Spatial data mining (SDM) refers to extracting and “mining” the hidden, implicit, valid, novel and interesting spatial or non-spatial patterns or rules from a large amount of, incomplete, noisy, fuzzy, random, and practical spatial databases. At present, SDM mainly concentrates on the efficiency of mining algorithm itself. Another important issue-application in SDMhas not been paid much attention to. Geological data are typical spatial data, which include geological, geophysical, geochemical, and remote sensing data. Regional metallogenic prediction and quantitative resource assessment have been recognized as important when integrating multisource geological spatial data in recent years. The statistical and mathematical approaches that have developed recently for multi-resources geological spatial data integration include weights-of -evidence modelling, and the logistic regression modeling. The fuzzy logic model, artificial neural networks model and the fractal method have been applied in the prediction and assessment of mineral resources potential. These methods promote the effectiveness of mineral resource prospecting. However, under the shortage of mineral resources and the leap cost of prospecting and exploring, how to make full use of the existing multi-sources geo-spatial data to find deep mining information through massive uncertain multi-source geology spatial database and establish a rapid, efficient, intelligent method of mineral prediction to reduce the cost of mineral exploration and improve the efficiency and accuracy of mineral prediction is particularly important and urgent. These traditional statistical mathematical approaches and GIS spatial analysis technology is lack of intelligence and reasoning mechanism. In this paper, a geoglogy spatial data mining method for regional metallogenic prediction is proposed. The geology spatial data mining method for regional metallogenic prediction mainly includes continuous geological spatial data discretization, spatial relationship abstracting and attribute transforming, mining metallogenic association rules and quality assessment, comprehensive evaluation of association rules and mineral potential mapping. Firstly, the best ore-controlling variable and threshold were determined using proximity analysis of the weights-of-evidence model. On the basis of a correlation analysis among evidence variables, the authors selected vector ore-controlling data of stratum, unconformity, fault, regional geochemical data, remote-sensing mineralization information, bouguer gravity data, aeromagnetic data, and mineral occurrence for this experiment. Secondly, the cloud model was used to execute the uncertainty translation between continuous geology data and discrete geology data. Meanwhile, on the basis of apriori algorithm, the metallogenic association rules were abstracted and six uncertainty indexesconfidence, support, lift, coverage, leverage and interesting-were adopted to assess the quality of metallogenic association rules. Finally, the experiment of regional metallogenic prediction for iron deposits was performed in Eastern Kunlun, Qinghai province, China, using geoglogy spatial data mining method and weights-of-evidence model, respectively. All the data sets were derived from a multi-source geology spatial database (MGSDB) established by the Institute of Geo-Spatial Information Science and Technology, University of Electronic Science and Technology of China in 2009, which contains geological, geophysical, geochemical and remote sensing data. The algorithm is implemented with C# programming language. The study areas were divided into three parts to assist the interpreting of the potential distribution of iron deposits. We defined the three target zones as high potential areas, moderate potential areas and low potential areas, respectively. Using the weights-of-evidence model, low potential areas contain 32% of the total gold deposits, covering 75% of the total area; the high potential areas contain 48% of the total gold deposits, covering 10% of the total area; and the moderate potential areas contain 20% of the gold deposits, covering 15% of the total area; about 68% of known deposits are located in the moderate and high potential areas. Using the geology spatial data mining method in this paper, Low potential areas contain 8% of the total gold deposits, covering 75% of the total area; the high potential areas contain 72% of the total gold deposits, covering 10% of the total area; and the moderate potential areas contain 20% of the gold deposits, covering 15% of the total area. About 92% of known deposits are located in the moderate and high potential areas. The results indicate that the prediction accuracy of geology spatial data mining was obvious higher than that of weights-of-evidence model’s, the method is suitable for mineral resources potential assessment and its effectiveness is good.