Temporally and Distributionally Robust Optimization for Cold-Start Recommendation

Authors: Xinyu Lin, Wenjie Wang, Jujia Zhao, Yongqi Li, Fuli Feng, Tat-Seng Chua

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

Reproducibility Variable Result LLM Response
Research Type Experimental Substantial experiments on three real-world datasets validate the superiority of our temporal DRO in enhancing the generalization ability of coldstart recommender models. We conduct extensive experiments on three real-world datasets to answer the following research questions:
Researcher Affiliation Academia Xinyu Lin1, Wenjie Wang1*, Jujia Zhao1, Yongqi Li2, Fuli Feng3, Tat-Seng Chua1 1National University of Singapore 2The Hong Kong Polytechnic University 3Mo E Key Laboratory of Brain-inspired Intelligent Perception and Cognition, University of Science and Technology of China
Pseudocode Yes Algorithm 1: Training Procedure of TDRO
Open Source Code Yes We release our codes and datasets at https://github.com/Linxyhaha/TDRO/.
Open Datasets Yes We conducted experiments on three real-world datasets across different domains: 1) Amazon (He and Mc Auley 2016)... 2) Micro-video is a real-world industry dataset... 3) Kwai6 is a benchmark recommendation dataset provided with rich visual features. We release our codes and datasets at https://github.com/Linxyhaha/TDRO/.
Dataset Splits Yes For Amazon and Micro-video datasets, we split the interactions into training, validation, and testing sets chronologically at the ratio of 8:1:1 according to the timestamps.
Hardware Specification No The paper does not provide specific hardware details (e.g., GPU/CPU models, memory, or cloud instances) used for running the experiments.
Software Dependencies No The paper mentions software components and frameworks implicitly through citations (e.g., PyTorch can be inferred from common practices in ML), but does not provide specific version numbers for any software dependencies.
Experiment Setup Yes We consider two factors for the training of TDRO: 1) a worst-case factor to guarantee worst-case performance... and 2) a shifting factor to capture the shifting trend of item features, which utilizes a gradientbased strategy to emphasize the optimization towards the gradually popular item groups across time periods. where λ is the hyper-parameter to balance the strength between two factors. We define βe = exp(p e), where a later period e will have a higher weight and p > 0 is the hyperparameter to control the steepness. model parameters θ are updated based on the selected group, i.e., θt+1 = θt η θ Lj (θt), where η is the learning rate. where µ is the hyper-parameter to control the streaming step size.