HAPENS: Hardness-Personalized Negative Sampling for Implicit Collaborative Filtering Haoxin Liu Yong Shi Qingwei Lin ∗ Pu Zhao ∗ Mirror Xu Dongmei Zhang Si Qin shi.yong@microsoft.com qlin@microsoft.com liuhaoxinthu@gmail.com mirrorxu@microsoft.com dongmeiz@microsoft.com puzhao@microsoft.com Microsoft Bing, Beijing Microsoft Research, Beijing siqin@microsoft.com China China Microsoft Research, Beijing China ABSTRACT For training implicit collaborative fltering (ICF) models, hard neg- ative sampling (HNS) has become a state-of-the-art solution for ob- taining negative signals from massive uninteracted items. However, selecting appropriate hardness levels for personalized recommenda- tions remains a fundamental, yet underexplored, problem. Previous HNS works have primarily adjusted the hardness level by tuning a single hyperparameter. However, applying the same hardness level to each user is unsuitable due to varying user behavioral char- acteristics, the quantity and quality of user records, and diferent consistencies of models’ inductive biases. Moreover, increasing the number of hyperparameters is not practical due to the massive number of users. To address this important and challenging prob- lem, we propose a model-agnostic and practical approach called hardness-personalized negative sampling (HAPENS). HAPENS uses a two-stage approach: in stage one, it trains the ICF model with a customized objective function that optimizes its worst per- formance on each user’s interacted item set. In stage two, it utilizes these worst performances as personalized hardness levels with a well-designed sampling distribution, and trains the fnal model with the same architecture. We evaluated HAPENS on the collected Bing advertising dataset and one public dataset, and the comprehensive experimental results demonstrate its robustness and superiority. Moreover, HAPENS has delivered signifcant benefts to the Bing advertising system. To the best of our knowledge, we are the frst to study this important and challenging problem. CCS CONCEPTS • Information systems → Data mining. KEYWORDS Personalization; Collaborative Filtering; Negative Sampling; ACM Reference Format: Haoxin Liu, Pu Zhao, Si Qin, Yong Shi, Mirror Xu, Qingwei Lin, and Dong- mei Zhang. 2023. HAPENS: Hardness-Personalized Negative Sampling for ∗ Corresponding author Permission to make digital or hard copies of part or all of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for proft or commercial advantage and that copies bear this notice and the full citation on the frst page. Copyrights for third-party components of this work must be honored. For all other uses, contact the owner/author(s). WWW ’23 Companion, April 30–May 04, 2023, Austin, TX, USA © 2023 Copyright held by the owner/author(s). ACM ISBN 978-1-4503-9419-2/23/04. https://doi.org/10.1145/3543873.3584631 Implicit Collaborative Filtering. In Companion Proceedings of the ACM Web Conference 2023 (WWW ’23 Companion), April 30–May 04, 2023, Austin, TX, USA. ACM, New York, NY, USA, 5 pages. https://doi.org/10.1145/3543873. 3584631 1 INTRODUCTION As a critical technology for personalized recommendation, collab- orative fltering (CF)[8, 10, 17] learns user preferences through observed interactions between users and items. Compared to ex- plicit feedback such as ratings, implicit feedback such as clicks and purchases are more accessible. In this paper, we focus on im- plicit CF, which is an important research topic that has received much attention and is widely applied in real-world recommender systems[2, 5, 13, 23]. In addition to the natural scarcity of negative signals, efciency must be ensured due to the massive number of items. Therefore, real-world recommender systems have widely adopted negative sampling, which samples a subset of a user’s uninteracted items based on a certain strategy [6, 21, 23]. Attempts to improve the quality of negative samples can be mainly grouped into two types. Static Negative Sampling (SNS)[1, 12] adopts a static distribu- tion specifed before training. The distribution is based on some heuristic information, such as the degree of nodes[12]. On the other hand, Hard Negative Sampling (HNS) [22ś25] replaces the static distribution with a dynamic distribution that changes during model training. The goal of HNS is to fnd informative samples that are intuitively items that are mapped nearby but should be far apart. As shown in [22ś25], HNS methods signifcantly beneft recom- mendation performance by enhancing click-through rates It should be noted that samples that are excessively difcult may not only produce false negatives, but also impede the training of the model. Conversely, overly simple samples may not be infor- mative. As a result, the difculty level is a critical factor for the efectiveness of HNS but has received insufcient exploration. Pre- vious HNS methods have primarily adjusted the level of difculty by tuning a single hyperparameter. However, we argue that the same difculty level may not be suitable for all users. Factors such as the quantity and quality of the user’s interacted item set, as well as the consistency of the model’s inductive bias, will result in diferent requirements for the level of difculty. The overall optimal level of difculty may be too difcult or too easy for certain users. Therefore, it is imperative to propose a solution to the problem of personalized hard negative sampling (PHNS). However, the problem of PHNS is far from trivial. It is imprac- tical to set independent hyperparameters for each user, and even 376