Decoders for predicting author age, gender, location from short texts

Disclosed are systems and methods for predicting demographics about authors of social media. A dataset relates features from social media messages to known demographic information about authors of the social media messages. The social media messages and known demographics may be retrieved from different social media services. Adjustable weights are associated with alternative demographic values for the features. The dataset is used to predict unknown demographics of users based on the adjustable weights associated with the alternative demographic values for extracted features from social media messages. A confidence level may be output for a predicted demographic that corresponds to a difference between a sum of weights associated with the predicted value and another demographic value. The predicted demographic value may be output when the confidence level exceeds a threshold. The weights may be updated when the confidence level is below the threshold even when the prediction is correct.

View patent on

Patent #: 9,471,944 Issue Date: October 18, 2016