An intelligent matching recommendation method for exhibition supply and demand

Abstract

The application provides a kind of intelligent matching recommendation method for supply and demand of exhibition, it is related to data processing technical field, the method obtains the static profile, dynamic behavior and category attention characteristic vector of supply and demand, and respectively calculates static profile correlation coefficient P, dynamic behavior resonance index D and category attention characteristic vector coupling degree B.And the supply and demand potential energy evaluation model is built to calculate the first supply and demand potential energy matching degree G, and high potential matching pair is identified accordingly.For high potential matching pair, the method dynamically adjusts the basic weight of P, D and B, generates adaptive weight, and finally uses adaptive weight for weighted summation to generate the second matching degree M, and sorts and recommends based on M value.The application realizes the contextual adaptive adjustment of matching logic through multi-source data fusion and dynamic weight mechanism, improves the commercial value and accuracy of recommendation result, and improves the user decision efficiency.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, specifically to an intelligent matching and recommendation method for supply and demand sides in the exhibition industry. Background Technology

[0002] With the development of information technology, the traditional exhibition industry is accelerating its transformation towards digitalization and intelligence. Online virtual exhibitions and digital support platforms for offline physical exhibitions have become key tools for improving exhibition efficiency and promoting business cooperation. One of the core functions of these platforms is how to effectively match a massive number of suppliers (exhibitors) with demanders (visitors and buyers) with different needs to generate business opportunities. As a venue for high-density commercial information interaction, the digitalization process of exhibitions has generated massive amounts of data. Every online browsing, search, and collection by attendees, and even offline scanning and negotiation, constitutes a huge data flow. However, existing digital matching tools generally face a core dilemma when processing this data: they cannot effectively distinguish between "strong signals" and "weak noise" in the information, resulting in an extremely low signal-to-noise ratio for matching recommendations and a significant reduction in commercial value.

[0003] In the context of trade shows and exhibitions, user digital behavior is extremely complex. A large amount of "noise" behavior, such as unconscious page scrolling, random clicks out of curiosity, and accidental touches due to interface layout, is mixed in with "signal" behavior that truly represents the category's focus (such as precise keyword searches, repeated browsing of the same exhibitor, and scheduling meetings). Existing methods often simply quantify and accumulate this behavioral data, misjudging noise as signals, thus constructing a profile that deviates from the user's true needs. For example, a customer might receive frequent recommendations from exhibitors in an unrelated field simply due to a chance click, causing serious information interference.

[0004] The participant's identity, job level, corporate strength, and exhibition goals are key signals determining their commercial potential. The behavioral signals of a department director with purchasing decision-making power are far more valuable than those of an information gatherer. Existing methods fail to quantify and utilize this difference in "potential energy," placing all participants on the same level for evaluation. This results in key figures and core exhibitors—the "strong signal sources"—being drowned out by a large number of "weak signal sources." The strategic intentions of both supply and demand sides (such as "exploring new markets" versus "finding exclusive agents") are core signals determining the success of cooperation. Traditional technologies rely solely on vague industry labels for matching, failing to accurately measure the compatibility of these deeper intentions. This leads to a large number of seemingly product-matching but actually misaligned business goals being pushed to users as ineffective matches. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent matching and recommendation method for supply and demand sides in the exhibition industry, so as to solve the problems mentioned in the background art above.

[0006] To achieve the above objectives, specifically, this invention provides a method for efficient and accurate matching and personalized recommendations between supply and demand sides in exhibition scenarios through multi-source data fusion and dynamic weight adjustment. This method falls within the scope of business intelligence, customer relationship management, and e-commerce recommendation methods. This invention provides the following technical solution:

[0007] A smart matching and recommendation method for supply and demand sides in the exhibition industry, comprising the following steps:

[0008] S1. Obtain static profiles of both the supply and demand sides and dynamic behavior sequences of the demand side. The static profiles include preset identity tags, product and service descriptions, and category-related feature vector tags for both the supply and demand sides.

[0009] S2. For each supply and demand pairing, calculate: the static file correlation coefficient P based on the static file; the dynamic behavior resonance index D based on the dynamic behavior sequence of the demand side; and the category attention feature vector coupling degree B based on the category attention feature vector tags in the static file;

[0010] S3. By combining static files and dynamic behavior sequences, the first supply and demand potential energy matching degree is evaluated and quantified, denoted as , and a potential energy threshold is preset. When the first supply and demand potential energy matching degree is greater than the potential energy threshold, the supply and demand pairing is marked as the first mark.

[0011] S4. When the supply and demand pairing is identified as having a first label, the preset basic weights are dynamically adjusted to generate adaptive weights that are compatible with the current supply and demand pairing.

[0012] S5. Using adaptive weights, the static file correlation coefficient, dynamic behavior resonance index, and category attention feature vector coupling degree are weighted and summed to generate the final second matching degree. All suppliers are ranked according to the second matching degree, and suppliers with a second matching degree value higher than the preset matching threshold are recommended to the demander.

[0013] Furthermore, the static file correlation coefficient P is obtained by performing natural language processing on the product and service description text in the static files of both the supply and demand sides to extract and construct keywords, and obtain the supplier feature vector and the demand feature vector.

[0014] The cosine similarity algorithm is used to calculate the cosine value of the angle between the supplier's feature vector and the demander's feature vector, and the normalized cosine value is used as the static file association coefficient.

[0015] Furthermore, before constructing the supplier feature vector and the demand feature vector, a static file preprocessing step is also included: identifying the core exhibitors among the suppliers, denoted as ROI suppliers; dividing the suppliers into ROI suppliers and non-ROI suppliers; for ROI suppliers, using knowledge graph-based entity linking technology to enhance features and improve the semantic richness of their feature vectors; for non-ROI suppliers, only the term frequency inverse document frequency algorithm is used to construct the supplier feature vector to reduce computational complexity.

[0016] Furthermore, the dynamic behavioral resonance index is obtained as follows:

[0017] Preset basic behavior scores for various dynamic behavior types of the demand side to form a behavior value mapping table; dynamic behavior types include searching keywords, browsing exhibitor pages, collecting exhibitors, scanning QR codes, staying at on-site booths, and making appointments for negotiations;

[0018] For each behavior in the demander's dynamic behavior sequence, a basic behavior score is obtained, and the basic behavior score is weighted by a time decay index to calculate the immediate behavior score. The time decay index is negatively correlated with the time interval from the time the behavior occurred to the current time. All immediate behavior scores related to the supplier are accumulated to obtain the dynamic behavior resonance index D.

[0019] Furthermore, the category attention feature vector coupling degree B is obtained by constructing a category attention feature vector compatibility matrix, which predefines the compatibility scores between different supplier category attention feature vectors and different demander category attention feature vectors.

[0020] In the current supply and demand matching, obtain the supplier's supplier intent set and the buyer's buyer intent set, where each intent set contains at least one category interest feature vector label;

[0021] Each consumer category interest feature vector in the consumer intent set is paired with each supplier category interest feature vector in the supplier intent set to generate one or more intent combination pairs.

[0022] For each generated intent pair, based on the category attention feature vector compatibility matrix, find and determine its corresponding compatibility score, thereby obtaining a list containing the compatibility scores corresponding to the intent pair;

[0023] Select the highest compatibility score from the list as the coupling degree of the category-focused feature vector.

[0024] Furthermore, the first supply-demand potential energy matching degree is obtained by constructing a supplier potential energy assessment model and a demand potential energy assessment model respectively, wherein,

[0025] The supplier's potential is quantified based on its enterprise stage, the urgency of its exhibition goals, and the status of its new product launch; the buyer's potential is quantified based on its job level, purchasing decision-making power, and activity level during the exhibition, resulting in supplier potential value PEs and buyer potential value PED. Through a preset nonlinear function, the product of supplier potential value PEs and buyer potential value PED, as well as the absolute difference between the two, are combined to generate the first supply and demand potential matching degree. A preset potential threshold is set. When the first supply and demand potential matching degree is higher than the potential threshold, it is judged as a high-potential matching pair, and the corresponding supply and demand pair is first paired to form the first mark.

[0026] Furthermore, the steps to improve activity levels during the establishment period include:

[0027] S100. Count the first high-value behavior, the second high-value behavior, and the general interaction behavior generated by the demand side during the meeting period; wherein, the first high-value behavior corresponds to meeting reservation; the second high-value behavior corresponds to core interactions other than meeting reservation; and the general interaction behavior corresponds to browsing.

[0028] S200. Determine whether the count value of the first high-value behavior is not less than 1, or whether the count value of the second high-value behavior is not less than the first preset threshold; if so, determine the meeting activity level as the first preset score and end the determination.

[0029] S300. If the conditions of S200 are not met, then continue to determine whether the count value of the second high-value behavior is greater than 0 and less than the first preset threshold, or determine whether the count value of the second high-value behavior is 0 and whether the count value of the general interaction behavior is not less than the second preset threshold; if so, then determine the meeting activity level as the second preset score and end the determination.

[0030] S400. If neither S200 nor S300 is met, the meeting activity level will be set to the third preset score or zero, depending on whether the demander has engaged in any activity during the meeting period.

[0031] Furthermore, the basic weights include a first weight corresponding to the correlation coefficient of static archives, a second weight corresponding to the dynamic behavior resonance index, and a third weight corresponding to the coupling degree of category attention feature vector.

[0032] Further, it is detected whether the current supply and demand pair has a first label. If the current supply and demand pair has a first label, a recommendation enhancement strategy is triggered, including: based on the quantified value of the first supply and demand potential matching degree, an adjustment factor Δ is calculated and generated that is positively correlated with the first supply and demand potential matching degree value through a preset gain function; the first weight is added to the adjustment factor Δ to obtain the intermediate first weight; the third weight is added to the adjustment factor Δ to obtain the intermediate third weight; the adjustment factor Δ is subtracted from the second weight to obtain the intermediate second weight; the intermediate first weight, intermediate second weight, and intermediate third weight are treated as a whole and normalized to generate the final adaptive weight.

[0033] Furthermore, based on the calculated second matching degree, all candidate suppliers are sorted in descending order to generate the first recommendation sequence;

[0034] A preset matching threshold is set, and all suppliers with a second matching degree higher than the preset matching threshold are selected from the first recommendation sequence to form a second candidate pool sequence. It is then determined whether the number of suppliers in the second candidate pool sequence is greater than the preset number of recommendations N.

[0035] If so, select N suppliers from the top of the second candidate pool sequence as the final recommendation list;

[0036] If not, then all suppliers in the second candidate pool sequence will be used as the final recommendation list.

[0037] Compared with the prior art, the beneficial effects of the present invention are:

[0038] This invention establishes a more comprehensive evaluation system for supply and demand relationships by integrating four dimensions of evaluation indicators: static profiles, dynamic behavior, category attention feature vectors, and intrinsic potential energy. This avoids the one-sidedness and uncertainty caused by existing technologies that rely solely on keyword or behavioral data for matching.

[0039] This invention also improves the accuracy of the method in identifying core matching signals and reduces the interference of noise data generated by massive amounts of ordinary interactive behavior on the recommendation results by constructing a supply and demand potential assessment model to quantify the job level, decision-making power and high-value interaction behavior of exhibitors.

[0040] This invention also introduces an adaptive weight adjustment strategy based on potential energy matching degree, enabling dynamic configuration of the recommendation model's weight coefficients. This addresses the problems of fixed matching rules and the application of a single evaluation standard to all matching pairs in existing technologies. Furthermore, by constructing a supply and demand potential energy assessment model, this invention quantifies the exhibitors' job levels, decision-making power, and high-value interaction behaviors, improving the accuracy of identifying core matching signals and reducing the interference of noise data generated by massive amounts of ordinary interaction behaviors on the recommendation results.

[0041] This invention also improves the accuracy and operability of the final recommendation list by combining a dual screening mechanism of matching threshold filtering and recommendation quantity control, thereby alleviating the problem of users' screening burden and low decision-making efficiency caused by the broad and excessive number of recommendation results. Attached Figure Description

[0042] Figure 1 This is a schematic diagram illustrating the overall supply and demand matching of the exhibition industry under this invention.

[0043] Figure 2 This is a schematic diagram of the recommended list execution process of the present invention;

[0044] Figure 3 This is a schematic diagram of steps S1-S5 of the overall method of the present invention. Detailed Implementation

[0045] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0046] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0047] Example 1:

[0048] Please see Figures 1 to 3 This invention provides a technical solution: an intelligent matching and recommendation method for supply and demand sides in the exhibition industry, the specific steps of which include:

[0049] S1. Obtain static profiles of both the supply and demand sides and dynamic behavior sequences of the demand side. The static profiles include preset identity tags, product and service descriptions, and category-related feature vector tags for both the supply and demand sides.

[0050] S2. For each supply and demand pairing, based on the static archives, calculate the static archive correlation coefficient, denoted as P, which represents the degree of matching of basic information. The static archive correlation coefficient P is obtained by performing natural language processing on the product and service description text in the static archives of both the supply and demand sides to extract and construct keywords, and obtain the supplier feature vector and the demand feature vector.

[0051] The cosine similarity algorithm is used to calculate the cosine value of the angle between the supplier's feature vector and the demander's feature vector, and the normalized cosine value is used as the static file association coefficient P.

[0052] For each supply and demand match, a dynamic behavior resonance index, denoted as D, is calculated based on the dynamic behavior sequence of the demander to represent the intensity of the demander's immediate interest. Basic behavior scores are preset for various dynamic behavior types of the demander to form a behavior value mapping table. Dynamic behavior types include searching keywords, browsing exhibitor pages, collecting exhibitors' information, scanning QR codes, staying at on-site booths, and making appointments for negotiations.

[0053] For each behavior in the demander's dynamic behavior sequence, a basic behavior score is obtained, and the basic behavior score is weighted by a time decay index to calculate the immediate behavior score. The time decay index is negatively correlated with the time interval from the time the behavior occurred to the current time. All immediate behavior scores related to the supplier are accumulated to obtain the dynamic behavior resonance index D.

[0054] For each supply and demand pairing, based on the category interest feature vector labels in the static archive, the category interest feature vector coupling degree, which represents the consistency of the business goals of both parties, is calculated and denoted as B. The category interest feature vector coupling degree B is obtained by constructing a category interest feature vector compatibility matrix. The category interest feature vector compatibility matrix predefines the compatibility scores between different supplier category interest feature vectors and different demander category interest feature vectors.

[0055] In the current supply and demand matching, obtain the supplier's supplier intent set and the buyer's buyer intent set, where each intent set contains at least one category interest feature vector label;

[0056] Each consumer category interest feature vector in the consumer intent set is paired with each supplier category interest feature vector in the supplier intent set to generate one or more intent combination pairs.

[0057] For each generated intent pair, based on the category attention feature vector compatibility matrix, its corresponding compatibility score is found and determined, thereby obtaining a list containing the compatibility scores of all pair combinations.

[0058] The highest compatibility score is selected as the coupling degree B of the category-focused feature vector.

[0059] Step S2 aims to penetrate the surface of the text description and accurately measure the compatibility of supply and demand parties in terms of basic information such as products and services at the semantic level. Its implementation follows a layered feature enhancement and similarity calculation strategy.

[0060] First, all suppliers are identified based on their historical performance, industry standing, or exhibition participation level. Core exhibitors are then identified and marked as ROI suppliers. The rest are considered non-ROI suppliers. This stratification is a prerequisite for adopting different calculation strategies subsequently.

[0061] Before constructing the supplier feature vector and the demand feature vector, a static profile preprocessing step is also included: identifying the core exhibitors among the suppliers, denoted as ROI suppliers; dividing the suppliers into ROI suppliers and non-ROI suppliers; for ROI suppliers, using knowledge graph-based entity linking technology to enhance features and improve the semantic richness of their feature vectors; for non-ROI suppliers, only using the term frequency inverse document frequency (TF-IDF) algorithm to construct supplier feature vectors to reduce computational complexity.

[0062] The set of core suppliers and the set of non-core suppliers; for the core supplier set (ROI suppliers), such as a certain technology company; pursuing semantic depth, adopting entity linking technology based on knowledge graph.

[0063] Identify known named entities (such as product models, technical terms, brands, standards, etc.) from the static profile text of ROI suppliers. Link these entities to a pre-built industry knowledge graph. Extract rich semantic information such as higher-level concepts, related attributes, and synonyms from the knowledge graph. Merge the original text features with the enhanced semantic features from the knowledge graph to generate a final, high-dimensional, semantically rich feature vector. Original text: "Focusing on AI chip design, the core product is VisionX-Pro, with performance comparable to Nvidia-A100."

[0064] Entity Link: The entity [AI chip, Nvidia-A100] was identified.

[0065] Feature enhancement: AI chip → link to knowledge graph → obtain enhanced features [artificial intelligence, semiconductor, hardware].

[0066] Nvidia A100 → Link to Knowledge Graph → Gain Enhanced Features [GPU, AI Acceleration, High-Performance Computing].

[0067] Generate an enhanced vector (example): Vector(SUP-001) = {AI chip: 1, VisionXPro: 1, NvidiaA100: 1, Artificial intelligence: 1, Semiconductor: 1, Hardware: 1, GPU: 1, AI acceleration: 1, High-performance computing: 1...}; The vector contains the original words and extended semantic concepts; it enhances the expressive power of the core supplier feature vector, enabling the matching algorithm to "understand" deeper meanings, not just surface-level words. For example, it can understand that both "NvidiaA100" and "Ascend910" belong to the concept of "AI acceleration chip".

[0068] For non-core supplier groups (non-ROI suppliers), such as a certain server rack company, the original text states: "Produces and sells general-purpose server racks and standard network cabling products."

[0069] TF-IDF calculation: Word segmentation: [production, sales, general-purpose, server, cabinet, standard, network, cabling, product];

[0070] Assuming that "server" and "rack" are relatively important terms in the entire document library, the TF-IDF vector is generated (example): Vector(SUP-003) = {server: 0.55, rack: 0.48, cabling: 0.35, network: 0.32, general: 0.21...}; the vector consists of keywords and their weights, and does not contain deep semantics; while ensuring the accuracy of core entity matching, it reduces the computational overhead of processing massive long-tail entities, achieving a balance between computational efficiency and semantic depth.

[0071] Due to its high informational value and strong semantic complexity, this method employs entity linking technology based on knowledge graphs. This technology can link keywords in the product / service descriptions (such as "AI chip" and "cloud-native security") to standard entities in the knowledge graph, thereby obtaining semantically enhanced feature vectors containing rich contextual relationships and information on synonyms, hyponyms, and hypernyms. This makes the feature vectors no longer isolated words, but rather network nodes containing industry knowledge.

[0072] For non-ROI suppliers, to balance computational efficiency and effectiveness, this method employs the classic Term Frequency-Inverse Document Frequency (TF-IDF) algorithm. The TF-IDF algorithm is chosen because it is a mature technique in natural language processing, capable of effectively extracting the most representative keywords from the text while ensuring computational efficiency, thereby constructing high-quality feature vectors. Its technical principle is as follows:

[0073] The TF-IDF algorithm effectively extracts core keywords from text and assigns them corresponding weights, quickly constructing standardized feature vectors. After obtaining the supplier and buyer feature vectors, this method employs a cosine similarity algorithm to calculate the static file association coefficient P. This algorithm measures directional consistency by calculating the cosine of the angle between two vectors in multidimensional space, rather than an absolute numerical value. This is particularly suitable for text processing, effectively avoiding interference caused by varying text lengths and more purely reflecting the similarity of content themes. The final cosine value, after normalization, becomes the static file association coefficient P.

[0074] The calculation of the Dynamic Behavior Resonance Index D filters noise from the intensity of the demander's immediate interest, accurately capturing the demander's true and immediate interest signals for a specific supplier from a series of complex and continuous dynamic behaviors. A behavior value mapping table assigns differentiated base scores to different types of dynamic behaviors. For example, "schedule a meeting," a strong-intention behavior, has the highest score, followed by "collect exhibitors," and "browse pages" has a lower score. This constitutes the first weighting, based on the commercial value of the behavior. A time decay index is introduced. For each behavior, its immediate behavior score is the product of its base score and a decay coefficient. This decay coefficient is negatively correlated with the time interval from the time the behavior occurred to the current time. This means that recently occurring behaviors have a much higher weight than historical behaviors. This constitutes the second weighting, based on the timeliness of the behavior. This method accumulates all the immediate behavior scores generated by the demander for a particular supplier, and the final sum is the Dynamic Behavior Resonance Index D. By mapping behavioral value, this method can automatically filter or reduce the weight of low-value noise behaviors such as "unconscious swiping" and "random clicking," while amplifying the influence of high-value signal behaviors such as "booking" and "favoriting," thereby constructing an interest profile that is closer to the user's true intentions. The time decay mechanism ensures that the index D highly reflects the demander's "current" focus, avoiding being misled by outdated historical behavioral data, making the recommendation results more timely and responsive.

[0075] The calculation of the category focus feature vector coupling degree B is a deep signal decoding of business goal consistency. It is used to go beyond product-level matching and delve into the business strategy level to assess the potential for cooperation between supply and demand parties. The category focus feature vector compatibility matrix is ​​like an expert knowledge base, where industry experts define the compatibility scores between different category focus feature vectors. For example, the compatibility score between "supplier: seeking regional agents" and "demand: hoping to become a brand agent" will be very high, while the compatibility score with "demand: making a one-time project equipment purchase" will be very low. For a supply and demand match, this method obtains the respective category focus feature vector label sets. Then, the category focus feature vector label sets generate all possible intent combination pairs (Cartesian products), and each pair is checked against the compatibility matrix for the corresponding compatibility score. After obtaining the compatibility scores of all intent combination pairs, this method selects the maximum value as the final category focus feature vector coupling degree B. This strategy is based on a business insight: as long as there is a strong point of cooperation between the two parties, cooperation is likely to be achieved, and this strongest signal should be given priority. By transforming abstract and vague business objectives into quantifiable coupling scores, this method can "understand" whether the supply and demand sides are aligned at the strategic level, effectively avoiding a large number of ineffective recommendations where "the products match but the goals are misaligned." By prioritizing the matching of supply and demand sides with highly compatible category focus feature vectors, the recommendation list generated by this invention not only has high information relevance but also a potential business success rate far exceeding that of traditional methods, creating higher business value for exhibition platforms.

[0076] Step S2 helps resolve the "semantic noise" problem mentioned in the background technology, achieving purification and amplification of the basic archive signal. By "signal enhancement" of the ROI provider through the knowledge graph, this method can understand the deep connection between "smart retail solutions" and "commercial displays," avoiding the loss of important business opportunities due to incomplete keyword matching.

[0077] S3. By combining static files and dynamic behavior sequences, the first supply and demand potential energy matching degree G is evaluated and quantified. The first supply and demand potential energy matching degree G is used to characterize the strength and complementarity of the potential transaction motivations of the supply and demand sides, and a potential energy threshold is preset. When G is greater than the potential energy threshold, the first mark is made.

[0078] The first supply-demand potential energy matching degree G is obtained by constructing a supplier potential energy assessment model and a demand potential energy assessment model respectively, wherein,

[0079] The supplier's potential is quantified based on its enterprise stage, the urgency of its exhibition goals, and the status of its new product launch; the buyer's potential is quantified based on its job level, purchasing decision-making power, and activity level during the exhibition, resulting in supplier potential value PEs and buyer potential value PED. Through a preset nonlinear function, the product of the supplier potential value PEs and the buyer potential value PED, as well as the absolute difference between the two, is used to generate the first supply-demand potential matching degree G. A preset potential threshold is set. When the first supply-demand potential matching degree G is higher than the potential threshold, it is judged as a high-potential matching pair and marked as the first match.

[0080] After quantifying basic information (P), immediate interests (D), and business objectives (B), this invention introduces an innovative step S3 aimed at detecting and quantifying a deeper, often overlooked, implicit signal: the intrinsic potential and matching ability of supply and demand. The first supply-demand potential matching degree G calculated in this step is crucial for identifying "high-potential matching pairs" and triggering subsequent adaptive recommendation strategies. The core technology of this step lies in constructing a bidirectional potential assessment model based on multidimensional non-behavioral features, and integrating it through a nonlinear function to achieve a comprehensive judgment of matching potential.

[0081] Supplier Momentum (PEs): PEs comprehensively reflects key indicators that demonstrate a supplier's current market status and motivation for participating in the exhibition. Specifically, this method maps data from multiple dimensions, including the supplier's "company stage" (e.g., startup, growth, maturity), "urgency of exhibition goals" (e.g., seeking funding, launching new products, expanding channels), and "new product launch status" (e.g., whether there is a strategic new product debut), to a standardized PEs value based on pre-defined quantitative rules. A supplier in the growth stage with an urgent goal of expanding channels and launching a major new product will have a higher PEs value than a mature company that is only maintaining market exposure.

[0082] Buyer Potential (PEd): PED aims to quantify the strength of a buyer's purchasing power and willingness. It is mainly composed of "Job Level," "Purchasing Decision-Making Authority," and a key dynamic indicator, "Conference Activity Level." The first two are scored using tags in static profiles (such as CEO, Purchasing Director), while "Conference Activity Level" is dynamically calculated using a sophisticated logical hierarchy model.

[0083] The generated "first tag" serves as the "trigger signal" for the entire method to dynamically adjust the recommendation logic. It enables the method to identify when a more biased enhancement strategy should be adopted, which is the logical cornerstone for realizing the transition from static recommendation to intelligent, context-aware recommendation.

[0084] S4. When the first supply and demand potential matching degree G of the first marker is identified, the preset basic weights are dynamically adjusted to generate adaptive weights that are compatible with the current supply and demand pairing. The basic weights include the first weight corresponding to the static file correlation coefficient P, the second weight corresponding to the dynamic behavior resonance index D, and the third weight corresponding to the category attention feature vector coupling degree B.

[0085] If the current supply and demand pair has a first label, the recommendation enhancement strategy is triggered, including: based on the quantified value of the first supply and demand potential matching degree G, an adjustment factor Δ positively correlated with the value of the first supply and demand potential matching degree G is calculated using a preset gain function; the first weight is added to the adjustment factor Δ to obtain the intermediate first weight; the third weight is added to the adjustment factor Δ to obtain the intermediate third weight; the adjustment factor Δ is subtracted from the second weight to obtain the intermediate second weight; and the intermediate first weight, intermediate second weight, and intermediate third weight are treated as a whole and normalized to generate the final adaptive weight.

[0086] Traditional recommendation methods use fixed weights, meaning the evaluation criteria remain unchanged regardless of whether the pair is ordinary or a high-potential "golden pair." Step S4 of this invention completely breaks this limitation. When this method identifies a high-potential pair (i.e., obtains the "first label") through the potential matching degree G, it essentially establishes a "high-confidence" prior judgment. Based on this judgment, the method triggers a recommendation enhancement strategy, dynamically adjusting the weights. This signifies that the method has evolved from a calculator following static rules into an intelligent agent capable of adjusting its analytical focus according to contextual confidence. When this method confirms that a pair has high potential (high G value), the most rational decision is to place greater trust in "stable strong signals" that reflect its long-term value and strategic fit, while appropriately reducing the influence of volatile and potentially misleading "short-term noise signals." By increasing the weights of the static profile correlation coefficient P and the category attention feature vector coupling degree B, these two points are the cornerstones determining the success of high-potential collaborations. By reducing the weight of the dynamic behavioral resonance index D, this method effectively avoids the erroneous downgrading of high-value matches due to occasional, transient negative behavioral signals. For example, a buyer with decision-making power (high G value) may have less online activity (low D value) due to a busy schedule during the event. Traditional models would therefore underestimate this match, while this invention can penetrate the fog of such "behavioral noise" and still identify and recommend this high-value opportunity. Robustness here refers to the ability of this method to make correct judgments when faced with incomplete or interfering information. By reducing the reliance on the dynamic behavioral resonance index D when identifying high-potential matches, the recommendation results of this invention become more stable and reliable. This prevents a significant business opportunity that perfectly matches in terms of core profile, category focus feature vector, and identity potential from being easily dismissed simply because the buyer did not log in on a particular day or accidentally clicked on a few irrelevant exhibitor pages. This ensures that business opportunities are not missed due to short-term data fluctuations, and the "anti-interference" capability of the recommendation method is qualitatively improved.

[0087] S5. Using adaptive weights, the static file correlation coefficient P, dynamic behavior resonance index D, and category attention feature vector coupling degree B are weighted and summed to generate the final second matching degree M. All suppliers are sorted according to the second matching degree M, and suppliers with M values ​​higher than the preset matching threshold are recommended to the demander.

[0088] Based on the calculated second matching degree M, all candidate suppliers are sorted in descending order to generate the first recommendation sequence;

[0089] A preset matching threshold is set. From the first recommendation sequence, all suppliers with a second matching degree M higher than the preset matching threshold are formed into a second candidate pool sequence. It is then determined whether the number of suppliers in the second candidate pool sequence is greater than the preset number of recommendations N.

[0090] If so, select N suppliers from the top of the second candidate pool sequence as the final recommendation list;

[0091] If not, then all suppliers in the second candidate pool sequence will be used as the final recommendation list.

[0092] The final second matching degree M in step S5 is not a simple linear weighting of the three independent coefficients P, D, and B. Because its weights are adaptively adjusted in step S4, the M value itself implies the method's judgment of the "potential" of the current matching pair. For marked "high-potential matching pairs," the calculation of the M value focuses more on the fit between their basic profile P and business intent TCPB, and its score better reflects strategic-level matching. For ordinary matching pairs, the M value reflects the basic, behavioral, and intent aspects in a balanced way. Figure 3 The second matching degree M is a highly contextualized and dynamically synthesized final score, which more comprehensively and accurately explains the overall recommendation value of a supply-demand match than any single-dimensional or fixed-weight score. This step employs a dual screening logic of "filtering quality first, then controlling quantity," optimizing the final presentation of the recommendation results and addressing the core pain point of users being overwhelmed by a large number of low-quality recommendations in the background technology.

[0093] A preset matching threshold is established to create a "quality threshold": this is the bottom-line guarantee for recommendation quality. Only suppliers with an M value higher than this threshold are eligible to enter the candidate pool. This mechanism effectively filters out all "not good enough" options, ensuring that every recommendation received by the user has basic matching value, fundamentally eliminating the interference of "poor information," and guaranteeing the overall signal-to-noise ratio of the recommendation list.

[0094] Pre-setting the number of recommendations (N) achieves "quantity control": this reflects a deep understanding of user attention and cognitive load. While ensuring quality, by selecting the Top-N options, user attention is guided to the few highest-quality, most likely-to-succeed opportunities. This prevents users from getting lost among too many "good" options, improving decision-making efficiency. Simultaneously, when the number of high-quality matches is less than N, all matches will be recommended.

[0095] The final recommendation list output by step S5 is no longer a vaguely related list of exhibitors for users, but a curated list of "high-potential business opportunities" that has undergone rigorous screening based on product strength, interest, strategic intent, and business potential. The list has a controllable length and guaranteed quality, eliminating the need for users to painstakingly sift through massive amounts of information and allowing them to focus directly on the most valuable communications and negotiations, saving time and energy. The accuracy and depth of the recommendations increase the likelihood of users reaching substantive cooperation agreements with the recommended parties.

[0096] Please refer to Figure 1The multi-dimensional data collection and indicator quantification stage integrates the execution logic of steps S1 and S2, forming the data foundation of the entire intelligent matching method. In this stage, step S1 first captures and analyzes the static profiles of both supply and demand sides and the dynamic behavioral sequences of the demand side from multi-source heterogeneous data streams. This process transforms unstructured business descriptions and real-time user interaction behaviors into structured feature tags. Subsequently, for each potential supply-demand pairing, a quantitative evaluation is performed in parallel from three core dimensions, which is step S2. The static profile correlation coefficient P, representing the inherent fit of products and services, the dynamic behavioral resonance index D, measuring the resonance strength between the demand side's explicit intentions and the supply side's value proposition, and the category attention feature vector coupling degree B, assessing the alignment of both parties' business goals and strategies, are calculated respectively. This process symbolizes the construction of a map of static, dynamic, and intentional data for each potential business connection. Figure 3A holographic digital profile composed of multiple dimensions. Supply and demand potential assessment and high-potential identification, this stage precisely corresponds to the execution process of step S3. This method is no longer limited to surface data correlation, but delves into the potential business momentum behind the matching. This method comprehensively evaluates the supplier's urgency of participation in the exhibition, the status of new product launches, and the buyer's decision-making power, procurement cycle, and other deep factors to quantify and generate the first supply and demand potential matching degree G. This index aims to predict the "explosive potential" of a potential cooperation. Next, this method compares the calculated G value with a preset potential threshold. When the G value exceeds the threshold, this method will mark this supply and demand pairing. This is like setting up a highly sensitive "value radar" in a massive data stream, actively capturing and marking those "high-energy signals" that contain huge business potential, providing a basis for decision-making for subsequent resource allocation and precise intervention. Adaptive weight dynamic generation, this stage corresponds to the complete execution of step S4. When this method recognizes the "first mark" assigned in the previous stage, the core weight system of the recommendation algorithm will trigger a dynamic adaptive adjustment. This method no longer uses universal basic weights, but instead uses the quantified value of the first supply-demand potential matching degree as the key adjustment variable. A higher G value indicates stronger commercial potential. This method generates an adjustment factor through a preset gain function, strategically increasing the weights of the static profile correlation coefficient P and the category attention feature vector coupling degree B, while appropriately decreasing the weight of the dynamic behavior resonance index D. The introduction of this mechanism symbolizes a crucial leap from a static matching model that treats all suppliers equally to a context-aware, precisely controlled model that adapts to changing circumstances. The final matching degree calculation and precise recommendation constitute the core of the entire intelligent matching process's decision-making and output, and its logic fully covers step S5. This method first uses the weights (basic weights or adaptive weights) generated in the previous stage to perform a weighted summation of the static profile correlation coefficient P, the dynamic behavior resonance index D, and the category attention feature vector coupling degree B, converging into a single, commercially valuable final second matching degree M. Subsequently, this method uses this M value as the basis to sort all candidate suppliers in descending order. Finally, the sorted list is refined through a two-layer filtering funnel consisting of a matching threshold and a maximum recommendation quantity (N), ensuring that the final output recommendation list meets both high-quality standards and user information reception expectations, and accurately delivers the best business opportunities to the target demand.

[0097] Specific data examples are set at the 2024 International Smart Manufacturing Summit;

[0098] Buyer: Manager Li, Purchasing Director of a large automobile manufacturer.

[0099] Supplier: Y Technology, a startup focused on industrial AI visual inspection.

[0100] S1. Data Acquisition:

[0101] The following raw data about Manager Li and "Y Technology Company" was obtained.

[0102] Supplier: Static profile of "Y Technology Company";

[0103] Identity tags: [AI startup, computer vision, Series A funding stage, high-tech enterprise];

[0104] Product / Service Description: This product provides a deep learning-based industrial vision inspection method that enables real-time detection of micron-level surface scratches on automotive parts and has been successfully applied in the field of electronic assembly.

[0105] Category-focused feature vector tags: [Seeking sales leads, market expansion (automotive industry)];

[0106] Buyer: Manager Li - Static File;

[0107] Identity tags: [Purchasing Director, Automobile Manufacturing, Fortune 500];

[0108] Product / Service Description (Requirement Description): We are seeking a high-precision, high-efficiency automated visual defect detection solution for the group's new energy vehicle battery production line. The core requirement is to improve quality inspection efficiency and product yield.

[0109] Category-focused feature vector tags: [Precise procurement, technological cooperation];

[0110] Buyer: Manager Li - Dynamic Behavioral Sequence, specifically:

[0111] Action 1 (24 hours ago): Searched for the keyword "visual inspection" in the conference app.

[0112] Behavior 2 (8 hours ago): Browsed the homepage of a "smart manufacturing robot" company (not the supplier in this case).

[0113] Action 3 (2 hours ago): Saved the forum schedule titled "Application of AI Quality Inspection in the New Energy Industry".

[0114] Action 4 (1 hour ago): Scanned the QR code at the booth of "Y Technology Company".

[0115] S2. Calculate the static archive correlation coefficient P:

[0116] Process: This method performs NLP vectorization processing on the "product / service descriptions" of both parties.

[0117] Manager Li's demand feature vector: [Automotive, Battery, Production Line, Visual Inspection, High Precision, Defect Detection, Efficiency], hereinafter referred to as Vector A1;

[0118] The supplier feature vector of "Y Technology Company": [Deep learning, visual inspection, automotive parts, scratches, real-time detection], hereinafter referred to as vector A2;

[0119] This method, through analysis, generates the following weight vectors for Manager Li (vector A1) and "Y Technology Company" (vector A2) across the aforementioned seven dimensions. A higher weight value (0-1) indicates a stronger correlation between the item and its description.

[0120] Vector A1 (Manager Li's requirements): Mr. Li's requirements are very clear, so the core keyword has a high weight. However, although "battery" is mentioned, it may not be the primary target in this search for a visual inspection solution, so its weight is slightly lower.

[0121] The weights of vector A1 are as follows: A1 = [1.0, 0.6, 0.9, 1.0, 1.0, 1.0, 0.8];

[0122] Vector A2 (“Y Technology Company” supply, semantically enhanced):

[0123] The company is strongly associated with "automobiles" and "production lines," but these are not its only businesses, hence the weight is high but not 1.

[0124] It has no direct relation to "battery" and has a weight of 0.

[0125] "High precision" is its technical feature, but it may not meet Manager Li's expectations, so its weight is not 1.

[0126] A2=[0.9, 0, 0.7, 1.0, 0.8, 1.0, 0.7];

[0127] We will use these two precise vectors A1 and A2, and calculate them strictly according to the cosine similarity formula.

[0128] Formula: P = (A1·A2) / (||A||×||B||);

[0129] Calculate the vector dot product (A1·A2), which is the sum of the products of the corresponding dimensions of the two vectors.

[0130] A1·A2=(1.0×0.9)+(0.6×0)+(0.9×0.7)+(1.0×1.0)+(1.0×0.8)+(1.0×1.0)+(0.8×0.7)=0.9+0+0.63+1.0+0.8+1.0+0.56=4.89;

[0131] Calculate the magnitude (||A||) of vector A1; the magnitude is the square root of the sum of the squares of the vector's dimensions.

[0132] ;

[0133] Calculate the magnitude (||B||) of vector A2;

[0134] ;

[0135] Calculate the final static archive correlation coefficient P, and substitute the above calculation result into the final formula.

[0136] P = 4.89 / (2.410 × 2.105) = 4.89 / 5.073 ≈ 0.8500; after rounding to two decimal places, we get P = 0.85.

[0137] Using the cosine similarity algorithm, this method found that the two objects highly overlap in key dimensions such as [automotive, visual inspection, and defect detection]. Example data: P=0.85 (a value between 0 and 1, indicating high correlation);

[0138] The data source for the static profile correlation coefficient P is the fixed information filled in by both the supply and demand sides during registration. Examples include: the company's industry, size, region, technology field, and product catalog. Specifically, it's the most basic "hard condition" matching. It measures the degree of compatibility between the two parties in an objective, static dimension. Just like calculating the cosine similarity of two vectors, the dimensions of the vectors are the profile fields such as industry and size.

[0139] Calculate the dynamic behavioral resonance index D; this method uses a preset "behavioral value table" and a time decay function, the specific formula is as follows: ; (t is in hours) to quantify behavior.

[0140] The behavioral value mapping table is shown in Table 1 below:

[0141] Table 1: Behavioral Value Mapping Table

[0142]

[0143] Calculation: Behavior 1 (Search): Related to "Visual Detection," but not a direct interaction. Low contribution.

[0144] Behavioral Value Table: {Search: 2, Browse: 4, Save Schedule: 5, Scan Booth Code: 8};

[0145] Behavior 3 (Collection Schedule): Highly relevant to the topic. Moderate contribution.

[0146] Behavior 4 (scanning QR codes): This is a direct and strong-intent interaction with "Y Technology Company".

[0147] To reflect "immediacy," actions that occur earlier should have lower impact. We use an exponential decay function to simulate this process.

[0148] Immediate Behavior Score = Base Value Score × ;

[0149] Where t represents the number of hours since the action occurred. λ: Time decay coefficient. Set λ=0.05, this is a relatively gentle decay rate, meaning that actions taken 24 hours ago still have some value. e: The base of the natural logarithm, approximately equal to 2.718.

[0150] Its immediate behavior score = 8 × 0.9048 ≈ 7.24;

[0151] List Manager Li's behavioral sequence again and calculate its immediate behavioral score and relevance. Relevance is crucial, determining whether a particular behavior's score should be included in the matching calculation for "Y Technology Company." See Table 2 for details.

[0152] Table 2: Example of Instant Behavior Score Calculation for Each Behavior

[0153]

[0154] This method weights and sums all real-time behavior scores related to the "Y Technology Company" technology field, and then normalizes them.

[0155] Behavior 1 score: 0.60 × 0.8 = 0.48; Behavior 2 score: 2.68 × 0 = 0; Behavior 3 score: 4.52 × 0.9 = 4.07; Behavior 4 score: 7.61 × 1.0 = 7.61; Raw score DScore =0.48+0+4.07+7.61=12.16;

[0156] The raw score (e.g., 12.16) is not suitable for weighting with other indices (P and B, both between 0 and 1) in the final formula. Therefore, it needs to be normalized to the range of 0-1. This method sets a "maximum possible score" D. max This represents the score that an extremely active and highly matched user might achieve. (D) max This was derived through analysis of historical data. (D) max Preferred value setting D max =15.6. Formula: D=Raw DScore / D maxD = 12.16 / 15.6 ≈ 0.7795. After rounding to two decimal places, we get the Dynamic Behavioral Resonance Index: D = 0.78 (the normalized value indicates that Manager Li has a strong and immediate interest in this field and the company recently). The data source for the Dynamic Behavioral Resonance Index D is the real-time, continuous sequence of behaviors generated by the buyer during the exhibition. For example: the duration of browsing the supplier's homepage, which products were clicked, whether they initiated communication, whether they scheduled meetings, etc. This is an indicator for measuring "immediate interest" and "potential motivation." It reflects the result of the buyer "voting with their feet," a true expression of subjective intentions, and can correct or enhance the judgments of P and B.

[0157] Calculate the coupling degree B of the category-focused feature vector; this method queries the preset "category-focused feature vector compatibility matrix", as shown in Table 3 below:

[0158] Table 3: Example Table of Compatibility Matrix for Enterprise-Level Category Focus Feature Vectors

[0159]

[0160] The scoring criteria for enterprise-level category focus feature vector compatibility matrices are no longer simple "yes" or "no", but are based on the deep logic of business activities.

[0161] [Buyer: Bulk Purchase] vs. [Supplier: Major Client / Long-Term Partnership] (Score 1.0): This is the ideal business match. The buyer seeks stable supply and economies of scale, while the supplier seeks stable revenue and long-term relationships; their goals are perfectly aligned.

[0162] [Buyer: Seeking contract manufacturing] and [Supplier: Seeking OEM / ODM orders] (Score 1.0): A perfect match in definition, with both parties playing completely complementary roles in the value chain.

[0163] [Buyer: Technology Introduction] and [Supplier: Seeking Technology Partners] (Score 1.0): Both parties take technology cooperation as their core objective, and their intentions are highly aligned, whether it is licensing, joint development or technology transfer.

[0164] Moderate coupling (score 0.5-0.8):

[0165] [Buyer: Targeted Procurement] vs. [Supplier: Major Client / Long-Term Cooperation] (Score 0.6): The buyer may only be making a one-time purchase, while the supplier prefers to develop a long-term client relationship. There is a foundation for cooperation, but the long-term goals differ.

[0166] [Demand side: Finding channel / distribution partners] vs. [Supplier side: Launching new technology] (Score 0.8): The supplier's new technology needs channels to promote it, and the demand side's channels need competitive new products to fill them. This is a very strong synergy, but not a direct supply and demand relationship.

[0167] [Buyer: Market Research] vs. [Supplier: Brand / Marketing Promotion] (Score 0.7): The buyer wants to obtain information, and the supplier wants to convey information. In scenarios such as trade shows, this is an effective interaction, but it usually does not directly lead to a transaction.

[0168] Weak coupling (score 0.1-0.4):

[0169] [Buyer: Bulk Purchase] vs. [Supplier: Seeking Strategic Financing] (Score 0.2): A large purchase order (PO) can be strong evidence in a supplier's financing story, but it is not a direct capital injection. The connection is very indirect.

[0170] [Demand side: Seeking OEM manufacturing] and [Supplier side: Brand / Marketing promotion] (score 0.2): OEM manufacturers are usually behind-the-scenes heroes, and their business goals are not closely related to front-end brand promotion.

[0171] [Demand Side: Talent Recruitment] and [Supply Side: Any Intent] (generally low scores): The core of business matching platforms is B2B collaboration, not recruitment. While talent may be discovered during interactions, this is not the primary intention for either party to participate in the platform, resulting in extremely low compatibility.

[0172] Zero coupling or conflict relationship (score 0):

[0173] [Buyer: Targeted Procurement] vs. [Supplier: Finding Upstream Suppliers] (Score 0): The two sides are looking in different directions along the value chain. The buyer is the customer (downstream), while the supplier is looking for its own suppliers (upstream), and there is no overlap.

[0174] [Demand side: seeking contract manufacturing] and [Supplier side: seeking strategic financing] (score 0): For tech companies seeking financing (usually asset-light), there is almost no direct capital connection between asset-heavy demanders seeking contract manufacturing.

[0175] We further refine the category interest feature vectors that both parties have marked or identified by the algorithm in this method.

[0176] Buyer (Manager Li): Precision Procurement (Specific Product): Seeking high-precision visual inspection equipment for automotive battery production lines.

[0177] Technology Acquisition / Licensing: We are interested in the deep learning algorithms of "Y Technology Company" and are considering the possibility of future technology cooperation or licensing.

[0178] Supplier (“Y Technology Company”): Seeking Sales Leads (Short-Term): Looking to quickly identify potential clients with clear purchasing intentions during the trade show. Expanding Sales Channels: Seeking partners or agents in the automotive industry to expand market coverage.

[0179] This method systematically traverses intent combinations and consults Table 3, the enterprise-level category-focused feature vector compatibility matrix. It automatically combines each intent from the demand side with each intent from the supplier, and retrieves their compatibility score from the matrix. This process reveals all possibilities and strengths of cooperation between the two parties. The method uses the highest matching score as the coupling degree.

[0180] Combination 1: Targeted procurement by buyers and sales lead generation by suppliers;

[0181] Consult the matrix: Row 1 ("Precision Procurement"), Column 1 ("Short-Term Sales Leads").

[0182] Score: 0.9; In-depth logic: This is the most direct and classic business match. One party has a clear buying need, and the other party has a strong selling intention. The transaction path is short, and the goals are highly aligned, resulting in an extremely high compatibility score. This is the basis for the two parties to most likely begin dialogue immediately.

[0183] Combination 2: Targeted procurement by buyers and expansion of sales channels by suppliers;

[0184] Consult the matrix: Row 1 ("Precision Procurement"), Column 3 ("Expanding Sales Channels").

[0185] Score: 0.5; In-depth logic: Manager Li's procurement as an end-user does not perfectly align with "Y Technology Company's" goal of finding a channel partner. While a successful procurement case can boost a channel partner's confidence, Manager Li himself cannot become their sales channel. Therefore, this is a moderately strong indirect connection, rather than a direct target match.

[0186] Combination 3: The buyer introduces / licenses technology and the supplier seeks sales leads.

[0187] Consult the matrix: row 3 ("Technology Acquisition / Licensing"), column 1 ("Short-term Sales Leads").

[0188] Score: 0.4; In-depth logic: Technology acquisition is typically a long-term, highly complex strategic undertaking, involving multiple stages such as evaluation, negotiation, and integration. In contrast, the supplier's "short-term sales leads" aim for a quick transaction. The two differ in timeframe and decision-making level. Although technology licensing ultimately constitutes a form of sales, its nature is fundamentally different from product sales, resulting in lower compatibility.

[0189] Combination 4: The buyer introduces / licenses technology and the supplier expands sales channels;

[0190] Refer to the matrix: row 3 ("Technology Introduction / License"), column 3 ("Expanding Sales Channels").

[0191] Score: 0.3; In-depth logic: The connection between these two intentions is weaker. Technology importers typically want to use the technology in their own products or production, while channel expanders want to distribute mature products. Their roles and needs in the value chain are very different, making direct cooperation unlikely.

[0192] Aggregate the scores and determine the final B value; this method has obtained a list of compatibility scores for all possible combinations: [0.9, 0.5, 0.4, 0.3]. According to the algorithm rules, the category focus feature vector coupling degree B is taken as the maximum value among all combination scores, because the maximum value represents the strongest and most likely direction of cooperation between the two parties. B = max(0.9, 0.5, 0.4, 0.3);

[0193] Ultimately, B=0.90 (indicating a high degree of alignment between the core business objectives of both parties); the data source for the category focus feature vector coupling degree B is the set of "category focus feature vector labels" declared by both the supply and demand parties in the static archive.

[0194] Unlike the P-value, which measures the generalization of background similarity, the B-value focuses on the compatibility and complementarity of the two parties at the level of business objectives. Its calculation is not through simple label matching, but through a "category attention feature vector compatibility matrix" with built-in business logic, which traverses and deeply evaluates all possible combinations of intent (many-to-many) between the two parties.

[0195] Ultimately, the algorithm selects the highest compatibility score among all combinations as the B-value, which represents the strongest potential direction of cooperation between the two parties. Therefore, the B-value is a highly indicative core indicator that better reflects the feasibility of cooperation.

[0196] Calculate the first potential energy matching degree: Phase 1, input factor quantification. In order for the model to be computable, descriptive information such as "enterprise stage" and "job level" must be converted into standardized values ​​(e.g., between 0 and 1).

[0197] 1.1 The supplier potential (PEs) factor quantification table is shown in Table 4 below:

[0198] Table 4: Quantification of Supplier Potential (PEs) Factors

[0199]

[0200] 1.2 The quantitative table of demand potential energy (PEd) factor is shown in Table 5 below:

[0201] Table 5: Quantification Table of Demand Potential Energy (PEd) Factor

[0202]

[0203] The steps for checking session activity in Table 5 include:

[0204] S100, the count values ​​of the first high-value behavior, the second high-value behavior, and the general interaction behavior generated by the demand side during the meeting period; wherein, the first high-value behavior corresponds to meeting reservation; the second high-value behavior corresponds to core interactions other than meeting reservation; and the general interaction behavior corresponds to browsing;

[0205] S200. Determine whether the count value of the first high-value behavior is not less than 1, or whether the count value of the second high-value behavior is not less than a first preset threshold (e.g., 5). If so, the meeting activity level is determined to be the first preset score (e.g., 0.9), and the determination ends. First, determine whether the buyer has generated at least one first high-value behavior, or whether the number of times the buyer has generated a second high-value behavior has reached a high preset threshold (e.g., 5 times). Once this condition is met, regardless of other behaviors, this method immediately determines its activity level to the highest first preset score (e.g., 0.9), and terminates the subsequent determination.

[0206] S300: If the conditions of S200 are not met, then continue to determine whether the count value of the second highest value behavior is greater than 0 and less than the first preset threshold, or whether the count value of the second highest value behavior is 0 and whether the count value of general interaction behavior is not less than the second preset threshold; if so, then the meeting activity level is determined to the second preset score (e.g., 0.6), and the judgment ends. This method divides the client's meeting behavior into three value levels: the first highest value behavior (such as initiating or accepting meeting reservations, representing a very strong business intention), the second highest value behavior (such as collecting exhibitors, scanning QR codes, exchanging business cards, and other core interactions), and general interaction behavior (such as page browsing and document downloading). This method counts the values ​​of these three types of behavior separately. This method adopts a "short-circuit" cascading judgment logic, prioritizing the response to high-value signals: this ensures that users with clear negotiation intentions can be immediately identified as highly active. If the highest priority condition is not met, then proceed to this level of judgment. This method checks whether the second highest-value behavior has occurred (greater than 0 times but less than the threshold of 5), or whether, although there is no core interaction, the number of general interactions reaches a relatively high level (e.g., not less than the second preset threshold). If either condition is met, the activity level is determined to be a medium level with a second preset score (e.g., 0.6).

[0207] S400. If neither S200 nor S300 is met, the meeting activity level will be set to a third preset score (e.g., 0.3) or zero, depending on whether the demander has engaged in any activity during the meeting period.

[0208] Example of applying the activity level during the meeting: User 001: Booked 1 meeting and scanned the QR code 2 times. Judgment: Meets the "High Activity" rule 1 → High (0.9).

[0209] User 002: Scanned QR codes 3 times and sent messages 2 times. Judgment: Total core interactions are 3 + 2 = 5. Satisfies "High Activity" rule 2 → High (0.9).

[0210] User 003: Scanned the QR code twice and viewed the company homepage six times. Judgment: Does not meet the high activity level requirement. Core interaction count is 2, meeting the "medium activity level" rule 1 → Medium (0.6).

[0211] User 004: Viewed the company homepage 12 times, with no core interactions. Judgment: Does not meet the high activity level requirement. Meets the "Medium Activity Level" rule 2 → Medium (0.6).

[0212] User 005: Viewed the company homepage 3 times. Judgment: Does not meet any high or medium activity criteria → Low (0.3).

[0213] Phase Two, Supplier Potential Energy (PEs) Calculation Model: PEs = (Enterprise Stage Score × 0.4) + (Exhibition Target Score × 0.4) + (New Product Status Score × 0.2) Weighting Logic: The major stage the enterprise is in and its strategic goals are the fundamental factors that determine its motivation, and have the highest weight; the launch of a new product is a specific tactical action, and is considered as a bonus.

[0214] Demand potential energy (PEd) calculation model: PED = (job level score × 0.3) + (procurement decision-making power score × 0.5) + (meeting activity score × 0.2) Weight logic: actual decision-making power is the core of the core and has the highest weight; job level is an important reference for decision-making power; and on-site behavior reflects the current sense of urgency.

[0215] Supplier "Y Technology Company": Series A funding (0.9), expansion into new markets (0.9), and the launch of new products (1.0);

[0216] Buyer "Manager Li": Director (0.9), Key Influencer (0.8), High Activity (0.9);

[0217] Calculation steps:

[0218] Calculate PEs: PEs = (0.9 × 0.4) + (0.9 × 0.4) + (1.0 × 0.2) = 0.36 + 0.36 + 0.20 = 0.92;

[0219] Calculate PED: PED = (0.9 × 0.3) + (0.8 × 0.5) + (0.9 × 0.2) = 0.27 + 0.40 + 0.18 = 0.85;

[0220] Calculate the product: PEs×PEd=0.92×0.85=0.782; this represents the overall strength of the potential motivations of both the supply and demand sides.

[0221] Calculate the absolute difference: |PEs-PEd|=|0.92-0.85|=0.07; it represents the degree of equilibrium of the potential motivations of supply and demand, and the smaller the value, the more balanced the supply and demand.

[0222] The method generates the G value by using a pre-defined nonlinear function. This function "synthesizes" the two inputs mentioned above. A specific implementation of this function is as follows: Substitute the "product" and "absolute difference" into the formula for calculation:

[0223] The final value of G is obtained as follows: G = 0.884 × 0.9951 ≈ 0.88; the first supply and demand potential energy matching degree is G = 0.88;

[0224] Judgment and labeling based on potential energy threshold: Obtaining the preset potential energy threshold: The preset potential energy threshold G in this method is... threshold =0.70.

[0225] Execute the judgment logic: compare the calculation result with the threshold: G(0.88)>G threshold (0.70).

[0226] It should be added that: "potential energy threshold G" threshold The calibration method for “” is as follows:

[0227] Potential threshold G threshold This is the "diversion valve" of the dynamic recommendation strategy of this invention. Its value directly affects the computational efficiency and the upper limit of the recommendation effect of this method. There is a mutual constraint between the two:

[0228] Excessive G threshold This could lead to overly stringent screening criteria, making the method less effective at identifying potential business opportunities and potentially missing some suboptimal but still highly valuable matching pairs (referred to as high-potential underreporting), thus missing out on business opportunities.

[0229] Too low G threshold This would reduce the value of "high potential" recommendations, causing a large number of regular matching pairs to enter the enhanced recommendation channel. This not only increases unnecessary computational overhead but may also interfere with the accuracy of adaptive weight adjustment due to excessive noise (referred to as high potential false positives).

[0230] Therefore, G thresholdThe optimal value is not subjectively set, but is derived through the following offline backtesting calibration experiment of this method: Experimental data preparation: Prepare a complete dataset of one or more previous similar exhibitions. This dataset not only contains the static profiles and dynamic behavior sequences of all participants, but also must contain "real matching results" labels that have been manually annotated or obtained through subsequent business tracking. This label marks each supply and demand match in history as one of the following three categories: [Highly effective conversion] (e.g., reaching a procurement intention of hundreds of millions of yuan after the event), [Moderately effective] (e.g., both parties establish contact and conduct multiple rounds of technical exchanges), [Basically ineffective] (e.g., no substantial follow-up).

[0231] Experimental environment and tools: Data backtesting platform: An offline computing environment was built that can fully reproduce the matching algorithm of this invention.

[0232] The "True Matching Results" database stores the labeled data mentioned above, serving as the "gold standard" for evaluating algorithm performance. Calibration process: In offline data analysis, a potential energy threshold G is set. threshold The search range is [0.50, 0.90], with a step size of 0.01. For each G within the search range... threshold For candidate values, perform the following operations: On the backtesting platform, iterate through all supply-demand pairs in the historical dataset and calculate the first supply-demand potential matching degree G for each pair. Based on the current G... threshold Candidate values ​​are used to categorize all pairings into "high-potential matches determined by this method" and "regular matches determined by this method." The results of this method are compared with the "real match results" database, and the following two core metrics are calculated: Potential-MissRate: The proportion of pairings judged as "regular" by this method but actually belonging to the "high-conversion" category, out of the total number of "high-conversion" pairings. Potential-False-AlarmRate: The proportion of pairings judged as "high-potential" by this method but actually belonging to the "basically ineffective" category, out of the total number of "high-potential matches determined by this method."

[0233] Optimal value selection: Plotting the false negative rate and false positive rate as a function of G threshold The graph shows the changing trend. The final selected optimal value, G, is the one that minimizes the high-potential false positive rate while keeping the high-potential false negative rate below 5% (ensuring no major business opportunities are missed). threshold Value. After the above calibration, this embodiment finally determines G. threshold The preferred value is 0.70.

[0234] The criteria have been met. Completed: This method determines that the combination of "Y Technology Company" and "Manager Li" is a high-potential match.

[0235] A first tag is successfully applied to the matching pair for subsequent priority recommendation processing.

[0236] The first supply and demand potential matching degree G is the result of the fusion calculation of static files (the basis of P) and dynamic behavior sequences (the basis of D). G is not an independent dimension, but a "preliminary conclusion" derived from a comprehensive evaluation of information from multiple dimensions such as P and D.

[0237] Technical implications: The first supply and demand potential matching degree acts as a "commander-in-chief" or "diverter." It does not participate in the final precise score calculation, but rather in making a strategic judgment: whether the current pairing has "great potential and deserves detailed analysis," or "moderate potential and should be handled according to standard procedures." Its core task is to initiate or skip the adaptive weight adjustment of S4.

[0238] Continuing with the example of matching the demander "Manager Li" with the supplier "Y Technology Company," in the previous steps, we calculated the following core indicators: Static profile correlation coefficient P: 0.85; Dynamic behavioral resonance index D: 0.78; Category attention feature vector coupling degree B: 0.90; First supply and demand potential matching degree G: 0.88;

[0239] Meanwhile, this method presets the following parameters:

[0240] Basic weights: First weight (corresponding to P): wpbase=0.40; Second weight (corresponding to D): wdbase=0.40;

[0241] Third weight (corresponding to B): wbbase=0.20; (Note: The sum of the base weights is 1, representing that under normal matching, this method considers static profiles and dynamic behaviors to be equally important, with category-focused feature vectors as a supplement.) Potential threshold (G) threshold ): 0.70;

[0242] S4: Adaptive weight generation:

[0243] Step 4.1: Detect the first marker and trigger the recommended enhancement strategy;

[0244] This method first detects whether the current supply and demand pair has a "first marker".

[0245] Decision logic: Compare G with G threshold .

[0246] Calculate: G(0.88) > G threshold (0.70).

[0247] Conclusion: The conditions are met. This method determines this pairing as a high-potential match, successfully identifies the "first marker," and immediately triggers the recommendation enhancement strategy.

[0248] Step 4.2: Calculate the adjustment factor Δ; This method calculates the adjustment factor Δ based on the quantized value of G using a preset gain function. The design goal of the gain function is: when the value of G just exceeds the threshold, Δ is small; when the value of G is close to 1, Δ increases, but there is an upper limit to prevent excessive weight shift. This embodiment uses a scaled and translated variant of the Sigmoid function: ;

[0249] Where: Δmax (maximum adjustment): preset to 0.15. This limits the maximum value of weight adjustment, ensuring model stability. k (curve steepness): preset to 10. Used to control the rate of function growth near the threshold. e is the base of the natural logarithm.

[0250] Calculation process: GG threshold =0.88-0.70=0.18;-k×(GG threshold ) = -10 × 0.18 = -1.8;

[0251] e (-1.8) ≈0.1653; 1 / (1+0.1653)=1 / 1.1653≈0.8581; 0.8581-0.5=0.3581; 0.3581×2=0.7162;

[0252] Δ=Δmax×0.7162=0.15×0.7162≈0.1074;

[0253] The final adjustment factor Δ is rounded to three decimal places, resulting in Δ = 0.107.

[0254] Step 4.3: Calculate intermediate weights; This method adjusts the basic weights based on the strategy of "enhancing static matching (P) and category-focused feature vectors (B), and suppressing pure behavior (D)".

[0255] First intermediate weight (P): wp = wpbase + Δ = 0.40 + 0.107 = 0.507;

[0256] The third intermediate weight (B): wb = wbbase + Δ = 0.20 + 0.107 = 0.307;

[0257] The second intermediate weight (D): wd = wdbase - Δ = 0.40 - 0.107 = 0.293;

[0258] Step 4.4: Normalize to generate the final adaptive weights;

[0259] To ensure that the final sum of weights is 1, the intermediate weights are normalized.

[0260] Weighted sum: Sum = wp + wd + wb = 0.507 + 0.293 + 0.307 = 1.107;

[0261] Final adaptive weights: Wp = wp / Sum = 0.507 / 1.107 ≈ 0.458; Wd = wd / Sum = 0.293 / 1.107 ≈ 0.265;

[0262] Wb=wb / Sum=0.307 / 1.107≈0.277;

[0263] Conclusion: For the high-potential matching pair of "Manager Li" and "Y Technology Company", this method generates new adaptive weights {final weight for P: 0.458, final weight for D: 0.265, final weight for B: 0.277}. Compared with the base weights {base weight for P: 0.40, base weight for D: 0.40, base weight for B: 0.20}, the importance of static profile (P) and category focus feature vector (B) is increased, while the weight of dynamic behavior (D) is decreased.

[0264] S5: Final Match Calculation and Recommendation List Generation:

[0265] Step 5.1: Calculate the final second matching degree M;

[0266] This method uses adaptive weights to perform a weighted summation of P, D, and B.

[0267] Calculation formula: M=(P×Wp)+(D×Wd)+(B×Wb); Calculation process: M=(0.85×0.458)+(0.78×0.265)+(0.90×0.277)M=0.3893+0.2067+0.2493=0.8453;

[0268] The final second-degree match M is 0.845.

[0269] Step 5.2: Sort all candidate suppliers and generate the first recommendation sequence. To demonstrate the sorting and filtering process, this method assumes that the matching degree M between Manager Li and several other candidate suppliers has been calculated, as shown in Table 6 below:

[0270] Table 6: First Recommended Sequence

[0271]

[0272] This method sorts all candidate suppliers in descending order based on the M value to generate the first recommendation sequence:

[0273] 1. Y Technology Company (M=0.845); 2. Intelligent Robot Company (M=0.752); 3. Precision Measurement Equipment Manufacturer (M=0.748); 4. Industrial Software Solution Provider (M=0.700); 5. Data Service Provider (M=0.560);

[0274] Step 5.3: Generate a second candidate pool sequence based on the matching threshold filtering;

[0275] This method presets the following filtering parameters:

[0276] Match threshold threshold ): 0.80; The final matching threshold is denoted as Match. threshold This represents a preset minimum matching score used for filtering in the final recommendation list. Only suppliers with a second matching score M higher than this threshold will be presented to the demand side. In this embodiment, the preferred value is 0.80.

[0277] It should be added that "Match threshold" threshold The calibration method for "" is as follows: Matching threshold Matc hthreshold The value determines the quality and length of the recommendation list, and there is a mutual constraint between the two: an excessively high Match value... threshold This will generate a very short but high-quality recommendation list, improving recommendation precision, but may filter out some second-best options that users might also be interested in, leading to a decrease in recommendation recall.

[0278] Too low Match threshold This will generate a lengthy recommendation list, ensuring a high recall rate, but the list may contain a large number of unrelated options, reducing precision and increasing the user's filtering burden.

[0279] Therefore, the optimal value of this parameter was also obtained through offline backtesting calibration experiments based on a user satisfaction model:

[0280] Experimental setup: The same historical dataset and backtesting platform as described above were used. At this stage, the potential energy threshold G... threshold The value has been fixed at 0.70. Calibration process: In offline data analysis, set the matching threshold (Match). threshold The search range is [0.60, 0.95], and the step size is 0.01.

[0281] For each Match within the search range threshold Candidate values, perform the following operations: a. On the backtesting platform, run the complete matching recommendation algorithm (including the fixed G). threshold(a) Calculate the second match degree M between each demander and all suppliers in the historical dataset, and generate a complete sorted list. (b) Use the current Match threshold The candidate values ​​are truncated from the sorted list to generate the final simulated recommendation list. The simulated recommendation list is compared with the "real matching results" database to calculate the following two key business metrics: Precision@N: The proportion of suppliers in the recommendation list that fall under the "high conversion rate" or "moderate effectiveness" category. N is the preset upper limit for the number of recommendations. Recall: The proportion of suppliers in the recommendation list that fall under the "high conversion rate" or "moderate effectiveness" category out of all "real" effective matches for that customer.

[0282] Optimal value selection: Calculate each Match threshold The F1 score (F1-Score) corresponding to each candidate value is the harmonic mean of precision and recall. The final match selected maximizes the F1 score. threshold The value represents the optimal balance between precision and recall, ensuring recommendation quality without sacrificing too much coverage. After the above calibration, this embodiment finally determines the Match value. threshold The preferred value is 0.80.

[0283] The preset recommendation quantity N=3; this method selects all recommendations with an M value higher than Match from the first recommendation sequence. threshold Supplier with a value of (0.80). Screening results:

[0284] Y Technology Company (0.845>0.80) → Retain; Intelligent Robot Company (0.752<0.80) → Remove; Precision Measurement Equipment Manufacturer (0.748<0.80) → Remove; Industrial Software Solution Provider (0.700<0.80) → Remove; Data Service Provider (0.560<0.80) → Remove;

[0285] Second candidate pool sequence: Y Technology Company.

[0286] Step 5.4: Generate the final recommendation list based on the number of recommendations N; this method determines whether the number of suppliers in the second candidate pool sequence is greater than N. Judgment: The number of suppliers in the candidate pool (1) is not greater than (3). Execution logic: If not, then all suppliers in the second candidate pool sequence are used as the final recommendation list. Final recommendation list: The final supplier list recommended to Manager Li by this method is as follows:

[0287] Y Technology Company.

[0288] It should be noted that all calculation formulas in this application employ regression analysis, including but not limited to machine learning algorithms, to deeply analyze the collected parameters and identify their natural trends and interrelationships. Specialized software, such as Python's Scikit-learn library or the R language, is used to automatically generate mathematical models that match the data. Then, cross-validation and other methods are used to objectively evaluate the model performance, and continuous feedback and optimization are combined to ensure that the created formulas truly reflect the inherent laws of the data, thereby guaranteeing their effectiveness and accuracy. In all calculation formulas in this application, the parameters in each formula undergo dimensionless processing within a consistent range to ensure that different physical quantities are compared on the same scale; dimensionless processing techniques include, but are not limited to, min-max-normalization and Z-score standardization.

[0289] The algorithm of this invention is implemented as a Python script. Before executing the core logic, the program first executes a data loading module (e.g., using the widely used pandas library in Python) configured to read the aforementioned spreadsheet file and load its contents into the program's working memory (e.g., a DataFrame data structure). Subsequent algorithm steps will directly query and retrieve the required configuration parameters from this in-memory data structure.

[0290] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for intelligent matching and recommendation between supply and demand sides in the exhibition industry, characterized in that, The specific steps include: S1. Obtain static profiles of both the supply and demand sides and dynamic behavior sequences of the demand side. The static profiles include preset identity tags, product and service descriptions, and category-related feature vector tags for both the supply and demand sides. S2. For each supply-demand pairing, calculate: the static file correlation coefficient based on the static file; the dynamic behavior resonance index based on the dynamic behavior sequence of the demander; and the category attention feature vector coupling degree based on the category attention feature vector tags in the static file; S3. By combining static files and dynamic behavior sequences, the first supply and demand potential energy matching degree is evaluated and quantified, and a potential energy threshold is preset. When the first supply and demand potential energy matching degree is greater than the potential energy threshold, the supply and demand pairing is marked as the first pairing. S4. Based on the supply and demand pairing marked in the first step, dynamically adjust the preset basic weights to generate adaptive weights that are compatible with the current supply and demand pairing. S5. Using adaptive weights, the static file correlation coefficient, dynamic behavior resonance index, and category attention feature vector coupling degree are weighted and summed to generate the final second matching degree. All suppliers are ranked according to the second matching degree, and suppliers with a second matching degree value higher than the preset matching threshold are recommended to the demander.

2. The intelligent matching and recommendation method for supply and demand sides of exhibitions according to claim 1, characterized in that: The method for obtaining the static file correlation coefficient is as follows: natural language processing is performed on the product and service description text in the static files of both the supply and demand sides to extract and construct keywords, and finally obtain the supplier feature vector and the demand side feature vector. The cosine similarity algorithm is used to calculate the cosine value of the angle between the supplier's feature vector and the demander's feature vector, and the normalized cosine value is used as the static file association coefficient.

3. The intelligent matching and recommendation method for supply and demand sides of exhibitions according to claim 2, characterized in that: Before constructing the supplier feature vector and the demand feature vector, a static file preprocessing step is also included: identifying the core exhibitors among the suppliers, denoted as ROI suppliers; dividing the suppliers into ROI suppliers and non-ROI suppliers; for ROI suppliers, using knowledge graph-based entity linking technology for feature enhancement; for non-ROI suppliers, only using the term frequency inverse document frequency algorithm to construct the supplier feature vector.

4. The intelligent matching and recommendation method for supply and demand sides of exhibitions according to claim 1, characterized in that: The dynamic behavioral resonance index is obtained as follows: Preset basic behavior scores for various dynamic behavior types of the demand side to form a behavior value mapping table; dynamic behavior types include searching keywords, browsing exhibitor pages, collecting exhibitors, scanning QR codes, staying at on-site booths, and making appointments for negotiations; For each behavior in the demander's dynamic behavior sequence, a basic behavior score is obtained, and the basic behavior score is weighted by a time decay index to calculate the immediate behavior score. The time decay index is negatively correlated with the time interval from the time the behavior occurred to the current time. All immediate behavior scores related to the supplier are accumulated to obtain the dynamic behavior resonance index.

5. The intelligent matching and recommendation method for supply and demand sides of exhibitions according to claim 1, characterized in that: The method for obtaining the coupling degree of category-focused feature vectors is as follows: construct a category-focused feature vector compatibility matrix. The category-focused feature vector compatibility matrix predefines the compatibility scores between different supplier category-focused feature vectors and different demander category-focused feature vectors. In the current supply and demand matching, obtain the supplier's supplier intent set and the buyer's buyer intent set, where each intent set contains at least one category interest feature vector label; Each consumer category interest feature vector in the consumer intent set is paired with each supplier category interest feature vector in the supplier intent set to generate one or more intent combination pairs. For each generated intent pair, based on the category attention feature vector compatibility matrix, find and determine its corresponding compatibility score, thereby obtaining a list containing the compatibility scores of all intent pairs. Select the highest compatibility score from the list as the coupling degree of the category-focused feature vector.

6. The intelligent matching and recommendation method for supply and demand sides of exhibitions according to claim 1, characterized in that: The first method for obtaining the supply-demand potential energy matching degree is as follows: A supplier potential energy assessment model and a demand potential energy assessment model are constructed separately, wherein... The supplier's potential is quantified based on its enterprise stage, the urgency of its exhibition goals, and the status of its new product launch; the buyer's potential is quantified based on its job level, purchasing decision-making power, and activity level during the exhibition; thus, supplier potential values ​​PEs and buyer potential values ​​PED are obtained respectively; through a preset nonlinear function, the product of supplier potential values ​​PEs and buyer potential values ​​PED, as well as the absolute difference between the two, are combined to generate the first supply and demand potential matching degree; when the first supply and demand potential matching degree is higher than the potential threshold, it is judged as a high-potential matching pair, and the corresponding supply and demand pair is first paired to form the first mark.

7. The intelligent matching and recommendation method for supply and demand sides of exhibitions according to claim 6, characterized in that: The steps to build session engagement include: S100. Count the first high-value behavior, the second high-value behavior, and the general interaction behavior generated by the demand side during the meeting period; wherein, the first high-value behavior corresponds to meeting reservation; the second high-value behavior corresponds to core interactions other than meeting reservation; and the general interaction behavior corresponds to browsing. S200. Determine whether the count value of the first high-value behavior is not less than 1, or whether the count value of the second high-value behavior is not less than the first preset threshold; if so, determine the meeting activity level as the first preset score and end the determination. S300. If the conditions of S200 are not met, then continue to determine whether the count value of the second high-value behavior is greater than 0 and less than the first preset threshold, or determine whether the count value of the second high-value behavior is 0 and whether the count value of the general interaction behavior is not less than the second preset threshold; if so, then determine the meeting activity level as the second preset score and end the determination. S400. If neither S200 nor S300 is met, the meeting activity level will be set to the third preset score or zero, depending on whether the demander has engaged in any activity during the meeting period.

8. The intelligent matching and recommendation method for supply and demand sides of exhibitions according to claim 1, characterized in that: The basic weights include the first weight corresponding to the correlation coefficient of static archives, the second weight corresponding to the dynamic behavior resonance index, and the third weight corresponding to the coupling degree of category attention feature vector.

9. The intelligent matching and recommendation method for supply and demand sides of exhibitions according to claim 5, characterized in that: If the current supply and demand pair has a first label, the recommendation enhancement strategy is triggered, including: based on the quantified value of the first supply and demand potential matching degree, an adjustment factor Δ is calculated and generated that is positively correlated with the value of the first supply and demand potential matching degree through a preset gain function; the first weight is added to the adjustment factor Δ to obtain the intermediate first weight; the third weight is added to the adjustment factor Δ to obtain the intermediate third weight; the adjustment factor Δ is subtracted from the second weight to obtain the intermediate second weight; the intermediate first weight, intermediate second weight, and intermediate third weight are treated as a whole and normalized to generate the final adaptive weight.

10. The intelligent matching and recommendation method for supply and demand sides of exhibitions according to claim 9, characterized in that: Based on the calculated second matching degree, all candidate suppliers are sorted in descending order to generate the first recommendation sequence; A preset matching threshold is set, and all suppliers with a second matching degree higher than the preset matching threshold are selected from the first recommendation sequence to form a second candidate pool sequence. It is then determined whether the number of suppliers in the second candidate pool sequence is greater than the preset number of recommendations. If so, the preset number of suppliers are selected from the top of the second candidate pool sequence as the final recommendation list. If not, then all suppliers in the second candidate pool sequence will be used as the final recommendation list.

Images