Rainfall-induced landslide susceptibility assessment is currently considered an effective tool for landslide hazard assessment as well as for appropriate warning and forecasting. As part of the assessment procedure, a credible index weight matrix can strongly increase the rationality of the assessment result. This study proposed a novel weight-determining method by using random forests (RFs) to find a suitable weight. Random forest weights (RFWs) and eight indexes were used to construct an assessment model of the Dongjiang River basin based on fuzzy comprehensive evaluation. The results show that RF identified the elevation (EL) and slope angle (SL) as the two most important indexes, and soil erodibility factor (SEF) and shear resistance capacity (SRC) as the two least important indexes. The assessment accuracy of RFW can be as high as 79.71%, which is higher than the entropy weight (EW) of 63.77%. Two experiments were conducted by respectively removing the most dominant and the weakest indexes to examine the rationality and feasibility of RFW; both precision validation and contrastive analysis indicated the assessment results of RFW to be reasonable and satisfactory. The initial application of RF for weight determination shows significant potential and the use of RFW is therefore recommended.