Key Technologies For Cognitive Reasoning Networks From Structured Data To Graph Data

A cognitive reasoning network refers to a network that unconsciously retrieves cues related to an reasoning task,then fuses features of these cues and consciously reasons about the interpretable results based on the fused features.The usual data representation is in the form of structured data.As humans entered the Web 2.0 era,the data representation began to shift from structured data to graph data.Further study of data in the form of graph structure has relevance and significant value.Therefore,in cognitive reasoning network research,feature fusion and cognitive reasoning analysis of structured data and graph data,respectively,are two key technical issues in cognitive reasoning network research.The structured data represented by ADMET data,which measures whether a compound can be a drug candidate,is characterized by strong correlation,strong dependence,and multiple features,and the graph data represented by attribute network data has the characteristics of high dimensionality and strong correlation of node information.The fusion and cognitive reasoning of the information encoded in these characteristics can further uncover comprehensive and multi-level value information.Currently,although researchers have conducted a lot of research on structured data to graph data,however,there are two major challenges in ADMET data processing and citation network data processing.first,in the research of processing ADMET data,the correlation and strong dependency of data features often derive extremely complex and very critical determinants,which makes deeper mining of these information more difficult.In addition,the existing data analysis methods still have shortcomings in considering the feature extraction and feature fusion of such information and in the interpretability and rationality of the analysis results,which cannot well accomplish data processing to solve the practical problems in engineering activities.Second,in the studies dealing with citation network data,most of the existing research methods can only implicitly assign weights to different neighborhood nodes in the target nodes,which makes the interpretability of node embedding weak.On the other hand,most of the methods aggregate the information of local neighborhood nodes in the purest form of definition,which leads to the absence of certain node association information and is extremely detrimental to the efficient utilization of network information.To address the above existing problems,this thesis investigates a cognitive reasoning network model that can handle multi-level features and realize the fusion of multi-level features and cognitive reasoning analysis by using capsule network as the main line of technology.The research idea of this thesis is to study from ADMET nature classification task to citation network node embedding task in terms of research key technical tasks,and to carry out cognitive reasoning technology research from single-layer capsule network to double-layer capsule network in terms of technical route.The main research key technical issues and research contents are as follows.1.This thesis study the key technology of cognitive reasoning network for the classification task of ADMET nature,which is an important basis for dealing with practical application problems such as candidate drug screening,prediction test results and real-time decision making.First,to address the problem of underutilization of correlation and dependency of drug data,a novel feature encoder is proposed using the technique of computer storage of RGB images,which incorporates the correlation and interdependency between features as a consideration into the classification basis,making the experimental results more realistic and valid.Secondly,to address the problem of weak interpretability of ADMET nature classification results,a single-way cognitive reasoning mechanism is proposed based on dynamic routing algorithm,and the extracted features are used to achieve interpretable classification of ADMET by a top-down,result-to-cause feedback mechanism.2.Based on the previous section,this thesis further explore the key techniques of cognitive reasoning networks for citation network node embedding tasks to reduce the difficulty of complex network analysis tasks.On the one hand,to address the problem that existing node embedding methods can only implicitly assign different weights to different neighborhood nodes in the target node,which leads to weak interpretability of the results,a node density fusion algorithm is proposed to quantify node density information as node features and explicitly assign different weights to neighborhood nodes in the target node,so as to improve the effectiveness and interpretability of node embedding.On the other hand,to address the problem of underutilization of node association information in citation networks,a two-way cognitive reasoning mechanism for the node embedding problem is proposed based on the single-way cognitive reasoning mechanism,which represents the dimensions of embedding as probability vectors and optimizes the interpretability of node embedding.