事件社交网中基于有向标签图及用户反馈的活动推荐方法

由于基于事件的社交网络（EBSN）中的活动具有时效性，传统社交网络推荐算法无法适用于EBSN。此外，大多数算法忽略了能影响后续推荐质量的前用户是否接受活动的反馈意见。为此，提出一种EBSN中基于有向标签图及用户反馈的活动推荐方法。首先，将EBSN抽象为有向标签图，并抽取图节点及边的属性特征信息，构建有向图结构特征（DGSF）索引，该索引由节点属性特征索引、有向边属性特征索引以及时间特征索引构成，利用该索引对节点及边进行初次过滤。其次，提出基于DGSF索引的多属性候选集过滤策略，利用时间、节点的出入度、标签类型等特征的限制，实现对查询图候选集的进一步剪枝，避免冗余计算。然后，提出一种具有用户反馈的改进UCB（Upper Confidence Bound）活动推荐算法——EN_UCB，通过引入弹性网回归，根据多影响因素计算用户对活动的兴趣值，为用户推荐兴趣值高的活动，同时接收用户是否接受该活动的反馈，以优化后续用户的推荐。大量实验结果表明，EN_UCB算法的接受率高于TS（Thompson Sampling）、UCB以及eGreedy算法，遗憾率远远低于TS和eGreedy算法，且运行效率高于TS、UCB以及eGreedy算法，活动数越大，优势越明显。所提算法能有效实现EBSN上的在线活动推荐。

Activity recommendation method based on directed label graph and user feedback in event-based social network

Due to the timeliness of activities in Event-Based Social Network （EBSN）， the traditional social network recommendation algorithms cannot be applied to EBSN. In addition， most of the traditional recommendation algorithms ignore the feedback that can affect whether the previous users accept the recommendation， which influences subsequent recommendation quality. Therefore， an activity recommendation method based on directed label graph and user feedback in EBSN was proposed. Firstly， EBSN was abstracted into a directed label graph， and a Directed Graph Structure Feature （DGSF） index was construction by extracting the property feature information of nodes and edges to filter nodes and edges for the first time. DGSF index consists of node property feature index， directed edge property feature index and time feature index. Secondly， a multi-attribute candidate set filtering strategy based on DGSF index was proposed. By using the limits of time， in-degrees and out-degrees of nodes， and label types， the further pruning of the candidate sets was realized to avoid redundant computation. Thirdly， an improved UCB （Upper Confidence Bound） activity recommendation algorithm with user feedback was put forward， namely EN_UCB （Elastic Net UCB）. In EN_UCB， with the introduction of the elastic net regression， the interest values of the user to the activities were calculated according to many influencing factors， and the activities with high interest values were recommended to the user. At the same time， the feedback whether the user accepted the activities was received to optimize the subsequent user recommendation. Experimental results show that EN_UCB has the accept rate higher than TS （Thompson Sampling）， UCB and eGreedy， the regret rate far lower than TS and eGreedy， the running time superior to TS， UCB and eGreedy， and the larger the number of activities， the more obvious the advantages. The proposed method implements online activity recommendation in EBSN effectively.