Analysing Twitter and web queries for flu trend prediction

Table 4 Linear regression results.

		*"cold"* negative	*"cold"* positive
flu reg exp		0.837
expanded reg exp		0.801
Queries		0.571
Queries + flu reg exp		0.849
Queries + expanded reg exp		0.885
Naïve Bayes	Counts Probabilities	0.761 0.804	0.825 0.824
Naïve Bayes + Queries	Counts Probabilities	0.781 0.794	0.872 0.886
Naïve Bayes + expanded reg exp	Counts Probabilities	0.770 0.807	0.703 0.791
Naïve Bayes + expanded reg exp + Queries	Counts Probabilities	0.873 0.867	0.836 0.801

Pearson's correlation ratios between linear regression estimates and Influenzanet data. flu and expanded regular expressions correspond to the pattern for "gripe" (flu) word derivations and the complete pattern as shown in Table 1, respectively. The weekly relative frequency was calculated based on the number of positively classified tweets (counts) or on the probabilities given by the classifier (Eq. 2). Tweets referring to "cold " were used either as positive or negative data when training the classifier.

ISSN: 1742-4682