Optimization and Analysis of the pAp@k Metric for Recommender Systems

Authors: Gaurush Hiranandani, Warut Vijitbenjaronk, Sanmi Koyejo, Prateek Jain

ICML 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our analysis and experimental evaluation suggest that p Ap@k indeed exhibits a certain dual behavior with respect to partial AUC and precision@k. Moreover, the proposed methods outperform all the baselines in various applications. Taken together, our results motivate the use of p Ap@k for large-scale recommender systems with heterogeneous user-engagement. In this section, we present evaluation of our methods on sim ulated and real data.
Researcher Affiliation Collaboration 1University of Illinois at Urbana-Champaign, Illinois, USA 2Microsoft Research, Bengaluru, Karnataka, India.
Pseudocode Yes Algorithm 1 GD-p Ap@k-surr and Algorithm 2 Subgradient calculation for p Ap@k surrogates
Open Source Code Yes Source code: https://github.com/gaurush-hiranandani/pap-k
Open Datasets Yes Movie Recommendation (70K instances, 15.5K positives, 638 users, d = 90, k = 8 24): We use the Movielens 100K dataset (Harper & Konstan, 2015)... Citation Recommendation (142K instances, 21K positives, 2477 users, d = 157, k = 6 18): The task in the citation dataset (Budhiraja et al., 2020)... Image Recommendation (670K instances, 111K positives, 2498 users, d = 150, k = 5 25): We take the Behance dataset (He et al., 2016)...
Dataset Splits No For all the methods, including baselines, the learning rate and regularization parameters are cross validated on the set {10 4 , 2 10 4 , 5 10 4 , 10 3 , . . . , 0.5} and 10{ 3,...,1}, respectively. While cross-validation implies a splitting strategy, the paper does not provide specific details on the splits (e.g., number of folds, percentages for train/validation sets).
Hardware Specification No The paper does not provide any specific details about the hardware used for running the experiments, such as CPU or GPU models, or cloud computing instance types.
Software Dependencies No The paper does not specify any software dependencies with version numbers (e.g., Python 3.x, PyTorch 1.x, specific library versions) that would be needed for replication.
Experiment Setup Yes For all the methods, including baselines, the learning rate and regularization parameters are cross validated on the set {10 4 , 2 10 4 , 5 10 4 , 10 3 , . . . , 0.5} and 10{ 3,...,1}, respectively. We fx ηt = η/ t + 1 in our methods and use a regularized version of the surrogates by adding λkwk2 .