安徽土壤肥料信息系统中施肥决策的研究

研究了安徽土壤肥料信息系统中配方施肥决策的算法和模型中的参数。采用土壤养分校正系数法,得出了安徽省各地区土壤养分校正系数值及其与土测值之间的关系,从而根据土壤养分含量、作物种类和目标产量确定某一区域的施肥量,为土壤肥料信息系统中的施肥决策提供方法依据。     

Towards agent-based decision making in the electronic marketplace: interactive recommendation and automated negotiation

The bloom of electronic commerce has changed the whole outlook of traditional trading behavior. Through the Internet, different business entities can now easily interact with each other and then have their transactions within a minimum time. However, the advanced hardware facilities for Internet infrastructure and the entrancing Web sites are not the only decisive factors to guarantee a successful business in the electronic market. To enable the transactions, the problems that ensue during the complicated activities in electronic commerce have to be resolved. This paper presents an agent-based system to support two types of activities most related to the decision-making process in the Internet commerce. The first part of our system is for interactive recommendation; and the second part, for automated negotiation. Experiments have been conducted to evaluate their corresponding performance and the results show the efficiency of the proposed system and its potential toward the future electronic market.     

The Recommendation of Digital TV Program Based on Rankboost Algorithm

A Rankboost algorithm using for recommendation of digital TV program is proposed. The program feature model expressed by time, channel, main category and subcategory are founded. The rank function is calculated based on Rankboost algorithm and real user TV watching records, its AUC （Area under the curve） of ROC （Receiver operator characteristic） is the highest comparing with those got by simple statistic and Bayesian approach, the experimental recommendation results which are calculated using the above rank function conform user real actions well.     

基于经验分布和KL散度的协同过滤推荐质量评价研究

提出了一种基于经验分布和KL散度的协同过滤推荐质量评价方法RQE-EDKL（recommendation quality evaluation based on empirical distribution and KL divergence）。RQE-EDKL首先利用历史用户—商品数据生成不同商品数量下的商品历史使用概率分布；然后，利用该分布与各个协同过滤推荐方法得到的用户商品使用概率进行比较，计算其KL散度；最后，将KL散度最小的推荐结果视为最佳推荐结果并推送给用户。在TalkingData数据集上的实验结果表明，RQE-EDKL评价方法能够有效地在不同的推荐结果中选择更为切合用户真实需求的推荐结果，从而提高了协同过滤推荐的质量。     

基于深度神经网络和加权隐反馈的个性化推荐

改进的矩阵分解(SVD++)将用户和物品特征向量的内积作为用户对物品的评分,而内积无法捕捉用户与物品之间复杂的高阶非线性关系.此外,SVD++在融入用户隐式反馈时,未区分不同交互物品对于用户特征表达的贡献.针对上述问题,文中提出基于深度神经网络和加权隐反馈的推荐算法(DeepNASVD++),采用深度神经网络建模用户与物品之间的关系,使用注意力机制计算历史交互物品在建模用户隐式反馈时的权重.在公开数据集上的实验验证文中算法的有效性.     

Investigating online social media users’ behaviors for social commerce recommendations

Online social media create virtual communities and network platforms that people use to create, share, and exchange opinions, views and experiences. With social networks, social commerce not only relies on commerce, but online social media can also promote the sale of goods or services online. Many online operators have begun to use recommendation systems to analyze customer purchase history and identify individual products that customers may purchase. This enables the company to send product information to consumers to attract their attention. In addition, consumers have a higher purchase rate for recommended products based on consumer data. Based on a survey in Taiwan society, this study uses the questionnaire survey method to collect data on a relational database. This study investigates Taiwan online social media users’ behaviors using data mining methods, including clustering analysis and association rules. Clustering analysis is to investigate possible profiles of users and association rules are to find knowledge patterns and rules of user profiles, online social media usage motivation/preferences and social commerce behavior in order to generate social commerce recommendations in terms of social technology development in the modern society.     

基于社交网络信任关系的服务推荐方法

随着服务型计算的兴起,大量跨领域电子服务应运而生。用户要从众多服务中挑选出适合自己且可信的服务十分困难,因而提出高效的服务推荐算法十分必要。传统的协同推荐方法存在冷启动、数据稀疏以及实时性不好等问题,在评分数据较少时推荐效果不佳。为获得更好的推荐结果,文中在社交网络中使用信任传递机制,建立信任传递模型,由此获取任意用户间的信任度。另一方面,设计了相似性判定指标,凭借系统评分数据,求得用户间的偏好相似度。在得到用户间信任度和偏好相似度的基础上,根据社交网络的特性,动态结合两部分指标以获得综合推荐权重,再以此权重替代传统相似度衡量标准进行基于用户的协同过滤推荐。所提方法能在解决传统推荐算法问题的基础上进一步提升推荐效果,并以准确率、覆盖率为标准在Epinions数据集上进行验证,获得了较好的效果。     

融合主题模型和协同过滤的多样化移动应用推荐

随着移动应用的急速增长,手机助手等移动应用获取平台也面临着信息过载的问题.面对大量的移动应用,用户很难找到想到的或适合的应用,而另一方面长尾应用淹没在资源池中不被人所知.已有推荐方法多注重推荐准确率,忽视多样性,推荐结果中多是下载量高的应用,使得推荐系统的数据积累越来越偏向于热门应用,导致长期的推荐效果越来越差.针对此问题,本文首先改进了两个推荐方法,提出了将用户的主题模型和应用的主题模型与MF相结合的LDA_MF模型,以及将应用的标签信息和用户行为数据同时加以考虑的LDA_CF算法.为了结合不同算法的优点,在保证推荐准确率的条件下提升推荐结果的多样性,我们提出了融合LDA_MF、LDA_CF以及经典的基于物品的协同过滤模型的混合推荐算法.文章使用真实的大数据评测所提推荐算法,结果显示所提推荐方法能够得到推荐多样性更好且准确率高的结果.