The article presents the implementation of the «supervised learning» approach for identifying heavy-oil-bearing reservoirs. At the first stage, the analysis was carried out concerning the current state of the problem of information support for prospecting and exploration operations and rating of wells with the help of the Russian patent search system Exactus System and the American Free Patents Online system. It was established that the classical statistical analysis integrated with machine learning methods is the powerful tool for interpretation of the chemical variable rock structure using fairly-well representative sample. Input data reflect more than 40 physicochemical features of rock and fluid, and they were collected from 2012 to 2016. Comparative analysis of the features values by basic statistical characteristics is given, category, ordinal, discrete and continuous features are highlighted. The need is defined related to elimination of blowout control for six features, scaling, the transformation of two categorical features into discrete ones was performed. Fixed six features with a predominance of missing values were eliminated. The depth of location of the productive strata, which is more than 4 thousand meters, was established. Features of speed and duration of prospecting and exploration operations, the years of their beginning and end were synthesized. It was revealed that the most productive in terms of the number of detections of heavy-oil-bearing reservoirs was the year 2015. To improve the classification efficiency, two of the three classes of the target feature are summarized. A priori distribution of the target feature by classes was estimated (getting into the «other» class is 8 times more likely than getting into the «heavy-oil-bearing reservoir» class). Dimensions of the problem is reduced by the method of principal components, since the percentage of dispersion explained by the first two principal components is greater than 70. The problem of quadratic optimization with a soft gap is formulated. Classification is performed by the support vector machine method with a Gaussian Radial Basis Function and the regularization parameter equal to one. Classification accuracy is 93 %. The approach can be recommended for improving the estimate accuracy of the deposits rate at the initial development stage.