Abstract:

Aiming at the problems that the features extracted by the traditional deep forest algorithm are not complete, and the equal-power decision mechanism is easy to produce differences in the classification results, an improved DeepForest (IDF) model applied to inexact graph matching is proposed. Firstly, in the process of mining feature subsets, the methods of fusing moving windows and random moving windows are adopted. While the moving window scans the sample, a same size feature subset of the moving scanning window is randomly captured, and these form a new feature subset, which is used as the input of the cascade module. Secondly, in the iterative process of the cascading forest, the weight of decision result in the current forest is calculated. Compared with the upper level forest, the weight value is assigned to the current forest by the strategy rule of Min, and iteration is continued until the result meets the given threshold value by the model. Finally, training and testing are conducted on datasets such as MUTAG, PTC and COX2. The experimental results show that, compared with traditional deep forest algorithm, IDF fully considers the structural characteristics of the graph, and can effectively enhance the diversity of samples and the goodness of fit, and reduce the decision-making difference and the complexity of the model. It efectively improves the classification and recognition rate of the model.

Key words: inexact graph matching, deep forest, decision tree, weighted

摘要：

针对传统深度森林算法提取的特征不够完整,以及采取的等权决策机制对分类结果易产生差异性等问题,提出一种应用于非精确图匹配的改进DF模型（IDF）。首先,在挖掘特征子集的过程中,采取融合移动窗口和随机移动窗口的方式。在移动窗口扫描样本的同时,随机捕获一个与移动扫描窗口相同大小的特征子集,两者构成新的特征子集,从而将新特征子集作为级联森林模块的输入。其次,在级联森林的迭代过程中,计算当前森林的决策结果所占权重,并与上一级森林进行对比,采用Min的权值策略规则赋值给当前森林,逐次迭代直至结果满足模型所设定的阈值。最后,在MUTAG、PTC、COX2等数据集上进行了训练和测试。实验结果表明,相比于传统深度森林算法,IDF充分考虑了图的结构特征,能够有效增强样本的拟合优度及多样性,降低了级联模块中各子树的决策差异及模型的复杂度,有效提升了模型的分类识别率。

关键词: 非精确图匹配, 深度森林, 决策树, 加权