In this paper, we present our AISIA team's contribution to the task Insight for Wellbeing: Multimodal personal health lifelog data analysis at MediaEval 2020. From the data sets provided, we extracted diferent types of useful attributes for the problem: the timestamp information, the geographical data, sensor data, and the semantic features from images captured by users. We proposed an approach, namely A2QI, by applying machine learning models for estimating the local AQI score and level, including Support Vector Machine and Random Forest. We evaluated the experimental data sets using Randomized Search and K-Fold cross-validation. The test sets' evaluation shows that employing a machine learning approach with appropriate features can significantly improve accuracy.
In many countries worldwide, the prediction of air pollution is an
increasingly undeniably significant problem. It can impact
individuals and their wellbeing. In this study, we aim to use a machine
learning approach using insights from the lifelog data provided by
the organizer to predict the personal air pollution data as well as
the individual air quality data, as given in the task description [
Observing the three columns of 2.5, 2, and 3 in the training dataset for both tasks, one can see that many data points have zero value, are negative numbers, or are unreasonably large (e.g., −3000, −4900, etc.). Also, one can find a similar observation even in positive-valued data points. They are called anomalies or outliers, which have to be preprocessed before extracting features.
Now, let us consider an arbitrary column whose data needs to
have a preprocessing step. One can determine these outliers in two
cases: the first one includes zero and negative signed values, the
other includes positive outliers (which will be defined later). For
the positive outliers, we apply z-score method [
In this work, a data point whose z-score is larger than 3.0 is called an outlier. It is worth noticing that the mean value is computing based on the positive values only, intending to avoid the influence of negative valued data points whose absolute values are large.
After detecting all the anomalies, we replace them with the average of positive values via the reason mentioned above. 2.2
The problem consists of two subtasks. Each task asks for using a diferent data set. Nevertheless, they both include information about time, location, weather, and concentration values of contaminants related to AQI (e.g., 2 or 3). Therefore, our proposed feature extraction techniques in these data types can be applied to both data sets. Also, we calculated the necessary features from the image data given in the second task.
2.2.1 Timestamp features. From the given information about time, we extract timestamp features. Specifically, we survey the correlation between the time point that the data are collected and the corresponding AQI values and ranks that need to be predicted. These features include part of day () and is rush hour ( ).
To begin with, we deduce the feature. That is, we split a day’s 24-hour time into five groups, The “Early Morning" group is for the time from 5 AM to before 7 AM, the time from 7 AM to before noon is considered the “Morning" group, between noon and before 4 PM is “Afternoon" group, between 4 PM and before 8 PM is “Evening", and the remaining period between 8 PM and 5 AM is the “Night" group. From our observation, there is a noticeable increase in trafic density during the time of Morning and Evening groups, which leads to a high level of pollution caused by smoke from these means of transportation. Consequently, we expect there is a fluctuation in the data collected during these periods.
Also, we check whether a particular local measured time is a rush hour or not, which leads to extracting the second feature in the group of timestamps features, i.e., is rush hour ( ). In
SMAPE 0.32 0.52 0.32 0.52 detail, if that given point of time falls into one of these periods (7:00 AM to 9:00 AM) and (4:00 PM to 7:00 PM), it is called a rush hour. This feature is a development of the former (i.e., ). We will survey the periods when the trafic density reaches the highest peak, resulting in sharp growth of AQI values and ranks.
2.2.2 Location features. When surveying the factors afecting the level of pollution of a location, we consider the distance between that location and the nearest railway station, which is usually crowded with people and transports. Using the information about coordinates of a place, we extract the feature about the distance from that place to the chosen station. In this study, we use the Shibuya station (35◦ 39′N, 139◦ 42′E).
To compute the mentioned distance, we use the Haversine formula[
2.2.3 Semantic features. In the second task, we are provided the
data of images captured in diferent locations, which is the most
challenging data type in our opinion. Our approach is to investigate
if the number of cars, motorbikes, and the contrast of the images
can impact the level of pollution in that captured location. We used
SSD ResNet 50 (Retina Net 50)[
Also, we extract features related to the contrast of the images,
which can be highly correlated to the intensity of a given place’s
pollution. In detail, given a two-dimensional image of size ×
, we use RMS contrast formula[
Finally, it is worth noticing that in this study, we did not use the number of people as a feature related to image data, as the people appearing in the given images have been blurred for the sake of privacy. 3
After extracting the necessary information, we evaluated two
machine learning models using a Randomized Search with a 5-fold
cross-validation technique to optimize the model hyper-parameters
and avoid overfitting our training data. The two models we used
were Support Vector Machine (SVM) [
The results on test sets are presented in Table 1. In the first subtask, we can see that using Random Forest can achieve the best result in general with data collected by a walker. For predicting the AQI value, the results of MAE, RMSE, and SMAPE, in this case, are 12.74, 15.93, and 0.32, respectively. In the second task, the best result can be achieved by using SVM. For predicting PM2.5, the best performance in MAE, RMSE, and SMAPE are 3.49,3.76 and 0.15, respectively.
Also, if one can enhance the quality of the images captured in the data set and combine it with public weather data, the training results can be improved significantly. 4
As the authors, we would like to thank AISIA Research Lab to support our team and allow us to use their computational resources for this study. Also, we would like to give our thanks to the Organization Board of MediaEval 2020 competition and Task Organizer for providing us with data sets to conduct necessary experiments. Insight for Wellbeing: Multimodal personal health lifelog data analysis