Analyzing the Quality of Twitter Data Streams.

There is a general belief that the quality of Twitter data streams is generally low and unpredictable, making, in some way, unreliable to take decisions based on such data. The work presented here addresses this problem from a Data Quality (DQ) perspective, adapting the traditional methods used in relational databases, based on quality dimensions and metrics, to capture the characteristics of Twitter data streams in particular, and of Big Data in a more general sense. Therefore, as a first contribution, this paper re-defines the classic DQ dimensions and metrics for the scenario under study. Second, the paper introduces a software tool that allows capturing Twitter data streams in real time, computing their DQ and displaying the results through a wide variety of graphics. As a third contribution of this paper, using the aforementioned machinery, a thorough analysis of the DQ of Twitter streams is performed, based on four dimensions: Readability, Completeness, Usefulness, and Trustworthiness. These dimensions are studied for several different cases, namely unfiltered data streams, data streams filtered using a collection of keywords, and classifying tweets referring to different topics, studying the DQ for each topic. Further, although it is well known that the number of geolocalized tweets is very low, the paper studies the DQ of tweets with respect to the place from where they are posted. Last but not least, the tool allows changing the weights of each quality dimension considered in the computation of the overall data quality of a tweet. This allows defining weights that fit different analysis contexts and/or different user profiles. Interestingly, this study reveals that the quality of Twitter streams is higher than what would have been expected.