Decision　support　systems　imposed　a　challenge　for　efficiently　reasoning　through　utilizing　stored　unordered　rules.　The　Extended　Belief　Rule-Base　(EBRB)　system,　renowned　for　its　capability　of　modeling　data　with　vagueness,　incompleteness,　uncertainty　and　nonlinear　features,　requires　to　traverse　over　all　the　rules　for　the　execution　of　reasoning　and　consequently　suffers　reduced　inference　accuracy　and　efficiency.　To　improve　the　performance,　we　propose　a　framework　of　storing　and　retrieving　belief　rules　based　on　employing　an　optimized　structure　called　improved　Ball　tree.　The　framework　first　constructs　a　Ball　tree　index　according　to　the　distance　between　different　rules　in　the　metric　space,　via　using　the　k-means++　algorithm　rather　than　traditionally　employed　k-centers.　Then,　through　our　proposed　dynamic　threshold　radius　adjusting　method,　our　algorithm　finds　an　appropriate　dataset　threshold　and　accordingly　activates　more　related　rules.　Moreover,　the　number　of　retrieved　and　activated　irrelevant　rules　is　significantly　reduced,　resulting　in　consequently　improved　reasoning　accuracy　and　efficiency.　Lastly,　three　sets　of　experiments　were　carried　out　to　validate　our　algorithm　in　comparison　with　the　previous　EBRB　systems.　©　2020　Elsevier　B.V.