To thorough analyse the sources of heavy metal pollution in typical farmland soil around city, the soil of typical farmland in western Tianjin was taken as the object, and the geo-accumulation index (Igeo) was used to evaluate and analyse the spatial distribution of heavy metal pollution. The machine learning algorithm-support vector machine (SVM) was used to conduct source apportionment of heavy metals. The SVM model was combined with positive matrix factorization (PMF) receptor model to identify the characteristic elements based on field investigation, and heavy metal traceability analysis was carried out. The results showed that the maximum values of the seven heavy metal contents (except for Cd) did not exceed the risk screening values of the “Soil environmental quality-risk control standard for soil contamination of development land” (GB15618-2018). The average values of Hg and Cd were 4.95 times and 2.55 times comparing with the background values, which were the main risk elements in the study area. The SVM algorithm has good fitting coefficients R2 above 0.779 for all eight heavy metals. The source apportionment results obtained by PMF model combined with SVM algorithm are four main pollution sources, accounting for 10.3% and 34.4% (results of PMF and SVM, the same below) of coal source, 45.1% and 39.4% of pesticide and fertilizer source, 7.5% and 15.0% of mixed agricultural and transportation source, and 37.2% and 11.1% of natural source, respectively. The overall variable fitting coefficient R2 of both models is above 0.9, which can effectively predict heavy metal content. Based on the investigation of industrial enterprises, road conditions, and residential areas in the comprehensive research area, the main source of pollution should be pesticides and fertilizers, and the proportion of PMF model source apportionment contribution rate is more in line with the actual situation on site.