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...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.展开更多
A two-sided assembly line is typically found in plants producing large-sized products. Its advantages over a one-sided line and the difficulties faced in two-sided line balancing problems were discussed. A mathematica...A two-sided assembly line is typically found in plants producing large-sized products. Its advantages over a one-sided line and the difficulties faced in two-sided line balancing problems were discussed. A mathematical model for two-ALB problem was suggested. A modification of the “ranked positional weight” method, namely two-ALB RPW for two-ALB problems was developed. Experiments were carried out to verify the performance of the proposed method and the results show that it is effective in solving two-sided assembly line balancing problems.展开更多
文摘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.
基金Key Projectof Scientific and TechnologicalCommittee of Shanghai(No.0 3 11110 0 5 )
文摘A two-sided assembly line is typically found in plants producing large-sized products. Its advantages over a one-sided line and the difficulties faced in two-sided line balancing problems were discussed. A mathematical model for two-ALB problem was suggested. A modification of the “ranked positional weight” method, namely two-ALB RPW for two-ALB problems was developed. Experiments were carried out to verify the performance of the proposed method and the results show that it is effective in solving two-sided assembly line balancing problems.
