Scientific and technological achievement recommendation method, device, medium and system based on business behavior chain

Abstract

This invention discloses a method, device, medium, and system for recommending scientific and technological achievements based on business behavior chains, belonging to the field of scientific and technological achievement recommendation technology. The steps include: profiling the organization and constructing an internal business behavior chain; upon receiving a recommendation request, reading the profiling data, performing multi-path hybrid recall, with the recall ratio of each path dynamically adjusted according to the scenario, and merging the multi-path recall results into a recall candidate set; using shared embedding hybrid dual-tower coarse ranking to score and re-rank the recall candidate set, forming a coarse-ranked candidate set; constructing prompt word templates for the coarse-ranked candidate set, inputting the prompt words into a large language model to jointly generate a matching score and a natural language recommendation reason. This invention can achieve high-precision, high-efficiency, and interpretable scientific and technological achievement recommendation in large-scale organizational internal scientific and technological collaboration scenarios.

Description

Technical Field

[0001] This invention relates to the field of technology for recommending scientific and technological achievements, and more specifically, to a method, device, medium, and system for recommending scientific and technological achievements based on a business behavior chain. Background Technology

[0002] Under the national policy guidance of strengthening the leading role of enterprises in technological innovation, large enterprise groups, research institutes, and university alliances have continuously increased their R&D investment, accumulating a vast amount of internal scientific and technological achievements, including patents, proprietary technologies, processes and methods, software systems, experimental equipment, technical standards, and research papers. These achievements are widely distributed across subsidiaries, business units, laboratories, and partner organizations within the group, and generally suffer from common problems such as "dormant achievements, redundant investment, and inefficient collaboration." To improve the efficiency of internal technology asset reuse, accelerate the transformation of achievements, and avoid resource waste, it is urgent to build an intelligent technology achievement recommendation system for internal organizations. This system should be able to accurately match reusable and transformable internal achievements based on the technical capabilities and actual needs of each unit, promoting a collaborative mechanism of "innovation in one place, sharing across the whole."

[0003] However, scientific and technological achievements are characterized by high specialization, diverse expressions, and complex semantics. For example, "defect detection algorithm based on deep learning" and "industrial visual quality inspection system" do not directly overlap in keywords, but their technical essence is highly related; another example is a subsidiary whose business registration scope is broad, but its actual technical capabilities are concentrated in the field of "biomedical materials." Traditional recommendation methods are difficult to effectively model such semantic associations and true capability profiles.

[0004] Furthermore, large organizations (especially state-owned enterprises, central enterprises, and research institutions) have extremely high requirements for system interpretability, data security, and governance compliance. They typically prohibit reliance on external user behavior data (such as clicks and browsing) and must manage systems based on their internal authoritative classification systems. Therefore, there is an urgent need for an intelligent recommendation technology solution that integrates structured tags and semantic understanding, adapts to internal collaborative scenarios, and operates entirely on a closed-loop system based on proprietary data. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method, device, medium and system for recommending scientific and technological achievements based on business behavior chains, which can achieve high-precision, high-efficiency and interpretable recommendation of scientific and technological achievements in the context of scientific and technological collaboration within large organizations.

[0006] The objective of this invention is achieved through the following solution: A technology recommendation method based on business behavior chains includes: Profiling the organization and building internal business behavior chains; After receiving a recommendation request, read the profile data, perform multi-path hybrid recall, dynamically adjust the recall ratio of each path according to the scenario, and merge the recall results of multiple paths into a recall candidate set; The recall candidate set is scored and reordered using a shared embedded hybrid dual-tower coarse ranking method to form a coarse-ranked candidate set. For the coarsely ranked candidate set, a prompt word template is constructed, and the prompt words are input into a large language model to jointly generate a matching score and a natural language recommendation reason.

[0007] Furthermore, the process of profiling the organization and constructing the internal business behavior chain specifically includes the following sub-steps: High-value agency behavior is defined by extracting high-value agency behaviors related to the titles of scientific and technological achievements from organizations with historical behavior records. Then, a semantic encoder is used to generate semantic vectors, and an exponential time decay mechanism is introduced to fuse and generate a first demand vector of enterprises that reflects the organization's true technological interests. For organizations without historical behavior records, a second demand vector of enterprises is generated using the organization's own scientific and technological achievements or business scope text. The first demand vector and the second demand vector of enterprises are used as profile data storage. Specifically, the definition of high-value agency behavior refers to interactive behaviors in historical behavior records that have business intent and are accompanied by actual costs.

[0008] Furthermore, the execution of multi-path hybrid recall, where the recall ratio of each path is dynamically adjusted according to the scenario, and the results of multiple recall paths are merged into a recall candidate set, specifically includes the following sub-steps: Dense semantic recall: Using the enterprise's first demand vector or the enterprise's second demand vector as the query, a FAISS approximate nearest neighbor search is performed in the full database of scientific and technological achievement titles and abstracts to capture technological opportunities that are semantically similar but whose labels may not overlap, and to obtain the candidate set R1. Structured tag recall: Based on the industry tag list and professional tag list in the user profile, the results belonging to the same tag are matched by the inverted index to obtain the candidate set R2; High-value outcome pool recall: Dynamically maintain the set of high-quality outcomes defined by business rules to obtain the candidate set R3; The recall ratios of dense semantic recall, structured tag recall, and high-value result pool recall are dynamically adjusted according to the scenario, and finally merged into a recall candidate set. .

[0009] Furthermore, in the shared embedded hybrid dual-tower coarse ranking, the user tower and the result tower share the same set of learnable embedding tables; The learnable embedding table covers four categories of identifiers: industry, profession, region, and organization type. Specifically, it includes industry embedding table, profession embedding table, region embedding table, and organization type embedding table. The industry embedding table and the professional embedding table are shared by the user tower and the outcome tower to ensure that when a profile contains an industry label and the candidate outcome happens to belong to that industry, the two are represented in the vector space in a consistent manner, maximizing the effectiveness of the structured matching signal. The region embedding table and the organization type embedding table are used only by the user tower to model the impact of geographical collaboration convenience and organizational function differences on technology adoption preferences.

[0010] Furthermore, the step of scoring and re-ranking the recall candidate set using shared embedding hybrid dual-tower coarse ranking specifically includes the following sub-steps: Both the enterprise's primary demand vector and secondary demand vector, along with the deliverable title vector, adopt frozen semantic vectors. The user profile only stores an ID list, and an online table lookup is used to generate a structured representation on average. When constructing the comprehensive user representation, User Tower first dynamically constructs a structured representation based on the cached identifier list. For the industry tag list and the professional tag list, it queries the corresponding shared embedding table and performs an arithmetic average to obtain the industry profile vector and the professional profile vector. If the list is empty, it is filled with a zero vector. For the region identifier and the organization type identifier, it directly queries their exclusive embedding table to obtain the corresponding dense vector. The four structured vectors are concatenated with the frozen enterprise demand vector to form the final comprehensive user representation. When constructing a comprehensive representation of results, the results tower queries the shared embedded table for each candidate result with its industry and professional tags to obtain the corresponding industry and professional vectors, and then concatenates them with the frozen title semantic vectors and business rule features to form a comprehensive representation of results. The results themselves are not bound to the region or organization type. The user's comprehensive representation and the result's comprehensive representation are then concatenated and input into a multilayer perceptron (MLP) for nonlinear interaction modeling. The final output is a matching score between 0 and 1, which is used to rank the candidate results.

[0011] Furthermore, the construction of the business behavior chain specifically includes the following sub-steps: using the ID of the organizational sub-unit as the aggregation key, aggregating multiple types of high-value agency behaviors from the internal business system or results platform, sorting them in chronological order to form a behavior sequence list, and storing each element in the sequence list in the form of structured data.

[0012] Furthermore, the step of inputting prompt words into a large language model to jointly generate matching scores and natural language recommendation reasons specifically includes the following sub-steps: The constructed prompt word template includes enterprise profile, business behavior chain and result details. For each candidate result, a structured context is constructed based on the prompt word template, and deep semantic understanding and personalized reasoning are performed by calling a large language model. The large model outputs a normalized matching score, reflecting the overall fit and the natural language recommendation rationale.

[0013] A technology achievement recommendation device based on a business behavior chain includes a processor and a memory, wherein the memory stores a computer program that, when loaded by the processor, executes the method described in any of the preceding methods.

[0014] A computer-readable storage medium storing a computer program that, when loaded by a processor, executes the method described in any of the preceding claims.

[0015] A technology achievement recommendation system based on business behavior chain includes the technology achievement recommendation device based on business behavior chain as described above.

[0016] The beneficial effects of this invention include: This invention innovatively constructs a three-stage recommendation pipeline by integrating semantic recall, structured tag matching, and business rules, achieving high-precision and interpretable result recommendations in large-scale internal technology collaboration scenarios. Specifically, the system employs frozen pre-trained semantic vectors to ensure recall-ranking consistency; users and results share a learnable embedding table to strengthen industry / professional tag alignment; and a large language model is introduced in the fine-ranking stage to generate personalized matching scores and natural language recommendation reasons (such as "Your organization has previously applied hydrogen compressor technology; this result is its supporting sealing system") based on unit profiles, historical behavior, and result details, thereby significantly improving recommendation credibility. The cold-start problem is effectively mitigated through semantic generalization and a high-value result pool mechanism, ensuring that new results receive reasonable exposure. The overall architecture is lightweight, efficient in coarse-ranking, and focused in fine-ranking, balancing effectiveness, performance, and maintainability, and can support large-scale internal collaboration needs. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating the overall steps of the method of the present invention; Figure 2 This is a flowchart of the offline preparation stage in the method of this embodiment of the invention; Figure 3 This is a flowchart of the steps in the online recommendation stage of the method in an embodiment of the present invention. Detailed Implementation

[0019] All features disclosed in all embodiments of this specification, or steps in all methods or processes implied in the disclosure, may be combined and / or extended or replaced in any way, except for mutually exclusive features and / or steps.

[0020] Given the current situation, the inventors of this invention, after creative reflection, believe that existing technology recommendation solutions mainly rely on two types of solutions: rule-based matching methods based on keywords or tags, and static comparison methods based on keywords or shallow semantics. Further analysis revealed that these solutions face the following significant limitations in large-scale organizational scenarios: (1) Rule-based matching methods based on keywords or tags. This type of method extracts keywords from the title of the results or matches them based on preset industry / professional tags. For example, it directly associates the user's "new energy" field with the result tag "lithium battery". Problems include: lack of semantic generalization ability, unable to identify results with similar technical substance but different expressions (such as "federated learning framework" and "privacy-preserving distributed training system"); lagging tag system, making it difficult for emerging cross-technology (such as "carbon footprint accounting tool") to be accurately covered by traditional classification; rigid matching, only supporting precise tag matching, unable to handle multi-domain integration scenarios (such as "intelligent manufacturing + green and low carbon" composite technology).

[0021] (2) Static comparison method based on keywords or shallow semantics. Some systems have introduced pre-trained models such as BERT and BGE to encode the title of the results into dense vectors and recall them through vector similarity. The problems are: lack of business controllability; pure semantic matching results are not interpretable and are difficult to meet the organization's governance requirements of "recommendations must be verifiable and auditable"; cold start units perform poorly; newly established R&D centers or units without historical cooperation records cannot generate effective demand representations due to lack of behavioral data; ignore organizational governance rules and cannot forcibly inject policy-oriented content such as "key technologies promoted by the group" or "achievements verified by sister units"; the profile noise is large. If the demand is generated by relying on business registration information (such as "technology development, consulting, and services"), a large amount of irrelevant noise will be introduced, which will seriously deviate from the actual technology direction.

[0022] Furthermore, some existing general-purpose recommendation systems suffer from incompatibility issues. For example, collaborative filtering models commonly used by e-commerce and content platforms heavily rely on explicit user behavior (clicks, purchases, ratings), while large organizations generally lack such behavioral logs. At the same time, their "black box" decision-making characteristics conflict with organizations' emphasis on transparency, accountability, and the ability for human intervention.

[0023] In summary, existing technologies are either too rigid (relying solely on label matching) or too black-box (pure semantic vectors), neither of which can meet the comprehensive requirements of large organizations for accuracy, interpretability, policy compliance, and internal collaboration.

[0024] To address the aforementioned technical issues, in a preferred embodiment, this invention provides an intelligent recommendation scheme for scientific and technological achievements within large organizations. This scheme does not rely on explicit user feedback (such as ratings or favorites), but is driven by real high-value technological activities and a high-value enterprise achievement database that occur within the organization. This includes auditable business behaviors such as signing technology contracts, obtaining achievement contact information, and downloading technical documents, supplemented by proprietary research achievements and business registration information as a cold start mechanism. Through a unified technology classification system and a semantic encoding model with frozen parameters, a dynamic representation of enterprise needs is constructed, achieving end-to-end interpretable, traceable, and highly accurate internal technology collaboration matching. In the coarse-ranking stage, a hybrid dual-tower coarse-ranking architecture using shared learnable embeddings and frozen semantic vectors is proposed, improving efficiency. In the fine-ranking stage, a large language model is introduced to generate personalized matching scores and natural language recommendation reasons based on unit profiles, historical behavior, and achievement details, significantly improving recommendation credibility. The overall process includes the following steps: offline enterprise profile and behavior chain construction, semantic / tag / high-value pool three-way hybrid recall, shared embedding hybrid dual-tower coarse-ranking, and fine-ranking and re-ranking based on a large language model.

[0025] Specifically, in the implementation of offline enterprise profiling and behavior chain construction, this invention proposes a method for representing dynamic enterprise needs based on high-value agent behavior chains. In internal organizational collaboration scenarios lacking explicit user preference feedback (such as likes, favorites, ratings), interactive behaviors with clear business intent and associated actual costs (signing contracts, downloading files, obtaining contact information) are defined as high-value agent behaviors. Based on structured historical behavior chains, behaviors related to the titles of scientific and technological achievements are extracted. A semantic vector is generated through a parameter-frozen semantic encoder, and an exponential time decay mechanism based on days is introduced to fuse and generate an enterprise demand vector reflecting the enterprise's recent real technological interests. For enterprises with no behavior records, a fallback vector is generated by sequentially reverting to their own scientific and technological achievements and the cleaned business scope text, thus ensuring full coverage. In a more specific implementation, such as... Figure 1 , Figure 2 and Figure 3 As shown, the process includes the following: S1, Construct a two-level classification system and label the results. First, the organization's science and technology management department constructs a two-tiered classification system, specifically including industry and profession. The industry category covers the organization's main business areas, comprising approximately 10-20 categories (such as "high-end equipment," "biomedicine," and "clean energy"). The profession category further subdivides technical directions, comprising approximately 50-100 categories (such as "industrial robots," "mRNA delivery systems," and "solid-state hydrogen storage materials"). Each profession uniquely belongs to one industry.

[0026] Secondly, all internal scientific and technological achievements are structurally tagged. A large model prompts the engineering or fine-tuning classifier to input important information such as the achievement title, abstract, and technical description for tagging. The output is a standardized string "industry;profession", and the corresponding industry tag (industry_id) and profession tag (profession_id) are stored.

[0027] S2, offline calculation of enterprise profile and enterprise demand vector (daily T+1, where T+1 means that the latest enterprise profile and demand vector will be calculated on the next day (T+1) based on the complete data of the previous day (T day). For each sub-unit within the organization (including enterprises, laboratories, business units, etc., hereinafter referred to as "enterprises", each with a unique sub-unit ID), an offline task is executed once daily to construct a structured historical behavior chain at the enterprise level, and an enterprise demand vector is generated based on this behavior chain. This behavioral chain also serves as contextual input for the fine-grained ranking stage of the Large Language Model (LLM), achieving data reuse and semantic consistency. Specifically, it includes the following sub-processes: (1) Generate industry profiles and professional profiles To construct a multi-dimensional technical capability profile of the target organization, the system extracts its technical capability tag set and determines the industry tag list and professional tag list according to the following priority order: If the unit has historical records of technical activities, then iterate through all the scientific and technological achievements associated with the activities, extract the industry classification identifier (referred to as iid) and professional and technical field identifier (referred to as pid) corresponding to each achievement, and after deduplication, form an industry tag list and a professional tag list respectively. If the unit has no historical record of technical activities, select no more than 10 recent achievements from its own scientific and technological achievement database as the first completion unit, extract the iid and pid of these achievements, and form a corresponding list after deduplication. If neither of the above two types of data exists, then based on the business scope registered in the company's business registration information, basic IID and PID tags are automatically generated using preset keyword mapping rules. Finally, two structured data entries are saved as the company's industry profile and professional profile: The list of industry tags is denoted as "IndIDs" = [iid1, iid2, ..., iid...]. m ]; The professional tag list is denoted as "ProfIDs" = [pid1, pid2, ..., pid2]. n ].

[0028] Where m is a positive integer representing the total number of industry sectors associated with the unit, i.e., the number of duplicate elements in the list IndIDs; n is a positive integer representing the total number of professional and technical directions covered by the unit, i.e., the number of duplicate elements in the list ProfIDs. The specific values ​​of parameters m and n are dynamically determined by the data source identified by the above strategy.

[0029] (2) Constructing the enterprise's historical behavior chain Each behavior record in the enterprise historical behavior chain is stored in a unified structured format, using the enterprise's unique identifier (enterprise_id) as the aggregation key. It includes the following fields: The type field indicates the behavior type and is an enumeration value, including "contract signing," "contact information acquisition," or "download"; the item_id field is the unique identifier of the associated technical achievement or document; the title field records the title of the achievement or document, such as "Intelligent Tower Crane Anti-Collision Algorithm Description.pdf"; and the datetime field records the specific timestamp of the behavior.

[0030] The behavior chain is persistently stored in a NoSQL database as a JSON array, with its storage key following a unified naming convention: behavior_chain:{enterprise_id}. For example, the behavior chain for enterprise ID ENT_102456 corresponds to the key behavior_chain:ENT_102456. During the online recommendation phase, the system can directly read this behavior chain based on the input enterprise_id to construct an enterprise demand vector, which is then provided as contextual information to the Large Language Model (LLM) to support matching score scoring and interpretable reasoning in the fine-grained ranking phase.

[0031] (3) Generating enterprise demand vectors based on multi-source signals

[0032] The enterprise demand vector is a 768-dimensional dense vector (encoded using the BAAI / bge-large-zh-v1.5 model) used to represent the semantic interest of enterprises in technological achievements at the current stage. This vector is not dependent on explicit user input but is automatically generated based on historical enterprise technology activity data and updated once daily in offline tasks. The system determines the data sources according to the following priority order.

[0033] ①Based on the enterprise's historical behavior chain If the enterprise has historical behavioral chains, then the enterprise's demand vector is generated by acquiring data from the enterprise's last 10 behavioral chains, specifically including: Step 1: Semantic Encoding. Let the first step be... Item of behavior record ( ,in The corresponding technical achievement title is , here The title of the achievement (e.g., "Intelligent Support System for Ultra-Deep Foundation Pit V2.0") uses a pre-trained Chinese semantic encoding model (such as BAAI / bge-large-zh-v1.5) to... Encode the vector to output a dense vector of fixed dimensions. ,Right now: ; in: Indicates the first The semantic vector of the technological achievement associated with each behavior is a 768-dimensional real column vector; Indicates the first The title of the relevant technological achievement; BGE represents a bidirectional encoder model based on the Transformer architecture. Its structure consists of a multi-layer self-attention mechanism and a feedforward network stacked together. In this scheme, the encoding model has fixed parameters and does not participate in any fine-tuning or training, ensuring semantic space consistency. Let represent a 768-dimensional Euclidean space, which is a set of 768-dimensional vectors whose components are all real numbers.

[0034] Step 2: Calculate the time decay weight. From the enterprise's historical behavior chain... Extract the occurrence date from the timestamp field (datetime) of each behavior record. Let the current system processing time be... Then the time interval for ; To reflect the objective law that a company's interest in historical behavior naturally diminishes over time, a time decay weight is introduced. Its definition is: ; in: Indicates the first The relative contribution of each behavior to the representation of firm demand at the current moment; when (i.e., the behavior has just occurred) This indicates that the action has the greatest impact; Increase Monotonically decreasing reflects the business common sense that "recent actions are more important"; The decay coefficient is determined by the half-life. (That is, the time required for the intensity of interest to decay to half of its initial value, in days) Determine: ; This design ensures that: when hour, That is, the first The contribution of each behavior decays to 50% of its initial value, which is consistent with the intuitive semantics of "half-life".

[0035] Step 3: Generate the final enterprise demand vector by weighted averaging. The average of all action vectors is normalized by time decay weights: .

[0036] ②Based on proprietary scientific and technological achievements If the enterprise has no activity record, but has registered achievements in the technology achievement database as the primary completion unit, then select the ≤10 most recent self-owned achievements, and generate a database using their titles following the same process as described above (including time decay, based on the achievement registration date). .

[0037] ③Based on the scope of business operations If a company has neither a behavioral chain nor its own achievements, the business scope field (e.g., "architectural engineering design; technology development, consulting, and services; promotion of new energy technologies") is extracted from its business registration information. Text cleaning is then performed, removing general legal statements (e.g., "Projects requiring approval by law can only be operated after approval by relevant departments"), filtering non-technical expressions (e.g., "sales," "agency," "investment"), and retaining technology-related keyword segments (e.g., "intelligent construction," "hydrogen energy storage and transportation," "BIM platform development"). The cleaned text is then concatenated into a single string and input into a semantic encoder to generate a company demand vector, without time decay (due to the lack of a time dimension).

[0038] After completing the organizational profiling, the embodiment of this invention specifically designs a three-way recall fusion mechanism of semantic generalization, tag matching, and strategic fallback. It proposes a hybrid recall scheme of semantics / tags / high-value pool, with the system executing the three-way recall strategy in parallel to form a complementary and synergistic candidate generation system, specifically including: Dense semantic recall (generalized discovery path): Using enterprise demand vectors as queries, FAISS approximate nearest neighbor search is performed in the full database of scientific and technological achievement title / abstract vectors to capture technological opportunities with similar semantics but not necessarily overlapping labels, supporting cross-professional and cross-domain innovation associations. Structured tag recall (precise matching path): Based on the industry tag list and professional tag list in the user profile, the results belonging to the same tag are matched through the inverted index to ensure that the recommendation results strictly fall within the user's technical capability boundary and meet the realistic constraints of engineering implementation; High-value achievement pool recall (strategic safety net): Dynamically maintain a high-quality achievement set defined by business rules, including technologies that are key technologies promoted by the group, achievements that have won awards at the provincial or ministerial level or above, achievements newly registered in the past 30 days, and highly reusable achievements that have been adopted by multiple units. This ensures that strategic intentions and emerging technologies are not buried due to data sparsity, provides a definite exposure channel for new achievements and key technologies, and breaks through the cold start and long tail dilemma.

[0039] The three-way recall ratio is dynamically adjusted according to the scenario (typical configuration: 60% semantic, 35% tag, and 5% high-value pool) to ensure a balance between breadth, accuracy, and governance. In a more specific implementation, the following sub-processes are executed: Upon receiving a recommendation request for a specific sub-unit (such as an enterprise, laboratory, or business unit), the system retrieves its offline pre-computed profile data from the feature storage, specifically including: Enterprise demand vector (Used for semantic recall); The list of industry tags is denoted as "IndIDs" = [iid1, iid2, ..., iid...]. m ]; The professional tag list is denoted as "ProfIDs" = [pid1, pid2, ..., pid2]. n ].

[0040] Based on the above input, design and implement the three-way recall strategy in Table 1 below: Table 1

[0041] In the specific implementation of the shared embedding hybrid dual-tower coarse-ranking scheme, this embodiment of the invention proposes a hybrid dual-tower coarse-ranking architecture that uses shared learnable embeddings and frozen semantic vectors. The user tower and the outcome tower share the same set of industry / professional learnable embedding tables, while the enterprise demand vector and outcome title vector use frozen BGE semantic vectors. The user profile only stores a list of IDs (industry, profession, region, organization type), and generates a structured representation online by looking up the table and averaging the results. This achieves strict alignment between users and outcomes in the label space, avoiding representational bias. Freezing the semantic vectors ensures consistency with the FAISS recall space, solving the problem of "recalling but not ranking," specifically including the following sub-processes: To achieve accurate ranking while ensuring high throughput performance, this invention employs a hybrid dual-tower neural network architecture in this stage to uniformly score and re-rank no more than 1000 candidate scientific and technological achievements output by the recall module. The core design idea is to use high-dimensional semantic information as a fixed anchor point to maintain recall-ranking consistency, model structured organizational attributes through end-to-end learnable embeddings, and ensure strict alignment between the user-side and achievement-side representations in the classification label space. Specifically, this includes: (1) Feature input system User-side features are derived from daily offline enterprise profile caches, including: The industry tag list (industry_id list) reflects the technology fields in which the company has long been involved; A list of professional tags (profession_id list) describes their core technical expertise; Region (region_id) represents the geographical region to which the enterprise belongs; Organization type (org_type_id) identifies its functional attributes (such as engineering bureau, research institute, design institute, etc.); Enterprise demand vector ( (This is generated from the company's recent technical needs text using a semantic encoder, and is used to capture dynamic interests.)

[0042] The features of the achievements are derived from the scientific and technological achievements metadata database, including: The title semantic vector (title_emb) is generated from the title of the work using the same semantic encoder; Industry tag (industry_id) and professional tag (profession_id) describe the technical category to which the achievement belongs; The business rule feature set includes binary or numerical signals such as whether it belongs to the group's key promotion technologies, whether it is a newly registered achievement in the past 90 days, whether it has been signed by multiple units in the past 180 days, and whether it has won provincial or ministerial-level science and technology awards. All semantic vectors generated by BGE keep their parameters frozen during training and inference and do not participate in gradient updates.

[0043] (2) Shared learnable embedding mechanism The system maintains a unified set of learnable embedding tables, covering four categories of identifiers: industry, profession, region, and organization type. Among them, the industry embedding table and the profession embedding table are shared by the user tower and the result tower, ensuring that when a user profile contains a certain industry label and the candidate result happens to belong to that industry, the two representations in the vector space are completely consistent, thereby maximizing the effectiveness of the structured matching signal.

[0044] The region embedding table and organization type embedding table are only used in the user tower to model the impact of geographical collaboration convenience and organizational function differences on technology adoption preferences. Each embedding dimension is reasonably set according to the category cardinality, such as 32 dimensions for industry, 64 dimensions for profession, 16 dimensions for region, and 8 dimensions for organization type.

[0045] (3) User representation construction During online inference, the user tower first dynamically constructs a structured representation based on the cached list of identifiers. For the industry tag list and the professional tag list, the corresponding shared embedding table is queried, and a simple arithmetic average is performed to obtain the industry profile vector and the professional profile vector; if the list is empty, it is filled with a zero vector; for the region identifier and the organization type identifier, their respective embedding table is queried directly to obtain the corresponding dense vector; the above four structured vectors are concatenated with the frozen enterprise demand vector to form the final comprehensive user representation. This process does not require pre-storing the embedding results, relies only on lightweight in-memory table lookup and averaging operations, has extremely low computational overhead, and supports embedding table version iteration without regenerating the profile cache.

[0046] (4) Construction of results representation For each candidate outcome, the outcome platform queries the shared embedding table using its industry and professional tags to obtain the corresponding industry and professional vectors. These vectors are then concatenated with the frozen title semantic vectors and business rule features to form a comprehensive representation of the outcome. The outcome itself is not bound to any region or organization type to maintain the universality of the technological assets and their ability to be promoted across regions.

[0047] (5) Integrating scoring and training strategies The user's comprehensive representation and the outcome's comprehensive representation are concatenated and input into a lightweight multilayer perceptron (MLP) for nonlinear interaction modeling. The final output is a matching score between 0 and 1, used to rank candidate outcomes. Model training employs a multi-level positive sample strategy to fully mine genuine intention signals during internal technological collaboration. Specifically, positive samples include the following three types of behavior, assigned different confidence weights according to business depth: High confidence positive sample: The company signs a formal technology contract with the technology provider, indicating that the technology assessment has been completed and the implementation phase has begun; Zhongzhixin's positive example: Enterprises actively obtain the contact information of the results contact person (such as by clicking the "Contact Information" button), reflecting a clear willingness to connect; Low-confidence positive samples: Enterprises downloading relevant documents (such as technical specifications, demonstration videos, and patent documents) represent initial interest exploration.

[0048] During training, the aforementioned behaviors can be uniformly modeled as positive samples, and their differences can be reflected through a weighted loss function or a hierarchical sampling strategy: for example, the weight of a contract sample is set to 1.0, obtaining contact information is set to 0.6, and file download is set to 0.3. If a pairwise ranking loss (such as BPR) is used, all positive samples are considered valid interactions, and negative samples are sampled from results that have never exhibited any of the above behaviors, prioritizing "hard negative examples" that appear in the recall path but have not been interacted with. All semantic vectors (including enterprise demand vectors and result title vectors) have their parameters frozen during training, and only the industry, profession, region, organization type embedding tables and MLP network parameters participate in gradient updates. This design ensures that the semantic space in the coarse-ranking stage is strictly aligned with the FAISS vector library in the recall stage, effectively avoiding inconsistencies between recall and ranking caused by representation drift.

[0049] This design achieves triple collaboration. Semantic and structural collaboration: frozen semantics ensures generalization discovery capabilities, while learnable embeddings enhance business-specific accuracy. User and outcome collaboration: shared embedding tables eliminate representational bias and improve tag matching reliability. Organization and technology collaboration: regional and organizational type features allow recommendation results to adapt to the implementation conditions and functional needs of different units. The overall architecture features efficient storage, fast inference, and stable training, thereby significantly improving the conversion intention rate and user satisfaction.

[0050] In the specific implementation of fine ranking and re-ranking based on a large language model, this invention proposes a large model fine ranking and personalized reason output scheme based on behavioral chains. For the coarse-ranked candidate set, the LLM input unit profile, historical behavioral chains (contracts / downloads / contact records), and achievement details are called, and the matching score and natural language recommendation reason (e.g., "Your organization has used X, and this achievement is a matching Y") are jointly output. High-value agency behaviors (non-click behaviors) of enterprises are transformed into interpretable context, generating causal and personalized reasons, significantly improving user trust and conversion intention, avoiding black-box recommendations, and meeting organizational requirements for algorithm transparency. The enterprise-level behavioral chain is combined with LLM reasoning to achieve highly reliable recommendations in B2B scenarios. Specifically, this includes the following sub-processes: To ensure both recommendation accuracy and high interpretability, this stage constructs large language model prompts based on enterprise profiles, natural language behavioral chains, and structured metadata of candidate results. LLM then performs semantic matching scoring and generates interpretable reasons for the coarsely ranked Top 50 results, ultimately outputting a highly credible Top 10 recommendation, specifically including: (1) Matching scoring and ranking based on large language model and generation of recommendation reasons For each candidate result The system constructs a structured context, calls large language models (such as Qwen, ChatGLM, etc.) deployed privately by the organization, and performs deep semantic understanding and personalized reasoning.

[0051] The input context includes: Company details: Company name, industry profile, professional profile, region, organization type (e.g., engineering bureau, research institute); Historical Behavior Chain: Retrieves the complete behavior chain of the enterprise from KV storage (interaction records such as technology contracts signed within the past 180 days, obtaining contact information, downloading technical documents, etc.), and arranges them in reverse chronological order; Candidate achievement details: Structured metadata including title, abstract, industry / specialty, whether it is a key technology for promotion, whether it has won awards, whether it has been applied in practice, and the organization providing the achievement. For example: [Company Details] Company Name: xxxxxx Co., Ltd. Industry: Housing construction, infrastructure, intelligent construction Specializations: Super high-rise construction, deep foundation pit engineering, BIM technology application Region: xx Province, xx City Organization type: Engineering Bureau [Historical Behavior] Your organization I downloaded the "Intelligent Tower Crane Anti-Collision Real-time Early Warning System V3.0" on January 10, 2025. I contacted the organization that produced the "Intelligent Monitoring of Underground Space" project on December 5, 2024. On November 18, 2024, the "Intelligent Support System V2.0 for Ultra-Deep Foundation Pit" contract was signed. [Candidate Results] Title: Intelligent Behavior Recognition Platform for High-Altitude Workers Abstract: Based on multimodal visual perception and rule engine, this system can identify high-risk behaviors such as not wearing a safety belt, illegal climbing, and fatigued operation in real time, and supports linkage alarm with IoT devices on construction sites.

[0052] Industry: Intelligent Construction, Safe Production Majors: Computer Vision, Engineering Safety Management Is it a key technology to be promoted? Yes (One of the Group's Top Ten Key Promotion Achievements in 2024) Award received: Yes (First Prize for Innovative Application of Intelligent Construction, Ministry of Housing and Urban-Rural Development, 2024) Is it being implemented in practice? Yes Results provided by: xxxxx Research Institute [Task Instructions] Based on the above information, please determine the degree to which this achievement matches the company's technological needs, and: (1) Output a matching score of 0–1; (2) Use a sentence in natural Chinese to explain the reason for the recommendation, highlighting the technological synergy or business value.

[0053] The large model outputs a normalized matching score. This reflects the overall suitability and a natural language recommendation reason, for example: "Your organization signed the 'Intelligent Support System V2.0 for Ultra-Deep Foundation Pit' on October 15, 2024, and downloaded the 'Intelligent Tower Crane Anti-Collision Algorithm Description' on June 20, 2024. This achievement, the 'Intelligent Recognition Platform for High-Altitude Worker Behavior,' is highly aligned with your current focus on intelligent construction safety monitoring." This mechanism not only improves the quality of sorting, but also enhances user trust through explainability, significantly increasing the willingness to engage with the platform in the future.

[0054] (2) Integrating LLM and coarse-ranking score reordering To synergistically leverage the semantic understanding capabilities of the large language model and the structured discrimination advantages of the coarse-ranking model, a weighted linear fusion strategy is adopted to calculate the final score: ; in: The matching score output by the large language model; The raw scores output by the hybrid dual-tower coarse-ranking model have embedded structured signals such as industry matching, organization type, regional attributes, key promotion indicators, and transaction activity. Configurable fusion weights are used to adjust the relative contributions of LLM and coarse ranking. Typical initial values ​​are set to 0.7–0.9, reflecting an emphasis on the depth understanding capabilities of LLM.

[0055] The final output includes the reordered results and personalized recommendation reasons generated by LLM.

[0056] It should be noted that, within the scope of protection defined in the claims of this invention, the following embodiments can be combined and / or extended or replaced in any logical manner from the above specific embodiments, such as the disclosed technical principles, disclosed technical features or implicitly disclosed technical features.

[0057] Example 1 A technology recommendation method based on business behavior chains includes: Profiling the organization and building internal business behavior chains; After receiving a recommendation request, read the profile data, perform multi-path hybrid recall, dynamically adjust the recall ratio of each path according to the scenario, and merge the recall results of multiple paths into a recall candidate set; The recall candidate set is scored and reordered using a shared embedded hybrid dual-tower coarse ranking method to form a coarse-ranked candidate set. For the coarsely ranked candidate set, based on the input profile, business behavior chain, and result details of the large language model, the large language model outputs a matching score and a natural language recommendation reason.

[0058] Example 2 Based on Example 1, the process of profiling the organization and constructing the internal business behavior chain specifically includes the following sub-steps: High-value agency behavior is defined by extracting high-value agency behaviors related to the titles of scientific and technological achievements from organizations with historical behavior records. Then, a semantic encoder is used to generate semantic vectors, and an exponential time decay mechanism is introduced to fuse and generate a first demand vector of enterprises that reflects the organization's true technological interests. For organizations without historical behavior records, a second demand vector of enterprises is generated using the organization's own scientific and technological achievements or business scope text. The first demand vector and the second demand vector of enterprises are used as profile data storage. Specifically, the definition of high-value agency behavior refers to interactive behaviors in historical behavior records that have business intent and are accompanied by actual costs.

[0059] Example 3 Based on Example 2, the multi-path hybrid recall is performed, with the recall ratio of each path dynamically adjusted according to the scenario, and the results of the multi-path recall are merged into a recall candidate set. This specifically includes the following sub-steps: Dense semantic recall: Using the enterprise's first demand vector or the enterprise's second demand vector as the query, a FAISS approximate nearest neighbor search is performed in the full database of scientific and technological achievement titles and abstracts to capture technological opportunities that are semantically similar but whose labels may not overlap, and to obtain the candidate set R1. Structured tag recall: Based on the industry tag list and professional tag list in the user profile, the results belonging to the same tag are matched by the inverted index to obtain the candidate set R2; High-value outcome pool recall: Dynamically maintain the set of high-quality outcomes defined by business rules to obtain the candidate set R3; The recall ratios of dense semantic recall, structured tag recall, and high-value result pool recall are dynamically adjusted according to the scenario, and finally merged into a recall candidate set. .

[0060] Example 4 Based on Example 2, in the shared embedded hybrid dual-tower coarse ranking, the user tower and the result tower share the same set of learnable embedding tables; The learnable embedding table covers four categories of identifiers: industry, profession, region, and organization type. Specifically, it includes industry embedding table, profession embedding table, region embedding table, and organization type embedding table. The industry embedding table and the professional embedding table are shared by the user tower and the outcome tower to ensure that when a profile contains an industry label and the candidate outcome happens to belong to that industry, the two are represented in the vector space in a consistent manner, maximizing the effectiveness of the structured matching signal. The region embedding table and the organization type embedding table are used only by the user tower to model the impact of geographical collaboration convenience and organizational function differences on technology adoption preferences.

[0061] Example 5 Based on Example 4, the step of scoring and re-ranking the recall candidate set using shared embedding hybrid dual-tower coarse ranking further includes the following sub-steps: Both the enterprise's primary demand vector and secondary demand vector, along with the deliverable title vector, adopt frozen semantic vectors. The user profile only stores an ID list, and an online table lookup is used to generate a structured representation on average. When constructing the comprehensive user representation, User Tower first dynamically constructs a structured representation based on the cached identifier list. For the industry tag list and the professional tag list, it queries the corresponding shared embedding table and performs an arithmetic average to obtain the industry profile vector and the professional profile vector. If the list is empty, it is filled with a zero vector. For the region identifier and the organization type identifier, it directly queries their exclusive embedding table to obtain the corresponding dense vector. The four structured vectors are concatenated with the frozen enterprise demand vector to form the final comprehensive user representation. When constructing a comprehensive representation of results, the results tower queries the shared embedded table for each candidate result with its industry and professional tags to obtain the corresponding industry and professional vectors, and then concatenates them with the frozen title semantic vectors and business rule features to form a comprehensive representation of results. The results themselves are not bound to the region or organization type. The user's comprehensive representation and the result's comprehensive representation are then concatenated and input into a multilayer perceptron (MLP) for nonlinear interaction modeling. The final output is a matching score between 0 and 1, which is used to rank the candidate results.

[0062] Example 6 Based on Example 2, the construction of the business behavior chain specifically includes the following sub-steps: using the ID of the organizational sub-unit as the aggregation key, aggregating multiple types of high-value agency behaviors from the internal business system or results platform, sorting them in chronological order to form a behavior sequence list, and storing each element in the sequence list in the form of structured data.

[0063] Example 7 Based on Example 2, the step of inputting prompt words into a large language model to jointly generate matching scores and natural language recommendation reasons specifically includes the following sub-steps: The constructed prompt word template includes enterprise profile, business behavior chain and result details. For each candidate result, a structured context is constructed based on the prompt word template, and deep semantic understanding and personalized reasoning are performed by calling a large language model. The large model outputs a normalized matching score, reflecting the overall fit and the natural language recommendation rationale.

[0064] Example 8 A technology achievement recommendation device based on a business behavior chain includes a processor and a memory. The memory stores a computer program, which, when loaded by the processor, executes the method described in any one of Embodiments 1 to 7.

[0065] Example 9 A computer-readable storage medium storing a computer program, which, when loaded by a processor, executes the method described in any one of Embodiments 1 to 7.

[0066] Example 10 A technology achievement recommendation system based on business behavior chain includes the technology achievement recommendation device based on business behavior chain described in Example 8.

[0067] The units described in the embodiments of the present invention can be implemented in software or hardware, and the described units can also be located in a processor. The names of these units do not necessarily limit the specific unit itself.

[0068] According to one aspect of the present invention, a computer program product or computer program is provided, the computer program product or computer program including computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and executes the computer instructions, causing the computer device to perform the methods provided in the various optional implementations described above.

[0069] In another aspect, embodiments of the present invention also provide a computer-readable medium, which may be included in the electronic device described in the above embodiments; or it may exist independently and not assembled into the electronic device. The computer-readable medium carries one or more programs, which, when executed by the electronic device, cause the electronic device to perform the methods described in the above embodiments.

Claims

1. A method for recommending scientific and technological achievements based on business behavior chains, characterized in that, include: Profiling the organization and building internal business behavior chains; After receiving a recommendation request, read the profile data, perform multi-path hybrid recall, dynamically adjust the recall ratio of each path according to the scenario, and merge the recall results of multiple paths into a recall candidate set; The recall candidate set is scored and reordered using a shared embedded hybrid dual-tower coarse ranking method to form a coarse-ranked candidate set. For the coarsely ranked candidate set, a prompt word template is constructed, and the prompt words are input into a large language model to jointly generate a matching score and a natural language recommendation reason.

2. The technology achievement recommendation method based on business behavior chain according to claim 1, characterized in that, The process of profiling the organization and constructing the internal business behavior chain specifically includes the following sub-steps: High-value agency behaviors are defined, and high-value agency behaviors with related scientific and technological achievement titles are extracted from organizations with historical behavior records. Then, a semantic encoder is used to generate semantic vectors, and an exponential time decay mechanism is introduced to fuse and generate a first demand vector of enterprises that reflects the organization's true technological interests. For organizations without historical behavior records, a second demand vector of enterprises is generated using the organization's own scientific and technological achievements or business scope text. The first demand vector of enterprises and the second demand vector of enterprises are used as profile data storage. The definition of high-value agency behavior specifically refers to interactive behaviors in historical behavior records that have business intent and are accompanied by actual costs.

3. The method for recommending scientific and technological achievements based on business behavior chains according to claim 2, characterized in that, The process of performing multi-path hybrid recall, where the recall ratio of each path is dynamically adjusted according to the scenario, and merging the recall results from multiple paths into a recall candidate set, specifically includes the following sub-steps: Dense semantic recall: Using the enterprise's first demand vector or the enterprise's second demand vector as the query, a FAISS approximate nearest neighbor search is performed in the full database of scientific and technological achievement titles and abstracts to capture technological opportunities that are semantically similar but whose labels may not overlap, and to obtain the candidate set R1. Structured tag recall: Based on the industry tag list and professional tag list in the user profile, the results belonging to the same tag are matched by the inverted index to obtain the candidate set R2; High-value outcome pool recall: Dynamically maintain the set of high-quality outcomes defined by business rules to obtain the candidate set R3; The recall ratios of dense semantic recall, structured tag recall, and high-value result pool recall are dynamically adjusted according to the scenario, and finally merged into a recall candidate set. .

4. The technology achievement recommendation method based on business behavior chain according to claim 2, characterized in that, In the shared embedded hybrid dual-tower coarse-ranking, the user tower and the result tower share the same set of learnable embedding tables. The learnable embedding table covers four categories of identifiers: industry, profession, region, and organization type. Specifically, it includes industry embedding table, profession embedding table, region embedding table, and organization type embedding table. The industry embedding table and the professional embedding table are shared by the user tower and the outcome tower to ensure that when a profile contains an industry label and the candidate outcome happens to belong to that industry, the two are represented in the vector space in a consistent manner, maximizing the effectiveness of the structured matching signal. The region embedding table and the organization type embedding table are used only by the user tower to model the impact of geographical collaboration convenience and organizational function differences on technology adoption preferences.

5. The technology achievement recommendation method based on business behavior chain according to claim 4, characterized in that, The step of scoring and re-ranking the recalled candidate set using shared embedding hybrid dual-tower coarse ranking further includes the following sub-steps: Both the enterprise's primary demand vector and secondary demand vector, along with the deliverable title vector, adopt frozen semantic vectors. The user profile only stores an ID list, and an online table lookup is used to generate a structured representation on average. When constructing a comprehensive user representation, the user tower first dynamically constructs a structured representation based on the list of identifiers in the cache. For the industry tag list and the professional tag list, it queries the corresponding shared embedding table respectively and performs an arithmetic average to obtain the industry profile vector and the professional profile vector. If the list is empty, it is filled with a zero vector; For region identifiers and organization type identifiers, their dedicated embedding tables are directly queried to obtain the corresponding dense vectors; the four structured vectors are concatenated with the frozen enterprise requirement vectors to form the final comprehensive user representation; When constructing a comprehensive representation of results, the results tower queries the shared embedded table for each candidate result with its industry and professional tags to obtain the corresponding industry and professional vectors, and then concatenates them with the frozen title semantic vectors and business rule features to form a comprehensive representation of results. The results themselves are not bound to the region or organization type. The user's comprehensive representation and the result's comprehensive representation are then concatenated and input into a multilayer perceptron (MLP) for nonlinear interaction modeling. The final output is a matching score between 0 and 1, which is used to rank the candidate results.

6. The method for recommending scientific and technological achievements based on business behavior chains according to claim 2, characterized in that, The construction of the internal business behavior chain specifically includes the following sub-steps: using the ID of the organizational sub-unit as the aggregation key, aggregating multiple types of high-value agency behaviors from the internal business system or results platform, sorting them in chronological order to form a behavior sequence list, and storing each element in the sequence list in the form of structured data.

7. The technology achievement recommendation method based on business behavior chain according to claim 2, characterized in that, The step of inputting prompt words into a large language model to jointly generate matching scores and natural language recommendation reasons specifically includes the following sub-steps: The constructed prompt word template includes enterprise profile, business behavior chain and result details. For each candidate result, a structured context is constructed based on the prompt word template, and deep semantic understanding and personalized reasoning are performed by calling a large language model. The large model outputs a normalized matching score, reflecting the overall fit and the natural language recommendation rationale.

8. A technology achievement recommendation device based on a business behavior chain, characterized in that, It includes a processor and a memory, wherein the memory stores a computer program that, when loaded by the processor, executes the method as described in any one of claims 1 to 7.

9. A computer-readable storage medium, characterized in that, A computer program is stored in a readable storage medium, which, when loaded by a processor, executes the method as described in any one of claims 1 to 7.

10. A technology achievement recommendation system based on business behavior chains, characterized in that, This includes the technology achievement recommendation device based on business behavior chain as described in claim 8.

Images