近十年来，协作学习在教育研究中越来越受关注并且在实践中广泛应用。交互是协作学习的基本活动单元。传统的研究大多从言语行为的视角对交互过程进行分析，这种分析方法与交互效果难以建立直接的联系。本研究把协作学习系统看成是信息系统，关注这个系统中的信息流，主要利用基于信息流的方法分析面对面协作学习的交互过程。本研究的主要内容包括如下四部分： 第一，制定基于信息流的面对面协作学习交互分析方法的操作规范，明确其操作步骤,包括绘制初始知识网络图、切分信息流、计算信息流所映射的知识网络图的属性并分析这些属性与交互效果的关系。同时开发专门的分析工具进行交互分析。另外，采择了能够预测协作学习交互效果的若干指标并设计其算法，包括目标知识网络图的总激活量、聚焦水平、深度、广度、度分布指数、网络中心度、网络结构熵、密度、度平均值、平均路径长度、知识点分布熵、激活生成树的带权路径、主题增长速度、激活量标准差、平均激活量等。 第二，采集了121个面对面协作学习交互的样本对基于信息流的协作学习交互分析方法进行实证研究。结果表明目标知识网络图的总激活量、聚焦水平、深度、度分布指数、网络结构熵分别能够预测交互效果。这几个属性中预测力最高的是“总激活量”、“聚焦水平”，分别能够解释交互效果的23%和24.8%的变异。另外，“聚焦水平”、“度分布指数”、“深度”三种属性联合预测，能够解释交互效果的36.9%。总激活量和交互效果之间的非线性关系更加拟合贝塔朗菲增长曲线模型。 第三，从时间和空间角度深入分析面对面协作学习交互过程中信息流的特征。从时间角度分析发现聚焦点随时间呈现回溯、转移的特征，知识建构的过程随时间呈现非线性特征。就空间特征而言，从小组层面分析可知不同交互效果的小组所形成的知识网络图存在差异性，交互效果好的小组和较差的小组在知识误建构的比例方面存在显著性差异，并且知识误建构的比例与交互效果呈负相关。交互效果好的样本与初始图的相似度显著高于交互效果较差的样本。通过挖掘最大频繁子图可以发现围绕相同主题交互的所有小组共同的关注点。通过对各个小组主题的分析可以发现主题分布比例的不同。从个体层面分析得出个体的贡献度采用其激活目标子图的总激活量衡量，而且交互效果好的个体的贡献度显著高于交互效果差的个体。通过对369名个体的学习效果进行分析，发现不同个体对目标知识网络图总激活量的20%与其小组整体对目标知识网络图总激活量的80%之和可以预测个体的学习效果，能够解释21.3%的变异。另外还发现，交互效果好与差的小组在信息类型和信息的表征方式方面都没有显著性差异。 最后，对于已有的基于言语行为的分析方法进行了检验，发现协作学习不存在所谓的 “阶段论” 的特征；聚焦水平可以预测小组绩效，但不是在30%-40%就可以预测，没有统一的模式可循；交互效果好与差的小组在言行方面没有显著性差异。这样从另外一面证明了基于信息流的协作学习交互分析方法的有效性。 本研究的核心在于建立基于信息流的面对面协作学习交互分析方法并尝试用这种方法分析交互过程。主要贡献是通过大样本证明了基于信息流的面对面协作学习交互分析方法不仅可以分析交互过程，而且能够有效预测协作学习效果。这种从信息流的视角出发并以知识网络图为样本开展研究的尚不多见。基于信息流的面对面协作学习交互分析方法相对客观，操作性强，未来将成为一种有前景的分析方法。

In the past decade, collaborative learning has been paid more and more attention and it has been widely used in the practice. Interactions are basic unit of collaborative learning. Privious studies focused on coding discourse transcripts into speech acts. This approach cannot predict group performance. This study considers the collaborative learning as a complex information system and information flows are identified as the analysis focus. The study has developed an innovative analytic method based on information flows and explored how to analyze interactions with this method from the following four dimensions:Firstly, we specified steps, rules of this new method. Then the analytic tool has been developed to help us to draw an initial map, code and segment information flows in the interaction process and compute attributes of information flows. We selected some indicators that can predict group performance and designed algorithms, including the sum of the quantity of activation, levels of convergence, depth, breadth, degree distribution, centrality, density, average path length, network structures entropy, knowledge distribution entropy, weighted path length of spanning tree of targeting knowledge network etc. Secondly, we collected 121 samples to verify this method. The results indicate that the quantity of activation and levels of convergence of targeting knowledge network can predict group performance, which can explain for 23% and 24.8% of variance respectively. The depth can reflect a deeper understanding of the underlying subject matter and the degree distribution can indicate knowledge structures. The network structures entropy can reflect the heterogeneity of knowledge network. They can predict group performance. In addition, levels of convergence, the depth and the degree distribution can account for 36.9% of variance. So, we can also select these three indicators to predict group performance.Thirdly, we analyzed characteristics of information flows from the temporal and spatial dimensions. The results show that foci transfer and retrospect over time. The process of knowledge construction has the characteristic of non-linearity. From the spatial perspective, knowledge network maps of different groups have differences. There is significant difference in the proportion of mis-constructed knowledge between more successful groups and less successful groups. The proportion of mis-constructed knowledge is negative correlation with group performance. Compare more successful groups with less successful groups, there is significant difference in the similarity between the sample map and the initial map. The contribution of member can be mearsured by the sum of the quantity of activation of targeting knowledge network. Contributions of more successful members are significant higher than less successful members. The performance of individual can be calculated by the sum of 20% of quantity of activation of each individual and 80% of group’s quantity of activation. The common concers can be found by analyzing the maximum frequent graph. Furthermore, there are no differences on information types and representations between more successful groups and less successful groups.Finally, the method of coding discourse into speech acts is examined by us. The results indicate that there is no staged characteristic in collaborative learning. Levels of convergence can predict group performance, however, group performance cannot be predicted based on what happened in the first 30-40% of a discussion. There is no difference on speech acts between more successful groups and less successful groups. Thus we verify the effectiveness of interaction analysis method based on information flow from another perspective.The main idea is to verify the method of analyzing interactions based on information flows in face-to-face collaborative learning. The primary contribution of this paper is the methodology for analysis of interactions. This method can not only predict group performance but also anayze interations elaborately. The IIS-map-based analysis method discards the subjectivity and arbitrary nature of value judgments when coding discussions and gives fully detailed analysis of interactions. This method will become a new analysis approach with high objectivity and strong feasibility. It will be widely applicable in the future.