Presentation is loading. Please wait.

Presentation is loading. Please wait.

Model-Agnostic Counterfactual Reasoning for Eliminating Popularity Bias in Recommender System 汇报人:张海超.

Similar presentations


Presentation on theme: "Model-Agnostic Counterfactual Reasoning for Eliminating Popularity Bias in Recommender System 汇报人:张海超."— Presentation transcript:

1 E-mail: haichao.zhang22@student.xjtlu.edu.cn Model-Agnostic Counterfactual Reasoning for Eliminating Popularity Bias in Recommender System 汇报人:张海超

2 INFORMATION https://dl.acm.org/doi/10.1145/ 3447548.3467289 KDD 为数据挖掘领域的顶级会议, 中国 计算机学会 (CCF) 评级为 A 类会议 1.Abstract 2.Problem definition 3.Methodology 4.Experiment 5.Related work 6.Conclusion and future work 7.Others ( Implementation Details )

3 ABSTRACT 1.item’s intrinsic property is the cause of mistakenly assigning it a higher ranking score 2.To eliminate popularity bias, it is essential to answer the counterfactual question that what the ranking score would be if the model only uses item property. To this end, we formulate a causal graph to describe the important cause-effect relations in the recommendation process 推荐系统中存在的问题:推荐系统中存在 Popularity Bias 问题 The general aim of the recommender system is to provide personalized suggestions to users, which is opposed to suggesting popular items 常规训练模型会导致马太效应( making popular items be more frequently recommended and become even more popular. )。 本文做的事儿:因果关系的角度来探讨推荐系统中的流行度偏差问题 (如何消除 bias )

4 问题引入 the frequency distribution of items is never even in the interaction data, exposure mechanism, word-of-mouth effect, sales campaign, item quality Majority of interactions are occupied by a small number of popular items. This makes the classical training paradigm biased towards recommending popular items, falling short to reveal the true preference of users 长尾交互数据使得最终的推荐 结果不能揭露用户真实偏好

5 研究现状 these methods lack fine-grained consideration of how item popularity affects each specific interaction, and a systematic view of the mechanism of popularity bias. 流行项上的交互总是向下加权,不管流行项更符合用户的偏好,盲目地将推荐者 推向长尾 现有方法的局限性: 现有去除偏置算法: The idea is to downweigh the influence from popular items on recommender training : 1. re-weighting their interactions in the training loss 2. Add unbiased data balanced training data 3. disentangling user and item embeddings

6 提出观点 文章观点: the key of eliminating popularity bias is to understand how item popularity affects interaction. 为了进行受欢迎程度的去偏置,我们只需从总体排名得分中扣除反 事实世界中( user-item matching is discarded , and an interaction is caused by item popularity and user conformity. )的 排名得分

7 定义和符号表示 users U = {1, 2, · · ·, } item set with items V = {1, 2, · · ·, }. user-item interaction matrix:

8 PRELIMINARIES can be calculated from the values of its ancestor nodes, which is formulated as: the total effect (TE) of = on is defined as: 总因果效应可以被分解为自然直接效应 (NDE) 和总间接效应 (TIE) ,分别表示通 过直接路径 (I-Y) 和间接路径 (I-K-Y) 产生的因果效应 TE can be decomposed into natural direct effect (NDE) and total indirect effect (TIE)

9 PRELIMINARIES 1. 当前大部分推荐系统,只关注用户 - 物品匹配因素如图 (a) 所示,而忽略了物品流行度如何影响交互 概率, 2. 如果两个物品对于一个用户的匹配度相同,流行程度越大的物品越有可能被该用户知道并消费, 因 此文章在因果图中加入 I -> Y 的因果关系,如图 2(b) 所示; 3. 另一方面,流行物品对不同用户的影响可能有所不同,例如,有些用户更喜欢使用受欢迎的物品, 而有些用户则不喜欢,因此文章在因果图中加入 U -> Y 的因果关系,如图 2(c) 所示。 综上可知,我们认为消除流行度偏差的关键是消除从 I 和 U 到 Y 的直接影响,使推荐系统能关注于 学习用户 - 物品之间的匹配程度特征 ( K ) 来进行推荐。 ——>

10 MACR 模型 - 框架 This framework can be implemented over any existing recommender models that follow the structure of & → → by simply adding a user module () and an item module () 这个框架遵循上页中的因果图 代表着传统推荐系统,用物品和用户表示作为输 入,并反映了项目在多大程度上匹配用户的偏好。 建模物品受欢迎程度的影响,通常越受欢迎的物 品对推荐的影响越大。 显示了在不管用户物品是否匹配的情况下,用 户 将会在多大程度上与物品交互,兴趣广泛容 易交互的用户的 Y 更高 三个分支融合得到最终的预测评分

11 MACR 模型 – 损失函数 applies additional supervision over ˆ and ˆ. Lo a recommendation loss such as the BCE loss Similar as, and are also recommendation losses

12 MACR 模型 – 去偏置推理 我们希望只关注 ,这需要用到我们前面介绍的 TIE 的方法,消除掉 U, I 对 Y 的直接影响。 1. NDE : ranking score without consideration of user-item matching 2. TE: 是 U, I , K 都是事实取值的情况下,相对于这些值都是基准值(对照组,这里采 用均值)时,对 Y 的整体影响。 3. TIE : eliminating popularity bias can be realized by reducing from E �

13 MACR 模型 – 去偏置推理 ( =, =, =, ) = ˆ ∗ ( ˆ ) ∗ ( ˆ ) ( =, =, = ∗, ∗ ) = ∗ ( ˆ ) ∗ ( ˆ ) 在 MACR 模型框架中得出: 其中 c 表示当 K= 时候 的值(也就是用户、物品 embedding 都 是均值时, 的值 ) The key difference of the proposed counterfactual inference and normal inference is using TIE to rank items rather than TE TIE = TE – NDE =

14 EXPERIMENTS 研究问题 数据集 Baseline

15 实验结果 - DOES MACR OUTPERFORM EXISTING DEBIASING METHODS? 1.In all cases, the average improvement of MACR_MF over MF on the five datasets is 153.13% in terms of HR@20 , the improvement of MACR_LightGCN over LightGCN is 241.98% 2. In most cases, LightGCN performs worse than MF 图 1 中也可以看出, the average recommendation frequency of popular items on LightGCN is visibly larger than MF. 这主要是由于 LightGCN 中嵌入的传播操作,受欢迎的商品的影响在用户 - 商品交互图上被传播,进一步放大了受欢迎 的偏差。然而,在我们的 MACR 框架中, MACR_LightGCN 的性能优于 MACR_MF 。这表明我们的框架可以大大减轻 流行偏见。 3. In terms of datasets, the improvements over the Globo dataset are extremely large Globo is a large-scale news dataset, and the item popularity distribution is particularly skewed. 流行度偏差较大,所以提升明显, 也进一步说明了模型的有效性

16 实验结果 - DOES MACR OUTPERFORM EXISTING DEBIASING METHODS? 4. Reg (正则化方式降低短尾权重,提高长尾推荐)表现不佳 : 简单地降低流行 项目的权重 5. CausE (使用去偏置的训练集)表现不佳: 依赖于去偏训练集, 这个训练 集一般很小, 效果不明显 6. BS (训练中学习有偏差的分数) and IPW (反向倾斜权重,重新加权样本) methods can alleviate the bias issue to a certain degree 7. DICE (兴趣和物品流行度分解为两组嵌入) achieved the best results among the baselines. 也从侧面表明物品受欢迎度是影响交互的重要原因

17 实验结果 - HOW DO DIFFERENT HYPER-PARAMETER AFFECT THE RECOMMENDATION PERFORMANCE? The hyper-parameter as formulated in Eq. (9) controls the degree to which the intermediate matching preference is blocked in prediction. · 适当增加 c 有利于提高推荐性能, · lightGCN 的变化比 MF 更敏感,也说明 lightGCN 更容易收到流行度偏差的影响

18 实验结果 -HOW DO DIFFERENT COMPONENTS IN OUR FRAMEWORK CONTRIBUTE TO THE PERFORMANCE? four special cases : 1&2 : MACR_MF w/o user (item) branch, where user (or item) branch has been removed 3&4 : MACR_MF w/o (), where we just simply remove () to block the effect of user (oritem) branch on training 结论 1.item popularity bias has more influence than user conformity on the recommendation. 2.the significance of further fusing the item and user influence in the prediction.

19 实验结果 -HOW DOES MACR ELIMINATE THE POPULARITY BIAS? Fig8. our methods indeed reduce the recommendations frequency of popular items and recommend more items that are less popular What is recall rate ? If an item appears times in the test data, its item recall is the proportion of it being accurately recommended to test users Compare to Fig8. and Fig9 : The most popular item group has the greatest recall increase, but our methods in Figure 8 show the recommendations frequency of popular items is reduced. It means that traditional recommender systems (MF, LightGCN) are prone to recommend more popular items to unrelated users due to popularity bias.

20 实验结果 -HOW DOES MACR ELIMINATE THE POPULARITY BIAS? The horizontal axis means the user groups with a certain number of interactions. The right vertical axis is the value of the polyline, which corresponds to ( ˆ ). with the increase of the occurrence frequency of users in the dataset, the sigmoid scores of them also increase. This indicates that the user’s activity is consistent with his/her conformity level.

21 收获 1. 推荐系统常用的评价指标: HR 、 NDCG 、 MRR https://blog.csdn.net/shiaiao/article/details/109004341 2. 真实世界中无法解决的问题,或者 无法提取的参数, 可以考虑从反事实 世界中寻求另外一种解决思路

22


Download ppt "Model-Agnostic Counterfactual Reasoning for Eliminating Popularity Bias in Recommender System 汇报人:张海超."

Similar presentations


Ads by Google