•We propose the data-driven linguistic multi-attribute decision making.•We develop a data-driven method to learn personalized individual semantics.•We present a case study based on two real datasets.•We make a comparison with existing methods to justify the proposed model. In parallel with the development of information and network technology, large amounts of data are being generated by the Internet, and data-driven methodologies are now often being used in decision-making. Recent studies have investigated personalized individual semantics (PIS) in various decision-making contexts to model a fact that words mean different things to different people. However, few studies have investigated PIS in the context of multi-attribute decision-making (MADM). In MADM, in addition to multi-attribute linguistic information, pre-existing classification of the alternatives is always present, which have not been considered in prior research. Most previous studies have simply demonstrated the feasibility of PIS methods with numerical examples using small-scale models, and not with realistic datasets. Therefore, in this study, we propose a data-driven learning model to analyze the PIS of decision makers to support a multi-attribute decision-making model that considers pre-existing classification of the alternatives. Specifically, we first propose a PIS multi-attribute learning function to define a general computation form for comprehensive evaluation of the value of alternatives. Then, considering this pre-existing classification of the alternatives, a PIS learning model is constructed by analyzing the relations between calculated values of alternatives and corresponding class assignments to obtain personalized numerical scales of linguistic terms for a decision maker. Finally, we present a case study based on two datasets and a comparison with other methods to justify the feasibility of the proposed model.