Cross Pairwise Ranking for Unbiased Item Recommendation

Abstract

Most recommender systems optimize the model on observed interaction data, which is affected by the previous exposure mechanism and exhibits many biases like popularity bias. The loss functions, such as the mostly used pointwise Binary Cross-Entropy and pairwise Bayesian Personalized Ranking, are not designed to consider the biases in observed data. As a result, the model optimized on the loss would inherit the data biases, or even worse, amplify the biases. For example, a few popular items take up more and more exposure opportunities, severely hurting the recommendation quality on niche items -known as the notorious Mathew effect.In this work, we develop a new learning paradigm named Cross Pairwise Ranking (CPR) that achieves unbiased recommendation without knowing the exposure mechanism. Distinct from inverse propensity scoring (IPS), we change the loss term of a sample -we innovatively sample multiple observed interactions once and form the loss as the combination of their predictions. We prove in theory that this way offsets the influence of user/item propensity on the learning, removing the influence of data biases caused by the exposure mechanism. Advantageous to IPS, our proposed CPR ensures unbiased learning for each training instance without the need of setting the propensity scores. Experimental results demonstrate the superiority of CPR over state-of-the-art debiasing solutions in both model generalization and training efficiency. The codes are available at https://github.com/Qcactus/CPR. CCS CONCEPTS• Information systems → Retrieval models and ranking; Recommender systems.

On the Theories Behind Hard Negative Sampling for Recommendation

“…They assign high sampling probability to top-ranked negative items, accounting for model status. There are also some fine-grained negative sampling methods [23,32,34,42]. Empirical experiments verify the effectiveness and efficiency of HNS.…”

Dually Enhanced Propensity Score Estimation in Sequential Recommendation

“…The following representative sequential recommendation models were chosen as the baselines: STAMP [18] which models the long-and short-term preference of users; GRU4Rec+ [32] is an improved version of GRU4Rec with data augmentation and accounting for the shifts in the inputs; BERT4Rec [31] employs an attention module to model user behaviors and trains with unsupervised style; FPMC [25] captures users' preference by combing matrix factorization with first-order Markov chains; DIN [43] applies an attention module to adaptively learn the user interests from their historical behaviors; BST [7] applies the transformer architecture to adaptively learn user interests from historical behaviors and the side information of users and items; LightSANs [11] is a lowrank decomposed SANs-based recommender model. We also chose the following unbiased recommendation models as the baselines: UIR [29] is an unbiased recommendation model that estimates the propensity score using heuristics; CPR [33] is a pairwise debiasing approach for exposure bias; UBPR [27]is an IPS method for nonnegative pair-wise loss. DICE [41]: A debiasing model focused on the user communities.…”

Mitigating Popularity Bias in Recommendation: Potential and Limits of Calibration Approaches

“…Unbiased, and thus more accurate, recommendations are also the focus in [128]. In this work, Wan et al propose a modified loss function, named "cross pairwise" loss.…”