基于神经网络的新闻推荐算法能够从纷繁复杂的新闻中筛选出符合用户偏好的新闻,对于提升用户获取信息的效率以及阅读新闻的体验具有重要的意义.现有的新闻推荐方法不仅在用户偏好的准确建模方面取得了显著的效果,同时通过识别新闻数据中的虚假关联(例如用户性别与特定新闻类别之间的联系)开展了无偏新闻推荐的初步尝试.然而用户的新闻点击行为是一系列复杂认知行为相互作用之后的决策结果,仅对有偏信息进行直接建模,简化了有偏信息和用户行为之间的复杂交互关系,忽略了用户行为背后的复杂认知因素的影响,导致无偏推荐效果难以满足实际需求.为了解决该问题,本文提出一种全新的从众性感知的因果去偏新闻推荐方法CADNN(Conformity-A ware Debiased Neural News Recommendation).具体而言,本文提出用户的点击行为包括以下三方面:用户偏好与新闻的匹配度,新闻的流行度,以及用户的从众性特质共同决定的观点.用户的从众性越低,新闻的流行度越小,相对应的点击行为越能反映出用户的真实偏好.基于该观点,本文首先对用户的历史点击数据进行了详细的统计分析,从中提取出反映用户行为的关键特征,根据统计分析结果构建能够描述用户偏好、新闻流行度和从众性特质之间关系的因果图模型,通过因果图模型更准确地理解和描述用户点击行为背后的驱动因素.在此基础上,本文提出一种全新的从众性建模方法,同时从用户的从众性特质和新闻的流行度两个角度进行更符合实际情况的从众性建模;与此同时本文利用现有的新闻推荐方法计算用户偏好和新闻内容的匹配度.最终在对三方面因素充分建模的情况下,实现更高质量的无偏新闻推荐.最后,本文在公开的数据集MIND上进行了充分的实验验证,并构建了一个用于无偏测试的挑战集,旨在评估推荐系统在无偏环境中的鲁棒性,与已有的最先进的基准方法相比,本文提出的CADNN方法在AUC、MRR、NDCG@5和NDCG@10四个指标上分别平均提高了 1.76％、2.37％、3.06％和2.16％,充分证明了 CADNN方法的有效性.本文同时提供了 CADNN的实现代码及相关数据,用于支持推动相关领域的研究.
Conformity-Aware Debiased Neural News Recommendation with Causal Reasoning
Neural news recommendation algorithms can filter out the news that meets the users'preference from the complicated news,which is important for improving the efficiency of the users'access to information and the news reading experience.It is of great significance to improve the efficiency of users'information access and news reading experience.Existing news recommendation methods have not only achieved remarkable results in accurately modeling users'preferences,but also carried out preliminary attempts of unbiased news recommendation by identifying spurious correlations in news data(e.g.the association between user gender and specific news categories).However,user clicking behavior is the result of a series of complex cognitive factors.Directly modeling biased information oversimplifies the complex interaction between biased information and user behavior,disregarding the impact of complex cognitive factors behind the user behavior,which indicates the challenging achievement of unbiased recommendation effectiveness and difficulty makes it hard to meet actual requirements.To address the issue,we propose a novel method named CADNN(Conformity-Aware Debiased Neural News recommendation)to realize higher quality unbiased news recommendation.Specifically,we argue that user clicking behavior is determined by the following three factors:the match score between user preference and news,news popularity,and user conformity level.The lower the user conformity level and news popularity are,the more the corresponding clicking behavior reflects the user real preference.Based on this viewpoint,this paper first conducts a detailed statistical analysis of users'historical clicking data,extracting key features that reflect user behaviors.These features are used to construct a causal diagram model,which describes the intricate relationships between user preferences,news popularity,and user conformity level.The causal diagram provides a comprehensive framework for understanding how these factors interact to influence user clicks.Building upon this model,we propose a novel crowd modeling method that realistically incorporates both users'conformity level and the popularity of news.This method aims to differentiate between clicks driven by genuine user interest and those influenced by the popularity of the news,thereby addressing the bias in traditional recommendation systems.We can better capture users'true interests by calculating the match between user preferences and news content using existing recommendation techniques.To achieve unbiased news recom-mendations,we ensure thorough modeling of the three factors-user preferences,news popularity,and user conformity level.This multifaceted approach leverages causal intervention to provide higher-quality recommendations that reflect users'genuine interests more accurately.Our method is rigorously validated through extensive experiments on the publicly available MIND dataset.Additionally,we constructed a challenge dataset specifically designed for unbiased testing to evaluate the robustness of our recommender system in various scenarios.The results show that our CADNN method significantly outperforms existing state-of-the-art benchmark methods.Specifically,CADNN improves the average of four key metrics:AUC,MRR,NDCG@5,and NDCG@10 by 1.76％,2.37％,3.06％,and 2.16％,respectively.These improvements under-score the effectiveness of CADNN and its potential for practical applications.Furthermore,this paper provides the implementation code and related data for CADNN,making it accessible for further research and application in related fields.
NETL
NSTL
万方数据
