Recommender　systems　frequently　encounter　the　challenge　of　popularity　bias:　the　interaction　data　exhibit　a　pronounced　skew　or　a　long-tailed　distribution　across　items.　This　imbalance　can　cause　models　to　disproportionately　recommend　popular　items,　deviating　from　accurately　reflecting　users’　true　preferences.　While　existing　methods　often　employ　propensity-based　unbiased　learning　or　causal　inference　to　tackle　data　imbalance.　However,　they　overlook　the　imbalance　issue　among　user　groups,　potentially　neglecting　the　influence　of　popularity　on　preference　representation　for　different　user　groups.　In　addressing　this　issue,　our　work　categorizes　users　into　two　groups:　those　who　are　intended　for　popularity　(Popularity-intended)　and　those　who　are　not　(Popularity-unintended).　We　propose　a　novel　framework　called　RPB4Rec　(Relieving　Popularity　Bias　for　Recommendation)　designed　to　perform　data　augmentation　at　the　user　group　level,　relieving　popularity　bias　stemming　from　imbalanced　interactions　among　user　groups.　For　each　actual　user–item　interaction,　RPB4Rec　generates　augmented　interactions,　which　are　then　utilized　in　a　contrastive　learning　paradigm.　To　precisely　modulate　the　influence　of　augmented　data　on　model　training,　we　incorporate　an　adaptive　temperature　coefficient　τ　for　contrastive　learning,　which　is　dynamically　adjusted　based　on　the　training　loss.　To　validate　the　effectiveness　of　our　proposed　method,　we　conducted　extensive　evaluations　using　three　diverse　datasets:　the　Yahoo!　R3　dataset,　the　Coat　Shopping　dataset,　and　the　Kuairand　video　dataset.　Our　comprehensive　experiments　across　these　real-world　datasets　validate　the　superiority　of　our　approach.　Specifically,　the　experimental　results　show　that　RPB4Rec　can　outperform　the　state-of-the-art　method　by　average　5%　in　Recall@20　and　NDCG@20,　indicating　that　RPB4Rec　is　highly　effective　in　relieving　popularity　bias　in　recommender　systems.　©　2024　Elsevier　B.V.