| 摘要: |
| 上下文广告与用户兴趣及网页内容相匹配,可增强用户体验并提高广告点击率.而广告收益与广告点击率直接相关,准确预测广告点击率是提高上下文广告收益的关键.目前,上下文广告推荐面临如下问题:(1) 网页数量及用户数量规模很大;(2) 历史广告点击数据十分稀疏,导致点击率预测准确率低.针对上述问题,提出一种基于联合概率矩阵分解的因子模型AdRec,它结合用户、广告和网页三者信息进行广告推荐,以解决数据稀疏时点击率预测准确率低的问题.算法复杂度随着观测数据数量的增加呈线性增长,因此可应用于大规模数据. |
| 关键词: 推荐算法 联合概率矩阵分解 上下文广告 准确率 数据稀疏 |
| DOI：10.3724/SP.J.1001.2013.04238 |
| 分类号: |
| 基金项目:国家自然科学基金(60873243) |
|
| Using Unified Probabilistic Matrix Factorization for Contextual Advertisement Recommendation |
|
TU Dan-Dan, SHU Cheng-Chun, YU Hai-Yan
|
|
Institute of Computing Technology, The Chinese Academy of Sciences, Beijing 100190, China
|
| Abstract: |
| Combining user interests with visited web page contents to perform contextual advertising enhances the user experience and adds more ad clicks, increasing revenue. The key issue is to improve the prediction accuracy of click rates for advertisements. The crucial challenges of the advertisement recommendation algorithm are the scalability on large number of users and web page contents, and the low prediction accuracy resulting from data sparsity. When data is large and sparse, the accuracy and efficiency of the traditional recommendation algorithms is poor. This paper proposes a factor model called AdRec. Based on the Unified Probability Matrix Factorization (UPMF), the model addresses the data sparsity problem by combining features of users, advertisements and web page contents to predict the click rate with higher accuracy. In addition, the computational complexity of our algorithm is linear with respect to the number of observed data, and scalable to very large datasets. |
| Key words: recommendation algorithm unified probabilistic matrix factorization contextual advertising accuracy sparse data |
