基于动态情境感知的W5模型研究

摘要

Social network apps, such as Twitter, Sina Micro-blog and etc. have provided mass context-information associated with user IDs (who), check-in time (when), GPS coordinates (where), topic (what) and incentive (why) of tweets (5W for short) for location based services. The availability of such data received from users offers a good opportunity to study the user’s behavior and preference. In this paper, we propose a W5 probabilistic model to exploit such data with context by jointly probability to discover users’ dynamic behaviors from temporal, spatial and activity aspects. Our work is applied to prediction for user and location. Experimental results on two real-world datasets show that W5 model is effective in discovering users’ spatial-temporal prediction, and outperforms state-of-the art baselines, such as W4, on accuracy [UP: (GT: 3.75%, ST: 6.54%) and LP: (GT: 8.7%, ST: 20.6%)] at aspects of user prediction and location prediction based on Geo-Text (GT) and Sina-Tweets (ST).%Twitter、Sina Micro-blog等社交网络应用为基于位置的服务提供了大量的情境信息，如用户ID(who)、签到时间(when)、GPS坐标(where)、微博内容主题词(what)和微博内容诱因词(why)等，简称5W。它们为用户的行为和偏好研究提供了契机。该文提出了基于5W动态情境感知信息的W5概率模型，并采用包含情境信息的联合概率分布分别从时间、空间和活动等方面挖掘用户动态行为，用于用户和位置的预测。该文实验基于两个数据集：Geo-text(GT)和Sina-tweets(ST)，在数据集上进行了用户预测(UP)和位置预测(LP)实验。实验结果表明，W5模型在UP和LP两方面准确率均高于W4模型。同时，W5模型在时间误差和空间距离误差两方面也取得了较好的性能。