基于不确定性估计的推荐系统多任务学习去偏方法

De⁃biasing method for multi⁃task learning in recommender systems based on uncertainty estimation

推荐系统在互联网应用中扮演重要的角色,它的核心任务是将最相关的物品推荐给用户,然而,由于推荐系统通常在大规模、稀疏和高维的数据集上运行,因此存在严重的偏差问题.近年来,多任务学习成为解决推荐系统偏差的有效方法,它可以同时学习多个相关任务,充分利用数据集的内在结构和相关性,研究人员最近还提出全空间反事实的转化率预测,利用逆倾向得分和双重鲁棒方法对推荐算法的效果进行估计.然而,通过理论分析发现,倾向性分数估计不准确和插值误差往往会导致预估偏差,这在实践中经常发生,影响了推荐的准确性和可靠性.由此,引入不确定性估计,结合多任务学习,通过计算每个用户反馈数据的概率分布来衡量数据的可靠程度,减轻模型在数据稀疏或数据噪声较大时的过拟合问题,有效地提高系统的泛化能力并减少偏差.实验结果表明,加入不确定性估计的多任务学习可以更好地适应不确定性的环境,在推荐系统中有广阔的应用前景.

Recommender systems play an important role in internet applications,with the core task of recommending the most relevant items to users.However,the large⁃scale,sparse,or high⁃dimensional datasets often lead to serious bias problems.In recent years,multi⁃task learning has become an effective method for addressing bias in recommender systems,allowing multiple related tasks to be learned simultaneously,thus fully utilizing the intrinsic structure and correlation of the datasets.Researchers recently proposed entire space counterfactual conversion rate prediction,which uses inverse propensity score and doubly robust methods to estimate the performance of recommendation algorithms.However,theoretical analysis has revealed that inaccurate propensity score estimation and interpolation errors often lead to estimation bias,which frequently occurs in practice,thereby affecting the accuracy and reliability of recommendations.We therefore introduce uncertainty estimation,combining multi⁃task learning to measure the reliability of feedback data by computing the probability distribution for each user,to mitigate model overfitting in sparse or noisy data and effectively improve system generalization to reduce bias.Experimental results show that multi⁃task learning with uncertainty estimation can better adapt to uncertain environments and has broad prospects in recommender systems.