tangent surpris recommend algorithm kensuk onuma hanghang tong christo faloutso recommend system find item relev older choic user focu surpris queri user bore usual genr item book movi hobbi recommend relat beaten path possibl lead genr book movi hobbi defin autom seemingli selfcontradict request introduc tangent recommend algorithm solv problem main idea tangent envis problem node select graph give high score node connect older choic time connect unrel choic method carefulli design paramet free effect fast illustr benefit tangent experi synthet real data set show tangent make reason surpris horizon broaden recommend fast scalabl easili exist fast algorithm graph node proxim 