The root cause of software security lies in the source code developed by software developers， but with the continues increasing size and complexity of software， it is costly and difficult to perform vulnerability detection only manually， while the existing code analysis tools have high false positive rate and false negative rate. Therefore， an automatic vulnerability detection method based on Relational Graph Convolution Network （RGCN） was proposed to further improve the accuracy of vulnerability detection. Firstly， the program source code was transformed into CPG containing syntax and semantic information. Then， representation learning was performed to the graph structure by RGCN. Finally， a neural network model was trained to predict the vulnerabilities in the program source code. To verify the effectiveness of the proposed method， an experimental validation was conducted on the real-world software vulnerability samples， and the results show that the recall and F1-measure of vulnerability detection results of the proposed method reach 80.27% and 63.78% respectively. Compared with Flawfinder， VulDeepecker and similar method based on Graph Convolution Network （GCN）， the proposed method has the F1-measure increased by 182%， 12% and 55% respectively. It can be seen that the proposed method can effectively improve the vulnerability detection capability.
Relation extraction aims to extract the semantic relationships between entities from the text. As the upper-level task of relation extraction， entity recognition will generate errors and spread them to relation extraction， resulting in cascading errors. Compared with entities， entity boundaries have small granularity and ambiguity， making them easier to recognize. Therefore， a relationship extraction method based on entity boundary combination was proposed to realize relation extraction by skipping the entity and combining the entity boundaries in pairs. Since the boundary performance is higher than the entity performance， the problem of error propagation was alleviated； in addition， the performance was further improved by adding the type features and location features of entities through the feature combination method， which reduced the impact caused by error propagation. Experimental results on ACE 2005 English dataset show that the proposed method outperforms the table-sequence encoders method by 8.61 percentage points on Macro average F1-score.
The discovery of community structures using urban big data is an important research direction in urban computing. Effective representation of the structural characteristics of the communities in the "15-minute living circle" can be used to evaluate the facilities around the living circle communities in a fine-grained manner， which is conducive to urban planning as well as the construction and creation of a livable living environment. Firstly， the urban community structure oriented to "15-minute living circle" was defined， and the structural characteristics of the living circle communities were obtained by representation learning method. Then， the embedding representation framework of the living circle community structure was proposed， in which the relationship between the Points Of Interest （POI） and the residential area was determined by using the travel trajectory data of the residents， and a dynamic activity map reflecting the travel rules of the residents at different times was constructed. Finally， the representation learning to the constructed dynamic activity map was performed by an auto-encoder to obtain the vector representations of the potential characteristics of the communities in the living circle， thus effectively summarizing the community structure formed by the residents’ daily activities. Experimental evaluations were conducted using real datasets for applications such as community convenience evaluation and similarity metrics in living circles. The results show that the daily latent feature expression method based on POI categories is better than the weekly latent feature expression method. Compared to the latter， the minimum increase of Normalized Discounted Cumulative Gain （NDCG） of the former is 24.28% and the maximum increase of NDCG is 60.71%， which verifies the effectiveness of the proposed method.
Aiming at the problem of poor recognition of sentencing circumstances in adjudication documents caused by the lack of labeled data， low quality of labeling and existence of strong logicality in judicial field， a sentencing circumstance recognition model based on abductive learning named ABL-CON （ABductive Learning in CONfidence） was proposed. Firstly， combining with neural network and domain logic inference， through the semi-supervised method， a confidence learning method was used to characterize the confidence of circumstance recognition. Then， the illogical error circumstances generated by neural network of the unlabeled data were corrected， and the recognition model was retrained to improve the recognition accuracy. Experimental results on the self-constructed judicial dataset show that the ABL-CON model using 50% labeled data and 50% unlabeled data achieves 90.35% and 90.58% in Macro_F1 and Micro_F1， respectively， which is better than BERT （Bidirectional Encoder Representations from Transformers） and SS-ABL （Semi-Supervised ABductive Learning） under the same conditions， and also surpasses the BERT model using 100% labeled data. The ABL-CON model can effectively improve the logical rationality of labels as well as the recognition ability of labels by correcting illogical labels through logical abductive correctness.
The combination of blockchain and supply chain is a popular research topic in recent years. The advantages of blockchain such as data traceability， tamper proof and distributed storage can guarantee good data security for supply chain， while the autonomy property of blockchain also provides possibility of supply chain autonomy. The autonomy of blockchain mainly depends on consensus mechanism， but the existing consensus mechanism is difficult to realize good support for supply chain autonomy. To solve the above problems， an election-based consensus mechanism based on Delegated Proof of Stake （DPoS） was proposed， and on this basis， a self-made framework of supply chain based on blockchain was constructed， namely Election-based Supply Chain （ESC）. In ESC， the credit score of a node was first calculated according to the smart contract activities participated in by this node. Then， from the perspective of game theory， the influences of node active degree and credit score on stake under ESC were analyzed. Finally， theorem proving and simulation experiments verify that the proposed mechanism has a good incentive effect on nodes and can effectively inhibit the maximum transaction cost paid by rational nodes，and the inhibition increasing with the increase of the number of delegates.
The pre-training language models used for text representation have achieved high accuracy on various text classification tasks， but the following problems still remain： on the one hand， the category with the largest posterior probability is selected as the final classification result of the model after calculating the posterior probabilities on all categories in the pre-training language model. However， in many scenarios， the quality of the posterior probability itself can provide more reliable information than the final classification result. On the other hand， the classifier of the pre-training language model has performance degradation when assigning different labels to texts with similar semantics. In response to the above two problems， a model combining posterior probability calibration and negative example supervision named PosCal-negative was proposed. In PosCal-negative model， the difference between the predicted probability and the empirical posterior probability was dynamically penalized in an end-to-and way during the training process， and the texts with different labels were used to realize the negative supervision of the encoder， so that different feature vector representations were generated for different categories. Experimental results show that the classification accuracies of the proposed model on two Chinese maternal and child care text classification datasets MATINF-C-AGE and MATINF-C-TOPIC reach 91.55% and 69.19% respectively， which are 1.13 percentage points and 2.53 percentage points higher than those of Enhanced Representation through kNowledge IntEgration （ERNIE） model respectively.
In the ride-hailing platform， matching is a core function，and the platform needs to increase the number of matched orders as much as possible. However， the demand distribution of ride-hailing is usually extremely uneven， and the starting points or end points of orders show the characteristic of high concentration in some time periods. Therefore， an incentive mechanism with early warning was proposed to encourage drivers to take orders across regions， thus achieving the purpose of rebalancing the platform cross-regional transport capacity. The order information was analyzed and processed in this strategy， and an early warning mechanism of transport capacity in adjacent regions was established. To reduce the number of unmatched orders in the region during the period of tight transport capacity and improve the platform utility and passenger satisfaction， drivers in adjacent regions were encouraged to accept cross-regional orders when regional transport capacity was tight. Experimental results on instances show that the proposed rebalancing mechanism improves the average utility by 15% and 38% compared with Greedy and Surge mechanisms， indicating that the cross-regional transport capacity rebalancing mechanism can improve the platform revenue and driver utility， rebalance the supply-demand relationship between regions to a certain extent， and provide a reference for the ride-hailing platform to balance the supply-demand relationship macroscopically.
On the MOOC platform， there may be multiple versions of videos for one course，in order to recommend a MOOC video that satisfies the learning interest of the student，it is necessary to analyze the relationship between student interests and video contents. For this purpose， a video recommendation model named Mrec was proposed based on meta-path attention mechanism. For one thing， the Heterogeneous Information Network （HIN） was used to describe the relationships between learners and MOOC resources， and then meta-path was used to express the interaction between students and videos. For another， the attention mechanism was used to capture the influences of the characteristics of students， videos and meta-paths on learning interest. Specifically， the Mrec model was composed of two layers of attention mechanism： the first layer was the node attention layer， the node own characteristics were weightely combined with neighbor characteristics， and the feature representations of entities under different meta-paths were obtained by multi-head attention； the second layer was the path attention layer， in which the feature representations of entities learned under the guidance of different meta-paths were integrated to capture the feature representations of entities under different interests， and the learned users and video entities were put into Multi-Layer Perceptron （MLP） to obtain the prediction scores for top-K recommendation. Experimental results on MOOCCube and MOOCdata datasets show that Mrec outperforms the comparison methods in terms of Hit Ratio （HR）， Normalized Discounted Cumulative Gain （NDCG）， Mean Reciprocal Ranking （MRR） and Area Under receiver operating characteristic Curve （AUC）.
