新世纪的人才需求变化推动了教育的改革发展，在2002年美国联邦教育部主持制订的“21世纪技能”框架中，复杂问题解决（CPS）是提倡重点培养的重要能力之一。包含相互冲突的多个目标是CPS测量系统的重要特征。自CPS领域开创以来，测量方法的不完善局限了对其所包含的认知过程的探讨，多目标平衡能力作为重要的CPS维度之一尚未得到充分的验证。2012年，大型国际比较测验“国际学生评估项目(PISA)”对学生的CPS能力进行了大规模的施测，并生成了记录学生作答过程的log数据文件，使得对解题过程中所反映的多目标平衡能力进行分析成为可能。本研究通过对该数据库中，一道包含三个相冲突目标的题目“车票”的log数据进行分析，对CPS中的多目标平衡能力进行探讨。 本研究将log数据中记录的单个行为操作合并并编码为不同的问题解决方案，研究一对每个方案是否追求各个目标进行识别和标注，以追求各目标的方案在所有方案中所占比重作为解题者对各目标的追求程度。以各目标追求程度的交互作用来代表多目标追求过程，采用logistic回归分析来探讨多目标追求对题目成绩的预测，采用线性回归分析来探讨多目标追求对CPS能力水平的预测，采用k-means聚类分析来探讨不同的目标追求优先级与CPS表现的关系。在此基础上，研究二进一步将连续的方案对编码为不同的多目标平衡行为：正确平衡、目标丢失、评估错误和成本过高。其中，正确平衡代表正确的多目标平衡行为，后三者分别代表明确目标子能力不足、评估目标子能力不足和目标最优化子能力不足所导致的错误的多目标平衡行为。分别采用logistic回归分析、线性回归分析和k-means聚类分析，对解题者的多目标平衡行为与CPS表现的关系进行探讨，研究多目标平衡中各子能力的影响作用。 研究一结果表明，多目标追求能够十分显著地预测题目成绩（β= -0.470~0.728, p < 0.05 ~ p < 0.001），及CPS能力水平（β= 0.110, p < 0.001；β=0.042, p < 0.05）。解题者的目标优先级与任务目标的重要性吻合程度较高的组有更好的CPS表现。研究二结果表明，评估错误和目标丢失能够非常显著地负向预测题目得分（β= -1.005, p < 0.001; β= -0.464, p < 0.001），正确平衡能够非常显著地正向预测题目得分（β= 0.497, p < 0.001）。成本过高能够较为显著地正向预测题目得分（β= 0.180, p = 0.05）。评估错误和正确平衡能够显著地负向预测CPS能力水平（β= -0.110, p < 0.001; β= -0.048, p < 0.05）。可以根据多目标平衡的三个子能力的差异区分不同的解题者。 本研究对CPS的测量理论和实践都具有丰富的意义。在理论上，验证了多目标平衡中评估目标优先级在CPS过程中的重要作用。从实践上，验证了log数据分析对于揭示CPS中包含的认知过程的重要作用，说明了大规模教育测验的数据挖掘潜力。

The change in the demand for talent in the new century has promoted the reform and development of education. In the “21st Century Skills” framework developed by the U.S. Federal Ministry of Education in 2002, Complex Problem Solving (CPS) is one of the important competencies advocated for promoting training. Since the creation of the CPS field, the inadequateness of the measurement method has limited the discussion of the cognitive processes CPS contains. Among them, the multi-goal balancing process mentioned in the CPS cognitive theory has not been fully verified. In 2012, The Program for International Student Assessment (PISA) conducted a large-scale test of students' CPS competencies and generated substantive log data files to record student responses. This study analyzed the log data of a multi-goal item in the database and discusses the multi-goal balancing in CPS. In study 1, the single operation recorded by the log data is merged and coded as problem solutions, and through the analysis of the behavior of students' efforts to achieve each goal, the effort of each goal in each solution is marked respectively, and the ratio of the efforts of each goal in all the solutions represents the degree of goal efforts. Using the interaction of each goal efforts to represent multi-goal pursuit, the logistic regression analysis was used to predict the item accuracy by the multi-goal pursuit process. The linear regression analysis was used to predict the proficiencies of the PISA by the multi-goal pursuit process. The K-means clustering analysis was used to explore the effect of the results of different goal priorities on the performance of the CPS. In study 2, we further encode continuous two solutions into different goal balancing behaviors: correct trade-offs, inadequate goals, erroneous trade-offs and over-cost. Among them, the former represents a correct multi-goal balancing behavior, and the latter three represent the wrong multi-goal balancing behavior resulting from the lack of goal clearity, insufficient ability to evaluate goal, and insufficient ability to optimize. Using the students with any multi-goal balancing behavior as the sample, logistic regression analysis and linear regression analysis were used to predict the item accuracy and the PISA proficiencies by multi-goal balancing. The k-means cluster analysis was used to explore the different types of multi-goal balancing behaviors and to study the influence of each process in the multi-goal balancing. The results of study 1 show that the ability of multi-goal pursuit could significantly predict the subject achievement(β= -0.470~0.728, p < 0.05 ~ p < 0.001), and the proficiency of CPS (β= 0.110, p < 0.001; β=0.042, p < 0.05). Students with a higher degree of match between goal priorities and task goals had better CPS performance. The results of study 2 showed that the erroneous trade-offs and insufficiency of goals could significantly negatively predict the item accuracy (β= -1.005, p < 0.001; β= -0.464, p < 0.001). The correct trade-off could be very significantly positive predictive the item accuracy (β= 0.497, p < 0.001). The over-cost could significantly negatively predict the item accuracy (β= 0.180, p = 0.05). Erroneous trade-offs and correct trade-offs could significantly negatively predict CPS proficiency levels (β= -0.110, p < 0.001; β= -0.048, p < 0.05). Different students could be distinguished based on the differences in the three sub-processes of multi-goal balancing. The above results are of great significance to the measurement theory and practice of CPS. Theoretically, it validated the important role of multi-goal balancing in the CPS process, and illustrated that multi-goal balancing is the comprehensive embodiment of three sub-processes: identifying and maintaining multiple goals, evaluating goal priorities, and goal optimization, in which the evaluation of the goal's priority is the most important sub-process. In practice, the important role of log data analysis in revealing the cognitive processes involved in CPS was verified, and the potential of data mining in large-scale educational tests was explored.