To use salmon protein hydrolysates as food ingredients and to mask the fish odor, Maillard reactions were associated with enzymatic production of hydrolysates. The study explored an original approach based on regression trees (RT) and random forest (RF) methodologies to predict hydrolysate odor profiles from volatile compounds. An experimental design with four factors: enzyme/substrate ratio, quantity of xylose, hydrolysis and cooking times was used to create a range of enzymatic hydrolysates. Twenty samples were submitted to a trained panel for sensory descriptions of odor. Hydrolysate volatile compounds were extracted by means of Headspace Solid Phase MicroExtraction (HS-SPME) and analyzed using gas chromatography/mass spectrometry (GC-MS). The results showed that RT and RF methodologies can be useful tools for predicting an entire sensory profile from volatile compounds. Four main volatile compounds made it possible to separate hydrolysates into five groups according to their specific sensory profile. 2,5-dimethylpyrazine, 1-hydroxy-2-propanone and 3-hydroxy-2-pentanone were identified as the main predictors of the roasted odor, whereas methanethiol was associated with a mud odor. These results also suggest the appropriate process conditions for obtaining a typical roasted odor.