Fig 1.
The 167 townships of Guangzhou City.
Fig 2.
Weekly dengue case count of Guangzhou from January 2015 to September 2019.
July 1 to November 30 was determined as the annual outbreak period of Guangzhou in this study.
Fig 3.
Meteorological data in Guangzhou City.
(A) Weekly mean temperature of an arbitrarily selected township from January 2015 to September 2019. (B) Weekly cumulative rainfall of an arbitrarily selected township from January 2015 to September 2019. (C) Weekly mean temperature and (D) weekly cumulative rainfall of all townships within the city during the week of September 12–18, 2016.
Table 1.
Records of one cellphone user in the mobile phone data.
Fig 4.
The framework of the Node2Vec model.
Based on a certain sampling strategy, many random walks can be generated on the graph. Treating the nodes as “words” and the random walks as “sentences,” the embeddings of the nodes can be learned by feeding these “sentences” into the Word2Vec model.
Fig 5.
Sampling strategy of random walks in the Node2Vec model.
The walk just transitioned from node <t> to node <v> and is now evaluating its next step out of node <v>. The transition probability from <v> to any one of its neighbors (i.e., <t>, <x>, <y>, and <z>) depends on the search bias a and the edge weight between them (thicker lines indicate larger edge weights).
Fig 6.
Framework of intra-urban dengue forecasting approach.
Common features were extracted from the dengue case data, meteorological data, and population data, while the interaction features were learned from the mobile phone data. The interaction features were combined with the common features to enhance the models (i.e., SVM, LASSO, and ANN) for L-week ahead dengue forecasting.
Table 2.
Common features extracted for each township.
Fig 7.
Exponential smoothing (αs = 0.25) applied to the time series of weekly dengue case count of two townships.
Fig 8.
Predicted smoothed case counts and observed case counts of three randomly selected townships in Guangzhou during the validation period.
The smoothed case counts were predicted by the 1-week ahead SVM-based model using both common features and interaction features.
Fig 9.
Predicted smoothed case counts of all townships during two different weeks.
The smoothed case counts were predicted by the 1-week ahead SVM-based model using both common features and interaction features.
Fig 10.
Performance comparison of models with and without interaction features based on the Pearson correlation coefficient of predicted and actual smoothed case counts.
(A) SVM-based forecasting. (B) LASSO-based forecasting. (C) ANN-based forecasting.
Fig 11.
Performance comparison of the SVM-based models with and without interaction features based on the hit rate metric.
Fig 12.
Performance comparison of the LASSO-based models with and without interaction features based on the hit rate metric.
Fig 13.
Performance comparison of the ANN-based models with and without interaction features based on the hit rate metric.