基于多任务学习的位置倾向性得分预测算法

曹泽麟; 徐君; 董振华; 文继荣

doi:10.7544/issn1000-1239.202110853

基于多任务学习的位置倾向性得分预测算法

曹泽麟^1,,
徐君^{1, 2, ,},
董振华³,
文继荣^{1, 2}

1.
中国人民大学高瓴人工智能学院　北京　100872
2.
大数据管理与分析方法研究北京市重点实验室（中国人民大学）　北京　100071
3.
华为技术有限公司诺亚方舟实验室　广东深圳　518129

基金项目: 国家重点研发计划项目（2019YFE0198200）；国家自然科学基金项目（61872338,61832017）；北京高校卓越青年科学家计划项目（BJJWZYJH012019100020098）

详细信息

作者简介:
曹泽麟: 1998年生.博士研究生.主要研究方向为数据挖掘和推荐系统

徐君: 1979年生.博士，教授，博士生导师.主要研究方向为信息检索、互联网搜索、机器学习和大数据分析

董振华: 1984年生.博士，研究员.主要研究方向为推荐系统、反事实学习和移动计算

文继荣: 1972年生.博士，教授，博士生导师.主要研究方向为互联网大数据管理、信息检索、数据挖掘和机器学习

通讯作者:
徐君（junxu@ruc.edu.cn）

中图分类号: TP311
计量
- 文章访问数: 252
- HTML全文浏览量: 27
- PDF下载量: 186
出版历程
- 收稿日期: 2021-08-19
- 修回日期: 2022-02-10
- 网络出版日期: 2023-02-10
- 刊出日期: 2022-12-31

Prediction of the Positional Propensity Scores Based on Multi Task Learning

Cao Zelin^1,,
Xu Jun^{1, 2, ,},
Dong Zhenhua³,
Wen Jirong^{1, 2}

1.
Gaoling School of Artificial Intelligence, Renmin University of China, Beijing 100872
2.
Beijing Key Laboratory of Big Data Management and Analysis Methods (Renmin University of China), Beijing 100071
3.
Huawei Noah’s Ark Lab, Shenzhen, Guangdong 518129

Funds: This work was supported by the National Key Research and Development Program of China (2019YFE0198200), the National Natural Science Foundation of China (61872338,61832017), and the Beijing Outstanding Young Scientists Program (BJJWZYJH012019100020098).

摘要

摘要:
用户搜索时产生的点击数据分布，在不同的搜索场景下存在较大差异. 现有算法如融合上下文的位置模型（contextual position based model, CPBM）往往只通过单个模型预测多种场景下的位置倾向性得分，不可避免地降低了模型在不同场景下的预测准确性，影响去除位置偏置的效果.基于上述问题提出一种基于多任务学习的多门专家混合位置倾向性得分预测模型（multi-gate contextual position based model, MCPBM），在CPBM模型的基础上加入信息筛选结构，解决了多场景数据联合训练时预测准确性不佳的问题. 同时，为了缓解不同任务收敛速度不一致的问题，提出了指数加权平均权重动态调整算法，在加速模型训练的同时提升了模型整体预测性能. 实验结果表明提出的MCPBM模型在多场景数据联合训练时，预测准确性优于传统的CPBM；在使用MCPBM模型去除位置偏置后，基于生成的无偏数据训练得到的排序模型，在AvgRank排序指标上有1%~5%的提升.
- 位置偏置 /
- 无偏排序 /
- 逆概率加权 /
- 多任务学习 /
- 倾向性得分预测
Abstract:
Users’ click data distribution during search is quite different in different search scenarios.The existing methods such as CPBM (contextual position based model) only predict the positional propensity score in multiple scenarios through single model, which inevitably reduces the prediction accuracy in different scenarios and affects the effect of removing position bias. In this work, A MCPBM (multi-gate contextual position based model) based on multi-task learning is proposed. In this model, the information filtering structure is added to CPBM model to solve the problem of poor prediction accuracy during joint training on multi-scene data. At the same time, in order to alleviate the problem that the convergence speed of different tasks is inconsistent. We propose an exponentially weighted average dynamic adjustment algorithm, which speeds up MCPBM training and improves the overall prediction performance of MCPBM. The experimental results show that MCPBM model proposed in this paper is better than traditional CPBM model in prediction accuracy when multi-scene data is jointly trained. After using MCPBM model to remove the position bias in the training data , the ranking model obtained by training on the generated unbiased data promotes the AvgRank ranking metric of test data by 1%–5%.
- position bias /
- unbiased learning-to-rank /
- inverse probability weighting /
- multi-task learning /
- propensity score prediction

HTML全文

图 1 随机投放下的位置点击率

Figure 1. Position click rate in random traffic

下载: 全尺寸图片幻灯片

图 2 多门专家混合位置倾向性得分预测模型及其Block模块结构

Figure 2. Multi-gate contextual position based model and its stucture of Block module

下载: 全尺寸图片幻灯片

图 3 指数加权平均在4场景数据训练时错误变化曲线

Figure 3. Error rate curve of exponential weighted average in 4 scene data training

下载: 全尺寸图片幻灯片

图 4 指数加权平均在3场景数据训练时错误变化曲线

Figure 4. Error rate curve of exponential weighted average in 3 scene data training

下载: 全尺寸图片幻灯片

图 5 权重相等在4场景数据训练时错误变化曲线

Figure 5. Error rate curve of equal weight in 4 scene data training

下载: 全尺寸图片幻灯片

图 6 Uncert在3场景数据训练时错误变化曲线

Figure 6. Error rate curve of Uncert in 3 scene data training

下载: 全尺寸图片幻灯片

图 7 Uncert在4场景数据训练时错误变化曲线

Figure 7. Error rate curve of Uncert in 4 scene data training

下载: 全尺寸图片幻灯片

表 1 双场景数据联合训练和独立训练时测试集错误率

Table 1 Error Rate on the Test Set During Joint Training and Independent Training Under Dual Scene Data

数据集	训练方式	双场景1		双场景2		双场景3
数据集	训练方式	θ = 0.1	θ = 0.3	θ = 0.1	θ = 0.6	θ = 0.1	θ = 10
Yahoo	联合	2.448	2.341	2.183	2.574	4.043	12.429
Yahoo	独立	3.969	2.761	3.969	2.143	3.969	5.459
MQ2007	联合	1.824	2.660	1.892	4.434	1.964	34.714
MQ2007	独立	2.226	3.223	2.226	4.431	2.226	32.302

下载: 导出CSV

表 2 单任务学习模型的超参数设置

Table 2 Hyperparameter Settings of Single-Task Learning Model

超参数	取值
批次大小	{16,32,64}
学习率	[1E–4, 2E–4]
优化器	Adam
学习率递减	[50,100]

下载: 导出CSV

表 3 多任务学习模型的超参数设置

Table 3 Hyperparameter Settings of Multi-Task Learning Model

超参数	取值
批次大小	{16,32,64}
学习率	[1E–4, 3E–4]
优化器	Adam
任务权重比例 $\gamma$	[0.5, 0.7]
任务权重平滑度 S	[1, 3]
学习率递减	[30,50]

下载: 导出CSV

表 4 3场景数据下模型测试集错误率

Table 4 Error Rate on the Test Set Under Three Scene Data

数据集	预测模型	3场景1			3场景2
数据集	预测模型	θ = 0.1	θ = 0.3	θ = 0.6	θ = 0.1	θ = 0.6	θ = 10
Yahoo	LE	2.052	2.120	3.418	2.052	3.418	19.547
	PBM	0.458	1.419	2.909	0.458	2.909	20.220
	CPBM	3.969	2.761	2.143	3.969	2.143	5.459
	MCPBM	0.449	0.717	0.974	0.798	1.092	5.000
MQ2007	LE	1.503	1.574	2.955	1.503	2.955	34.543
	PBM	1.167	1.729	3.568	1.167	3.568	35.718
	CPBM	2.226	3.223	4.431	2.226	4.431	32.302
	MCPBM	1.125	1.485	2.872	1.991	2.431	29.767

下载: 导出CSV

表 5 3场景数据下测试集的AvgRank情况

Table 5 AvgRank on the Test Set Under Three Scene Data

数据集	预测模型	3场景
数据集	预测模型	θ = 0.1	θ = 0.3	θ = 0.6
Yahoo	LE	18.34	18.28	18.44
	PBM	18.27	18.25	18.43
	CPBM	18.42	18.32	18.39
	PAL	18.30	18.29	18.33
	MCPBM	18.26	18.25	18.36
	Click	18.51	18.47	18.57
MQ2007	LE	16.86	16.84	16.79
	PBM	16.84	16.86	16.80
	CPBM	16.88	16.91	16.82
	PAL	18.28	17.59	17.83
	MCPBM	16.82	16.83	16.74
	Click	16.89	17.16	16.96

下载: 导出CSV

参考文献(30)

[1]	Chen Jiawei, Dong Hande, Wang Xiang, et al. Bias and debias in recommender system: A survey and future directions [J]. arXiv preprint, arXiv: 2010.03240, 2020
[2]	Joachims T, Swaminathan A, Schnabel T. Unbiased learning-to-rank with biased feedback[C] //Proc of the 10th ACM Int Conf on Web Search and Data Mining. New York: ACM, 2017: 781−789
[3]	Hu Ziniu, Wang Yang, Peng Qu, et al. Unbiased lambdamart: An unbiased pairwise learning-to-rank algorithm[C] //Proc of the 28th World Wide Web Conf. New York: ACM, 2019: 2830−2836
[4]	Agarwal A, Zaitsev I, Wang Xuanhui, et al. Estimating position bias without intrusive interventions[C] //Proc of the 12th ACM Int Conf on Web Search and Data Mining. New York: ACM, 2019: 474−482
[5]	Fang Zhichong, Agarwal A, Joachims T. Intervention harvesting for context-dependent examination-bias estimation[C] //Proc of the 42nd Int ACM SIGIR Conf on Research and Development in Information Retrieval. New York: ACM, 2019: 825−834
[6]	Broder A. A taxonomy of web search[J]. ACM Special Interest Group on Information Retrieval Forum, 2002, 36(2): 3−10
[7]	Dupret G E, Piwowarski B. A user browsing model to predict search engine click data from past observations[C] //Proc of the 31st Annual Int ACM SIGIR Conf on Research and Development in Information Retrieval. New York: ACM, 2008: 331−338
[8]	Li Pengcheng, Li Runze, Da Qing, et al. Improving multi-scenario learning to rank in e-commerce by exploiting task relationships in the label space[C] //Proc of the 29th ACM Int Conf on Information & Knowledge Management. New York: ACM, 2020: 2605−2612
[9]	Sheng Xiangrong, Zhao Liqin, Zhou Guorui, et al. One model to serve all: Star topology adaptive recommender for multi-domain CTR prediction [J]. arXiv preprint, arXiv: 2101.11427, 2021
[10]	Yuan Bowen, Hsia J Y, Yang Mengyuan, et al. Improving ad click prediction by considering non-displayed events[C] //Proc of the 28th ACM Int Conf on Information and Knowledge Management. New York: ACM, 2019: 329−338
[11]	Yuan Bowen, Liu Yaxu, Hsia J Y, et al. Unbiased ad click prediction for position-aware advertising systems[C] //Proc of the 14th ACM Conf on Recommender Systems. New York: ACM, 2020: 368−377
[12]	Wang Xuanhui, Golbandi N, Bendersky M, et al. Position bias estimation for unbiased learning to rank in personal search[C] //Proc of the 11th ACM Int Conf on Web Search and Data Mining. New York: ACM, 2018: 610−618
[13]	Chuklin A, Markov I, Rijke M. Click models for web search[J]. Synthesis Lectures on Information Concepts, Retrieval, and Services, 2015, 7(3): 1−115
[14]	Chandar P, Carterette B. Estimating clickthrough bias in the cascade model[C] //Proc of the 27th ACM Int Conf on Information and Knowledge Management. New York: ACM, 2018: 1587−1590
[15]	Thung K H, Wee C Y. A brief review on multi-task learning[J]. Multimedia Tools and Applications, 2018, 77(22): 29705−29725
[16]	Vandenhende S, Georgoulis S, Van Gansbeke W, et al. Multi-task learning for dense prediction tasks: A survey[J/OL]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021[2021-12-09]. https://arxiv.org/pdf/2004.13379.pdf
[17]	吴宾,娄铮铮,叶阳东. 一种面向多源异构数据的协同过滤推荐算法[J]. 计算机研究与发展,2019,56(5):1034−1047 doi: 10.7544/issn1000-1239.2019.20180461 Wu Bin, Lou Zhengzheng, Ye Yangdong. A collaborative filtering recommendation algorithm for multi-source heterogeneous data[J]. Journal of Computer Research and Development, 2019, 56(5): 1034−1047 (in Chinese) doi: 10.7544/issn1000-1239.2019.20180461
[18]	Zhao Zhe, Hong Lichan, Wei Li, et al. Recommending what video to watch next: A multitask ranking system[C] //Proc of the 13th ACM Conf on Recommender Systems. New York: ACM, 2019: 43−51
[19]	Chen Zhao, Ngiam J, Huang Yanping, et al. Just pick a sign: Optimizing deep multitask models with gradient sign dropout [J]. arXiv preprint, arXiv: 2010.06808, 2020
[20]	Chapelle O, Chang Yi. Yahoo! Learning to rank challenge overview[C/OL] //Proc of the 2010 Int Conf on Yahoo! Learning to Rank Challenge. Cambridge, MA: JMLR, [2011-12-09]. http://proceedings.mlr.press/v14/chapelle11a
[21]	Qin Tao, Liu Tieyan. Introducing LETOR 4.0 datasets [J]. arXiv preprint, arXiv: 1306.2597, 2013
[22]	Ma Jiaqi, Zhao Zhe, Yi Xinyang, et al. Modeling task relationships in multi-task learning with multi-gate mixture-of-experts[C] //Proc of the 24th ACM SIGKDD Int Conf on Knowledge Discovery & Data Mining. New York: ACM, 2018: 1930−1939
[23]	Caruana R. Multitask learning[J]. Machine Learning, 1997, 28(1): 41−75 doi: 10.1023/A:1007379606734
[24]	Jacobs R A, Jordan M I, Nowlan S J, et al. Adaptive mixtures of local experts[J]. Neural Computation, 1991, 3(1): 79−87 doi: 10.1162/neco.1991.3.1.79
[25]	Chen Zhao, Badrinarayanan V, Lee C Y, et al. Gradnorm: Gradient normalization for adaptive loss balancing in deep multitask networks[C] //Proc of the 35th Int Conf on Machine Learning. Cambridge, MA: JMLR, 2018: 794−803
[26]	Liu Shikun, Johns E, Davison A J. End-to-end multi-task learning with attention[C] //Proc of the 32nd of the IEEE/CVF Conf on Computer Vision and Pattern Recognition. Piscataway, NJ: IEEE, 2019: 1871−1880
[27]	Kendall A, Gal Y, Cipolla R. Multi-task learning using uncertainty to weigh losses for scene geometry and semantics[C] //Proc of the 31st IEEE Conf on Computer Vision and Pattern Recognition. Piscataway, NJ: IEEE, 2018: 7482−7491
[28]	Ai Qingyao, Bi Keping, Luo Cheng , et al. Unbiased learning to rank with unbiased propensity estimation[C] //Proc of the 41st Int ACM SIGIR Conf on Research and Development in Information Retrieval. New York: ACM, 2018: 385−394
[29]	Joachims T. Optimizing search engines using clickthrough data[C] //Proc of the 8th ACM SIGKDD Int Conf on Knowledge Discovery and Data Mining. New York: ACM, 2002: 133−142
[30]	Guo Huifeng, Yu Jinkai, Liu Qing, et al. PAL: A position-bias aware learning framework for CTR prediction in live recommender systems[C] //Proc of the 13th ACM Conf on Recommender Systems. New York: ACM, 2019: 452−456

施引文献(3)

期刊类型引用(2)

1.	刘阳，鲁圆圆，郭成城. 基于优先级的数据中心任务优化调度算法设计. 计算机仿真. 2025(01): 497-500+507 . 百度学术
2.	骆海霞. 基于递推估计的Web前端偶发任务能耗感知方法. 黑龙江工业学院学报(综合版). 2023(10): 115-120 . 百度学术