Differentially Private Link Prediction with Protected Connections

Authors: Abir De, Soumen Chakrabarti63-71

AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental 6 Experiments We report on experiments with five real world datasets to show that DPLP can trade off privacy and the predictive accuracy more effectively than three standard DP protocols (Geng and Viswanath 2015; Machanavajjhala, Korolova, and Sarma 2011; Mc Sherry and Talwar 2007). In the extended draft (De and Chakrabarti 2020), we provide additional details about the experiments and more experimental results.
Researcher Affiliation Academia Abir De, Soumen Chakrabarti Indian Institute of Technology Bombay {abir, soumen}@cse.iitb.ac.in
Pseudocode Yes Algorithm 1 DPLP: Learns A based on a non-private LP algorithm A and then uses it to recommend K nodes to u.
Open Source Code Yes Code: https://www.cse.iitb.ac.in/ abir/codes/dplp.zip
Open Datasets Yes Datasets. We use five networks: Facebook (Leskovec and Mcauley 2012), USAir (Batagelj and Mrvar 2006), Twitter (Leskovec and Mcauley 2012), Yeast (Von Mering et al. 2002), PB (Ackland et al. 2005).
Dataset Splits No Finally, we sample 80% of | nbr(q)| neighbors and 80% of nbr(q) non-neighbors and present the resulting graph Gsampled to A the non-private base LP, and subsequently to A the protected-pair differentially private LP which is built on A. The paper mentions a "20% held-out set" for evaluation, implying a test set, but does not explicitly describe a separate validation split or its purpose.
Hardware Specification No The paper does not provide any specific details about the hardware (e.g., CPU, GPU models, or memory) used for running the experiments.
Software Dependencies No The paper does not specify versions for any software dependencies, such as programming languages, libraries, or frameworks used for implementation.
Experiment Setup Yes We set the fraction of protected edges σprot = 0.3, the privacy leakage ϵp = 0.1 and the number of recommended nodes K = 30. The above loss function encourages that the (noisy) transformed score of a neighbor g exceeds the corresponding score of a non-neighbor b by at least (a tunable) margin ϱ. τ is a temperature hyperparameter.