The　transient　temperature　rise　in　the　active　region　in　AlGaN/GaN　high　electron　mobility　transistors　(HEMTs)　is　measured　using　an　electrical　method.　The　original　data　are　smoothed　and　denoised　by　a　nonparametric　fitting　algorithm,　called　locally　weighted　scatterplot　smoothing　(LOWESS).　The　thermal　time-constant　spectrum　is　extracted　to　analyze　the　physical　structure　of　the　heat-conduction　path　in　AlGaN/GaN　HEMTs.　The　thermal　time-constant　spectra　extracted　using　the　LOWESS　algorithm　are　richer　and　the　RC　network　obtained　is　greater　compared　with　those　with　the　traditional　denoising　method　(multi-exponential　fitting).　Thus,　the　analysis　of　the　heat-flow　path　is　more　precise.　The　results　show　that　the　LOWESS　nonparametric　fitting　algorithm　can　remove　noise　from　measured　data　better　than　other　methods　and　can　retain　the　subtle　variation　tendency　of　the　original　discrete　data.　The　thermal　time-constant　spectra　extracted　using　this　method　can　describe　the　subtle　temperature　variations　in　the　AlGaN/GaN　HEMT　active　region.　This　will　help　researchers　to　precisely　analyze　the　layer　composition　of　the　heat-flow　path.　©　2015　Chinese　Institute　of　Electronics.