Phone-based Ambient Temperature Measurement with a New Confidence-based Truth Inference Model

What is it about?

Monitoring the indoor temperature can help saving of energy and improve the comfort level. Smartphone, as a ubiquitousdevice, can be an additional data source to provide the ambient temperature estimation. However, the estimation resultssometimes can be unreliable due to the different phone using states. How to integrate multiple estimation results in one areato get a more accurate prediction result is still a challenge. In this work, we proposed one phone-based ambient temperaturemeasurement system which contains two models. The first temperature prediction model takes easily accessible phone statefeatures as inputs and outputs ambient temperature prediction with a confidence value. The second truth inference modeltakes multiple prediction results with confidences as inputs and outputs a referred final prediction result. Our temperatureprediction model reaches 0.253◦C with MAE in our testing set. We also proved by transfer learning our model can be usedin other new type of phones. We evaluate the truth inference model in our testing dataset and it reaches 0.128◦C, whichoutperforms the state-of-the-art truth inference algorithms. We believe this work can contribute to energy conservation andprovide new ideas for crowdsourcing.

Featured Image

Why is it important?

(1) We proposed one new truth inference model named CBTS to integrate multiple prediction results. Thetree structure enables the model deal with the crowdsourcing groups of dynamic number. (2) The truth inference model we proposed is based on the generated confidence value of every prediction.To our knowledge, there are only a few truth inference algorithms take into account the confidence ofgiven answers. Moreover, our confidence is self generated according the information of collected features,which gets rid of the subjectivity of artificial calibration and reduce the cost of crowdsourcing. (3) Our temperature estimation model can not only give a prediction of ambient temperature, but also couldgive a confidence value of this prediction, which, to our knowledge, is not implemented in any previoustemperature prediction models. The generated confidence value could help us better integrate multipleestimation results. (4) Almost all the previous related works call for the information of CPU utility to do a temperature predic-tion. However, this kind of information has been forbidden to be obtained in most of smartphones. Ourprediction model do not need the CPU utility information and all the 9 features are easy to access in almostall mobile operating systems.