多任务学习技术在混合推荐中的应用研究与实现

发布时间：2018-04-05 15:00

本文选题：推荐算法　切入点：协同过滤　出处：《浙江大学》2012年硕士论文

【摘要】：互联网内容和商业化的不断发展,一方面带给用户方便的娱乐、消费方式,另一方面也带给了用户严重的信息过载问题。用户受困于过量的信息、过量的商品选择,将要花费更多的时间来浏览备选信息,却有可能因为受限的信息处理能力做出更差的选择。搜索引擎在一定程度上缓解了信息过载,但需要用户非常明确的知道自己的选择,并不能帮助用户做决策。推荐系统正是为了帮助用户决策而产生。本文主要对如何提升推荐系统的预测准确性进行深入研究。我们将此处的混合推荐定义为多个推荐预测器的混合,本文认为,因为不同预测器模型的基本假设和侧重点不同,将他们混合将有助于最终的预测性能。首先我们建立多个经典协同过滤算法模型,并且根据特定应用环境进行时效扩展,之后将这些算法模型进行统一的、用户无关的混合。观察到这相较于独立预测器有性能提高,但是对于不同用户性能差异很大。于是我们提出利用多任务学习机制,充分考虑不同用户的个性特征,建立起用户相关的混合模型,主要表现为对于不同用户具有不同的预测器权重向量。为减少计算复杂度,我们还进行了用户聚类,并考察对性能的影响。我们详细描述了实验设计。实验证明,进行区分用户的混合模型建模有助于预测准确性的提高。但是考虑更有商业价值的覆盖率／命中率考察指标时,发现我们的算法并没有实质性的进步。这启示我们进行下一步的研究。
[Abstract]:With the continuous development of Internet content and commercialization, on the one hand, it brings users convenient entertainment and consumption mode, on the other hand, it also brings serious information overload problems to users.Users are trapped in excessive information, excessive choice of goods, will spend more time to browse alternative information, but may make worse choices because of limited information processing capacity.The search engine alleviates the information overload to some extent, but needs the user to know his choice very clearly, can not help the user to make the decision.Recommendation systems are created to help users make decisions.This paper focuses on how to improve the prediction accuracy of recommendation system.We define the hybrid recommendation here as a mixture of multiple recommendation predictors. In this paper we think that mixing different predictors will contribute to the final prediction performance because of the different basic assumptions and different emphases of different predictor models.Firstly, we build several classical collaborative filtering algorithms and extend them according to the specific application environment. Then we combine these models into a unified, user-independent mixture.It is observed that the performance is improved compared with the independent predictor, but the performance varies greatly for different users.So we propose to use the multi-task learning mechanism to fully consider the personality characteristics of different users and establish a user-related hybrid model which is mainly represented by different predictor weight vectors for different users.To reduce computational complexity, we also perform user clustering and investigate the effect on performance.We describe the experimental design in detail.It is proved by experiments that the mixed model can improve the prediction accuracy.But when we consider the coverage / hit ratio of more commercial value, we find that our algorithm has not made substantial progress.This enlightens us to carry on the next research.
【学位授予单位】：浙江大学
【学位级别】：硕士
【学位授予年份】：2012
【分类号】：TP391.3

【相似文献】