协同过滤是目前解决信息过载问题的主要方法之一,然而其推荐的多样性不足,且在冷启动场景下推荐效果较差.提出了基于用户偏好和动态兴趣的多样性推荐方法 DRMUD(A Diversified Recommendation Method Based on User Preference and Dyna...协同过滤是目前解决信息过载问题的主要方法之一,然而其推荐的多样性不足,且在冷启动场景下推荐效果较差.提出了基于用户偏好和动态兴趣的多样性推荐方法 DRMUD(A Diversified Recommendation Method Based on User Preference and Dynamic Interest).首先通过对用户历史反馈数据分析用户的多样性偏好,得出用户的多样倾向度;然后引入时间衰减函数,动态调整用户的历史评分数据;最后将矩阵分解和项目疲劳函数相结合,并加入多样倾向度调节两者所占比重.当新用户加入系统时,通过网格索引为其产生最信任邻居,新用户缺失的反馈信息由最信任邻居代替.实验结果表明,DRMUD算法有效缓解了用户冷启动问题,并能在保证准确率的前提下提高推荐结果的多样性.展开更多
This paper intuitively examines the dynamic behavior of two highly relevant interest rates in China. The first one is the government rate, which is decided and published by the central bank and can be simulated by pur...This paper intuitively examines the dynamic behavior of two highly relevant interest rates in China. The first one is the government rate, which is decided and published by the central bank and can be simulated by pure jump process. Estimation of the jump intension is given out. And by different robustness test, it keeps stable. The jump size has met the condition to make interest rate within reasonable bounds and shown some meaning of economic cycle behavior. The second one is the market rate, which is estimated by spline approximation based on the transaction data of government bonds. Several models, including Vasicek model, Vasicek-GARCH （1,1） model, CIR model, and CIR-GARCH（1,1）, are empirically tested and the best performance is done by the Vasicek-GARCH（1,1） model. Furthermore, the estimate bias problem due to the near unit root process is tested and evidenced by both traditional methods and GPH test. Impact of government rate on market rate is finally checked and analyzed.展开更多
针对目前在微博推荐领域主要使用单一向量表示用户兴趣且缺乏对兴趣之间复杂关系的捕捉能力,导致用户兴趣表示不全面,推荐准确性较低的问题,提出了基于双重注意力机制的多兴趣动态路由微博推荐算法(multi-interest network with dynamic...针对目前在微博推荐领域主要使用单一向量表示用户兴趣且缺乏对兴趣之间复杂关系的捕捉能力,导致用户兴趣表示不全面,推荐准确性较低的问题,提出了基于双重注意力机制的多兴趣动态路由微博推荐算法(multi-interest network with dynamic routing microblogging recommendation algorithm based on dual attention mechanism,MINDDouAtt),用于提高用户兴趣的表征能力。首先,通过动态路由从用户行为数据中提取多个兴趣胶囊,并将这些兴趣胶囊输入到自注意力机制中以对不同兴趣胶囊之间的关联信息进行交叉学习,提高兴趣的表征能力。然后,通过引入标签感知注意力机制来调节不同兴趣胶囊之间的重要性,以更好地满足用户的个性化推荐需求。实验表明,MINDDouAtt算法在亚马逊图书、天猫和微博数据集上的S HR@10值相较于最好的对比模型分别提升了33.66%、10.49%、9.60%。该算法能够在电子商务等领域为用户提供更准确和个性化的推荐结果。展开更多
推荐系统的准确度经常受到各类偏差的影响,流行度偏差是影响推荐准确度的重要因素之一。传统的纠偏方法主要基于项目属性,通过引入惩罚因子来抑制热门项目的推荐,未考虑用户兴趣和时间的影响。针对此问题,提出了基于项目流行度和用户动...推荐系统的准确度经常受到各类偏差的影响,流行度偏差是影响推荐准确度的重要因素之一。传统的纠偏方法主要基于项目属性,通过引入惩罚因子来抑制热门项目的推荐,未考虑用户兴趣和时间的影响。针对此问题,提出了基于项目流行度和用户动态兴趣的自适应纠偏方法(Adaptive Popularity and Dynamic Interest,APDI)。结合因果图从项目流行度和用户个性化两个方面综合分析影响流行度偏差的主要因素,根据项目质量、从众效应、用户兴趣对时间的敏感度不同,计算相应时间间隔内项目流行度与用户动态兴趣的综合评分,更加有效地降低流行度偏差;通过指数加权移动平均的方法,根据时间衰减程度对用户当前兴趣的影响来计算用户兴趣评分,以捕捉用户的短期兴趣偏好。在3个真实数据集上验证了所提方法的有效性,实验结果表明,APDI有效提高了推荐的准确度、召回率和归一化折损累计增益。展开更多
As living standards improve,the demand for artworks has been escalating,transcending beyond the realm of mere basic human necessities.However,amidst an extensive array of artwork choices,users often struggle to swiftl...As living standards improve,the demand for artworks has been escalating,transcending beyond the realm of mere basic human necessities.However,amidst an extensive array of artwork choices,users often struggle to swiftly and accurately identify their preferred piece.In such scenarios,a recommendation system can be invaluable,assisting users in promptly pinpointing the desired artworks for better service design.Despite the escalating demand for artwork recommendation systems,current research fails to adequately meet these needs.Predominantly,existing artwork recommendation methodologies tend to disregard users’implicit interests,thereby overestimating their capability to articulate their preferences in full and often neglecting the nuances of their diverse interests.In response to these challenges,we have developed a weighted artwork correlation graph and put forth an embedding-based keyword-driven artwork search and recommendation methodology.Our approach transforms the keywords that delineate user interests into word embedding vectors.This allows for an effective distinction between the user’s core and peripheral interests.Subsequently,we employ a dynamic programming algorithm to extract artworks from the correlation graph,thereby obtaining artworks that align with the user’s explicit keywords and implicit interests.We have conducted an array of experiments using real-world datasets to validate our approach.The results attest to the superiority of our method in terms of its efficacy in searching and recommending artworks.展开更多
文摘协同过滤是目前解决信息过载问题的主要方法之一,然而其推荐的多样性不足,且在冷启动场景下推荐效果较差.提出了基于用户偏好和动态兴趣的多样性推荐方法 DRMUD(A Diversified Recommendation Method Based on User Preference and Dynamic Interest).首先通过对用户历史反馈数据分析用户的多样性偏好,得出用户的多样倾向度;然后引入时间衰减函数,动态调整用户的历史评分数据;最后将矩阵分解和项目疲劳函数相结合,并加入多样倾向度调节两者所占比重.当新用户加入系统时,通过网格索引为其产生最信任邻居,新用户缺失的反馈信息由最信任邻居代替.实验结果表明,DRMUD算法有效缓解了用户冷启动问题,并能在保证准确率的前提下提高推荐结果的多样性.
文摘This paper intuitively examines the dynamic behavior of two highly relevant interest rates in China. The first one is the government rate, which is decided and published by the central bank and can be simulated by pure jump process. Estimation of the jump intension is given out. And by different robustness test, it keeps stable. The jump size has met the condition to make interest rate within reasonable bounds and shown some meaning of economic cycle behavior. The second one is the market rate, which is estimated by spline approximation based on the transaction data of government bonds. Several models, including Vasicek model, Vasicek-GARCH （1,1） model, CIR model, and CIR-GARCH（1,1）, are empirically tested and the best performance is done by the Vasicek-GARCH（1,1） model. Furthermore, the estimate bias problem due to the near unit root process is tested and evidenced by both traditional methods and GPH test. Impact of government rate on market rate is finally checked and analyzed.
文摘针对目前在微博推荐领域主要使用单一向量表示用户兴趣且缺乏对兴趣之间复杂关系的捕捉能力,导致用户兴趣表示不全面,推荐准确性较低的问题,提出了基于双重注意力机制的多兴趣动态路由微博推荐算法(multi-interest network with dynamic routing microblogging recommendation algorithm based on dual attention mechanism,MINDDouAtt),用于提高用户兴趣的表征能力。首先,通过动态路由从用户行为数据中提取多个兴趣胶囊,并将这些兴趣胶囊输入到自注意力机制中以对不同兴趣胶囊之间的关联信息进行交叉学习,提高兴趣的表征能力。然后,通过引入标签感知注意力机制来调节不同兴趣胶囊之间的重要性,以更好地满足用户的个性化推荐需求。实验表明,MINDDouAtt算法在亚马逊图书、天猫和微博数据集上的S HR@10值相较于最好的对比模型分别提升了33.66%、10.49%、9.60%。该算法能够在电子商务等领域为用户提供更准确和个性化的推荐结果。
文摘推荐系统的准确度经常受到各类偏差的影响,流行度偏差是影响推荐准确度的重要因素之一。传统的纠偏方法主要基于项目属性,通过引入惩罚因子来抑制热门项目的推荐,未考虑用户兴趣和时间的影响。针对此问题,提出了基于项目流行度和用户动态兴趣的自适应纠偏方法(Adaptive Popularity and Dynamic Interest,APDI)。结合因果图从项目流行度和用户个性化两个方面综合分析影响流行度偏差的主要因素,根据项目质量、从众效应、用户兴趣对时间的敏感度不同,计算相应时间间隔内项目流行度与用户动态兴趣的综合评分,更加有效地降低流行度偏差;通过指数加权移动平均的方法,根据时间衰减程度对用户当前兴趣的影响来计算用户兴趣评分,以捕捉用户的短期兴趣偏好。在3个真实数据集上验证了所提方法的有效性,实验结果表明,APDI有效提高了推荐的准确度、召回率和归一化折损累计增益。
文摘As living standards improve,the demand for artworks has been escalating,transcending beyond the realm of mere basic human necessities.However,amidst an extensive array of artwork choices,users often struggle to swiftly and accurately identify their preferred piece.In such scenarios,a recommendation system can be invaluable,assisting users in promptly pinpointing the desired artworks for better service design.Despite the escalating demand for artwork recommendation systems,current research fails to adequately meet these needs.Predominantly,existing artwork recommendation methodologies tend to disregard users’implicit interests,thereby overestimating their capability to articulate their preferences in full and often neglecting the nuances of their diverse interests.In response to these challenges,we have developed a weighted artwork correlation graph and put forth an embedding-based keyword-driven artwork search and recommendation methodology.Our approach transforms the keywords that delineate user interests into word embedding vectors.This allows for an effective distinction between the user’s core and peripheral interests.Subsequently,we employ a dynamic programming algorithm to extract artworks from the correlation graph,thereby obtaining artworks that align with the user’s explicit keywords and implicit interests.We have conducted an array of experiments using real-world datasets to validate our approach.The results attest to the superiority of our method in terms of its efficacy in searching and recommending artworks.
