The knowledge mapping in the field of wind control is very different from the common knowledge mapping in the way it is constructed and applied, and this presentation has three main aspects:
1. Knowledge maps in the field of wind control spectrum
2. Application of wind control knowledge maps
3. Development and outlook of knowledge mapping in the field of wind control
Share the guest xu de wa, wing payments, director general
Editor of the beijing teacher training university
I'm sorry
"datafun."
01
Knowledge mapping for wind control
What is wind control
Wind control, risk control, is closely related to our business landscape. The three main business scenarios currently covered by the wings are the payment scene, the electricity business scene and the credit scenario. We will face different risks in these three areas。
One of the most important concerns in the payment scene is the risk of money-laundering, whereby illicits may complete a series of money-laundering operations through our payment platforms, such as cash, gambling, money transfers, etc., and the protection of personal accounts, such as the embezzlement of payment accounts, bank card theft, etc。
Some of the most frequent issues in the electric business landscape include the arbitrage of the wool parties (where platform operators conduct marketing activities, the wool parties maliciously usurp and profit from marketing resources, including through billing or call-type arbitrage), as well as phenomena such as false businesses, false transactions and so forth, which cause some loss to companies or users。
The most commonly encountered in the credit landscape are intermediary frauds, such as the intermediary village of loan fraud or internet lending, as well as applications from individuals with apparent credit irregularities and fraud, which are now at risk in the business landscape。
2. Current situation in the professional black market
At the 315th evening, black production was revealed to the navy, as part of the entire black production system. The technology of black production today is very comprehensive and the division of labour is very clear, and it will be upstream with the provision of a series of so-called black production productivity tools, such as cat/carpools, cloud phones, equipment farms, as well as a platform for the delivery of registered sms authentication codes, a team that specializes in the deployment of a single key, controls multiple accounts, and a detailed division of labour among the different players in the entire black industry chain. According to our experience of confrontation, black production is not only very diligent but also very professional, sometimes starting at 6 p. M. It may have started at 2 or 3 a. M. The day before. At the same time, some of the fraudulent tools and scripts provided by black production not only provide complete version management and maintenance, but also support a series of upgrades and optimizations to provide a full “post-sales service”. According to the study, the entire black industry has already grown to more than hundreds of millions and more than 1. 6 million practitioners. Public reports estimate the losses of black production to be nearly hundreds of millions of dollars, with the appearance of clustering, regionalization and a clear division of labour in the industrial chain。
3. Wind control systems for wing payments
To counter black production, we have a complete system of wind control, of which mapping is part. The bottom level is the collection of data stored in multiple modes, including user behaviour, user-registered image information, completed text information, user operational transaction data and a range of data such as tripartite operators, industry lists, etc. With bottom-up data, an independent wind-controlled role engine, including the wind-controlled ai engine and an early warning module for analytical disposal, is built on the basis of a complete management system that manages the knowledge base of indicator models and rules, thus achieving risk management throughout the life cycle of users before, during and after。
In particular, ai technology is concerned with the following aspects. First, the identification of face attacks, as well as the detection of scenes such as photo synthesis, tampering, etc., should be carried out using visual and deceptive cv technology, based on information such as identity cards and business licences. Second, there are many instances where there are no sample labels in the electric business scene, and this requires identification using unsupervised methods. Third, monitored learning needs to be used in the context of precision identification of various types of known risks and user credit assessment. Furthermore, as regulatory requirements become more sophisticated, we will work with different companies based on federal learning techniques. And the most important is knowledge mapping。
4. Wing payments construction
The wings paid for the completion of the knowledge mapping from three dimensions, for payment operations, and currently stocked at the point level of 100 billion. It's almost a billion-dollar volume. The third part of communications fraud, because it involves communications data, is amazing in volume, with a stock of 10 billion in short-term cycles
5. Main differences between wing payment mapping and general knowledge mapping
In the construction of the wing pay wind-controlled knowledge map, the difference with the so-called industry-wide knowledge map is mainly in three ways。
1 build paradigm
Financial institutions have two distinct features, one strong schema and the other top-down construction. The entire schema definition of the schema is defined as a pre-defined relationship between the subject and the need to maintain, based on the needs for use and the actual business scene, e. G., clearly defined prior to the construction of the payment spectroscopy, and then started to construct the schematics, so it is a top-down system, less related to the need for physical extraction, physical alignment, relationship reasoning, etc. In mass data during the construction of the generic knowledge spectroscopy。
2 data volume
The financial mapping, both in terms of the volume of stocks and the increasing volume, is striking, and it is highly time-sensitive. Financial scenarios leave much of the decision-making time for wind control at milliseconds, in short tens of milliseconds, and long at about 100 milliseconds, during which it is necessary to complete the entire wind-control decision-making chain, not just the reasoning of the knowledge map, so there is a very high demand for timeliness。
3 data quality
Because financial institutions are subject to strong regulation, unlike the universal mapping that requires the internet to crawl over data, the quality of the data is much higher, the problem of unusual data missing is less likely, and there are no problems similar to physical alignment or physical diversity in the construction of the maps. And that's some of the obvious differences that we have between the drawings and the general drawings in the area of wind control。
02
Application of wind-controlled knowledge maps
There are two main categories of users of cloud-based knowledge maps, operational and algorithms. The knowledge map of wings paid has gone through several generations, the first generation based on a commercial version of neo4j, but when the volume of data reached a level, it was evident that bottlenecks were encountered and that the structure of the commercial version did not adequately meet the actual high-mixing and distributed demand. Thus, in designing version 2. 0, efforts were made to resolve the problem of version 1. 0 and to construct cloud-based knowledge maps for the payment of wings。
The first feature of the most important component for operational staff is visualization of the exploration, which visualizes the full use of accounts, equipment, letters, etc。
1. Application structure for the wing payment chart
First, at the data level, mapping collects its own data and accessible third-party data, including operators and some open data, as a base for bottom data. A set of capabilities, such as task movement, indicator management and knowledge modelling, can be achieved throughout the mapping platform. The effective functioning of the platform is premised on storage systems. The storage system for wing payments has two main characteristics, one is distributed and the second is capable of storing data in multiple forms, and the integration of multi-modular models will significantly enhance the work of wind control。
On the basis of the management platform and the bottom storage platform, the platform supports the application tools for customization and optimization of performance, such as those for task analysis for nodes analysis, path analysis and offline long cycles. In addition, the platform supports the atlas。
The algorithms may be divided into two broad categories: those of traditional migraines, such as louvain, wcc, label dissemination, etc.; and those of partial depth learning, such as gcn, gat, graph sage, for example, where the platform provides a complete library of algorithms to serve such scenarios as money-laundering, fraud audits, and exploitation of business relationships。
One of the important functions for business students is visualization exploration, where the entity to be consulted, such as mobile phone number, equipment name, business name, etc., can be entered directly into the platform, thus accurately locating the entity and excavating the full set of connections to the node by reaching out to the point layer. Based on practical experience, visualization can significantly increase the efficiency of analysis by allowing business specialists and non-technical students to visualize the overall distribution of issues to be analysed at first sight, and also by providing a clear picture of the overall picture。
2. Core functions of the wing payment chart
Visualization in large data volume
A lot of tricks are needed to get visualization done, especially in large data volumes. For example, how to deal with larger nodes, and how to optimise when reaching out to multiple layers of multiple nodes with uncontrolled factors. Visualization exploration includes both online exploration and offline missions, which can be offline when the focus is on a long-cycle mission。
2 multimodular data integration
The traditional chart database focuses on the storage of map data. However, in addition to graphic data, there is also an important part of the current scene, such as the image information submitted in the business application, which, in the case of an attack by an intermediary, has some relevance to other nodes on the map, such as the ip of the operating device, the wifi name, etc., and the different background images of the applicant are sometimes clearly identified in the same scene. By combining multiple-mode data, we can integrate image information and maps, which, when used by operators, can be judged very quickly as a possible intermediary group, thus avoiding barriers to use across systems and levels。
3. Quality control
For the role of managing the mapping, it is important to understand the size and operational status of the current picture as a whole, and the platform provides very clear and visual quality monitoring. As mentioned earlier, because of the number of long-cycle tasks analysed in the spectrograph, the results cannot be interactively obtained in real time, the task movement function provides a complete back-office movement capability that allows the task to be prioritized and the performance and completion of historical tasks to be viewed, and also allows the results of single run runs to be returned to the spectro or imported into other systems to facilitate the learning of model tasks from downstream machines。
4. Visual drag and search
In practical use, it has been found that technical classmates usually perform queries by writing cypher, but it is not realistic to require each business colleague to master cypher, and the platform provides a drag point selection function to assist business colleagues in their ability to achieve complex mapping queries. Users can define the mode they want to look at, such as the fact that an account has entered certain equipment and that there are transactions between this and other accounts, or limit the time of registration of the account. These complex search conditions allow for more complex queries to be made in the form of drag-and-truck model matching, so that operational staff will not be completely technocratic when dealing with and analysing issues, nor will they be able to improve efficiency by simply clicking on the top floor of the map for visualization. Such missions combine the platform's movement function to achieve real-time or off-line feedback on results。
Data role control
For financial institutions, the control of authority is important, as different departments have different access to the data that they use, which needs to be isolated; maps that can be viewed by different actors, and the same mapping can be fine-tuned to the nodes and relationships displayed by different users. In addition, functions that can be used by different users, such as the possibility of editing maps, modifications to schema, and the possibility of achieving a secure isolation through the control of privileges, meet business requirements。
3. Graphical system for the distribution of wing payments
Three types of methods are common in mapping algorithms。
The first category consists of subgraph-based excavations, whether through louvain, wcc or otherwise, of such suspected fraud syndicates, particularly in the context of such large-scale attacks, such as money-laundering, arbitrage, which can be based on single-point excavations。
The second category is mEtapath, based on model matching, the business experience accumulated by business personnel can help to identify some unusual structures, such as the three-way transfer of users to the business, and such a single-link model also has business risks. This long-link matching approach does not support the traditional sql database at three degrees or more. Long-link excavations of maps have a clear advantage。
The third method is gnn, which combines mapping structural and attribute data through a neurological network methodology, thus enhancing model effects, with a typical scenario based on sub-charting-- - group fraud, such as the black broker of credit fraud and the wool party。
A very large group can be seen in the first case chart by analysing purple nodes and several red nodes. In the case of a traditional approach, a one-time count of a single node could be achieved through sql. However, this makes it difficult to understand the overall risk scenario from a global perspective and to estimate the overall risk scale. Through the gnn method, a conviction rate of more than 99 per cent can be achieved for certain types of money-laundering offences, and there has been a marked increase in the reported amount。
The second is the case of intermediary fraud, where in individual financial operations intermediaries package individuals who are less creditworthy users so that they can apply for letters of credit, and where intermediaries profit from clients' applications for loans by cutting their bells. In this case, the combination of multi-mode data makes it possible to discover very intuitively that the nodes are from an intermediary group in the same place, thus completing the group's excavation. This was also the case during the actual visit, when the operator did business with the user, without informing the client of the high-value gifts and only with a small return to the user, and after leaving the gift to the operator. For users, such acts already constitute fraud. Such acts have resulted in more complex models, such as the entire order business for the user by the operator,a and the transfer to the user by the other operator,b of the preferential amount of the promise. The two operators have an intersection of equipment and a correlation between the agency's subordination, and the pattern of abstracting such scenes can be identified by means of cypher or model matching, thus making effective use of the experience of operational experts and improving their accuracy. Interception of brokering fraud can be achieved by deploying real-time surveillance。
The third case uses the gnn methodology. Gnn has the advantage of combining both structural and nodal information. The details on the figure are amplified by the fact that the entire financial chain is a long-link return and that 10 per cent of the entire financial flow will be depleted by layer, a more typical way of dealing with certain types of money-laundering offences. Thus, in designing the programme, first, there is a predefined list of suspicious basic structures of expansion, such as long chains, cross-rings, and basic structural patterns of centralization of funds transferred as part of the mapping nodes expansion information; secondly, additional information on the attributes of some users, such as the loss of funds in money-laundering cases, is a very visible feature, so that information on the loss of funds is also included in the mapping as an attribute; and thirdly, the integration of the patterns of behaviour of users, the analysis of the users of money-laundering is clearly similar to those of their groups and the pattern of behaviour of the normal users, the chart also achieves the representation of this information, the nodal information in the final map and the structural information of the entire picture, and the excavation of the site of the intermediary fraud syndicate is achieved。
03
Future outlook
The computational framework of the neurological network has evolved considerably in recent years, and large-scale knowledge mapping can be built on frameworks such as ddl or pyg. The wings pay for the current version of the volume in the order of tens of millions, supporting training in the mapping depth learning model of the sub-model. In the context of the widespread interest in chatgpt, there is a discussion about the future of the map。
1. Data dimension
Early attention was focused on data storage, with emphasis on how to design more efficient distributed storage and thus improve query performance. Through these years of efforts, the current mainstream technology programmes have met industry needs. Now we are more concerned about whether, with the emergence of large models such as gpt, the knowledge mapping itself will be reconstructed. The current pattern of graphics is based more on experience or rules, but with the validation of models such as gpt, we find that gpt itself has found a way to store a large amount of the knowledge in its own network structure, and in a sense it has been able to store it. The way in which such large-language models can achieve knowledge storage can be combined with the existing technological architecture of the current knowledge map to increase the efficiency of its construction and flexibility in its use, which may require future focus。
2. System real-time requirements
Second, in the application of the actual financial scene, the time available for wind-control analysis is very short, so the efficiency of real-time data input and real-time response in the field of wind-control is very high, which is also the focus of future optimization and exploration。
3. Application scenarios
While there are some more mature distributed training framework products available on the market, there are still efficiencies that need to be optimized and upgraded; moreover, for the integration of multi-isomeral data, the current mapping training is based on data from extended structural data and feature fairs, how to integrate future multi-modular data such as image data, even audio-visual, into knowledge maps, and training to integrate multi-isomer data will be the way forward. Moreover, when the inter-agency federation of unions controls, for example, for data security reasons, the current federation is unable to obtain information about the drifting relationship, which is also a work being explored by the wing payments and regulatory authority on the basis of the federal identification of intermediaries and fraud。
Moreover, since the financial sector is a well-regulated institution, there are requirements for the safety, privacy and fairness of the model system as a whole, and how to balance these issues in model design is some of the direction that future thinking will take。
04
