Resource recommendation method, device, equipment and medium based on dynamic portrait of small enterprises

Abstract

The present application relates to a resource recommendation method, device, equipment and medium based on dynamic portraits of small enterprises. The method extracts key events from multi-dimensional data of enterprises and standardizes them to form an initial event sequence, calculates importance scores based on frequency and weight, and adaptively filters to generate a core event set; calculates the dependency strength between events based on timestamps and semantic features, and constructs a causal event relationship map; uses graph neural networks and gated recurrent units to learn event embedding and time series features to generate dynamic portrait vectors; analyzes the impact weights of historical events through the attention mechanism to predict future demand distribution; optimizes resource urgency and matching in combination with enterprise capacity constraints to generate a forward-looking recommendation list; triggers incremental map updates and model retraining through prediction deviations to form a closed loop. Ultimately, dynamic capture of development trajectories and accurate prediction of resource needs are achieved, significantly improving recommendation timeliness and business adaptability.

Description

Technical Field

[0001] The present invention belongs to the field of data processing technology, and in particular relates to a resource recommendation method, device, equipment and medium based on dynamic portraits of small businesses. Background Art

[0002] In the current enterprise services landscape, accurately matching development resources for small and medium-sized enterprises (SMEs) remains a significant challenge. Traditional resource recommendation methods rely on static enterprise profiles, often constructed from historical financial data or fixed tag systems, making it difficult to capture the dynamic trajectory of a company's development. This is especially true for rapidly iterating SMEs, where key events such as financing cycles, product updates, and market strategies are highly sequential and sudden. Existing technologies have significant limitations when handling these sequential events: First, isolating single event types prevents the identification of deep causal relationships between them; second, the lack of a dynamic screening mechanism for core influencing factors within an event sequence makes it prone to noise interference.

[0003] Furthermore, most demand forecasting models rely heavily on manually defined rules or large-scale historical samples, making them inadequate for the dynamic business landscape of startups. When a company is experiencing rapid growth, its resource needs adjust rapidly based on market feedback. However, traditional recommendation systems, due to lagging profile updates, often experience resource mismatches. These include a disconnect between the recommendation cycle and the company's actual development stage, or a mismatch between resource types and the company's current capabilities.

[0004] Although graph neural networks and time series modeling technologies have been gradually applied to enterprise analysis in recent years, existing attempts still have three flaws: first, the timestamp information of event data has not been effectively converted into time-dependent features; second, the construction of dynamic profiles is separated from the resource recommendation process, resulting in the inability of prediction results to guide decision-making in real time; finally, the system lacks a closed-loop feedback mechanism, and cannot actively correct the model when actual business deviates from the prediction.

[0005] The above defects together make the recommendation results lack of foresight, making it difficult to support agile decision-making of small and medium-sized enterprises in complex environments. Summary of the Invention

[0006] Based on this, it is necessary to provide a resource recommendation method, device, equipment and medium based on dynamic portraits of small businesses to address the above technical problems.

[0007] In a first aspect, this application provides a resource recommendation method based on dynamic profiles of small businesses, including:

[0008] S1. Based on the enterprise multidimensional dataset, identify key event types and standardize the original timestamps of events corresponding to each key event type to obtain an initial event sequence; wherein the enterprise multidimensional dataset includes industrial and commercial data, financing data, product data, and market data;

[0009] S2. Calculate importance scores based on the event frequencies and type weights in the initial event sequence, and filter the importance scores using an adaptive threshold to obtain a core event set.

[0010] S3. Based on the timestamps and business semantic features of each event in the core event set, calculate the inter-event dependency strength matrix; and based on the inter-event dependency strength matrix, establish directed edges through causal reasoning to obtain an event relationship graph;

[0011] S4. Using the event relationship graph as input, we learn event embedding features through a graph neural network, and then process the event embedding features with a gated recurrent unit to obtain a dynamic enterprise portrait vector.

[0012] S5. Based on the enterprise dynamic portrait vector, the impact weights of historical events are calculated using a time series model with an attention mechanism to obtain a future demand probability distribution that includes the probabilities of various types of demand.

[0013] S6. Calculate resource urgency based on the probability distribution of future demand, and construct an objective function based on resource urgency and the matching degree between resources and demand. Solve the objective function under the constraints of enterprise capabilities to obtain a forward-looking resource recommendation list.

[0014] S7. By comparing the deviation between the probability distribution of future demand and the actual development data, the incremental update of the event relationship graph and the parameter retraining of the time series model are triggered.

[0015] In a second aspect, the present application also provides a resource recommendation device based on dynamic portraits of small businesses, including:

[0016] The event processing module is used to identify key event types based on enterprise multidimensional datasets, and to standardize the original timestamps of events corresponding to each key event type to obtain an initial event sequence. The enterprise multidimensional dataset includes industrial and commercial data, financing data, product data, and market data.

[0017] Importance scoring and filtering module, which is used to calculate importance scores based on the frequency and type weight of events in the initial event sequence, and filter the importance scores through adaptive thresholds to obtain the core event set;

[0018] The event relationship construction module is used to calculate the inter-event dependency strength matrix based on the timestamps and business semantic features of each event in the core event set. Based on the inter-event dependency strength matrix, it establishes directed edges through causal reasoning to obtain an event relationship graph.

[0019] The enterprise portrait generation module uses the event relationship graph as input, learns event embedding features through a graph neural network, and then processes the event embedding features with a gated recurrent unit to obtain a dynamic enterprise portrait vector.

[0020] The demand distribution calculation module is used to calculate the impact weights of historical events based on the enterprise dynamic profile vector through a time series model with an attention mechanism to obtain the future demand probability distribution that includes the probability of each type of demand;

[0021] The resource recommendation module calculates resource urgency based on the probability distribution of future demand and constructs an objective function based on resource urgency and the matching degree between resources and demand. The objective function is solved under the constraints of enterprise capabilities to obtain a forward-looking resource recommendation list.

[0022] The update optimization module is used to trigger incremental updates of the event relationship graph and parameter retraining of the timing model by comparing the deviation between the probability distribution of future demand and the actual development data.

[0023] In a third aspect, the present application also provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, it implements a resource recommendation method based on dynamic portraits of small businesses as in the first aspect.

[0024] In a fourth aspect, the present application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a resource recommendation method based on dynamic portraits of small businesses as in the first aspect.

[0025] The above-mentioned resource recommendation method, device, equipment and medium based on the dynamic portrait of small enterprises extract key events from the enterprise's multi-dimensional data and perform standardized processing to form an initial event sequence, calculate the importance score based on the event frequency and type weight, and adaptively filter to obtain the core event set, then calculate the dependency strength between events based on the timestamp and business semantic features, and construct an event relationship graph through causal reasoning; then use the graph neural network to learn the event embedding features, combine the gated recurrent unit to model the time series features to generate the enterprise dynamic portrait vector; on this basis, use the time series model with an attention mechanism to analyze the impact weight of historical events to predict the probability distribution of future demand; calculate the resource urgency and matching degree based on the distribution, and perform multi-objective optimization solution under the enterprise capacity constraints to generate a forward-looking resource recommendation list; finally, by comparing the deviation between the predicted results and the actual data, trigger the graph incremental update and model retraining to form an adaptive closed loop, realize the dynamic capture of the enterprise development trajectory and the accurate prediction of resource demand, and significantly improve the timeliness and business adaptability of the recommendation results. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] In order to more clearly illustrate the technical solutions in the embodiments of the present application or related technologies, the following briefly introduces the drawings required for use in the embodiments or related technical descriptions. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0027] Figure 1 A flowchart of a resource recommendation method based on dynamic portraits of small businesses provided by the present invention;

[0028] Figure 2 A schematic diagram of a process for generating a dynamic enterprise portrait vector in an optional embodiment of the present invention;

[0029] Figure 3 A structural diagram of a resource recommendation device based on dynamic portraits of small businesses provided by the present invention. DETAILED DESCRIPTION

[0030] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0031] refer to Figure 1 , which presents a flow chart of a resource recommendation method based on a dynamic portrait of a small business provided by this application, the method comprising the following steps:

[0032] S1. Based on the enterprise multidimensional dataset, identify key event types, standardize the original timestamps of events corresponding to each key event type, and obtain an initial event sequence; wherein the enterprise multidimensional dataset includes industrial and commercial data, financing data, product data, and market data.

[0033] Specifically, the first step is to construct a multidimensional enterprise dataset, which comprehensively covers multiple dimensions, including industrial and commercial data, financing data, product data, and market data, to fully reflect the operating conditions of small businesses. Industrial and commercial data encompasses enterprise registration information, such as company name, registered address, registered capital, business scope, legal representative, and equity structure. This data can be legally obtained and compiled from the information platforms of industrial and commercial administrative departments. Financing data includes key information such as the company's past financing rounds, financing amount, financing time, investor background, and investment proportion. This data is collected and aggregated through financial data service agencies, company-specific disclosures, and industry financing dynamics monitoring channels. Product data covers the name, category, technical parameters, functional characteristics, iteration update time, and market feedback evaluation of the company's main products or services. This data is integrated through the company's internal product management system, market research agency reports, and user feedback data. Market data covers various aspects of the company's industry market size, competitive landscape, market share changes, marketing channel expansion, customer group characteristics, and geographical distribution. This data is formed by integrating reports released by third-party market research institutions, industry statistics, and the company's own sales business data.

[0034] Based on this multidimensional dataset, data mining and analysis algorithms are used to identify key event types closely related to corporate development, such as successful financing, major product upgrades, and entry into new markets. This can be achieved, for example, through pre-trained language models.

[0035] For the original timestamps of events corresponding to each key event type, since the time formats of different data sources are different, they need to be standardized and converted into a predefined standard time format, such as the ISO 8601 format, so as to construct an initial event sequence arranged in chronological order and in a unified format.

[0036] S2. Calculate the importance score based on the event frequency and type weight in the initial event sequence, filter the importance score through the adaptive threshold, and obtain the core event set.

[0037] Specifically, the frequency of an event reflects the level of activity of the event within the life cycle of the enterprise. For example, if product iteration events occur frequently, it indicates that the enterprise's products are in a stage of rapid updating. Type weights can be pre-set based on expert experience or historical data analysis. Different key event types have different degrees of impact on the development of the enterprise. For example, the weight of a financing event is usually higher than that of a regular organizational structure adjustment event within the enterprise. Through a specific weighted calculation formula, such as normalizing the frequency of an event and multiplying it by the corresponding type weight, and then adding the sum, the importance score of each event is obtained. Next, an adaptive threshold filtering mechanism is adopted. The threshold is not a fixed value, but is dynamically determined based on the overall score distribution of the current event set. For example, the threshold is set to the range of the current score mean plus or minus several times the standard deviation, filtering out non-core events below the threshold, and screening out a set of core events that have a significant impact on the enterprise's resource needs, to ensure that subsequent analysis focuses on key business activities.

[0038] S3. Based on the timestamps and business semantic features of each event in the core event set, the inter-event dependency strength matrix is ​​calculated; and based on the inter-event dependency strength matrix, directed edges are established through causal reasoning to obtain an event relationship graph.

[0039] Specifically, the timestamp information of each event is extracted, which is the key basis for measuring the time interval between events and constructing temporal relationships; for mining the business semantic features of events, such as the scale of funds corresponding to financing events, the industry influence of investors, the product function innovations associated with product events, the expansion of target user groups, and other attribute characteristics that can reflect the business connotation of events. Based on the above information, the dependency strength matrix between events is calculated, and statistical analysis and association rule mining algorithms are used to quantitatively analyze the strength of the chronological dependency relationship between different events. For example, there is a direct causal dependency between the first product development event and the subsequent product launch event, and the dependency strength is relatively high. Then, with the help of causal reasoning algorithms, such as the causal inference method based on Bayesian networks, the directed edge connections between events are determined based on the numerical relationship in the dependency strength matrix, and a complete event relationship map is constructed to intuitively present the development context and internal connections of key enterprise events, providing a structured representation for in-depth understanding of the dynamic development logic of the enterprise.

[0040] S4. Using the event relationship graph as input, we learn event embedding features through the graph neural network, and then use the gated recurrent unit to process the event embedding features to obtain the enterprise dynamic portrait vector.

[0041] Specifically, using a constructed event relationship graph as input, a graph neural network (GNN) is introduced to learn the nodes (events) and their edges (dependencies) in the graph. Using a message-passing mechanism, the GNN propagates and aggregates features on graph-structured data, learning the embedded features of each event and capturing both the event's own attributes and its associations with surrounding events. Furthermore, a gated recurrent unit (GRU) is incorporated. Because event sequences have time series characteristics, the GRU can effectively process temporal dependencies within the sequence, memorize the impact of previous events on the enterprise's status, and dynamically update the enterprise's profile. Specifically, the event embedding feature sequence output by the GNN is input into the GRU network. The GRU's hidden state is updated with time steps, gradually integrating the temporal dynamic information of events. Ultimately, a dynamic enterprise profile vector is generated. This vector dynamically and in real time reflects the comprehensive characteristics of the enterprise at different stages of development, integrating multi-dimensional information such as event content, occurrence sequence, and interrelationships.

[0042] S5. Based on the enterprise dynamic portrait vector, the impact weight of historical events is calculated through a time series model with an attention mechanism to obtain the future demand probability distribution that includes the probability of each type of demand.

[0043] Specifically, time series models, such as long short-term memory networks (LSTMs) or the time series module in Transformers, are used to process time series data of events from a company's historical development. The introduction of the attention mechanism aims to allow the model to automatically learn the different impact weights of historical events on a company's future resource needs, rather than treating all historical events uniformly. For example, a recent financing event may have a higher impact on subsequent resource expansion needs than a similar event from a more distant period. Through model training, the impact weight of each historical event is calculated, and then a future demand probability distribution is predicted, which includes the probabilities of various types of demand. This covers different dimensions, such as the probability of funding demand, the probability of talent recruitment demand, and the probability of marketing resource demand, providing a forward-looking probabilistic quantitative basis for enterprise resource planning.

[0044] S6. Calculate resource urgency based on the probability distribution of future demand, and construct an objective function based on resource urgency and the matching degree between resources and demand; solve the objective function under the constraints of enterprise capabilities to obtain a forward-looking resource recommendation list.

[0045] Specifically, resource urgency is calculated based on the probability distribution of future demand. This can involve a comprehensive consideration of factors such as the time sensitivity of different resource demand types and the intensity of market competition. For example, if funding demand is approaching a peak in maturing debt repayments, its urgency is higher; the demand for specific high-end talent also increases when the current talent market is in short supply. Based on resource urgency and predefined resource-demand matching assessment rules, such as the matching between funding resources and funding needs in terms of amount and investor preferences, and the matching between talent resources and talent needs in terms of professional skills and experience requirements, an optimization objective function is constructed. This objective function aims to maximize the utility of resource allocation. Within the constraints of enterprise capabilities, such as the enterprise's current financial capacity and the upper limit of the organizational structure to accommodate talent, mathematical optimization algorithms such as linear programming and integer programming are used to solve the objective function. The result is a forward-looking resource recommendation list, providing enterprises with accurate and timely resource acquisition guidance.

[0046] S7. By comparing the deviation between the probability distribution of future demand and the actual development data, the incremental update of the event relationship graph and the parameter retraining of the time series model are triggered.

[0047] Specifically, a feedback optimization mechanism is established to compare the predicted probability distribution of future demand with the actual collected data on subsequent enterprise development. Actual development data can be obtained through continuous monitoring of enterprise operating systems, market feedback channels, and other channels. Deviation calculations can use indicators such as mean square error and absolute error to quantify the difference between the two. When the deviation exceeds the preset threshold, the incremental update process of the event relationship map is triggered, and the newly occurring events and their associations with existing events are added to the map, keeping the map tracking the enterprise's development dynamics in real time. At the same time, the parameter retraining process of the time series model is initiated, and the model parameters are adjusted using new actual data to optimize the model's prediction accuracy for future demand, ensuring that the entire resource recommendation method can adapt to the dynamic changes in the enterprise's environment, maintain the effectiveness and reliability of the recommendation results, and form a resource recommendation system with continuous optimization and closed-loop feedback.

[0048] The above-mentioned resource recommendation method based on the dynamic portrait of small enterprises extracts key events from the enterprise's multi-dimensional data and standardizes them to form an initial event sequence, calculates the importance score based on the event frequency and type weight, and adaptively filters to obtain the core event set, then calculates the dependency strength between events based on timestamps and business semantic features, and constructs an event relationship graph through causal reasoning; then uses graph neural networks to learn event embedding features, combines gated recurrent units to model time series features to generate enterprise dynamic portrait vectors; on this basis, uses a time series model with an attention mechanism to analyze the impact weights of historical events to predict the probability distribution of future demand; calculates the resource urgency and matching degree based on the distribution, and performs multi-objective optimization under the constraints of enterprise capabilities to generate a forward-looking resource recommendation list; finally, by comparing the deviation between the predicted results and the actual data, triggers the incremental update of the graph and the retraining of the model to form an adaptive closed loop, realizing the dynamic capture of the enterprise development trajectory and the accurate prediction of resource needs, significantly improving the timeliness and business adaptability of the recommendation results.

[0049] In an optional embodiment, S2 includes the following steps:

[0050] S21. According to the event type, an impact weight coefficient is assigned to the event in the initial event sequence as a type weight.

[0051] Specifically, after the initial event sequence is constructed, impact weight coefficients, or type weights, are assigned to the events in the sequence based on their type. The different type weights can be determined based on expert experience, industry practices, and a comprehensive consideration of the degree of impact on the company's development. For example, a successful financing event significantly boosts the company's development and thus receives a higher weight; whereas a routine internal meeting event, which has a relatively small impact on overall resource demand, receives a lower weight. In this way, each event is assigned an impact weight coefficient that reflects its potential impact.

[0052] S22. Calculate the importance score of each event in the initial event sequence based on the impact weight coefficient and the frequency of occurrence. The calculation formula for the importance score is:

[0053] ;

[0054] in, is the i-th event, represents the importance score of the i-th event, α and β are preset parameters, Representing an event The number of occurrences within the set time window, is the impact weight coefficient of the i-th event.

[0055] Specifically, after assigning type weights, the importance score of each event is calculated based on the frequency of events. Frequency refers to the number of times an event occurs within a set time window. This indicator reflects the activity level of the event and the intensity of its impact on the company's recent development. For example, if multiple product iteration events occur consecutively within a month, it means that the company is actively promoting product optimization and the activity level of such events is high. , combining the impact weight coefficient and frequency of occurrence. α and β are preset parameters that balance the relative contributions of frequency and type weight in the importance calculation. Taking the logarithm prevents excessively high and rapid increases in importance scores due to excessive frequency, resulting in more reasonable and smooth calculation results. This formula quantitatively assesses the importance of each event at the current stage of a company's development.

[0056] S23, adjust the threshold θ according to the importance scores of all events in the initial event sequence, and filter Events, get the core event set; among them, , is the average importance score, is the standard deviation of importance rating, is the preset coefficient.

[0057] Specifically, after obtaining the importance scores of all events in the initial event sequence, we need to filter out the core events that have a significant impact on the enterprise's resource needs. First, calculate the average importance score of all events in the initial event sequence. and standard deviation These two statistics can reflect the central tendency and dispersion of importance scores. Then, according to the formula To adjust the threshold, It is a preset coefficient used to control the strictness of the screening. When the threshold is determined, the importance score of each event is compared with the threshold to screen out the events that meet the threshold. These events are the core event set. This screening method can dynamically adapt to the importance and distribution characteristics of different enterprises, ensuring that the selected core events can accurately reflect the current key development trends and resource demand priorities of the enterprise.

[0058] In an optional embodiment, S3 includes the following steps:

[0059] S31, for any event in the core event set The dependency strength is calculated to obtain the inter-event dependency strength matrix; the calculation formula of dependency strength is:

[0060] ;

[0061] in, Representing an event and events The strength of the dependence between Representing an event timestamp, Representing an event timestamp; is the time decay rate, which is used to control the time decay rate; Representing an event and The business semantic similarity between them is calculated by the cosine similarity of word vectors.

[0062] Specifically, first, for any two events in the core event set and The dependence strength is calculated by the formula , comprehensively considers the time factor and business semantic factors between two events. Among them, the time factor is expressed by the time decay rate To quantify, The setting can effectively control the decay rate of time effects, so that the dependency between events gradually weakens as the time difference increases. This is because the time span between events is often inversely proportional to their direct correlation. The semantic similarity between two events is converted into numerical values ​​using word embedding technology and calculated using cosine similarity. This step organically combines time and business semantic factors to construct an inter-event dependency strength matrix, providing a foundation for subsequent causal relationship analysis.

[0063] S32, when the event and events When the target condition is met, the slave event is established Pointing to events Directed causal edges , and obtain the causal edge set; wherein, the target condition is And Granger causality test probability , and is the preset value.

[0064] Specifically, the core of this step is to set target conditions to screen out significant causal relationships, requiring events to and Dependence strength Must be greater than a preset threshold To ensure that only those event pairs with strong dependencies are considered to have causal relationships. At the same time, the Granger causality test probability is also introduced. To further verify the rationality of the causal relationship, only when this probability is greater than another preset threshold Time, event and The causal relationship between them is considered reliable. By jointly screening these two conditions, a reliable causal edge set can be established, in which each causal edge Both represent events About the event significant causal influence.

[0065] S33. With the core event set as the node set and the causal edge set as the edge set, a directed graph data structure is constructed to obtain an event relationship graph.

[0066] Specifically, a directed graph data structure is constructed using the events in the core event set as a node set and the directed causal edges in the causal edge set as an edge set, resulting in an event relationship graph. This graph not only intuitively displays the causal relationships between core events but also provides a structured data foundation for subsequent graph-based in-depth analysis and prediction. Each node in the graph represents a key event, and each directed edge indicates the causal relationship between events. This structured representation enables the system to more effectively perform graph analysis, such as path analysis and community detection, thereby providing enterprise decision makers with clearer and more intuitive business insights and development trend predictions.

[0067] refer to Figure 2 In an optional embodiment, S4 includes the following steps:

[0068] S41. Convert each event in the core event set into a feature vector through a word embedding model; combine the feature vectors of all events to form a node feature matrix.

[0069] Specifically, after obtaining the core event set, we first need to convert the feature vector of each event, which can be achieved through the word embedding model. is mapped to a fixed-dimensional feature vector The word embedding model can capture the semantic information of words in event descriptions, converting discrete text data into continuous vector representations, so that events with similar semantics are closer in the vector space. The feature vectors of all events are combined to form a node feature matrix , each row of the matrix corresponds to the feature vector of an event, providing the basic data input format for subsequent graph neural network processing.

[0070] S42. Extract the adjacency matrix of the event relationship graph, and normalize the adjacency matrix to obtain a normalized adjacency matrix.

[0071] Specifically, the adjacency matrix A of the event relationship graph is a key data structure that represents the connection relationship between events. The elements in the adjacency matrix reflect the degree of association between events, and non-zero elements indicate that there are pairs of events that are associated. However, the original adjacency matrix may contain values ​​of different magnitudes, which will affect the numerical stability and convergence speed of subsequent graph convolution operations. Therefore, it is necessary to normalize the adjacency matrix. Normalization methods can include strategies such as row normalization or symmetric normalization, which adjust the elements of the adjacency matrix to be in a suitable numerical range, such as normalizing the sum of the elements in each row to 1, or using other normalization formulas to ensure numerical stability and consistency. The normalized adjacency matrix can be used more effectively for graph convolution operations, ensuring that information is transmitted more balanced and accurately in the graph.

[0072] S43. Perform graph convolution operation on the normalized adjacency matrix and the node feature matrix, and output the event embedding vector.

[0073] Specifically, the graph convolution operation is one of the core steps in graph neural networks, which aims to combine the feature information of events and the structural relationship between them to generate a richer event embedding representation. The normalized adjacency matrix is ​​multiplied by the node feature matrix. This step actually aggregates the feature information of each event and its neighboring events. In this way, each event can integrate the features of its neighboring events, thereby capturing the local dependencies between events. In actual calculations, graph convolution operations can involve multiple iterations and nonlinear transformations to gradually extract higher-level feature representations. For example, taking the node feature matrix and the normalized adjacency matrix For input, execute Layer Graph Convolution: Finally, the graph convolution operation outputs the output event embedding vector set , is a set of core events. Each event embedding vector contains comprehensive information about each event and its surrounding events, providing a more comprehensive feature basis for subsequent time series processing.

[0074] S44. According to the time sequence of the initial event sequence, the corresponding event embedding vector is input into the gated recurrent unit, and the time-step hidden state vector is output.

[0075] Specifically, the Gated Recurrent Unit (GRU) is a neural network structure specifically designed to process sequential data, effectively capturing temporal dependencies within sequences. In this step, the corresponding event embedding vectors are sequentially input into the GRU according to the chronological order of the initial event sequence. The GRU internally contains structures such as update gates and reset gates. These gating mechanisms dynamically control the flow of information, determining which information should be retained, updated, or forgotten. In this way, the GRU progressively updates its hidden state, generating a hidden state vector at each time step. This hidden state vector not only contains information about the current event but also incorporates sequence information from all previous events, thereby forming a temporal model of the dynamic development process of the enterprise. The hidden state vector at each time step is effectively the GRU's comprehensive representation of the current event and its historical events, capturing the dynamic characteristics of events as they evolve over time.

[0076] S45. Integrate all time-step hidden state vectors to construct a dynamic enterprise portrait vector.

[0077] Specifically, after obtaining the latent state vectors for all time steps, these vectors are integrated into a single vector that comprehensively represents the dynamic characteristics of the enterprise, namely the enterprise dynamic profile vector. This integration method can use a simple concatenation operation, sequentially connecting all latent state vectors into a long vector, or adopt more complex strategies, such as weighted summation of latent state vectors, where the weights can be determined based on the importance of the time step or other business rules. The construction of the enterprise dynamic profile vector aims to condense the key events and evolution of the enterprise at different periods into a comprehensive feature representation, providing a compact and information-rich foundation for subsequent resource recommendation and decision analysis. This vector can reflect the latest development status of the enterprise in real time and integrate the impact of historical events, providing strong support for the prediction and analysis of enterprise resource needs.

[0078] In an optional embodiment, S5 includes the following steps:

[0079] S51. Perform attention calculation based on the time step hidden state vector corresponding to each historical time step in the enterprise dynamic portrait vector and the time step hidden state vector at the current moment to obtain the attention weight corresponding to each historical time step; the formula for attention calculation is:

[0080] ;

[0081] in, For the The attention weight corresponding to the time step, is a learnable vector, is the learnable weight matrix, For the The time-step hidden state vector corresponding to the time step is is the hidden state vector of the current time step, Indicates that and Perform splicing.

[0082] Specifically, this step aims to determine the degree of influence of the time-step hidden state vector corresponding to each historical time step on the current state of the enterprise through attention calculation, that is, the attention weight. Attention calculation is performed based on the time-step hidden state vector corresponding to each historical time step in the enterprise dynamic portrait vector and the time-step hidden state vector at the current moment. The attention calculation formula used in this step is: .in, Indicates the The attention weight corresponding to the time step, is a learnable vector, is the learnable weight matrix, It is The time-step hidden state vector corresponding to the time step is is the hidden state vector at the current time step, Indicates that and This formula is achieved by concatenating the hidden state vectors of the historical time step and the current time step, multiplying them with the weight matrix, and then passing them through the tanh activation function and The linear transformation of the vector is performed, and finally the softmax function is applied to normalize the output weights to obtain the attention weights for each historical time step. This attention mechanism can dynamically focus on the historical information most relevant to the current moment, providing a basis for subsequent context vector generation.

[0083] S52. Based on the attention weight, the time-step hidden state vectors corresponding to all historical time steps are weighted summed to obtain the context vector. The weighted summation formula is:

[0084] ;

[0085] in, is the context vector, is the number of time steps.

[0086] Specifically, after obtaining the attention weight of each historical time step, this step integrates the time step hidden state vectors corresponding to all historical time steps through weighted summation to obtain the context vector. The specific calculation formula is: .in, is the context vector, is the number of time steps. This context vector comprehensively considers the hidden state vector and its corresponding attention weight at each historical time step, highlighting important historical information while suppressing irrelevant information. It provides a comprehensive and focused historical context representation for subsequent prediction of future demand probability distributions.

[0087] S53: Input the context vector into a multi-layer perceptron and output the probability distribution of future demand.

[0088] Specifically, the multilayer perceptron (MLP) possesses nonlinear mapping capabilities, enabling it to transform complex patterns in context vectors into probabilistic predictions of a company's future resource needs. Through pre-training, the MLP learns the mapping from historical context to future demand probabilities, providing companies with quantitative predictions of the probability of different resource demands. These probability distributions provide a basis for subsequent resource recommendation decisions, helping companies plan and allocate resources in advance to meet future development needs.

[0089] In an optional embodiment, S6 includes the following steps:

[0090] S61. Based on the urgency mapping formula, the probability value in the future demand probability distribution is converted into resource urgency. The resource urgency mapping formula is:

[0091] ;

[0092] in, for The resource urgency of the type of demand, for Preset weights for type requirements; for The probability value of the type demand in the future demand probability distribution, represents the maximum probability value in the future demand probability distribution.

[0093] Specifically, the core of this step is to convert the probability value in the future demand probability distribution into resource urgency, so as to more intuitively reflect the urgency of different types of resource demands. To achieve this conversion. express The resource urgency of the type of demand, yes The preset weight of the type of demand reflects the importance of the demand type in the enterprise strategy; yes The probability value corresponding to the type of demand in the probability distribution of non-demand represents the possibility of the demand occurring; The maximum probability value in the probability distribution of unrequired resources is used to normalize the probability values. Normalization ensures that the urgency of different demand types is comparable. The product form in the formula comprehensively considers the importance and probability of a demand, and the resulting resource urgency effectively indicates the types of demands that the company should prioritize.

[0094] S62: Obtain the matching degree between each resource and each requirement by calculating the dot product or cosine similarity between the embedding vector of each resource and the embedding vector of each requirement.

[0095] Specifically, this step aims to calculate the degree of match between each resource and each requirement, providing a quantitative basis for subsequent resource recommendations. The calculation method includes finding the dot product or cosine similarity between the embedding vector of each resource and the embedding vector of each requirement. The embedding vector is obtained by mapping the feature information of resources and requirements into the same semantic space, and can capture the characteristics of resources and requirements in multiple dimensions. Dot product or cosine similarity can measure the similarity between two vectors. The dot product reflects the correlation between the two vectors in direction, and the cosine similarity measures the size of the angle between the two vector directions. The larger the value, the higher the match, and vice versa. In this way, the degree of match between each resource and each requirement can be obtained, providing quantitative indicators for subsequent resource recommendations, helping enterprises to accurately locate suitable resources to meet their needs.

[0096] S63. Based on the enterprise capacity constraints, solve the objective function constructed by resource urgency and matching degree to obtain a forward-looking resource recommendation list; the expression of the objective function is:

[0097] ;

[0098] in, is the preset weight coefficient, For resources and demand The matching degree, A collection of resources.

[0099] Specifically, in the process of constructing the objective function and solving it to generate a forward-looking resource recommendation list, it is first necessary to construct the objective function based on the enterprise's capacity constraints. Enterprise capacity constraints include capital limits, manpower limits, management complexity limits, etc. These constraints define the feasible solution space for the enterprise in resource allocation. The objective function is constructed by resource urgency and matching degree, and the expression is .in, It is a preset weight coefficient used to balance the relative importance of urgency and matching in the objective function; For resources and demand The degree of matching reflects the degree to which resources meet demand. Solving this objective function aims to find a set of resource allocation solutions that maximizes the combined benefits of urgency and matching while satisfying the enterprise's capacity constraints. This solution can employ mathematical optimization techniques such as linear programming and integer programming. The resulting forward-looking resource recommendation list will guide enterprises to prioritize resources that both meet urgent needs and efficiently match resources to support their future development.

[0100] The above-mentioned resource recommendation method based on the dynamic portrait of small enterprises extracts key events from the enterprise's multi-dimensional data and standardizes them to form an initial event sequence, calculates the importance score based on the event frequency and type weight, and adaptively filters to obtain the core event set, then calculates the dependency strength between events based on timestamps and business semantic features, and constructs an event relationship graph through causal reasoning; then uses graph neural networks to learn event embedding features, combines gated recurrent units to model time series features to generate enterprise dynamic portrait vectors; on this basis, uses a time series model with an attention mechanism to analyze the impact weights of historical events to predict the probability distribution of future demand; calculates the resource urgency and matching degree based on the distribution, and performs multi-objective optimization under the constraints of enterprise capabilities to generate a forward-looking resource recommendation list; finally, by comparing the deviation between the predicted results and the actual data, triggers the incremental update of the graph and the retraining of the model to form an adaptive closed loop, realizing the dynamic capture of the enterprise development trajectory and the accurate prediction of resource needs, significantly improving the timeliness and business adaptability of the recommendation results.

[0101] It should be understood that, although the various steps in the flowcharts involved in the various embodiments described above are displayed in sequence according to the instructions of the arrows, these steps are not necessarily executed in sequence in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction on the execution of these steps, and these steps can be executed in other orders. Moreover, at least a portion of the steps in the flowcharts involved in the various embodiments described above can include multiple steps or multiple stages, and these steps or stages are not necessarily executed and completed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a portion of steps or stages in other steps.

[0102] Based on the same inventive concept, the embodiments of the present application also provide a device for implementing the aforementioned method for recommending resources based on dynamic small business profiles. The solution provided by this device is similar to the solution described in the aforementioned method. Therefore, the specific limitations of one or more embodiments of the device for recommending resources based on dynamic small business profiles provided below can be found in the above-mentioned limitations of the method for recommending resources based on dynamic small business profiles, and will not be repeated here.

[0103] In an exemplary embodiment, Figure 3 As shown, a resource recommendation device 30 based on a dynamic portrait of a small business is provided, comprising:

[0104] The event processing module 31 is used to identify key event types based on the enterprise multidimensional data set, standardize the original timestamps of events corresponding to each key event type, and obtain an initial event sequence; wherein the enterprise multidimensional data set includes industrial and commercial data, financing data, product data and market data.

[0105] The importance scoring and filtering module 32 is used to calculate the importance scores according to the event occurrence frequency and type weight in the initial event sequence, and filter the importance scores through an adaptive threshold to obtain a core event set.

[0106] The event relationship construction module 33 is used to calculate the inter-event dependency strength matrix based on the timestamps and business semantic features of each event in the core event set; and based on the inter-event dependency strength matrix, establish directed edges through causal reasoning to obtain an event relationship graph.

[0107] The enterprise portrait generation module 34 is used to take the event relationship graph as input, learn event embedding features through the graph neural network, and then combine the gated recurrent unit to process the event embedding features to obtain the enterprise dynamic portrait vector.

[0108] The demand distribution calculation module 35 is used to calculate the impact weight of historical events based on the enterprise dynamic portrait vector through a time series model with an attention mechanism to obtain the future demand probability distribution that includes the probability of each type of demand.

[0109] The resource recommendation module 36 is used to calculate resource urgency based on the probability distribution of future demand, construct an objective function based on resource urgency and the matching degree between resources and demand; solve the objective function under the constraint of enterprise capacity to obtain a forward-looking resource recommendation list.

[0110] The update optimization module 37 is used to trigger the incremental update of the event relationship graph and the parameter retraining of the time series model by comparing the deviation between the future demand probability distribution and the actual development data.

[0111] Optional importance scoring and filtering modules include:

[0112] The event weight assignment unit is used to assign an impact weight coefficient to the events in the initial event sequence according to the event type, as a type weight.

[0113] The event score calculation unit is used to calculate the importance score of each event in the initial event sequence based on the impact weight coefficient and the occurrence frequency. The calculation formula of the importance score is:

[0114] ;

[0115] in, is the i-th event, represents the importance score of the i-th event, α and β are preset parameters, Representing an event The number of occurrences within the set time window, is the impact weight coefficient of the i-th event.

[0116] The core event screening unit is used to adjust the threshold θ according to the importance scores of all events in the initial event sequence, and screen events that meet Events, get the core event set; among them, , is the average importance score, is the standard deviation of importance rating, is the preset coefficient.

[0117] Optionally, event relationship building blocks include:

[0118] Dependency strength calculation unit, used to calculate the dependency strength of any event in the core event set. The dependency strength is calculated to obtain the inter-event dependency strength matrix; the calculation formula of dependency strength is:

[0119] ;

[0120] in, Representing an event and events The strength of the dependence between Representing an event timestamp, Representing an event timestamp; is the time decay rate, which is used to control the time decay rate; Representing an event and The business semantic similarity between them is calculated by the cosine similarity of word vectors.

[0121] Causal edge establishment unit, used when events and events When the target condition is met, the slave event is established Pointing to events Directed causal edges , and obtain the causal edge set; among them, the target condition is And Granger causality test probability , and is the preset value.

[0122] The graph construction unit is used to construct a directed graph data structure with the core event set as the node set and the causal edge set as the edge set to obtain an event relationship graph.

[0123] Optional enterprise portrait generation module includes:

[0124] The event vector conversion unit is used to convert each event in the core event set into a feature vector through the word embedding model; and combine the feature vectors of all events to form a node feature matrix.

[0125] The adjacency matrix processing unit is used to extract the adjacency matrix of the event relationship graph and normalize the adjacency matrix to obtain a normalized adjacency matrix.

[0126] The graph convolution operation unit is used to perform graph convolution operations on the normalized adjacency matrix and the node feature matrix, and output an event embedding vector.

[0127] The time series processing unit is used to input the corresponding event embedding vector into the gated recurrent unit according to the time order of the initial event sequence, and output the time-step hidden state vector.

[0128] The portrait vector construction unit is used to integrate all time-step hidden state vectors to construct a dynamic portrait vector of the enterprise.

[0129] Optionally, the demand distribution calculation module includes:

[0130] The attention weight calculation unit is used to perform attention calculation based on the time step hidden state vector corresponding to each historical time step in the enterprise dynamic portrait vector and the time step hidden state vector at the current moment to obtain the attention weight corresponding to each historical time step; the formula for attention calculation is:

[0131] ;

[0132] in, For the The attention weight corresponding to the time step, is a learnable vector, is the learnable weight matrix, For the The time-step hidden state vector corresponding to the time step is is the hidden state vector of the current time step, Indicates that and Perform splicing.

[0133] The context vector generation unit is used to perform weighted summation of the time-step hidden state vectors corresponding to all historical time steps based on the attention weight to obtain the context vector; the weighted summation formula is:

[0134] ;

[0135] in, is the context vector, is the number of time steps.

[0136] The demand distribution prediction unit is used to input the context vector into the multi-layer perceptron and output the future demand probability distribution.

[0137] Optionally, the resource recommendation module includes:

[0138] The resource urgency calculation unit is used to convert the probability value in the future demand probability distribution into resource urgency based on the urgency mapping formula. The resource urgency mapping formula is:

[0139] ;

[0140] in, for The resource urgency of the type of demand, for Preset weights for type requirements; for The probability value of the type demand in the future demand probability distribution, represents the maximum probability value in the future demand probability distribution.

[0141] The resource requirement matching degree calculation unit is used to obtain the matching degree between each resource and each requirement by calculating the dot product or cosine similarity between the embedding vector of each resource and the embedding vector of each requirement.

[0142] The resource recommendation solving unit is used to solve the objective function constructed by resource urgency and matching based on enterprise capacity constraints to obtain a forward-looking resource recommendation list; the objective function expression is:

[0143] ;

[0144] in, is the preset weight coefficient, For resources and demand The matching degree, A collection of resources.

[0145] An embodiment of the present application further provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps in the above-mentioned method embodiments when executing the computer program.

[0146] An embodiment of the present application further provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps in the above-mentioned method embodiments are implemented.

[0147] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to the partial description of the method embodiments. The device embodiments described above are merely illustrative, wherein the components described as separate parts may or may not be physically separated, and the parts displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the disclosed solution. A person of ordinary skill in the art can understand and implement it without expending creative work.

[0148] The above-described embodiments merely represent several implementation methods of the embodiments of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the patent application. It should be noted that a person skilled in the art may make various modifications and improvements without departing from the concept of the embodiments of the present application, and these modifications and improvements fall within the scope of protection of the embodiments of the present application.

Claims

1. A resource recommendation method based on dynamic portraits of small businesses, characterized in that: The method comprises: S1. Based on an enterprise multidimensional dataset, identify key event types and normalize the original timestamps of events corresponding to each key event type to obtain an initial event sequence; wherein the enterprise multidimensional dataset includes industrial and commercial data, financing data, product data, and market data; S2. Calculating importance scores based on the occurrence frequencies and type weights of events in the initial event sequence, and filtering the importance scores using an adaptive threshold to obtain a core event set; S3. Calculate an inter-event dependency strength matrix based on the timestamp and business semantic features of each event in the core event set; and establish directed edges based on the inter-event dependency strength matrix through causal reasoning to obtain an event relationship graph; S4. Using the event relationship graph as input, learning event embedding features through a graph neural network, and then processing the event embedding features with a gated recurrent unit to obtain a dynamic enterprise portrait vector; S5. Based on the enterprise dynamic portrait vector, the impact weights of historical events are calculated using a time series model with an attention mechanism to obtain a future demand probability distribution that includes the probabilities of various types of demand. S6. Calculating resource urgency based on the future demand probability distribution, constructing an objective function based on the resource urgency and the matching degree between resources and demand; solving the objective function under the enterprise capacity constraint to obtain a forward-looking resource recommendation list; S7. By comparing the deviation between the future demand probability distribution and the actual development data, triggering the incremental update of the event relationship graph and the retraining of the parameters of the time series model.

2. The method according to claim 1, characterized in that The S2 includes: S21. Allocate an impact weight coefficient to the event in the initial event sequence according to the event type, as the type weight; S22. Calculate the importance score of each event in the initial event sequence based on the impact weight coefficient and the occurrence frequency; the calculation formula of the importance score is: ； in, is the i-th event, represents the importance score of the i-th event, α and β are preset parameters, Representing an event The number of occurrences within the set time window, is the impact weight coefficient of the i-th event; S23, adjusting the threshold θ according to the importance scores of all events in the initial event sequence, screening events, and obtain the core event set; wherein, , is the average importance score, is the standard deviation of importance rating, is the preset coefficient.

3. The method according to claim 2, characterized in that The S3 includes: S31, for any event in the core event set The dependency strength is calculated to obtain the inter-event dependency strength matrix; the calculation formula of the dependency strength is: ； in, Representing an event and events The strength of the dependence between Representing an event timestamp, Representing an event timestamp; is the time decay rate, which is used to control the time decay rate; Representing an event and The business semantic similarity between them is calculated by word vector cosine similarity; S32, when the event and events When the target condition is met, the slave event is established Pointing to events Directed causal edges , and obtain the causal edge set; wherein, the target condition is And Granger causality test probability , and is the preset value; S33. Using the core event set as a node set and the causal edge set as an edge set, a directed graph data structure is constructed to obtain the event relationship graph.

4. The method according to claim 3, characterized in that The S4 includes: S41, converting each event in the core event set into a feature vector using a word embedding model; combining the feature vectors of all events to form a node feature matrix; S42, extracting the adjacency matrix of the event relationship graph, and normalizing the adjacency matrix to obtain a normalized adjacency matrix; S43, performing a graph convolution operation on the normalized adjacency matrix and the node feature matrix, and outputting an event embedding vector; S44, inputting the corresponding event embedding vector into the gated recurrent unit according to the time sequence of the initial event sequence, and outputting a time-stepped hidden state vector; S45. Integrate all the time-step hidden state vectors to construct the enterprise dynamic portrait vector.

5. The method according to claim 4, characterized in that The S5 includes: S51. Perform attention calculation based on the time step hidden state vector corresponding to each historical time step in the enterprise dynamic portrait vector and the time step hidden state vector at the current moment to obtain the attention weight corresponding to each historical time step; the formula for the attention calculation is: ； in, For the The attention weight corresponding to the time step, is a learnable vector, is the learnable weight matrix, For the The time-step hidden state vector corresponding to the time step is is the hidden state vector of the current time step, Indicates that and Perform splicing; S52: Based on the attention weight, perform weighted summation on the time-step hidden state vectors corresponding to all historical time steps to obtain a context vector; the weighted summation formula is: ； in, is the context vector, is the number of time steps; S53: Input the context vector into a multi-layer perceptron, and output the future demand probability distribution.

6. The method according to any one of claims 1 to 5, characterized in that The S6 includes: S61. Convert the probability value in the future demand probability distribution into the resource urgency based on an urgency mapping formula. The resource urgency mapping formula is: ； in, for The resource urgency of the type of demand, for Preset weights for type requirements; for The probability value of the type demand in the future demand probability distribution, represents the maximum probability value in the future demand probability distribution; S62. Obtain the matching degree between each resource and each requirement by calculating the dot product or cosine similarity between the embedding vector of each resource and the embedding vector of each requirement; S63. Based on the enterprise capability constraint, solve the objective function constructed by the resource urgency and the matching degree to obtain the forward-looking resource recommendation list; the expression of the objective function is: ； in, is the preset weight coefficient, For resources and demand The matching degree, A collection of resources.

7. A resource recommendation device based on dynamic portraits of small businesses, characterized in that: The device comprises: An event processing module is configured to identify key event types based on a multidimensional enterprise dataset, and to normalize the original timestamps of events corresponding to each key event type to obtain an initial event sequence; wherein the multidimensional enterprise dataset includes business data, financing data, product data, and market data; An importance scoring and filtering module, configured to calculate importance scores based on the frequency and type weights of events in the initial event sequence, and filter the importance scores using an adaptive threshold to obtain a core event set; An event relationship construction module is used to calculate an inter-event dependency strength matrix based on the timestamp and business semantic features of each event in the core event set; and to establish directed edges based on the inter-event dependency strength matrix through causal reasoning to obtain an event relationship graph; An enterprise portrait generation module is configured to take the event relationship graph as input, learn event embedding features through a graph neural network, and then process the event embedding features in combination with a gated recurrent unit to obtain an enterprise dynamic portrait vector; A demand distribution calculation module is used to calculate the impact weight of historical events based on the enterprise dynamic portrait vector through a time series model with an attention mechanism to obtain a future demand probability distribution that includes the probability of each type of demand; A resource recommendation module is configured to calculate resource urgency based on the probability distribution of future demand, construct an objective function based on the resource urgency and the matching degree between resources and demand, and solve the objective function under the constraints of enterprise capabilities to obtain a forward-looking resource recommendation list; The update optimization module is used to trigger the incremental update of the event relationship graph and the parameter retraining of the timing model by comparing the deviation between the future demand probability distribution and the actual development data.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the method according to any one of claims 1 to 6 is implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 6 is implemented.

Images