This paper presents an overview of our proposed methods for the two subtasks of MediaEval 2019 “Insights for Wellbeing”. The goal is to investigate the factors related to personal environmental health conditions (PEH). We model this as a regression problem where environmental factors measured by both physical sensors and psychological perception ratings are used as indicators for predicting the target values of objective or subjective PEH measures. A variety of models (e.g., GBDT and LSTM) have been employed to conduct the regressions. The experimental results indicate that objective PEH is mainly determined by the factors indicated by physical sensors and the temperature and humidity contribute the most. Subjective PEH is dominated by the perceptual factors collected from questionnaires and indicated by urban natures mined from a GIS dataset.
Plenty of studies have been carried out for investigating the
relationship between the environmental conditions and the human
wellbeing [
With SEPHLA, this paper presents our methods proposed for the two subtasks of MediaEval 2019 “Insights for Wellbeing”. The goal is to investigate the main factors related to personal environmental health conditions (PEH). In this study, PEH has been measured objectively by the PM 2.5 and subjectively by P-AQI (personal air quality index which is determined by a human rating system). We model this as a regression problem in which PEH measures are the targets while factors are the indicators. We have investigated a wide of factors including those measured by physical sensors (e.g., temperature, humidity, heart rate), those collected from questionnaires for measuring people’s psychological perceptions to the PEH, and urban natures extracted from OpenStreetMap1 (a third-party GIS dataset). A variety of models (e.g., GBDT and LSTM) have been employed to fulfill the regressions. 2 2.1
Since PEH focuses on a road-level area, the factors from city-level sensors are discarded, including N O2, O3 and N O. The remaining factors are collected from wearable devices, such as PM2.5, temperature, humidity as well as heart rate. For PM2.5, a calibration strategy, which replace those values shifting from mean with more than three times of variation with mean of its surrounding 30 values, is adopted to filter outliers. In total, there are 43,684 and 24,055 samples in training and testing set respectively. 2.2
In SEPHLA, the participants should answer 5 questions at each
checkpoint according to the data collection instruction. The
participants need to provide their subjective perception about the segment,
including quietness, calmness, fun, easy of walking and
crowdedness. To assign questionnaires to their corresponding checkpoints,
we use k-means to group them into clusters based on time
distribution and the number of clusters is equal to the number of
checkpoints. Finally, we have 307 and 197 valid questionnaires for
training and testing respectively and each questionnaire is
represented by a 5-dimensional vector. The label (i.e. P-AQI) of
questionnaires are directly derived from that of corresponding segments in
a multiple instance learning strategy [
Since questionnaire is inaccurate because of individual diference,
we make use of OSM to have a relatively accurate description about
the surrounding environment. In OpenStreetMap, each location is
described by a JSON file, which contains the description of
surrounding buildings, roads and scenery. We sample locations along the
routine by sliding window with stride as 20 meters and a circle with
radius as 25 meters is drawn around the location. Then we collect
all the urban nature description inside this circle and all the name
entities are extracted with Stanford CoreNLP [
We have submitted five runs for each task to explore the impact of diferent factors on PEH. 3.1
The five runs are listed as following:
• MEAN: The mean value of PM2.5 from the other
participants in the routine is calculated and the hidden values in
the replaced segment are filled with it. This is the baseline
run based on the assumption that the PM2.5 should not
largely fluctuate within a small region.
• LOCA: The hidden values is replaced with the PM2.5 value
of the nearest point from other participants. Note that the
map distance between two points is adopted rather than l2
distance. This run also bases on the same assumption of
MEAN but further narrow down the region area.
• GBDT: LightGBM [
• QUEST: As described in Section 2.2, the questionnaires are represented by a five dimensional vector with the same label as corresponding segment. • WEAR: The physical sensor factors, including temperature, humidity and PM2.5 are assigned with the same label as corresponding segment. • OSM: As described in Section 2.3, the locations inside a segment with urban nature are used as training samples. • WOMER: The results of WEAR and OSM are further fused by average voting. • QOMER: The results of QUEST and OSM are further fused by average voting.
The performance is listed in Table 3. WEAR is the worst run, which means that physical sensors actually basically provide no clues for P-AQI prediction. Meanwhile, QUEST and OSM achieve better performance. Questionnaire is easy to be afected by individual difference and OSM provides a better description of the environment. Hence, OSM achieves the best performance as single feature. In addition, by fusing diferent features, we further find that questionnaire provides complementary information to OSM and achieve the best performance among all the five runs. 4
While the findings in this study are interesting, however, they are not conclusive, in the way that they are made with a biased set of factors with a limited number of sensors and a limited coverage of participants. In the future, We will extend the study by including more sensors such as ECG, EEG, and blood sugar, and a larger coverage of participants across age, gender, occupation, and education.
This work was supported by the National Natural Science Foundation of China under Project 61906108.