From confirmation bias to echo-chambers: a data-driven approach