A master intelligent matching and screening system and method

Abstract

The application discloses a kind of talents intelligent matching screening system and method, it is related to electric business intelligent recommendation technical field, including: based on brand-person history cooperation data construction interaction matrix, through collaborative filtering calculation initial recommendation score, and introduction path concentration and long tail coverage degree judge whether there is historical path dependence;When there is path dependence, through brand community discovery and restart random walk method, get the deviation recommendation score of bias transmission;With task semantic vector and person state vector, calculate content consistency, crowd consistency and person stability, build ranking feature vector and input learning ranking model, output final matching score;Finally, through feedback record updates interaction matrix, refresh brand community and retrain ranking model, form closed loop iteration, solve the technical problem that recommended result is excessively dependent on existing popular person because head person history high-frequency cooperation path is repeatedly amplified, difficult to identify more suitable person.

Description

Technical Field

[0001] This invention relates to the field of e-commerce intelligent recommendation technology, and more specifically, to an expert intelligent matching and screening system and method. Background Technology

[0002] Currently, in influencer marketing platforms, the matching between brands and influencers primarily relies on collaborative filtering recommendation methods based on historical collaboration data. This method constructs a brand-influencer interaction matrix, calculates the similarity between brands, and recommends influencers from similar brands with past collaborations to the target brand. This technology can leverage historical collaboration experience to a certain extent for influencer selection, and has advantages such as data-driven approach and ease of implementation, making it widely used in e-commerce promotion, content marketing, and other fields.

[0003] The existing technology has the following shortcomings: The core logic of collaborative filtering is that influencers collaborating with similar brands are potential suitable influencers for the target brand. However, in the e-commerce field, similar brands in the same category often form homogeneous collaborations due to the traffic advantage of top influencers. This method does not deduplicate and compress evidence of similar collaborations during the dissemination process, leading to the repeated accumulation and amplification of evidence of collaborations with the same top influencer by multiple similar brands. This ultimately creates a historical path dependency phenomenon, which not only suppresses the exposure opportunities of mid-to-long-tail, low-frequency influencers, resulting in low efficiency in the utilization of influencer resources, but also ignores dynamic factors such as the content requirements of the current task, the target audience, and the influencer's recent operational status. This leads to a disconnect between the recommended results and the brand's actual marketing needs, resulting in high matching scores but low conversion rates, increasing the brand's marketing trial-and-error costs, and failing to meet the brand's need for precise and personalized influencer matching, making it difficult to achieve truly accurate matching. There is an urgent need for an intelligent matching method that can identify and eliminate historical biases and combine task semantics with influencer status, solving the problem of over-concentration of top influencers and achieving precise influencer selection based on task characteristics.

[0004] To address the above problems, this invention proposes a solution. Summary of the Invention

[0005] In order to overcome the above-mentioned defects of the prior art, embodiments of the present invention provide an expert intelligent matching and screening system and method to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A method for intelligent matching and filtering of influencers, including the following steps; Step S1: Construct a brand-influencer interaction matrix, calculate the cosine similarity between brands, filter similar brand sets to calculate initial recommendation scores and construct a head candidate set, calculate path concentration, divide low-frequency influencer sets according to the number of historically cooperating brands, calculate long-tail coverage, collect the path concentration and long-tail coverage distribution of similar historical tasks, determine the upper and lower thresholds of the box plot criterion, and determine whether historical path dependence is triggered by comparing the current task's path concentration with the upper threshold and long-tail coverage with the lower threshold, and decide whether to proceed to the next step. Step S2: Construct a similar brand graph based on the similar brand set and calculate the edge weights. Use a community detection algorithm to divide brand communities, perform row normalization on the number of brand-influencer collaborations, construct community-level influencer evidence, construct a three-layer propagation graph of target brand-brand community-influencer, calculate the transition strength from target brand to brand community, construct a transition matrix and use a random walk method with restart for iterative propagation, and extract the influencer layer node components as the bias-free recommendation score after convergence. Step S3: Construct a task semantic vector for the target brand task, construct an influencer state vector for the influencer, calculate the content consistency between the task and the influencer, calculate the audience consistency based on the target audience distribution and fan distribution, calculate the coefficient of variation based on the influencer's recent conversion rate sequence, construct influencer stability, and combine the biased recommendation score, content consistency, audience consistency, and influencer stability into a ranking feature vector. Use a gradient boosting ranking tree model to train and learn the ranking model, output the final matching score, and generate the final recommended influencer set. Step S4: Establish a set of influencer collaboration feedback records, update the number of collaborations in the interaction matrix according to the performance criteria, generate paired sorted samples based on the advantage relationship, incrementally update the sorting model, periodically refresh the brand community structure, introduce a verification mechanism for boundary samples, and continuously correct the model boundary based on the actual execution results.

[0007] In a preferred embodiment, step S1 includes the following: Build a brand-influencer interaction matrix, where the matrix elements are the historical number of collaborations between the brand and the influencer; Calculate the cosine similarity between the target brand and other brands, and select the top K brands with the highest similarity as the set of similar brands; Based on the set of similar brands, the initial recommendation score of each influencer is calculated by weighted average number of historical collaborations, and a top candidate set is constructed by sorting the scores. The path concentration is calculated as the proportion of the sum of the initial recommendation scores in the head candidate set to the sum of all initial recommendation scores. The low-frequency influencer group is divided according to the number of brands that have historically collaborated with the influencer group, and the proportion of low-frequency influencers in the top candidate group is calculated as the long-tail coverage. Collect the path concentration and long-tail coverage distribution of similar historical tasks to determine the upper and lower thresholds of the box plot criterion; Compare whether the path concentration of the current task is greater than the upper threshold and whether the long tail coverage is less than the lower threshold. If so, determine that historical path dependency has been triggered and proceed to step S2. Otherwise, use the initial recommended score directly.

[0008] In a preferred embodiment, step S2 includes the following: A similar brand graph is constructed based on a set of similar brands, where nodes represent similar brands and edge weights represent the cosine similarity between brands. The Louvain community discovery algorithm is used to divide similar brand images into brand communities; The number of brand-influencer collaborations is normalized. For each brand community, community-level influencer evidence is constructed based on the normalized number of collaborations between each brand within the community. A three-layer propagation graph containing the target brand, brand community, and influencers is constructed, and the transfer intensity from the target brand to each brand community is calculated. A transfer matrix is ​​constructed, and a random walk method with restart is used to iteratively propagate on the propagation graph. The iteration stops when the convergence condition is met, and the steady-state probability of the influencer layer node is extracted as the bias removal recommendation score.

[0009] In a preferred embodiment, step S3 includes the following: Construct task semantic vectors for the target brand task, including converting task information into text and inputting it into a pre-trained text embedding model; To construct a state vector for influencers, the process involves converting the influencer's most recent content and follower information into text and then inputting it into the same text embedding model. The cosine similarity between the task semantic vector and the expert's state vector is used as the content consistency. Based on the target audience distribution and influencer fan distribution for the target task, calculate the Jensen-Shannon divergence and convert it into audience consistency. The coefficient of variation is calculated based on the conversion rate sequence of the influencer's most recent collaborations and then converted into influencer stability. The ranking feature vector is composed of bias-free recommendation score, content consistency, audience consistency, and influencer stability. A gradient boosting sorting tree model is used to train the sorting feature vectors and output the final matching score. Sort by the final matching score to generate the final set of recommended influencers.

[0010] In a preferred embodiment, step S4 includes the following: Establish a feedback record set for collaborations with influencers to document the implementation status of collaborations; According to the performance criteria, when the feedback record meets the requirements of completion and release and no performance abnormality occurs, the corresponding number of collaborations in the interaction matrix is ​​incremented by one. Based on the performance of different experts in the same task, paired ranking samples are generated through advantage relationships; the learning ranking model is incrementally updated using the paired ranking samples. The brand similarity is periodically recalculated based on the updated interaction matrix, and the brand community structure is refreshed. For influencers whose final matching scores are near the boundary, a mechanism for manual review or small-scale trial deployment is introduced to correct the model boundary based on the review or trial deployment results.

[0011] A talent intelligent matching and screening system includes: an interactive diagnosis module, a community de-biasing module, a task correction module, and a feedback iteration module, with signal connections between the modules; Interactive Diagnosis Module: Constructs a brand-influencer interaction matrix, calculates the cosine similarity between brands, filters similar brand sets to calculate initial recommendation scores and constructs a head candidate set, calculates path concentration, divides low-frequency influencer sets based on the number of historically cooperating brands, calculates long-tail coverage, collects the path concentration and long-tail coverage distribution of similar historical tasks, determines the upper and lower thresholds of the box plot criterion, and determines whether historical path dependence is triggered by comparing the current task's path concentration with the upper threshold and long-tail coverage with the lower threshold, and decides whether to proceed to subsequent steps. Community bias removal module: Construct a similar brand graph based on a set of similar brands and calculate edge weights. Use a community discovery algorithm to divide brand communities. Perform row normalization on the number of brand-influencer collaborations. Construct community-level influencer evidence. Construct a three-layer propagation graph of target brand-brand community-influencer. Calculate the transition strength from target brand to brand community. Construct a transition matrix and use a random walk method with restart for iterative propagation. After convergence, extract the influencer layer node components as the bias removal recommendation score. Task Correction Module: Constructs a task semantic vector for the target brand task, constructs an influencer state vector for the influencer, calculates the content consistency between the task and the influencer, calculates the audience consistency based on the target audience distribution and fan distribution, calculates the coefficient of variation based on the influencer's most recent conversion rate sequence, constructs influencer stability, and combines the bias-free recommendation score, content consistency, audience consistency, and influencer stability into a ranking feature vector. It uses a gradient boosting ranking tree model to train and learn the ranking model, outputs the final matching score, and generates the final recommended influencer set. Feedback Iteration Module: Establishes a set of influencer collaboration feedback records, updates the number of collaborations in the interaction matrix based on performance criteria, generates paired ranking samples based on advantageous relationships, incrementally updates the ranking model, periodically refreshes the brand community structure, introduces a verification mechanism for boundary samples, and continuously corrects the model boundaries based on actual execution results.

[0012] The technical effects and advantages of the intelligent matching and screening method for experts proposed in this invention are as follows: This invention, while retaining the foundation of the brand-influencer interaction matrix and collaborative filtering recommendation, addresses the problem of repeated amplification of high-frequency historical collaboration paths of top influencers in existing technologies. It first uses path concentration and long-tail coverage for judgment, then weakens the repeated propagation of homogeneous historical evidence through brand community segmentation and community evidence compression, thereby reducing the over-reliance of recommendation results on existing popular influencers. Furthermore, by combining the semantics of the target task, the current state characteristics of influencers, and a learned ranking model, the biased candidate influencers are matched and corrected for the current task. This ensures that the selection results are no longer limited to influencers frequently selected in the past, but can more accurately identify influencers more suitable for the current brand task. Through real-world execution feedback, the interaction matrix, brand community structure, and ranking model are continuously updated, forming a closed-loop iterative mechanism, thereby improving the stability, adaptability, and feasibility of the influencer selection process. Attached Figure Description

[0013] Figure 1 This is a schematic diagram of the structure of the expert intelligent matching and screening method of the present invention.

[0014] Figure 2 This is a schematic diagram of a feedback iterative closed loop. Detailed Implementation

[0015] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0016] Example 1: Please refer to Figures 1-2 As shown, this invention discloses an intelligent matching and screening method for experts, including the following steps: Step S1: Construct a brand-influencer interaction matrix, calculate the cosine similarity between brands, filter similar brand sets to calculate initial recommendation scores and construct a head candidate set, calculate path concentration, divide low-frequency influencer sets according to the number of historically cooperating brands, calculate long-tail coverage, collect the path concentration and long-tail coverage distribution of similar historical tasks, determine the upper and lower thresholds of the box plot criterion, and determine whether historical path dependence is triggered by comparing the current task's path concentration with the upper threshold and long-tail coverage with the lower threshold, and decide whether to proceed to the next step. Step S2: Construct a similar brand graph based on the similar brand set and calculate the edge weights. Use a community detection algorithm to divide brand communities, perform row normalization on the number of brand-influencer collaborations, construct community-level influencer evidence, construct a three-layer propagation graph of target brand-brand community-influencer, calculate the transition strength from target brand to brand community, construct a transition matrix and use a random walk method with restart for iterative propagation, and extract the influencer layer node components as the bias-free recommendation score after convergence. Step S3: Construct a task semantic vector for the target brand task, construct an influencer state vector for the influencer, calculate the content consistency between the task and the influencer, calculate the audience consistency based on the target audience distribution and fan distribution, calculate the coefficient of variation based on the influencer's recent conversion rate sequence, construct influencer stability, and combine the biased recommendation score, content consistency, audience consistency, and influencer stability into a ranking feature vector. Use a gradient boosting ranking tree model to train and learn the ranking model, output the final matching score, and generate the final recommended influencer set. Step S4: Establish a set of influencer collaboration feedback records, update the number of collaborations in the interaction matrix according to the performance criteria, generate paired sorted samples based on the advantage relationship, incrementally update the sorting model, periodically refresh the brand community structure, introduce a verification mechanism for boundary samples, and continuously correct the model boundary based on the actual execution results.

[0017] In step S1, a brand-influencer interaction matrix is ​​constructed, cosine similarity between brands is calculated, a set of similar brands is selected to calculate the initial recommendation score and construct a head candidate set, path concentration is calculated, a low-frequency influencer set is divided based on the number of historically cooperating brands, long-tail coverage is calculated, the distribution of path concentration and long-tail coverage of similar historical tasks is collected, the upper and lower thresholds of the box plot criterion are determined, and by comparing the path concentration of the current task with the upper threshold and the long-tail coverage with the lower threshold, it is determined whether historical path dependence is triggered, and whether to proceed to the next step. Specific details include: An interaction matrix is ​​constructed using the historical collaboration relationship between brands and influencers. Initial recommendation results are obtained through a neighborhood-based collaborative filtering method. Two quantitative indicators that can clearly characterize historical path dependence, path concentration and long-tail coverage, are extracted from the initial recommendation results. The decision on whether to initiate subsequent de-biasing steps is based on the indicator results. Extract historical collaboration records. Each historical collaboration record includes the brand logo, influencer logo, and the number of collaborations. Let the brand collection be A collection of experts Then build a brand-influencer interaction matrix. ; where matrix elements Indicate brand With experts Number of historical collaborations; Conduct a similar brand search and select the brand In the interaction matrix R, the i-th row is denoted as the brand vector. For the target brand task The corresponding target brand Calculate its cosine similarity to other brands: ;in, For target brand With brand Similarity; For target brand The brand vector corresponding to the interaction matrix R; For the brand The brand vector corresponding to the interaction matrix R; Based on similarity Select the top K brands from largest to smallest to form the target brand. Similar brand collection Based on existing collaborative filtering methods, the historical collaboration intensity of similar brands with each influencer is propagated to the target brand, thus obtaining influencer data. Initial recommended score: ;in,

[0018] For experts The initial recommended score; For target brand A collection of similar brands; For target brand With brand Similarity; For the brand With experts Number of historical collaborations; All influencers were assigned their initial recommended scores. Sort by popularity from highest to lowest, and select the top K talents to form the head candidate set. Where K is a positive integer, and according to the common application scenarios in the e-commerce industry, the value of K ranges from 5 to 20; for tasks with larger brand scale and higher category segmentation, K∈[10,20] is used, and for tasks with smaller brand scale and new category expansion, K∈[5,10]. Those skilled in the art can flexibly adjust it according to the brand's category, marketing budget, and influencer pool size; based on this top candidate set, the path concentration is calculated: ;in, For the target brand mission Path concentration; D represents the set of the top K influencers after sorting by their initial recommendation scores; D is the complete set of influencers. For experts The initial recommended score; Path Concentration The meaning is: what percentage of the total initial recommendation scores for the current task are held by the top K influencers? If the percentage is too high, it means that the recommendation evidence obtained from similar brand promotion has been overly concentrated on a few top influencers, indicating a clear path amplification phenomenon. Path concentration alone is insufficient to explain the problem, as some tasks naturally favor a minority of influencers. Therefore, this implementation method introduces long-tail coverage to determine whether low-frequency influencers have been sexually suppressed. First, the number of brands each influencer has collaborated with historically is counted, denoted as . Then, based on the first quartile of the total number of brands that all influencers have historically collaborated with, Categorization of low-frequency experts ;in, A gathering of low-frequency enthusiasts; For experts Number of brands that have collaborated with historically; The first quartile of the total number of brands that all influencers have historically collaborated with; Calculate the coverage ratio of the head candidate set to low-frequency influencers, i.e., the long-tail coverage: ;in, For the target brand mission Long-tail coverage; This is the candidate set for the head; For low-frequency influencers, if the initial recommended head candidate set only contains high-frequency head influencers and has no low-frequency influencers, it indicates that the path dependence of collaborative filtering has severely suppressed long-tail influencers. The core function of this indicator is to determine the diversity of head recommendation results, rather than requiring the head candidate set to contain a large number of low-frequency influencers. When the long-tail coverage is <10%, it is judged as insufficient long-tail coverage, indicating that the influencer types in the head candidate set are highly homogeneous. Thresholds are automatically generated using box plot criteria from similar historical tasks. Specifically, data is collected related to the target brand task. For a set of historical tasks with similar purposes, their path concentration is distributed, and the third quartile is obtained. The interquartile range (IQR)(C) is used to derive the path concentration threshold: ;in, Threshold for path concentration is the third quartile of the distribution of path concentration for similar historical tasks; IQR(C) is the interquartile range of the distribution of path concentration for similar historical tasks. Similarly, by forming a distribution of the long-tail coverage of similar historical tasks, the first quartile can be obtained. The threshold for long-tail coverage is obtained by using the interquartile range IQR(L). ;in, This is the threshold for long-tail coverage; is the first quartile of the long-tail coverage distribution of similar historical tasks; IQR(L) is the interquartile range of the long-tail coverage distribution of similar historical tasks. When satisfied When the current task is determined to have historical path dependency, it needs to proceed to step S2; where, The path concentration of the current task; The threshold for path concentration; This represents the long-tail coverage of the current task. This is the threshold for long-tail coverage; Only when both excessive head concentration and insufficient long-tail coverage occur simultaneously is the existing collaborative filtering path considered to have evolved from normal recommendation to historical path dependency; if neither condition is met simultaneously, the existing propagation path is still within an acceptable range, and the initial recommendation score can be used directly. Output the results without requiring further complex processing; This step fully preserves the basic process of brand-influencer interaction matrix and collaborative filtering dissemination in the existing technology. By using the two results of path concentration and long-tail coverage, a clear judgment is made on whether the existing technology needs further modification. Subsequent steps will only be initiated if the existing technology does indeed have historical path dependence.

[0019] In step S2, a similar brand graph is constructed based on the similar brand set, and edge weights are calculated. A community detection algorithm is used to divide brand communities. Row normalization is applied to the number of brand-influencer collaborations. Community-level influencer evidence is constructed. A three-layer propagation graph of target brand-brand community-influencer is built. The transition strength from the target brand to the brand community is calculated. A transition matrix is ​​constructed, and a random walk method with restart is used for iterative propagation. After convergence, influencer layer node components are extracted as the bias-reduction recommendation score. Specific content includes:

[0020] It is activated only when the path concentration exceeds the upper threshold and the long-tail coverage is lower than the lower threshold. Multiple highly similar brands may repeatedly point to the same top influencer, causing the evidence of the same source to be accumulated multiple times, thus forming a historical high-frequency path amplification. Based on similar brand sets Building a similar brand image Each node corresponds to a similar brand, and the edge weight between any two similar brands is taken as their cosine similarity in the interaction matrix. Its edge weight is denoted as: ;in, Similar brands and The boundary weights between them; For the brand The brand vector corresponding to the interaction matrix R; For the brand The brand vector corresponding to the interaction matrix R; Similar brand images The Louvain community discovery algorithm is executed to set up similar brands. Divided into several brand communities, denoted as: ;in, To create a brand community; Let r be the r-th brand community; p is the number of brand communities; The reason for using brand communities instead of directly disseminating information to original similar brands is that brands within the same brand community often have historically chosen highly overlapping groups of influencers. If dissemination continues brand by brand, the repeated support from the same source will be mistaken for support from multiple independent sources. Therefore, it is necessary to compress the influencer evidence within the community first, and then disseminate the compressed community evidence to the target brand. To make the number of collaborations between different brands comparable, we first compared the brands. Normalizing the number of collaborations with influencers, we get: ;in, For the brand For experts Normalized cooperation intensity; For the brand With experts Number of historical collaborations; For the brand The maximum number of collaborations with all influencers in history; It should be noted that the core purpose of this normalization method is to eliminate the magnitude difference in the scale of cooperation between different brands. If a brand only cooperates with a few influencers, its maximum value is the number of times the brand's core influencers have cooperated. After normalization, the cooperation intensity of the brand can be mapped to the [0,1] interval, making the cooperation intensity of influencers of different brands comparable. For influencers with no cooperation records, the value after normalization is 0, which does not affect the construction of community-level influencer evidence. For each brand community With every expert Building community-level expert evidence: ;in, For brand community For experts Evidence of community-level influencer status; For the r-th brand community; For the brand For experts The normalized cooperation intensity means that if multiple brands in the community have a history of cooperation with the same influencer, the community's support for that influencer will increase. However, this increase is gradual and saturated, rather than a simple linear accumulation. This way, the problem of high-frequency path amplification caused by multiple similar brands repeatedly supporting the same influencer can be controlled at the community level, thereby reducing the phenomenon of top influencers being repeatedly promoted from the source.

[0021] Construct a three-tiered communication diagram: target brand - brand community - influencers. The first tier is the target brand node. The second layer consists of brand community nodes. The third layer is the expert node. The transfer intensity from the target brand to the brand community is given by the average similarity between the target brand and brands within the community: ;in, For target brand To the brand community The transfer intensity; For brand community Number of brand nodes; For target brand With brand Similarity; The conversion intensity of brand community reach is directly based on community-level influencer evidence. To perform global propagation learning, a transition matrix P is constructed for the above three-layer propagation graph, and a random walk method with restart is used to iteratively obtain the steady-state propagation result: ;in, Let be the node probability vector for the t-th iteration; Let P be the transpose of the transition matrix; The random walk preservation coefficient, ranging from 0.7 to 0.9, controls the propagation weight of historical paths during the random walk; in the e-commerce influencer matching scenario, ρ=0.8 is preferred. The convergence threshold ε is a very small positive real number, ranging from... - Preferred This ensures a balance between the stability of the iteration results and computational efficiency. This is the initial vector starting from the target brand node, where the value is 1 at the target brand node and 0 at the other nodes. When the convergence condition is met: When the iteration stops, the iteration continues; where, This is the convergence threshold; Represents the vector norm; After convergence, the component of the expert-level node in the steady-state probability vector is taken as the bias-reduction recommendation score: ;in, For experts The biased recommendation score; For the converged steady-state probability vector at the master node The component at the location; Compared with the initial recommendation score compared to, It no longer comes from the direct linear superposition of similar brands' evidence of influencers, but first compresses the same source evidence through the brand community, and then completes global propagation through random walk in the three-layer graph. Therefore, it can better reflect which influencers the historical evidence of the current task truly supports after eliminating the repeated amplification of high-frequency paths, and only modifies the propagation position that is most likely to cause bias. This step outputs the bias-free recommendation score. After obtaining the candidate influencer ranking results compressed by the brand community, step S3 will further answer another more crucial question: even though the historical evidence has been compressed, it is still necessary to determine whether these candidate influencers are truly suitable for the current target brand mission. Therefore, the output of step S2 will be used to construct the final matching score.

[0022] In step S3, a task semantic vector is constructed for the target brand task, and an influencer state vector is constructed for the influencer. The content consistency between the task and the influencer is calculated. The audience consistency is calculated based on the target audience distribution and fan distribution. The coefficient of variation is calculated based on the influencer's recent conversion rate sequence. Influencer stability is constructed. The bias-free recommendation score, content consistency, audience consistency, and influencer stability are combined into a ranking feature vector. A gradient boosting ranking tree model is used to train and learn the ranking model. The final matching score is output, and the final set of recommended influencers is generated. Specific content includes: The bias-correction recommendation score obtained in this step Essentially, it is still a graph propagation result based on historical cooperation relationships. Although it eliminates historical amplification, it still does not directly answer whether these experts are suitable for the current task. Therefore, a learning ranking model is introduced, which uses the biased recommendation score as a historical prior feature and combines the consistency results between the current task and the current state of the experts to directly learn the final matching score. Target brand mission Constructing task semantic vectors The target brand task should include at least four types of information: product category, price range, target audience, and content style. These four types of information are then uniformly transcribed into structured text descriptions and input into a pre-trained text embedding model to obtain the task semantic vector. ;in, For the target brand mission Task semantic vector For text embedding models; The task description text should include information such as product category, price range, target audience, and content style. It should be noted that the pre-trained text embedding model used in this invention is the Sentence-BERT model, specifically using bert-base-chinese pre-trained weights. The model input dimension is 768, and the output is a 768-dimensional semantic vector. The format of the structured text description is "Product Category: XXX; Price Range: XXX; Target Audience: XXX; Content Style: XXX" or "Recent Content Category: XXX; Copywriting Style: XXX; Fan Profile: XXX; Collaboration Category: XXX". The model training corpus consists of publicly available text data in the field of e-commerce influencer marketing and historical text data of brand-influencer collaborations. Those skilled in the art can directly fine-tune the model without retraining. For experts Constructing the expert's state vector Specifically, the system reads the titles, summaries, product category tags, and fan profile information of commercial content posted by influencers in their recent observation window, and transcribes these into influencer status description text. This text is then input into a pre-trained text embedding model to obtain the influencer status vector. ;in, For experts The expert's state vector; A text embedding model that uses the same semantic vectors as the task. For experts Status description text; In obtaining the task semantic vector And expert state vector Then, calculate the consistency of the content: ;in, For the target brand mission With experts Content consistency; For task semantic vectors; This represents the state vector of the expert. Content consistency is used to characterize the similarity between the content appeal of the current task and the influencer's recent commercial expressions. However, content consistency alone is insufficient for screening, because while some influencers may have a matching content style, their fan base may not align with the brand's target audience. Therefore, audience consistency is also assessed by setting target brand tasks. The target population distribution vector is Expert The fan distribution vector is The consensus of a population is defined as follows: ;in, For the target brand mission With experts Consistency of the population; This represents the distribution vector of the target population for the objective task. For experts The fan distribution vector; Jensen-Shannon divergence; A higher degree of audience consistency indicates that the influencer's fan distribution is closer to the target audience of the current brand task. Considering that if an influencer's recent status fluctuates too much, even if the content and audience are consistent, it may be difficult to consistently undertake tasks. Therefore, influencer stability is introduced, assuming the influencer... The conversion rate sequence in several recent business collaborations is as follows: Its coefficient of variation is: ;in, For experts Coefficient of variation of recent business cooperation conversion rate sequences; Indicates standard deviation; This represents the mean; Constructing expert stability based on the coefficient of variation: ;in, For experts Stability; For experts Coefficient of variation of recent business cooperation conversion rate sequences; Four feature results were generated that can be directly used for ranking learning: biased recommendation score Content consistency Consistency of the population And the stability of experts The above four results are combined to form a feature vector: ;in, For the target brand mission With experts The sorted feature vector; During the model training phase, a ranking model is trained using historical task samples. Each historical task sample contains task features, candidate expert features, and collaboration result labels. A gradient boosting ranking tree model is preferably used to perform ranking learning on each candidate expert. The output format is as follows: ;in, For the target brand mission For experts The final matching score; T is the number of sorting trees; Let t be the sorting tree; For the target brand mission With experts The sorted feature vector; It should be noted that the gradient boosting ranking tree model in this invention specifically adopts the XGBoost ranking model, the loss function is LambdaRank, and the model hyperparameters are set as follows: tree depth is 6-10, learning rate is 0.01-0.1, number of iterations is 100-200, and number of leaf nodes is 20-40. The model training samples are divided into training set, validation set, and test set in a 7:2:1 ratio. The ranking AUC value of the validation set is used as the model convergence criterion. Training stops when the AUC value does not improve for 10 consecutive iterations. The labels of the training samples are the actual conversion effect after brand-influencer collaboration, divided into three levels: excellent, good, and poor. For example, 2% and 5% are used for grading to guide the model to learn the ranking weights. During training, the sorting tree model uses actual collaboration results from historical tasks to form sorting labels, and learns model parameters by maximizing the quality of the head sort. This approach, which removes bias from the recommendation score, is effective. To eliminate historical path dependence, the historical collaboration evidence score of the expert is used as a historical prior feature and entered into the model. Therefore, step S3 no longer requires manually setting the historical evidence weight or task evidence weight, and the model automatically learns the weight ratio of each feature. The model outputs the final matching score. Then, all the candidate talents will be categorized according to... Sort the brands from highest to lowest and select the top K as the target brands. The final recommended expert set is the final selection result formed after two stages of processing: path bias removal and current task learning and ranking. The process of transforming historical collaborative propagation results into current task matching results is completed. Step S2 addresses the problem of over-amplification of historical high-frequency paths, while step S3 addresses how to further select truly suitable experts for the current task after bias removal. Step S3 outputs the final matching score. In step S4, the results will be iteratively updated based on actual execution feedback.

[0023] In step S4, a set of influencer collaboration feedback records is established. The number of collaborations in the interaction matrix is ​​updated based on performance criteria. Paired ranking samples are generated based on advantage relationships. The ranking model is incrementally updated. The brand community structure is periodically refreshed. A verification mechanism is introduced for boundary samples. The model boundary is continuously corrected based on actual execution results. Specific details include: Target brand mission Feedback records are established for each influencer in the final recommended influencer set. Each feedback record includes brand identification, influencer identification, whether the post was completed, actual exposure achievement rate, actual conversion achievement rate, and whether any fulfillment abnormalities occurred. All feedback records are combined into a feedback record set F. Update the interaction matrix R. Since the interaction matrix R uses historical collaboration counts, this step uses a clear fulfillment criterion to determine whether to include the collaboration in the interaction matrix. Specifically, when a feedback record simultaneously meets the conditions of completion of publication and no fulfillment anomalies, the collaboration is recorded as a valid collaboration, and the collaboration count between the target brand and the influencer in the interaction matrix is ​​incremented by one. ;in, For the updated interaction matrix elements; For the target brand before the update With experts The number of historical collaborations; 1 indicates a new valid collaboration. If a feedback record is not published or a fulfillment error occurs, the interaction matrix R will not be updated. Incremental updates are performed on the learned ranking model in step S3, and ranking labels are generated using dominance relationships for the same target brand task. The two people who have already completed their tasks and If the expert Simultaneously meeting the following conditions: its exposure achievement rate is not lower than that of influencers. Its conversion rate is no less than that of influencers. And its performance status is no lower than that of an expert. And in at least one of the above three aspects, they are significantly better than the experts. Then it is written as: ;in, Indicating the target brand mission Below, expert Superior to experts ; Based on the above advantages, pairs of sorting samples can be generated from the same task and added to the training set of the sorting model learned in step S3. The sorting tree model is periodically retrained, and the model update process is more in line with the logic of correcting the sorting boundary with the actual execution results. The brand community structure in step S2 is periodically refreshed. When the number of newly added valid cooperation records reaches the preset update batch, the similarity between brands is recalculated based on the updated interaction matrix R, and the Louvain community discovery algorithm is re-executed on the new similar brand graph to obtain a new set of brand communities. As brands continue to collaborate with new influencers, the similarity relationships between brands will also change. If the brand community is not updated for a long time, the community compression structure in step S2 will gradually become distorted. Therefore, step S4 refreshes the brand community periodically to ensure that the path correction capability in step S2 remains effective. It should be noted that the preset update batch for periodic refresh is 500-2000 valid cooperation records, with 1000 records preferred for daily marketing scenarios on e-commerce platforms; if the brand belongs to the category of non-standard products or has a high frequency of KOL cooperation, 500 records can be used, while for standard products or categories with low cooperation frequency, 2000 records can be used. A monthly refresh mechanism can also be set up, and either one can be implemented to ensure that the brand community structure matches the latest cooperation data. To avoid the data becoming completely fixed on a few top performers after multiple rounds of tasks, this step introduces a boundary sample verification mechanism. Specifically, it verifies the final matching score in step S3. The samples of influencers close to the current selection boundary are manually reviewed or tested on a small scale. If the review results show that some boundary samples have a small number of historical collaborations but perform well in actual tasks, their feedback records will be amplified in the next round of updates through the interaction matrix and ranking model. Conversely, if some historically high-frequency influencers are still selected but their performance results are consistently poor, their relative advantage in the ranking model will gradually decrease due to the reduction of advantageous relationship samples. This mechanism does not change the main algorithm of steps S1 to S3, but continuously corrects the model boundary through the real business closed loop. It should be noted that the boundary sample determination criteria in this invention are: influencers whose final matching scores fall within the range of "the lower limit of the final recommended influencer set selection score ± 10%"; where the lower limit of the selection score is the lowest matching score of the top K influencers. If there are fewer than 5 influencers in this range, the 5 influencers with the lowest scores are taken as boundary samples. The verification method is manual review of the matching degree between the influencer's content and the task. The trial launch mechanism is for brands and influencers to conduct short-term collaborations with a small budget, and the decision on whether to include them in the core recommendation pool is based on the conversion effect of the trial launch. Through this step, a closed-loop iteration of interaction matrix update - brand community refresh - ranking model retraining is completed. The historical cooperation foundation in step S1 can continuously introduce the latest effective cooperation records. The brand community compression structure in step S2 can be updated as brand relationships evolve. The learning ranking boundary in step S3 can be continuously corrected based on the actual execution results. Thus, the expert intelligent matching and screening method described in this invention forms a complete closed-loop process with consistent terminology and direct implementation.

[0024] This invention discloses an intelligent matching and screening system for experts, comprising: an interactive diagnosis module, a community bias correction module, a task correction module, and a feedback iteration module, with signal connections between the modules; Interactive Diagnosis Module: Constructs a brand-influencer interaction matrix, calculates the cosine similarity between brands, filters similar brand sets to calculate initial recommendation scores and constructs a head candidate set, calculates path concentration, divides low-frequency influencer sets based on the number of historically cooperating brands, calculates long-tail coverage, collects the path concentration and long-tail coverage distribution of similar historical tasks, determines the upper and lower thresholds of the box plot criterion, and determines whether historical path dependence is triggered by comparing the current task's path concentration with the upper threshold and long-tail coverage with the lower threshold, and decides whether to proceed to subsequent steps. Community bias removal module: Construct a similar brand graph based on a set of similar brands and calculate edge weights. Use a community discovery algorithm to divide brand communities. Perform row normalization on the number of brand-influencer collaborations. Construct community-level influencer evidence. Construct a three-layer propagation graph of target brand-brand community-influencer. Calculate the transition strength from target brand to brand community. Construct a transition matrix and use a random walk method with restart for iterative propagation. After convergence, extract the influencer layer node components as the bias removal recommendation score. Task Correction Module: Constructs a task semantic vector for the target brand task, constructs an influencer state vector for the influencer, calculates the content consistency between the task and the influencer, calculates the audience consistency based on the target audience distribution and fan distribution, calculates the coefficient of variation based on the influencer's most recent conversion rate sequence, constructs influencer stability, and combines the bias-free recommendation score, content consistency, audience consistency, and influencer stability into a ranking feature vector. It uses a gradient boosting ranking tree model to train and learn the ranking model, outputs the final matching score, and generates the final recommended influencer set. Feedback Iteration Module: Establishes a set of influencer collaboration feedback records, updates the number of collaborations in the interaction matrix based on performance criteria, generates paired ranking samples based on advantageous relationships, incrementally updates the ranking model, periodically refreshes the brand community structure, introduces a verification mechanism for boundary samples, and continuously corrects the model boundaries based on actual execution results.

[0025] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0026] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0027] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and inventive constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0028] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0029] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0030] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for intelligent matching and filtering of influencers, characterized in that, Includes steps; Step S1: Construct a brand-influencer interaction matrix, calculate the cosine similarity between brands, filter similar brand sets to calculate initial recommendation scores and construct a head candidate set, calculate path concentration, divide low-frequency influencer sets according to the number of historically cooperating brands, calculate long-tail coverage, collect the path concentration and long-tail coverage distribution of similar historical tasks, determine the upper and lower thresholds of the box plot criterion, and determine whether historical path dependence is triggered by comparing the current task's path concentration with the upper threshold and long-tail coverage with the lower threshold, and decide whether to proceed to the next step. Step S2: Construct a similar brand graph based on the similar brand set and calculate the edge weights. Use a community detection algorithm to divide brand communities, perform row normalization on the number of brand-influencer collaborations, construct community-level influencer evidence, construct a three-layer propagation graph of target brand-brand community-influencer, calculate the transition strength from target brand to brand community, construct a transition matrix and use a random walk method with restart for iterative propagation, and extract the influencer layer node components as the bias-free recommendation score after convergence. Step S3: Construct a task semantic vector for the target brand task, construct an influencer state vector for the influencer, calculate the content consistency between the task and the influencer, calculate the audience consistency based on the target audience distribution and fan distribution, calculate the coefficient of variation based on the influencer's recent conversion rate sequence, construct influencer stability, and combine the biased recommendation score, content consistency, audience consistency, and influencer stability into a ranking feature vector. Use a gradient boosting ranking tree model to train and learn the ranking model, output the final matching score, and generate the final recommended influencer set. Step S4: Establish a set of influencer collaboration feedback records, update the number of collaborations in the interaction matrix according to the performance criteria, generate paired sorted samples based on the advantage relationship, incrementally update the sorting model, periodically refresh the brand community structure, introduce a verification mechanism for boundary samples, and continuously correct the model boundary based on the actual execution results.

2. The expert intelligent matching and screening method according to claim 1, characterized in that, Build a brand-influencer interaction matrix, where the matrix elements are the historical number of collaborations between the brand and the influencer; Calculate the cosine similarity between the target brand and other brands, and select the top K brands with the highest similarity as the set of similar brands; Based on the set of similar brands, the initial recommendation score of each influencer is calculated by weighted average of the number of historical collaborations, and a top candidate set is constructed by sorting the scores.

3. The expert intelligent matching and screening method according to claim 2, characterized in that, The path concentration is calculated as the proportion of the sum of the initial recommendation scores in the head candidate set to the sum of all initial recommendation scores. The low-frequency influencer group is divided according to the number of brands that have historically collaborated with the influencer group, and the proportion of low-frequency influencers in the top candidate group is calculated as the long-tail coverage. Collect the path concentration and long-tail coverage distribution of similar historical tasks to determine the upper and lower thresholds of the box plot criterion; Compare whether the path concentration of the current task is greater than the upper threshold and whether the long tail coverage is less than the lower threshold. If so, determine that historical path dependency has been triggered and proceed to step S2. Otherwise, use the initial recommended score directly.

4. The expert intelligent matching and screening method according to claim 1, characterized in that, A similar brand graph is constructed based on a set of similar brands, where nodes represent similar brands and edge weights represent the cosine similarity between brands. The Louvain community discovery algorithm is used to segment similar brand images to obtain brand communities.

5. The expert intelligent matching and screening method according to claim 4, characterized in that, The number of brand-influencer collaborations is normalized. For each brand community, community-level influencer evidence is constructed based on the normalized number of collaborations between each brand within the community. A three-layer propagation graph containing the target brand, brand community, and influencers is constructed, and the transfer intensity from the target brand to each brand community is calculated. A transfer matrix is ​​constructed, and a random walk method with restart is used to iteratively propagate on the propagation graph. The iteration stops when the convergence condition is met, and the steady-state probability of the influencer layer node is extracted as the bias removal recommendation score.

6. The expert intelligent matching and screening method according to claim 1, characterized in that, Construct task semantic vectors for the target brand task, including converting task information into text and inputting it into a pre-trained text embedding model; To construct a state vector for influencers, the process involves converting the influencer's most recent content and follower information into text and then inputting it into the same text embedding model. The cosine similarity between the task semantic vector and the expert's state vector is used as the content consistency. Based on the target audience distribution and influencer fan distribution for the target task, calculate the Jensen-Shannon divergence and convert it into audience consistency. The coefficient of variation is calculated based on the conversion rate sequence of the influencer's most recent collaborations and then converted into influencer stability.

7. The expert intelligent matching and screening method according to claim 6, characterized in that, The ranking feature vector is composed of bias-free recommendation score, content consistency, audience consistency, and influencer stability. A gradient boosting sorting tree model is used to train the sorting feature vectors and output the final matching score. Sort by the final matching score to generate the final set of recommended influencers.

8. The expert intelligent matching and screening method according to claim 1, characterized in that, Establish a feedback record set for collaborations with influencers to document the implementation status of collaborations; According to the performance criteria, when the feedback record meets the requirements of completion and release and no performance abnormality occurs, the corresponding number of collaborations in the interaction matrix is ​​incremented by one. Based on the performance of different experts in the same task, paired ranking samples are generated through advantage relationships; the learning ranking model is incrementally updated using paired ranking samples.

9. The expert intelligent matching and screening method according to claim 8, characterized in that, The brand similarity is periodically recalculated based on the updated interaction matrix, and the brand community structure is refreshed. For influencers whose final matching scores are near the boundary, a mechanism for manual review or small-scale trial deployment is introduced to correct the model boundary based on the review or trial deployment results.

10. A talent intelligent matching and screening system, used to implement the talent intelligent matching and screening method according to any one of claims 1-9, characterized in that... ； Interactive Diagnosis Module: Constructs a brand-influencer interaction matrix, calculates the cosine similarity between brands, filters similar brand sets to calculate initial recommendation scores and constructs a head candidate set, calculates path concentration, divides low-frequency influencer sets based on the number of historically cooperating brands, calculates long-tail coverage, collects the path concentration and long-tail coverage distribution of similar historical tasks, determines the upper and lower thresholds of the box plot criterion, and determines whether historical path dependence is triggered by comparing the current task's path concentration with the upper threshold and long-tail coverage with the lower threshold, and decides whether to proceed to subsequent steps. Community bias removal module: Construct a similar brand graph based on a set of similar brands and calculate edge weights. Use a community discovery algorithm to divide brand communities. Perform row normalization on the number of brand-influencer collaborations. Construct community-level influencer evidence. Construct a three-layer propagation graph of target brand-brand community-influencer. Calculate the transition strength from target brand to brand community. Construct a transition matrix and use a random walk method with restart for iterative propagation. After convergence, extract the influencer layer node components as the bias removal recommendation score. Task Correction Module: Constructs a task semantic vector for the target brand task, constructs an influencer state vector for the influencer, calculates the content consistency between the task and the influencer, calculates the audience consistency based on the target audience distribution and fan distribution, calculates the coefficient of variation based on the influencer's most recent conversion rate sequence, constructs influencer stability, and combines the bias-free recommendation score, content consistency, audience consistency, and influencer stability into a ranking feature vector. It uses a gradient boosting ranking tree model to train and learn the ranking model, outputs the final matching score, and generates the final recommended influencer set. Feedback Iteration Module: Establishes a set of influencer collaboration feedback records, updates the number of collaborations in the interaction matrix based on performance criteria, generates paired ranking samples based on advantageous relationships, incrementally updates the ranking model, periodically refreshes the brand community structure, introduces a verification mechanism for boundary samples, and continuously corrects the model boundaries based on actual execution results.

Images