Traditional Japanese wooden houses are sustainable and vernacular in nature; however, energy consumption owing to poor insulation is inevitable in the winter. These wooden houses in Japan are an integral part of historical landscapes and are protected from the potential risks of retrofitting. To continue using these houses without bearing additional energy burdens, it is necessary to better understand their actual thermal performance. Therefore, this study proposes a combined approach to assess the thermal performance of traditional Japanese-style rooms using in-situ thermal measurements and machine learning predictive methods. The in-situ thermal measurements evaluated the actual thermal environment of traditional Japanese-style rooms with and without air conditioners (AC) during winter. Similarly, two machine learning models, support vector regression (SVR) and random forest regression (RFR) were developed to predict the indoor temperature based on the external factors and their adjacent rooms temperature. The result revealed strong thermal interaction between adjacent rooms, likely driven by heat transfer dynamic. To further investigate these interactions, thermal images of each room were analyzed after the AC was in use. The result revealed the presence of hot and cold spot within the gap of sliding fusuma (thick paper made door) and ranma (transom or indoor ventilation), indicating significant heat transfer between the AC and non-AC rooms. Consequently, the prediction model for the AC room temperature was notably influenced by the thermal conditions of adjacent non-AC room. Thus, this study proposes a simple yet reliable approach that enables resident to prioritize thermal improvements for efficient and cost-effective renovations.