census data made available by folktables (Ding et al., 2021).
Dataset Splits No The paper mentions splitting data for testing ('0.2 of them for testing') but does not explicitly state a validation set or describe a train/validation/test split.
Hardware Specification No The paper does not specify the hardware used for running the experiments (e.g., specific GPU/CPU models, memory, or cloud instance types).
Software Dependencies No The paper mentions 'Python toolkit Surprise' and 'folktables' but does not provide specific version numbers for these software components or for Python itself.
Experiment Setup Yes Then the selected service updates their parameter with the gradient of the loss on the user s data with step size ηt = 1/t. At each time step, we sample a user x at random from the data described above. We assign this user to one of k services according to bounded rational with parameters ζ. For MSGD, we illustrate results of different total number of services k = 2, 4, 6. for each k, we use compute the average accuracy after T = 2000 k total timesteps and plot the average over 3 trials.