Recognizing Unknown Activities Using Semantic Word Vectors and Twitter Timestamps

Moe Matsuki, Sozo Inoue,
International Symposium on Applied Engineering and Sciences (SAES2016)
(Not Available)
(Not Available)
In the common method of activity recognition, supervised machine learning are used from the pairs of activity labels and sensor data as a training dataset. However, these methods generate the problem that unknown activity labels which do not appear in the training dataset cannot be estimated.
To overcome the problem, we propose a zero-shot learning with the word vectors constructed from semantic word vectors and Twitter timestamps.
The zero-shot learning can estimate the classes which do not appear in the training dataset, by utilizing additional source data, such as text data.
Our approach uses semantic word vectors generated by the tool word2vec as additional source, which can generate the semantic vectors of the words by using the Skip-gram model. Moreover, to make the relation between the sensor data and semantic words vectors more correlated, we add the Twitter timestamps, which are the times of activities in a day, to the semantic words vectors.
To evaluate the proposed method, we evaluated whether we could estimate unknown activity classes, with the sensor data set collected from 20 households for 4 months, along with the user-generated labels using the web system which can estimate, modify, and add new activity types.
As a result, the proposed method could even estimate unknown activity classes.
Moreover, by utilizing Twitter timestamps and semantic word vectors from the Japanese Wikipedia in word vectors, the method could estimate 9 unknown activity classes.

Data Files

No records to display