End-to-end lightweight policy pushing method and device, equipment and storage medium

Abstract

The application discloses an end-to-end lightweight policy pushing method and device, equipment and storage medium, relates to the technical field of computer information and intelligent recommendation, and the method comprises the following steps: acquiring enterprise key index data from a data source of an enterprise end, dividing the life cycle stages of the enterprise, and outputting a fixed-length enterprise stage state vector for each life cycle stage; an independent policy state vector is maintained for each policy, and the dimensions of the policy state vector are aligned with the dimensions of the enterprise stage state vector; a bidirectional GRU network of a preset time length sliding window is constructed to extract an enterprise stage evolution trend vector, and the extracted enterprise stage evolution trend vector is spliced with the enterprise stage state vector to obtain an enterprise complete representation vector; the cosine similarity between the enterprise complete representation vector and each policy state vector is calculated respectively, and the to-be-pushed policy is determined based on the calculation result and is pushed. The application can realize accurate policy pushing.

Description

Technical Field

[0001] This application relates to the fields of computer information technology and intelligent recommendation technology, specifically to an end-to-end lightweight policy push method, apparatus, device and storage medium. Background Technology

[0002] In recent years, with the intensive rollout of policies supporting inclusive finance, technological innovation, and industrial upgrading, accurately matching massive amounts of policy information to eligible enterprises has become a crucial issue in the enterprise services sector. Currently, policy delivery systems primarily employ the following solutions: (1) Policy recommendation system based on keyword and tag matching: By performing NLP (Natural Language Processing) word segmentation and tag extraction on policy texts, an enterprise tag system (industry, scale, region, etc.) is established, and policy recommendation is achieved through bidirectional tag matching. This type of solution relies on a large amount of manual participation in tag maintenance, and the unstructured features and implicit semantics of policy texts are difficult to be fully captured, resulting in limited matching accuracy. (2) Semantic matching recommendation system based on deep learning: By constructing enterprise feature vectors and converting them into semantic embeddings, semantic unit division and context-aware vector generation are performed on policy texts. Combined with multi-view representation mechanism, a comprehensive matching score is generated. This type of solution has achieved good results at the semantic level, but the model size is large and the inference delay is high, making it difficult to meet the lightweight requirements of large-scale real-time push scenarios. (3) Event-driven self-iterative matching system: Subscribe to enterprise declaration process events through event triggers, and dynamically optimize matching rules by combining feature vector generator and rule engine. However, this type of solution relies on the accumulation of enterprise declaration data, has poor effect in the cold start stage, and lacks the ability to model the evolution of policy cycles.

[0003] Therefore, the existing policy delivery methods have the following problems: 1. The separation of enterprise and policy cycles: Existing solutions usually match enterprise characteristics and policy characteristics in the same space, lacking decoupled modeling of their independent evolution cycles, which makes it impossible to effectively identify the temporal misalignment between the enterprise growth period and the policy release window. 2. High real-time synchronization cost: Policy status (such as application progress, remaining quota, deadline) needs to be frequently synchronized and refreshed. Traditional solutions rely on high-frequency polling or full recalculation, resulting in high consumption of computing resources and high response latency. 3. Redundant model dimensions and high deployment costs: In pursuit of matching accuracy, existing deep learning solutions typically employ high-dimensional embedding spaces and complex network structures, resulting in a large number of model parameters, making them difficult to deploy on lightweight terminals or edge nodes. 4. Lack of stage discrimination ability: There is a lack of explicit modeling of the life cycle stage of the policy (such as the warm-up period, peak period, and closing period), the push strategy is rigid, and it is impossible to dynamically adjust the push priority according to the timeliness of the policy. Summary of the Invention

[0004] This application provides an end-to-end lightweight policy delivery method, apparatus, device, and storage medium, which can achieve accurate policy delivery.

[0005] In a first aspect, embodiments of this application provide an end-to-end lightweight policy delivery method, the end-to-end lightweight policy delivery method comprising: Connect to enterprise-side data sources to obtain key enterprise indicator data, so as to divide the enterprise into life cycle stages, and output a fixed-length enterprise stage state vector for each life cycle stage. Acquire policy information and maintain an independent policy state vector for each policy, aligning the dimensions of the policy state vector with the dimensions of the enterprise stage state vector in a low-dimensional space; A bidirectional GRU network with a preset duration sliding window is constructed to extract the enterprise stage evolution trend vector. The extracted enterprise stage evolution trend vector is then concatenated with the enterprise stage state vector to obtain the complete enterprise representation vector. Calculate the cosine similarity between the complete representation vector of the enterprise and the state vector of each policy, and determine the policy to be pushed based on the calculation results.

[0006] In conjunction with the first aspect, in one implementation, the method of connecting to the enterprise-side data source acquires key enterprise indicator data to divide the enterprise into lifecycle stages, and outputs a fixed-length enterprise stage state vector for each lifecycle stage, specifically including: Connect to enterprise-side data sources to acquire key enterprise indicator data; A lightweight temporal classifier is used to divide the enterprise into life cycle stages and output the corresponding enterprise stage state vectors by combining the acquired key enterprise indicator data. Each stage of a company's life cycle corresponds to a fixed-length enterprise stage state vector.

[0007] In conjunction with the first aspect, in one implementation method, The key indicators for the company include revenue growth rate, patent growth rate, employee size change rate, changes in financing rounds and amounts, and industry prosperity index. The life cycle stages include the startup stage, growth stage, maturity stage, and transformation stage.

[0008] In conjunction with the first aspect, in one implementation method, The factors influencing the update of the policy status vector include the policy release time and deadline, remaining application quota / number of companies that have applied, policy type, and the level of the issuing unit; The update process for the policy state vector specifically includes: Based on the timeliness sensitivity of policies, a lightweight snapshot comparison is performed periodically, and only the incremental change part of the policy state vector is differentially calculated. When the change exceeds a preset threshold, the policy state vector is updated; otherwise, the original policy state vector is maintained. Specifically, when an emergency change is detected through the asynchronous event listener interface of the configured external policy data source API, the policy state vector is updated immediately.

[0009] In conjunction with the first aspect, in one implementation, the construction of a bidirectional GRU network with a preset duration sliding window to extract the enterprise stage evolution trend vector specifically includes: Obtain the monthly key performance indicator (KPI) data sequence of the enterprise for a set number of past months; The forward GRU network is used to encode the monthly enterprise key indicator data sequence to obtain the forward hidden state sequence, and the backward GRU network is used to encode the monthly enterprise key indicator data sequence to obtain the backward hidden state sequence. By concatenating the last forward hidden state in the forward hidden state sequence with the first backward hidden state in the backward hidden state sequence, and then performing linear projection, the enterprise stage evolution trend vector is obtained.

[0010] In conjunction with the first aspect, in one implementation, the step of calculating the cosine similarity between the complete enterprise representation vector and each policy state vector, and determining and pushing the policy to be pushed based on the calculation results, specifically includes: Calculate the cosine similarity between the complete enterprise representation vector and each policy state vector; Based on the calculation results, if the cosine similarity between the enterprise's complete representation vector and the current policy state vector is not less than the set matching threshold, then the policy corresponding to the current policy state vector is determined as the policy to be pushed and pushed.

[0011] In conjunction with the first aspect, in one implementation method, The end-to-end lightweight policy delivery method also includes a peak-shaving scheduling strategy; The off-peak scheduling strategy involves extracting and processing the enterprise stage evolution trend vector during off-peak business hours, and updating the policy state vector during the hourly hours of the day.

[0012] Secondly, embodiments of this application provide an end-to-end lightweight policy push device, the end-to-end lightweight policy push device comprising: The interface module is used to connect to enterprise-side data sources to obtain key enterprise indicator data, so as to divide the enterprise into life cycle stages and output a fixed-length enterprise stage state vector for each life cycle stage. The acquisition module is used to acquire policy information and maintain an independent policy state vector for each policy, aligning the dimensions of the policy state vector with the dimensions of the enterprise stage state vector in a low-dimensional space. The splicing module is used to construct a bidirectional GRU network with a preset time sliding window to extract the enterprise stage evolution trend vector, and splice the extracted enterprise stage evolution trend vector with the enterprise stage state vector to obtain the complete enterprise representation vector. The push module is used to calculate the cosine similarity between the complete representation vector of the enterprise and the policy state vectors, and to determine the policies to be pushed based on the calculation results and push them.

[0013] Thirdly, embodiments of this application provide an end-to-end lightweight policy push device, the end-to-end lightweight policy push device including a processor, a memory, and an end-to-end lightweight policy push program stored in the memory and executable by the processor, wherein when the end-to-end lightweight policy push program is executed by the processor, it implements the steps of the end-to-end lightweight policy push method described above.

[0014] Fourthly, embodiments of this application provide a computer-readable storage medium storing an end-to-end lightweight policy push program, wherein when the end-to-end lightweight policy push program is executed by a processor, it implements the steps of the end-to-end lightweight policy push method described above.

[0015] The beneficial effects of the technical solutions provided in this application include: By modeling the independent development stages of enterprises and the independent evolution cycles of policies separately, a decoupled backbone is formed to avoid matching deviations caused by the coupling of the two cycles. While ensuring the timeliness of push notifications, the computational overhead and communication frequency of policy status synchronization are significantly reduced, achieving lightweight synchronization with low response latency. While reducing the model dimensionality, trend vectors are introduced to enhance the ability to identify policy stages, thereby effectively improving the accuracy of push notifications. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the end-to-end lightweight policy delivery method of this application; Figure 2 This is a schematic diagram of the functional modules of the end-to-end lightweight policy delivery device of this application; Figure 3 This is a schematic diagram of the hardware structure of the end-to-end lightweight policy delivery device in this application. Detailed Implementation

[0017] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present application.

[0018] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0019] In the first aspect, the embodiments of this application provide an end-to-end lightweight policy push method. Specifically, it provides an end-to-end lightweight policy push method based on a dual-cycle decoupling architecture, a dynamic scheduling mechanism, and trend vector matching. It is applicable to enterprise services, government big data, policy intelligent matching, and precise recommendation scenarios. It adopts a dual-cycle decoupling architecture as the backbone, with the enterprise end and the policy end independently operating cycle identification and status tracking. In a low-dimensional embedding space, it drives push decisions through similarity matching.

[0020] In one embodiment, reference is made to Figure 1 , Figure 1 This is a flowchart illustrating the end-to-end lightweight policy delivery method of this application. Figure 1 As shown, the end-to-end lightweight policy delivery method includes: S1: Connect to the enterprise's data source to obtain key enterprise indicator data, so as to divide the enterprise into life cycle stages, and output a fixed-length enterprise stage state vector for each life cycle stage. S2: Obtain policy information and maintain an independent policy state vector for each policy, and align the dimensions of the policy state vector with the dimensions of the enterprise stage state vector in a low-dimensional space; S3: Construct a bidirectional GRU network with a preset duration sliding window to extract the enterprise stage evolution trend vector, and concatenate the extracted enterprise stage evolution trend vector with the enterprise stage state vector to obtain the complete enterprise representation vector. S4: Calculate the cosine similarity between the complete representation vector of the enterprise and the state vector of each policy, and determine the policy to be pushed based on the calculation results and push it.

[0021] Furthermore, in one embodiment, key enterprise indicator data is obtained by connecting to the enterprise's data source to divide the enterprise into lifecycle stages, and a fixed-length enterprise stage state vector is output for each lifecycle stage, specifically including: S101: Connects to enterprise data sources to acquire key enterprise indicator data, including revenue growth rate (quarterly year-on-year / quarter-on-quarter), patent growth rate (annual licensing trend), employee size change rate, changes in financing rounds and amounts, and industry prosperity index. S102: A lightweight temporal classifier is used to divide the enterprise into life cycle stages based on the acquired key enterprise indicator data and output the corresponding enterprise stage state vectors; the life cycle stages include the startup stage, growth stage, maturity stage, and transformation stage. Each stage of a company's life cycle corresponds to a fixed-length enterprise stage state vector.

[0022] Specifically, an enterprise stage identification model is constructed. By connecting with enterprise-side data sources, key enterprise indicator data is obtained. Then, a lightweight time series classifier is used to divide the enterprise into life cycle stages based on the obtained key enterprise indicator data and output the corresponding enterprise stage state vector.

[0023] Furthermore, in one embodiment, the factors influencing the updating of the policy status vector include the policy release date and deadline, the remaining application quota / number of companies that have applied, the policy type, and the issuing authority level. Specifically, an independent policy status vector is maintained for each policy, and the updating of the policy status vector is affected by the following factors: policy release date and deadline, remaining application quota / number of companies that have applied, policy type (funding, qualification, tax, etc.), and issuing authority level.

[0024] After obtaining the policy state vector, the dimensions of the policy state vector are aligned with the dimensions of the enterprise stage state vector in a low-dimensional space, laying the foundation for subsequent matching.

[0025] Furthermore, the update process for the policy state vector specifically includes: Based on the timeliness sensitivity of policies, lightweight snapshot comparisons are performed periodically, and differential calculations are only performed on the incremental changes in the policy state vector. When the change exceeds a preset threshold, the policy state vector is updated; otherwise, the original policy state vector is maintained. Specifically, when an urgent change is detected through the asynchronous event listening interface of the configured external policy data source API (Application Programming Interface), the policy state vector is updated immediately.

[0026] Specifically, in order to reduce the update frequency of policy state vectors, a dynamic scheduling and inspection mechanism is integrated on the policy tracking side, and lightweight snapshot comparison is triggered according to the three-level time sensitivity, as shown in Table 1 below.

[0027] Table 1

[0028] For the snapshot comparison mechanism in this application, differential calculation is only performed on the incremental change part of the policy state vector. When the change exceeds the preset threshold, the policy state vector is updated; otherwise, the original policy state vector is maintained. At the same time, an asynchronous event listening interface for the external policy data source API is configured. When an emergency change occurs (such as the policy expiring early or the quota being reduced sharply), the state synchronization is forcibly triggered beyond the scheduling cycle to ensure data reliability.

[0029] Furthermore, in one embodiment, a bidirectional GRU network with a preset duration sliding window is constructed to extract the enterprise stage evolution trend vector, specifically including: S301: Obtain the monthly key performance indicator data sequence of the enterprise for a specified number of past months; S302: The forward GRU network is used to encode the monthly enterprise key indicator data sequence to obtain the forward hidden state sequence, and the backward GRU network is used to encode the monthly enterprise key indicator data sequence to obtain the backward hidden state sequence. S303: The last forward hidden state in the forward hidden state sequence is concatenated with the first backward hidden state in the backward hidden state sequence, and the enterprise stage evolution trend vector is obtained by linear projection.

[0030] Specifically, a bidirectional GRU network with a 36-month sliding window is constructed to extract the enterprise stage evolution trend vector. First, the monthly key indicator data sequence of the enterprise over the past 36 months is obtained. The forward GRU network is used to encode the monthly key indicator data sequence to obtain the forward hidden state sequence. Then, the backward GRU network is used to encode the monthly key indicator data sequence to obtain the backward hidden state sequence. Then, the last forward hidden state in the forward hidden state sequence is concatenated with the first backward hidden state in the backward hidden state sequence. After linear projection, the enterprise stage evolution trend vector is obtained. Finally, the enterprise stage evolution trend vector is concatenated (or weighted and fused) with the enterprise stage state vector to obtain the complete enterprise representation vector.

[0031] Furthermore, in one embodiment, the cosine similarity between the complete enterprise representation vector and each policy state vector is calculated, and the policy to be pushed is determined and pushed based on the calculation results. Specifically, this includes: S401: Calculate the cosine similarity between the firm’s complete representation vector and each policy state vector; S402: Based on the calculation results, if the cosine similarity between the enterprise's complete representation vector and the current policy state vector is not less than the set matching threshold, then the policy corresponding to the current policy state vector is determined as the policy to be pushed and pushed.

[0032] Specifically, in the low-dimensional embedding space, the cosine similarity between the complete enterprise representation vector and each policy state vector is calculated, and a matching threshold is set. Based on the calculation results, if the cosine similarity between the complete enterprise representation vector and the current policy state vector is not less than the set matching threshold, the policy corresponding to the current policy state vector is determined as the policy to be pushed and pushed. The push results include a policy summary, matching reasons and suggested operation guidelines, and are output in an end-to-end manner.

[0033] Furthermore, in one embodiment, the end-to-end lightweight policy push method of this application also includes a peak-shaving scheduling strategy; the peak-shaving scheduling strategy is to extract and process the enterprise stage evolution trend vector during off-peak business hours (such as 2:00-5:00 am), and to update the policy status vector during the hourly hours of the day (every 2-3 hours).

[0034] The following example illustrates the end-to-end lightweight policy delivery method of this application.

[0035] Taking a real-world deployment scenario of a provincial-level enterprise service platform as an example, the platform connects to the operating data of more than 100,000 SMEs in the province, and releases approximately 50 to 200 new policies daily.

[0036] First, anonymized data on revenue, patents, and personnel of enterprises within the jurisdiction over the past three years were collected to construct a training set. A lightweight temporal classifier (such as XGBoost or a lightweight MLP) was used to output enterprise stage state vectors, with a dimension of 64.

[0037] Then, when each policy is registered, its structured fields (release time, expiration time, policy type, etc.) are parsed to obtain the initialized policy state vector.

[0038] Then, a dynamic scheduling strategy is configured to automatically assign a time sensitivity level based on the number of days remaining until the policy deadline, in order to update the policy state vector: if the deadline is within 7 days, it is level 1, and a snapshot check is triggered at 1:00 AM every day; if the deadline is between 8 and 30 days, it is level 2, and it is triggered once every 3 days; if the deadline is more than 30 days, it is level 3, and it is triggered once every 7 days; at the same time, the Webhook callback of the policy release platform is used as a backup for external API events.

[0039] Then, the enterprise stage evolution trend vector is generated. The GRU inference window is from 2:00 to 5:00 every day. The 36-month monthly enterprise key indicator data sequence of each enterprise is read in batches. The enterprise stage evolution trend vector is generated through bidirectional GRU network and concatenated with the enterprise stage state vector to obtain the complete enterprise representation vector.

[0040] Finally, policy push is performed. At each hourly interval (avoiding the GRU inference window), the policy database is traversed, the cosine similarity between the enterprise's complete representation vector and each policy state vector is calculated, and push tasks are generated for matching items that exceed the set matching threshold and sent to the enterprise through a message queue.

[0041] The end-to-end lightweight policy delivery method of this application models the independent development stage of enterprises and the independent evolution cycle of policies separately, forming a decoupled backbone to avoid matching deviations caused by the periodic coupling of the two. While ensuring the timeliness of the delivery, it significantly reduces the computational overhead and communication frequency of policy status synchronization, achieving lightweight synchronization with low response latency. While reducing the model dimensionality, it introduces trend vectors to enhance the ability to identify policy stages, thereby effectively improving the accuracy of the delivery.

[0042] Secondly, embodiments of this application also provide an end-to-end lightweight policy delivery device.

[0043] In one embodiment, reference is made to Figure 2 , Figure 2 This is a schematic diagram of the functional modules of the end-to-end lightweight policy delivery device of this application. Figure 2 As shown, the end-to-end lightweight policy push device includes: a docking module, an acquisition module, a splicing module, and a push module.

[0044] The docking module connects to enterprise-side data sources to obtain key enterprise indicator data, enabling the division of the enterprise into lifecycle stages and outputting a fixed-length enterprise stage state vector for each lifecycle stage. The acquisition module acquires policy information and maintains an independent policy state vector for each policy, aligning the dimensions of the policy state vector with the dimensions of the enterprise stage state vector in a low-dimensional space. The splicing module constructs a bidirectional GRU network with a preset duration sliding window to extract the enterprise stage evolution trend vector and splices it with the enterprise stage state vector to obtain the complete enterprise representation vector. The push module calculates the cosine similarity between the complete enterprise representation vector and each policy state vector, determines the policy to be pushed based on the calculation results, and pushes it.

[0045] Thirdly, embodiments of this application provide an end-to-end lightweight policy push device, which can be a personal computer (PC), laptop computer, server, or other device with data processing capabilities.

[0046] Reference Figure 3 , Figure 3This is a schematic diagram of the hardware structure of the end-to-end lightweight policy push device involved in the embodiments of this application. In the embodiments of this application, the end-to-end lightweight policy push device may include a processor, a memory, a communication interface, and a communication bus.

[0047] The communication bus can be of any type and is used to interconnect the processor, memory, and communication interface.

[0048] The communication interface includes input / output (I / O) interfaces, physical interfaces, and logical interfaces used for interconnecting devices within the end-to-end lightweight policy delivery device, as well as interfaces used for interconnecting the end-to-end lightweight policy delivery device with other devices (such as other computing devices or user equipment). Physical interfaces can be Ethernet interfaces, fiber optic interfaces, ATM interfaces, etc.; user equipment can be displays, keyboards, etc.

[0049] Memory can be various types of storage media, such as random access memory (RAM), read-only memory (ROM), non-volatile RAM (NVRAM), flash memory, optical storage, hard disk, programmable ROM (PROM), erasable PROM (EPROM), electrically erasable PROM (EEPROM), etc.

[0050] The processor can be a general-purpose processor, which can call the end-to-end lightweight policy push program stored in memory and execute the end-to-end lightweight policy push method provided in the embodiments of this application. For example, the general-purpose processor can be a central processing unit (CPU). The method executed when the end-to-end lightweight policy push program is called can be referred to in the various embodiments of the end-to-end lightweight policy push method of this application, and will not be repeated here.

[0051] Those skilled in the art will understand that Figure 3 The hardware structure shown does not constitute a limitation of this application and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0052] Fourthly, embodiments of this application also provide a computer-readable storage medium.

[0053] The present application stores an end-to-end lightweight policy push program on a computer-readable storage medium, wherein when the end-to-end lightweight policy push program is executed by a processor, it implements the steps of the end-to-end lightweight policy push method as described above.

[0054] The method implemented when the end-to-end lightweight policy push procedure is executed can be referred to in the various embodiments of the end-to-end lightweight policy push method of this application, and will not be repeated here.

[0055] The terms "comprising" and "having," and any variations thereof, in the specification, claims, and accompanying drawings of this application are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to such process, method, product, or apparatus. The terms "first," "second," and "third," etc., are used to distinguish different objects, etc., and do not indicate a sequence, nor do they limit "first," "second," and "third" to different types.

[0056] In the description of the embodiments of this application, terms such as "exemplary," "for example," or "for instance" are used to indicate examples, illustrations, or explanations. Any embodiment or design described as "exemplary," "for example," or "for instance" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of terms such as "exemplary," "for example," or "for instance" is intended to present the relevant concepts in a concrete manner.

[0057] In the description of the embodiments of this application, unless otherwise stated, " / " means "or". For example, A / B can mean A or B. The "and / or" in the text is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, A and / or B can mean: A exists alone, A and B exist simultaneously, and B exists alone. In addition, in the description of the embodiments of this application, "multiple" means two or more.

[0058] In some processes described in the embodiments of this application, multiple operations or steps are included in a specific order. However, it should be understood that these operations or steps may not be executed in the order they appear in the embodiments of this application, or they may be executed in parallel. The sequence number of the operation is only used to distinguish different operations, and the sequence number itself does not represent any execution order. In addition, these processes may include more or fewer operations, and these operations or steps may be executed sequentially or in parallel, and these operations or steps may be combined.

[0059] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes several instructions to cause a terminal device to execute the methods described in the various embodiments of this application.

[0060] The above are merely preferred embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A lightweight end-to-end policy delivery method, characterized in that, The end-to-end lightweight policy delivery method includes: Connect to enterprise-side data sources to obtain key enterprise indicator data, so as to divide the enterprise into life cycle stages, and output a fixed-length enterprise stage state vector for each life cycle stage. Acquire policy information and maintain an independent policy state vector for each policy, aligning the dimensions of the policy state vector with the dimensions of the enterprise stage state vector in a low-dimensional space; A bidirectional GRU network with a preset duration sliding window is constructed to extract the enterprise stage evolution trend vector. The extracted enterprise stage evolution trend vector is then concatenated with the enterprise stage state vector to obtain the complete enterprise representation vector. Calculate the cosine similarity between the complete representation vector of the enterprise and the state vector of each policy, and determine the policy to be pushed based on the calculation results.

2. The end-to-end lightweight policy delivery method as described in claim 1, characterized in that, The system connects to the enterprise's data source to obtain key enterprise indicator data, thereby dividing the enterprise into lifecycle stages. For each lifecycle stage, a fixed-length enterprise stage state vector is output, specifically including: Connect to enterprise-side data sources to acquire key enterprise indicator data; A lightweight temporal classifier is used to divide the enterprise into life cycle stages and output the corresponding enterprise stage state vectors based on the acquired key enterprise indicator data. Each stage of a company's life cycle corresponds to a fixed-length enterprise stage state vector.

3. The end-to-end lightweight policy delivery method as described in claim 2, characterized in that: The key indicators for the company include revenue growth rate, patent growth rate, employee size change rate, changes in financing rounds and amounts, and industry prosperity index. The life cycle stages include the startup stage, growth stage, maturity stage, and transformation stage.

4. The end-to-end lightweight policy delivery method as described in claim 1, characterized in that: The factors influencing the update of the policy status vector include the policy release time and deadline, remaining application quota / number of companies that have applied, policy type, and the level of the issuing unit; The update process for the policy state vector specifically includes: Based on the timeliness sensitivity of policies, a lightweight snapshot comparison is performed periodically, and only the incremental change part of the policy state vector is differentially calculated. When the change exceeds a preset threshold, the policy state vector is updated; otherwise, the original policy state vector is maintained. Specifically, when an urgent change is detected through the asynchronous event listener interface of the configured external policy data source API, the policy state vector is updated immediately.

5. The end-to-end lightweight policy delivery method as described in claim 1, characterized in that, The construction of a bidirectional GRU network with a preset duration sliding window to extract the enterprise stage evolution trend vector specifically includes: Obtain the monthly key performance indicator (KPI) data sequence of the enterprise for a set number of past months; The forward GRU network is used to encode the monthly enterprise key indicator data sequence to obtain the forward hidden state sequence, and the backward GRU network is used to encode the monthly enterprise key indicator data sequence to obtain the backward hidden state sequence. By concatenating the last forward hidden state in the forward hidden state sequence with the first backward hidden state in the backward hidden state sequence, and then performing linear projection, the enterprise stage evolution trend vector is obtained.

6. The end-to-end lightweight policy delivery method as described in claim 1, characterized in that, The step of calculating the cosine similarity between the complete enterprise representation vector and each policy state vector, and determining and pushing policies based on the calculation results, specifically includes: Calculate the cosine similarity between the complete enterprise representation vector and each policy state vector; Based on the calculation results, if the cosine similarity between the enterprise's complete representation vector and the current policy state vector is not less than the set matching threshold, then the policy corresponding to the current policy state vector is determined to be the policy to be pushed and pushed.

7. The end-to-end lightweight policy delivery method as described in claim 1, characterized in that: The end-to-end lightweight policy delivery method also includes a peak-shaving scheduling strategy; The off-peak scheduling strategy involves extracting and processing the enterprise stage evolution trend vector during off-peak business hours, and updating the policy state vector during the hourly hours of the day.

8. An end-to-end lightweight policy delivery device, characterized in that, The end-to-end lightweight policy delivery device includes: The interface module is used to connect to enterprise-side data sources to obtain key enterprise indicator data, so as to divide the enterprise into life cycle stages and output a fixed-length enterprise stage state vector for each life cycle stage. The acquisition module is used to acquire policy information and maintain an independent policy state vector for each policy, aligning the dimensions of the policy state vector with the dimensions of the enterprise stage state vector in a low-dimensional space. The splicing module is used to construct a bidirectional GRU network with a preset time sliding window to extract the enterprise stage evolution trend vector, and splice the extracted enterprise stage evolution trend vector with the enterprise stage state vector to obtain the complete enterprise representation vector. The push module is used to calculate the cosine similarity between the complete representation vector of the enterprise and the policy state vectors, and to determine the policies to be pushed based on the calculation results and push them.

9. An end-to-end lightweight policy delivery device, characterized in that, The end-to-end lightweight policy push device includes a processor, a memory, and an end-to-end lightweight policy push program stored in the memory and executable by the processor, wherein when the end-to-end lightweight policy push program is executed by the processor, it implements the steps of the end-to-end lightweight policy push method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores an end-to-end lightweight policy push program, wherein when the end-to-end lightweight policy push program is executed by a processor, it implements the steps of the end-to-end lightweight policy push method as described in any one of claims 1 to 7.

Images