In this paper, we introduce a method using multimodal data analytics to predict missing data collected by sensors. Our approach is to find data at the near-by location and time by using the time-filtering algorithm and incrementally scanning radius to replace missing data. The method is evaluated by using MediaEval 2019 Insight for wellbeing - subtask 1 dataset and evaluation metric. The results show that the proposed method works well and predict missing data with high accuracy.
Air pollution is proved to be a significant factor afect on
human beings [
The primary purpose of the proposed method is to define a hypothesis that can represent the associations among heterogeneous data and towards building a system that able to predict missing values in the provided dataset. The hypothesis points out that there is a strong association of heterogeneous data recording at the near-by location and time. Thus, we build the time filtering algorithm and radius-based increment scan policy to gather near-by data whose values can be used to predict missing data. The following (sub)sections describe in detail how to filter data and collect useful position information to predict missing data. 2.1
• Circling Time: This function is to collect all nearby-time data. We first cluster all given datasets into diferent groups so that each group has the same date and time (i.e., same day). Then only data happens within start _time and end_time are selected. It should be noted that start _time and end_time denote the time period when data missing. • Circling Position: In order to collect all near-bylocation data, we define the formula that calculates the distance of two coordinates. All data recorded within this distance are selected. The formula is deifned as follows: where: d: is the distance between the two points; r : is the radius of the sphere; α1, α2: latitude of point 1 and latitude of point 2 (in radians); β1,β2: longitude of point 1 and longitude of point 2 (in radians). The radius is set from 1m to 100m. 2.2
After running the circling time and circling position, we obtain the P M2.5 value of some nearest positions; we then calculate the Maximum, Minimum and Average of these values from a position that needs to predict. To optimizing the results, we incrementally increase the radius step by step from 1m to 20m at this time to scan all positions. According to our experience, we choose the ideal radius is 20m since within the 20m radius the predicted P M2.5 values reach the highest accuracy.
If within 20m radius, we cannot get any point, we will take a single nearest point in [21m, 100m]. If there is no point in [0m, 100m], set value for P M2.5 is zero and from thence we have build Algorithm 1. 3
The experimental results running on the training dataset is denoted in Table 1
Experimental results are evaluated based on optimized the Maximum, Minimum and Average precision. This result shows that although our proposed method is simple but it is efective. Our best run is run with Minimum. Because the Minimum value has noise very low, the value is more accurate than the other two methods (Maximum, Average). Nevertheless, there is not a big gap among submitted runs. 4
We report our work at the MediaEval 2019 Insight for Wellbeing task - subtask 1. We use time-filtering algorithm and radius-based increment policy to gather near-by location and time data towards predicting missing data. The results show that our solution has high accuracy.
Algorithm 1: Recovery PM2.5’s values from near-by location and time data 1 DataA: Merge all data in a group; 2 DataB: In DataA, retrieve all data in the period from starttime to endtime of data lost PM2.5; 3 DataC: A list coordinates of data lost PM2.5; 4 for each coordinate in DataC do 5 - initialization array(PM2.5) containing values of 6
PM2.5; - initialization array(coordinate) to store coordinate; while radius less than or equal hundred do for each coordinate in DataB do set d is distance coordinate in DataC and DataB; if d less than radius and coordinate not in array(coordinate) then - add value PM2.5 of coordinate B into array(PM2.5); - add coordinate into array(coordinate); else
do nothing if radius greater than twenty and number of element in array(PM2.5) greater than zero then calculator output for PM2.5; - get maximun value in array(PM2.5); - get average all values in array(PM2.5); - get minimun value in array(PM2.5); break loop on DataB and then break for radius loop, go to next coordinate in DataC; else
do nothing if radius equal hundred and number of element in array(PM2.5) equal zero then
set output value of PM2.5 is zero; else
do nothing [3] Peijiang Zhao and Koji Zettsu. 2018. Convolution Recurrent Neural Networks for Short-Term Prediction of Atmospheric Sensing Data. In 2018 IEEE GreenCom-CPSCom-SmartData). IEEE, 815–821.