1.
标题
· Recognizing Human Daily Activity Using Social Media Sensors and Deep Learning
· 基于深度学习识别社交媒体中的人类日常活动
2. 成果信息
· Gong, J., Li, R., Yao, H., Kang, X., & Li, S. (2019). Recognizing Human Daily Activity Using Social Media Sensors and Deep Learning. International Journal of Environmental Research and Public Health, 16(20), 3955.
· https://doi.org/10.3390/ijerph16203955
· This research was supported by the NSF of China (Grant No. 41801378, 61972365, 61673354, 61672474), the Fundamental Research Funds for the Central Universities, China University of Geosciences(Wuhan) (Grant No. CUGQY1948).
3. 成果团队成员
· 龚君芳(第一作者),博士,bat365官网登录入口,研究方向:空间统计与分析。Email: jfgong@cug.edu.cn。
· 李润佳,硕士生,中国地质大学(武汉)计算机学院。
· 姚宏,博士,教授,中国地质大学(武汉)计算机学院。
· 康晓军,博士,副教授,中国地质大学(武汉)计算机学院。
· 李圣文,博士,副教授,bat365官网登录入口。
4. 成果介绍
人类的日常活动类别,如运动、购物等,代表着个人的生活方式和模式,反映了个人习惯、生活方式和喜好,在人类健康和其他许多应用领域都具有重要价值。社交媒体是一个公共、开放的平台和工具,人们在这个平台上自由地交流、表达、分享产生了海量的数据。这些数据不仅带有时空信息,还带有丰富的属性信息,蕴含了一定的行为语义,为更精确地建模人类行为模式提供了新的可能性。现有研究大多仅考虑文本结构语义,或者仅考虑时空语义,或者时空语义加上简单的文本标注信息来进行人类活动类型判别,存在判别精细度与准确度不高的问题。为此,本研究在综合考虑文本语义、外部知识和时间信息的基础上,提出了一种基于深度学习的人类活动类型识别模型。此外,我们还构建了一个可用于训练和评估人类活动类型识别模型的数据集。实验结果表明,与传统方法相比,该模型显著提高了人类活动类别的识别精度。
Table 1
Activity |
Accuracy |
SVM |
TF-IDF |
ALSTM-BASIC |
ALSTM-DE |
ALSTM-TE |
ALSTM-DE-TE |
Eating food |
0.6502 |
0.4281 |
0.8019 |
0.8722 |
0.8738 |
0.8706 |
Beauty & spa |
0.6384 |
0.1040 |
0.8944 |
0.9200 |
0.9200 |
0.9200 |
Entertainment activity |
0.5376 |
0.2064 |
0.7888 |
0.7936 |
0.8000 |
0.8384 |
Travel |
0.5408 |
0.1504 |
0.7904 |
0.8368 |
0.8112 |
0.8352 |
Shopping |
0.4480 |
0.2304 |
0.7520 |
0.7632 |
0.8016 |
0.7888 |
Services |
0.4432 |
0.2112 |
0.7440 |
0.7360 |
0.7056 |
0.7296 |
Sports |
0.4064 |
0.5200 |
0.7536 |
0.7824 |
0.7712 |
0.7440 |
Treating & Health |
0.5072 |
0.8288 |
0.8912 |
0.8800 |
0.8736 |
0.8496 |
Car-related activities |
0.4384 |
0.8928 |
0.8176 |
0.8608 |
0.8400 |
0.8256 |
Nightlife |
0.4544 |
0.6864 |
0.8496 |
0.7792 |
0.8128 |
0.7744 |
Keep pets |
0.4624 |
0.4960 |
0.8656 |
0.8512 |
0.9040 |
0.8832 |
Engaged in education |
0.4080 |
0.7808 |
0.8112 |
0.7376 |
0.8288 |
0.8096 |
Religious activities |
0.5296 |
0.5856 |
0.8368 |
0.8624 |
0.8592 |
0.8736 |
Listening |
0.3215 |
0.5749 |
0.7629 |
0.7902 |
0.7929 |
0.7766 |
Overall accuracy |
0.4897 |
0.4753 |
0.8129 |
0.8199 |
0.8293 |
0.8242 |
(a) Accuracy (b) loss curve
Here, ALSTM-BASIC is the model composed of the post words embedding component and the LSTM component. ALSTM-DE is the model that includes the post words embedding component, dictionary embedding component and the LSTM component. ALSTM-TE is the model composed of the post words embedding component, temporal information embedding component and the LSTM component. ALSTM-DE-TE is the model composed of all the components