Tender information recommendation method and system based on responsibility chain matching

Abstract

The application discloses a tender information recommendation method and system based on a responsibility chain matching, relates to the technical field of intelligent recommendation and big data processing, and comprises the following steps: when a subscription rule is stored in a database, a subscription label system is constructed to configure each matching processor in a flow matching engine, flow matching analysis based on the matching processor is performed on tender information, a tender candidate set is obtained, three-dimensional vector conversion and multi-objective interaction-based non-dominated solution set searching are performed, a recommendation sequence is generated, and a distributed scheduling framework is deployed to perform timing and real-time push management, so that the technical problem that the matching efficiency is low, multi-objective recommendation is difficult to consider multiple parties and lacks effective semantic generalization and rule relaxation mechanism, cannot adapt to changing subscription requirements, and is difficult to meet the capture rate and timeliness requirements in different business scenarios in the prior art is solved, and flexible and accurate push management in different business scenarios is realized.

Description

Technical Field

[0001] This invention relates to the field of intelligent recommendation and big data processing technology, specifically to a method and system for recommending bidding information based on chain of responsibility matching. Background Technology

[0002] With the accelerated digital transformation of the bidding and procurement market, various bidding announcements and procurement information are growing exponentially. How to accurately and efficiently recommend business opportunities that match the business capabilities of enterprise users from massive, high-dimensional, and dynamically updated bidding data has become the core challenge of current bidding information service platforms.

[0003] Currently, most mainstream bidding information recommendation systems employ rule-based filtering and matching methods. These systems typically use fixed, static filtering rules and pre-defined linear matching processes to screen bidding information item by item. This approach cannot dynamically adjust based on the complexity of the subscription rules or historical matching efficiency, leading to wasted system computing resources, delayed matching responses, and a lack of diversity in evaluation, resulting in highly homogenized recommendation sets and limited usability for users. Furthermore, it cannot adaptively adjust based on the stringency of subscription conditions, thus exhibiting certain limitations.

[0004] In existing technologies, traditional bidding information has low matching efficiency due to its fixed matching order, and multi-objective recommendation is difficult to balance multiple factors and lacks effective semantic generalization and rule relaxation mechanisms. It cannot adapt to changing subscription needs and cannot meet the capture rate and timeliness requirements of different business scenarios. Summary of the Invention

[0005] This application provides a method and system for recommending bidding information based on chain of responsibility matching. It addresses the technical problems in existing technologies, such as the low efficiency of fixed matching order in traditional bidding information, the difficulty in considering multiple objectives in multi-objective recommendation, the lack of effective semantic generalization and rule relaxation mechanisms, the inability to adapt to changing subscription needs, and the difficulty in meeting the capture rate and timeliness requirements in different business scenarios.

[0006] In view of the above problems, this application provides a method and system for recommending bidding information based on chain of responsibility matching.

[0007] This application provides a method for recommending bidding information based on chain of responsibility matching. The method includes: when subscription rules are entered into the database, multi-dimensional feature extraction is performed to construct a subscription tag system; chain sorting and rule initialization are performed on each matching processor in the streaming matching engine to obtain a streaming matching engine with a chain of responsibility sequence configuration; wherein, if the subscription rule is modified, causing a change in condition density, a chain of responsibility reordering is triggered; the condition density is positively correlated with the number of keywords, the number of geographical restrictions, the number of type restrictions, and the number of excluded words in the subscription rule; by passing bidding information through the streaming matching engine, streaming matching analysis based on the matching processor is performed to obtain a bidding candidate set and perform three-dimensional vector transformation and non-dominated solution set search based on multi-objective interaction to generate a recommendation sequence; for the recommendation sequence, a distributed scheduling framework is deployed to perform timed and real-time push management; wherein, by judging the empty set of the recommendation sequence, if it is empty, the semantic extension matching processor is activated, and the rules are relaxed in the word vector space of the bidding domain based on counterfactual inference, and relaxed round by round until a non-empty recommendation sequence is generated, which replaces the recommendation sequence.

[0008] This application provides a bidding information recommendation system based on chain of responsibility matching. The system includes: an engine configuration unit: when subscription rules are entered into the database, multi-dimensional feature extraction is performed to construct a subscription tag system, and chain sorting and rule initialization are performed on each matching processor in the streaming matching engine to obtain a streaming matching engine with a chain of responsibility sequence configuration. If the subscription rule is modified, causing a change in condition density, a chain of responsibility reordering is triggered. The condition density is positively correlated with the number of keywords, the number of geographical restrictions, the number of type restrictions, and the number of excluded words in the subscription rule; a recommendation analysis unit: by passing bidding information through the streaming matching engine, streaming matching analysis based on the matching processor is performed to obtain a bidding candidate set and perform three-dimensional vector transformation and non-dominated solution set search based on multi-objective interaction to generate a recommendation sequence; and a push management unit: for the recommendation sequence, timed and real-time push management is performed by deploying a distributed scheduling framework. The system includes: by judging the empty set of the recommendation sequence, if it is empty, the semantic extension matching processor is activated, and the rules are relaxed in the word vector space of the bidding domain based on counterfactual inference. The rules are relaxed round by round until a non-empty recommendation sequence is generated, which replaces the original recommendation sequence.

[0009] One or more technical solutions provided in this application have at least the following technical effects or advantages: The bidding information recommendation method based on responsibility chain matching provided in this application embodiment performs multi-dimensional feature extraction to construct a subscription tag system when subscription rules are entered into the database. It then performs chain-like sorting and rule initialization on each matching processor in the streaming matching engine to obtain a streaming matching engine with a responsibility chain sequence configuration. If a change in condition density is caused by modification of the subscription rules, a responsibility chain reordering is triggered. By passing bidding information through the streaming matching engine, streaming matching analysis based on matching processors is performed to obtain a bidding candidate set. This set is then subjected to three-dimensional vector transformation and non-dominated solution set search based on multi-objective interaction to generate a recommendation sequence. For the recommendation sequence, the following steps are then performed: This application deploys a distributed scheduling framework to perform scheduled and real-time push management. This addresses the technical problems of existing technologies, such as low efficiency in fixed-order matching of traditional bidding information, difficulty in considering multiple aspects in multi-objective recommendation, lack of effective semantic generalization and rule relaxation mechanisms, inability to adapt to changing subscription needs, and difficulty in meeting the capture rate and timeliness requirements of different business scenarios. This application improves streaming filtering efficiency by dynamically rearranging the matching responsibility chain based on rule density, and generates balanced recommendations by combining three-dimensional vector multi-objective non-dominated ranking. When there are no results, the matching rules are relaxed round by round through counterfactual deduction using word vectors, thereby achieving flexible and accurate push management in different business scenarios. Attached Figure Description

[0010] Figure 1 This application provides a flowchart illustrating the bidding information recommendation method based on chain of responsibility matching.

[0011] Figure 2 This application provides a schematic diagram of the structure of a bidding information recommendation system based on chain of responsibility matching.

[0012] Figure labeling: Engine configuration unit 11, recommendation analysis unit 12, push management unit 13. Detailed Implementation

[0013] This application provides a bidding information recommendation method based on chain of responsibility matching to address the technical problems in existing technologies, such as low efficiency of fixed matching order for traditional bidding information, difficulty in considering multiple objectives in multi-objective recommendation, lack of effective semantic generalization and rule relaxation mechanisms, inability to adapt to changing subscription needs, and difficulty in meeting the capture rate and timeliness requirements of different business scenarios.

[0014] Example 1: As Figure 1 As shown, this application provides a method for recommending bidding information based on chain of responsibility matching, the method comprising: S1: When the subscription rules are entered into the database, multi-dimensional feature extraction is performed to construct a subscription tag system. The matching processors in the streaming matching engine are chained and the rules are initialized to obtain the streaming matching engine after the responsibility chain sequence is configured. If the subscription rule is modified and the condition density changes, the responsibility chain is reordered. The condition density is positively correlated with the number of keywords, the number of regional restrictions, the number of type restrictions, and the number of excluded words in the subscription rule.

[0015] In this embodiment, when a subscription rule is added to the database, multi-dimensional feature extraction is performed on the subscription rule to construct a structured subscription tag system. The multi-dimensional features include keyword density, keyword quantity, geographical coverage breadth, and type coverage. Keywords are core fields; for example, announcement identifier fields include announcement ID, announcement title, and announcement type; geographical fields include province, city, and district / county; and monetary fields include project budget and winning bid amount. The extracted keywords are then quantified and organized to form a multi-dimensional subscription tag system.

[0016] Subsequently, a subscription tag system based on multi-dimensional features is used to perform chained sorting and rule initialization on each matching processor in the streaming matching engine.

[0017] The streaming matching engine deploys multiple serially connected matching processors, including at least keyword matching processors, announcement type matching processors, and geographic range matching processors. First, the historical filtering rate of each matching processor is calculated from historical matching logs; this is the proportion of irrelevant bidding information removed by each matching processor. Then, the historical filtering rate is weighted and adjusted based on the condition density of the current subscription rule to obtain the expected filtering rate of each matching processor for the current subscription rule. The condition density is positively correlated with the number of keywords, geographic restrictions, type restrictions, and excluded words in the subscription rule.

[0018] Subsequently, the expected filtering rate is sorted in a chain based on the greedy principle: at each step, the unsorted matching processor with the highest current expected filtering rate is selected as the next node in the chain of responsibility. This process is iterated until all are sorted, the chain of responsibility sequence is determined, and each matching processor is initialized and synchronized to the streaming matching engine according to this sequence, resulting in the streaming matching engine configured with the chain of responsibility sequence.

[0019] Preferably, after the subscription rules are entered into the database, the system uses a database change data capture method to monitor log changes in the subscription rule database in real time. When a user modifies a subscription rule, the CDC component captures the change event, recalculates the modified condition density, and compares it with the original condition density to obtain the condition density variable. When the condition density variable exceeds a preset threshold, the system generates a rearrangement instruction, re-executes the chained sorting, generates a new responsibility chain sequence, and synchronizes it to the streaming matching engine; otherwise, the current responsibility chain sequence remains unchanged.

[0020] Furthermore, to obtain the streaming matching engine after the chain of responsibility sequence is configured, step S1 of this application includes: Multi-dimensional features are extracted from the subscription rules, including keyword density, keyword quantity, geographical coverage breadth, and type coverage. Historical filtering rates of each matching processor are statistically analyzed based on historical matching logs. These historical filtering rates are then weighted and corrected according to the condition density of the subscription rules to obtain the expected filtering rate. The expected filtering rate represents the degree to which each matching processor filters information related to the subscription rules. Based on the expected filtering rate, the matching processors are sorted and rules are initialized using a greedy algorithm to determine the responsibility chain sequence. This responsibility chain sequence is then synchronized to the streaming matching engine.

[0021] The keyword density is calculated as follows: (number of keywords + number of excluded words) / preset keyword limit; the geographical coverage is calculated as: (number of selected provinces + number of selected cities) / total number of provinces; the type coverage is calculated as: number of selected announcement types / total number of announcement types.

[0022] In this embodiment, when a subscription rule is added to the database, multi-dimensional feature extraction is first performed on the subscription rule, specifically including keyword density, keyword quantity, geographical coverage breadth, and type coverage. The number of keywords reflects the breadth of topics of interest to the user; keyword density is calculated as (keyword quantity + excluded word quantity) / preset keyword limit; the excluded word quantity is the total number of negative filter words set by the user; and the preset keyword limit is the maximum keyword capacity of a single rule preset by the system to prevent overly broad matching results.

[0023] For example, if a subscription rule sets 5 keywords and 3 exclusion words, with a preset upper limit of 20, then the keyword density is (5+3) / 20=0.4.

[0024] The calculation method for geographical coverage is (number of selected provinces + number of selected cities) / total number of provinces. When a user selects a province but does not specify any subordinate cities, the system defaults to covering the entire province and only counts the number of provinces. When a specific city is selected, the city is counted independently to accurately reflect the granularity of geographical focus. The calculation method for type coverage is the number of selected announcement types / total number of announcement types supported by the system. The system supports at least the following announcement types: tender announcements, change announcements, winning bid announcements, and termination announcements.

[0025] After completing multi-dimensional feature extraction, the system calls historical matching logs to calculate the historical filtering rate of each matching processor. The streaming matching engine deploys multiple serially connected matching processors, including at least keyword matching processors, announcement type matching processors, and geographic range matching processors. The historical filtering rate is calculated as follows: for each matching processor, the number of bidding information items filtered out by that processor in historical runs is divided by the total number of processes. A higher filtering rate indicates a stronger screening capability of that processor.

[0026] For example, if the keyword matching processor processes 10,000 bidding information entries and filters out 9,500, its historical filtering rate is 0.95.

[0027] Furthermore, the condition density of the current subscription rule is obtained. The condition density is used to characterize the strictness of the overall constraint of the subscription rule, and it is positively correlated with the number of keywords, the number of geographical restrictions, the number of type restrictions, and the number of excluded words.

[0028] In one embodiment, the condition density can be calculated by weighted summation of multi-dimensional features to assign weights based on the importance of the conditions. Subsequently, the historical filtering rates of each matching processor are weighted and corrected according to the condition density to obtain the expected filtering rate of each matching processor for the current subscribed rule, so that the historical average performance is adapted to the constraint features of the current rule. The higher the expected filtering rate, the higher the proportion of predictions that the processor will eliminate irrelevant information under the current rule.

[0029] Subsequently, the subscription tag system corresponding to the subscription rules is written into the corresponding matching processors. A greedy chain sort is then performed based on the expected filtering rate: each time, the matching processor with the highest expected filtering rate is selected as the next node in the responsibility chain from the unsorted matching processors, iterating until all sorting is complete, thus determining the responsibility chain sequence. The greedy sort places the matching processor with the strongest filtering capability at the head of the chain, utilizing short-circuit evaluation to terminate the processing of mismatched data in an early stage, reducing unnecessary computation by subsequent processors. After sorting is complete, each matching processor is initialized according to the responsibility chain sequence and synchronized to the streaming matching engine.

[0030] Furthermore, step S1 of this application includes: If the subscription rules are modified, determine whether the condition density has changed and determine the condition density variable; when the condition density variable is greater than a preset threshold, generate a rearrangement instruction, reorder the responsibility chain of the matching processor, generate a new responsibility chain sequence and synchronize it to the streaming matching engine.

[0031] In this embodiment, when a user modifies a subscription rule that has already been entered into the database, the database change data capture component acquires the subscription rule data before and after the modification in real time. The condition density before and after the modification is calculated according to the aforementioned multi-dimensional feature extraction method. The condition density variable is obtained by subtracting the condition density value before the modification from the condition density value after the modification.

[0032] The condition density variable is used to characterize the magnitude and direction of the change in the strictness of constraints before and after the subscription rule modification. A positive value indicates that the rule constraints are stricter, and a negative value indicates that the rule constraints are more lenient. The condition density variable is compared with a preset threshold. When the condition density variable is greater than the preset threshold, it indicates that the change in condition density caused by the rule modification has substantially affected the filtering efficiency distribution of each matching processor. If the current responsibility chain sequence is continued, it will lead to a decrease in matching performance. At this time, a rearrangement instruction is generated, and the greedy chain sorting operation based on the expected filtering rate of each matching processor is re-executed to generate a new responsibility chain sequence and synchronize it to the streaming matching engine to replace the current responsibility chain sequence.

[0033] When the conditional density variable is less than or equal to the preset threshold, it indicates that the impact of rule modification on the matching efficiency distribution is within an acceptable range, and the current responsibility chain sequence remains unchanged.

[0034] S2: By passing the bidding information through the streaming matching engine, streaming matching analysis based on the matching processor is performed to obtain the bidding candidate set, and three-dimensional vector transformation and non-dominated solution set search based on multi-objective interaction are performed to generate the recommendation sequence.

[0035] In this embodiment of the application, the bidding information is input into the streaming matching engine in the form of real-time data stream. The system calls each matching processor in sequence according to the aforementioned chain of responsibility sequence to perform streaming matching analysis. After each piece of bidding information enters the engine, it is matched and judged by each matching processor in sequence. If any level determines that it does not match, the matching process of the information is terminated. The short-circuit evaluation characteristic of the chain of responsibility is used to reduce invalid calculations. If all judgments pass, the information is added to the bidding candidate set.

[0036] After obtaining the bidding candidate set, a three-dimensional vector transformation is performed on each candidate bidding information. The three-dimensional vector consists of a relevance target score, a quality target score, and a diversity target score. The relevance target score reflects the degree of keyword matching between the candidate bidding information and the subscription rules; the quality target score reflects the timeliness, completeness, and monetary scale of the candidate information; and the diversity target score reflects the degree of dispersion of the candidate information in the tag dimension.

[0037] Subsequently, a multi-objective non-dominated solution set search is performed based on the three-dimensional vectors of each candidate information. A preference weight vector containing three weight components is set. Valid three-dimensional vectors that are not dominated by any other vector are identified through non-dominated sorting. Then, each objective component is weighted and summed according to the preference weight vector. A recommendation sequence is generated by sorting the comprehensive scores from high to low, and a mapping relationship between the preference weight vector and the recommendation sequence is established.

[0038] Furthermore, a three-dimensional vector transformation and a non-dominated solution set search based on multi-objective interaction are performed to generate a recommendation sequence. Step S2 of this application includes: For the bidding candidate set, each candidate bidding information is represented as a three-dimensional vector, which serves as the candidate three-dimensional vector. The three-dimensional vector consists of a relevance target score, a quality target score, and a diversity target score. Historical interaction data of the bidding candidate set is retrieved to mine implicit feedback information. Based on the implicit feedback information and the candidate three-dimensional vector, a non-dominated solution set search based on multi-objective interaction is performed to generate a recommendation sequence.

[0039] In this embodiment, for the candidate bidding information set, the system represents each candidate bidding information as a three-dimensional vector, which serves as the candidate three-dimensional vector. The three-dimensional vector consists of three dimensions: relevance target score, quality target score, and diversity target score. The relevance target score characterizes the degree of matching between the candidate bidding information and the user's subscription rules. It is calculated based on the number of subscription keywords matched by the candidate bidding information during the matching process, the preset weight coefficient of each matched keyword, and the location of the matched keywords. Specifically, when a subscription keyword is matched in the title of the bidding information, its weight contribution value is higher than when it is matched in the body text; when the matched keyword is a high-weight subscription term, its contribution value increases accordingly; the weight contribution values ​​of all matched keywords are summed and normalized to obtain the relevance target score.

[0040] For example, if the title of a candidate tender information contains the high-weight keyword "nitrogen production system" subscribed by the user, and the body text also contains the medium-weight keyword "air separation equipment", then the information will obtain a high relevance target score. Conversely, if the body text only contains low-weight keywords with low frequency, then the relevance target score will be correspondingly low.

[0041] The quality target score is used to characterize the quality level of the candidate bidding information itself. It includes a timeliness factor, an information completeness factor, and a monetary value factor. The timeliness factor is calculated based on the number of days between the date of the bidding announcement and the current time, with a higher score for more recent dates. The information completeness factor is a weighted cumulative evaluation based on whether the announcement contains key fields such as project budget, contact information, and attachments, with a higher score for more complete fields. The monetary value factor is logarithmically normalized based on the project budget or the winning bid amount, with a higher score for higher amounts, but logarithmic scaling is used to prevent extremely high-value projects from dominating the bidding process.

[0042] For example, if a candidate tender information is published on the same day, contains complete contact information and a budget amount, and the budget amount is high, then it will receive a higher quality target score.

[0043] Among them, the diversity target score is used to characterize the distribution characteristics of the industry tags of the candidate bidding information. The calculation method is to calculate the distribution information entropy of the industry tags carried by each candidate information in the bidding candidate set. The higher the information entropy value, the more dispersed the distribution of the candidate set in the industry tag dimension and the wider the coverage of the sub-fields. The lower the information entropy value, the more concentrated the candidate set is in a few tags and the higher the degree of homogeneity.

[0044] Through the comprehensive representation of the above three dimensions, each candidate bidding information is mapped to a coordinate point in three-dimensional space, i.e., a candidate three-dimensional vector, providing a data foundation for subsequent multi-objective optimization search.

[0045] Furthermore, historical interaction data from the candidate bidding list is retrieved to uncover implicit feedback information. This historical interaction data records various user actions during the recommendation process, including but not limited to: user clicks on bidding information in the recommended list and the corresponding browsing duration; user favorites or follows of bidding information; user downloads of bidding document attachments; and user participation in bidding activities.

[0046] Optionally, the implicit feedback information mining process specifically involves: pre-setting corresponding behavior weight coefficients for each behavior type to quantify the differences in the intensity of user interest indications among different behavior types. For example, the weight coefficient for bidding participation is higher than that for bookmarking, the weight coefficient for bookmarking is higher than that for clicking and browsing, and ignore or skip behaviors are set to negative weight coefficients. For each candidate bidding information, the frequency of various historical behaviors generated by the user is statistically analyzed. The frequency of each behavior is weighted and accumulated with its corresponding behavior weight coefficient, and a time decay factor is applied, i.e., recently occurring behaviors are given a higher decay weight, and older behaviors are given a lower decay weight, so that the implicit feedback information more accurately reflects the user's current interest state.

[0047] For example, if a user clicks on, favorites, and participates in a bid for nitrogen generation system bidding information multiple times in the past 30 days, then this implicit feedback indicates that the user has a strong positive interest signal in this category of information.

[0048] Subsequently, based on the mined implicit feedback information, the values ​​of each weight component in the multi-objective preference weight vector are adjusted to match the preference weight vector with the implicit interest tendencies reflected by users based on historical behavior. The candidate three-dimensional vectors in the bidding candidate set are traversed, and valid three-dimensional vectors not dominated by any other vector are identified through non-dominated sorting, forming a non-dominated solution set.

[0049] Subsequently, based on the adjusted preference weight vector, the three target components of each effective three-dimensional vector in the non-dominated solution set are weighted and summed to obtain a comprehensive score. The recommendation sequence is then generated by sorting the comprehensive scores from high to low.

[0050] Furthermore, a non-dominated solution set search based on multi-objective interaction is performed to generate a recommendation sequence. Step S2 of this application includes: A first preference weight vector based on multiple objectives is defined; each three-dimensional vector in the candidate three-dimensional vectors is traversed, and the effective three-dimensional vectors are determined by non-dominated sorting, wherein if the first three-dimensional vector is not inferior to the second three-dimensional vector on each objective and is superior to the second three-dimensional vector on at least one objective, the first three-dimensional vector is regarded as the dominant vector; according to the first preference weight vector, the effective three-dimensional vectors are weighted and summed to obtain the first recommendation sequence, wherein the recommendation sequence contains multiple mapping groups represented as preference weight vector-recommendation sequence.

[0051] Specifically, a first preference weight vector based on multiple objectives is defined. This preference weight vector contains three weight components corresponding to the relevance objective score, the quality objective score, and the diversity objective score, respectively. The sum of each weight component is 1, and the value of each weight component reflects the degree of emphasis on the three optimization objectives in the current recommendation scenario. That is, the first preference weight vector is the preference weight distribution corresponding to the first recommendation scenario.

[0052] Furthermore, the candidate three-dimensional vectors in the bidding candidate set are traversed, and valid three-dimensional vectors are identified through non-dominated sorting to form a non-dominated solution set. The determination rule for non-dominated sorting is as follows: for the first and second three-dimensional vectors among the candidate three-dimensional vectors, if the score of the first three-dimensional vector in each target dimension is not inferior to the score of the second three-dimensional vector in each target dimension, and the score of the first three-dimensional vector in at least one target dimension is superior to the score of the second three-dimensional vector in that dimension, then the first three-dimensional vector is determined to be the dominant one of the second three-dimensional vectors, and the second three-dimensional vector is correspondingly determined to be the dominated one.

[0053] After traversing all candidate 3D vectors, the candidate 3D vectors that are not dominated by any other candidate 3D vectors constitute a non-dominated solution set. Preferably, each candidate bidding information in the non-dominated solution set is at the optimal boundary in the 3D target space, meaning that it is impossible to improve the score of one target without sacrificing the score of another.

[0054] Subsequently, the implicit feedback information is incorporated as a constraint or preference-oriented factor into the screening process of the non-dominated solution set.

[0055] Specifically, the values ​​of each weight component in the preference weight vector are adjusted according to the implicit feedback information so that the preference weight vector reflects the implicit preference tendency of users based on historical behavior. Then, the three objective components of each effective three-dimensional vector in the non-dominated solution set are weighted and summed according to the adjusted preference weight vector to obtain the multi-objective comprehensive score of each candidate bidding information.

[0056] Subsequently, the candidate bidding information in the non-dominated solution set is sorted in descending order of comprehensive score to generate the first recommended sequence.

[0057] Similarly, using preference weight vectors for different recommendation scenarios, the above decision-making steps are executed to obtain the recommendation sequence corresponding to the scenario. Multiple sequences corresponding to the recommendation scenario and the recommendation sequence are then added to the recommendation sequence. Preferably, adaptive and rapid recommendation switching under different scenarios can be achieved.

[0058] Through the above methods, the system achieves collaborative optimization of multiple recommendation objectives, ensuring the relevance of recommendation results while taking into account quality and diversity, and the recommendation ranking results can adaptively match the personalized preferences reflected by users based on implicit behaviors.

[0059] Furthermore, in this application, the recommended sequence is judged to be an empty set. If it is an empty set, the semantic extension matching processor is activated, and the rules are relaxed in the word vector space of the bidding domain based on counterfactual inference. The rules are relaxed round by round until a non-empty recommended sequence is generated, which replaces the recommended sequence.

[0060] In this embodiment, an empty set judgment is performed on the generated recommendation sequence. If the recommendation sequence is not empty, it is directly used as the final recommendation result; if the recommendation sequence is empty, it indicates that no bidding information is matched under the current subscription rule, and the semantic extension matching processor is then activated.

[0061] Specifically, the semantic extension matching processor is pre-configured with a word vector space in the bidding and tendering field. The word vector space is constructed based on a large-scale bidding and tendering corpus. Each word is mapped to a vector representation in a high-dimensional space, and words with similar semantics have similar vector directions in this space.

[0062] The matching constraints of the subscription rules are relaxed in rounds based on a counterfactual reasoning mechanism: in each round of relaxation, the matching threshold is lowered or the keywords are expanded to a set of semantic synonyms according to a preset step size, i.e., a pre-set scale that can be relaxed in a single instance.

[0063] In this process, semantic synonyms are selected by calculating the cosine similarity between the subscription keyword vector and the candidate word vector in the word vector space, and candidate words with similarity exceeding a preset threshold are selected as extended words. After one round of relaxation is completed, streaming matching analysis and recommendation sequence generation are re-executed, and the newly generated recommendation sequence is checked for empty set again. If it is still empty set, the next round of relaxation is performed until a non-empty recommendation sequence is generated.

[0064] Preferably, the degree of constraint relaxation in each round is controlled by a preset step size to ensure a smooth and controllable relaxation process, avoiding deviation of the recommendation result from the user's original subscription intent due to excessive relaxation in a single instance. After generating a non-empty recommendation sequence, the original empty set recommendation sequence is replaced by this non-empty recommendation sequence as the final recommendation result output.

[0065] Furthermore, the steps in this application also include: If the tender candidate set is empty, a semantic expansion matching instruction is generated. Based on the semantic expansion matching instruction, the semantic expansion matching processor is activated, and a first-round rule relaxation process based on counterfactual inference is performed in the word vector space of the tendering domain. The matching processor in the streaming matching engine adjusts the rules to determine a first-round relaxed tender candidate set, wherein a single rule relaxation constraint is performed with a preset step size. A three-dimensional vector transformation and non-dominated solution set search are performed on the first-round relaxed tender candidate set to determine a first-round relaxed recommendation sequence. It is determined whether the first-round relaxed recommendation sequence is empty; if so, multiple rounds of relaxation processing are performed until the relaxed recommendation sequence is non-empty.

[0066] Among them, the diversity target score is the information entropy of the label distribution in the bidding candidate set; in the word vector space, semantic expansion is performed by calculating the cosine similarity between the n-dimensional word vectors based on the subscription keywords and the n-dimensional word vectors of the candidate matching words.

[0067] In this embodiment of the application, if the bidding candidate set is empty after the streaming matching engine has completed the chain of responsibility matching, it indicates that no bidding information has passed the screening of all matching processors under the current subscription rule conditions, and a semantic extension matching instruction is generated.

[0068] According to the semantic extension matching instruction, the semantic extension matching processor is activated, which has a word vector space in the bidding field pre-set.

[0069] In one specific implementation, the word vector space is obtained by pre-training with Word2Vec or BERT based on a large-scale bidding corpus. Each word is mapped to a vector representation in an n-dimensional vector space. Words with similar semantics have similar vector directions in this space. The value of n determines the representation accuracy and computational complexity of the word vectors. In one embodiment, n is 256-dimensional.

[0070] Subsequently, after the semantic expansion matching processor is activated, a round of rule relaxation processing based on counterfactual inference is performed in the word vector space.

[0071] The specific principle is as follows: hypothetically relax the matching constraints of the current subscription rule according to a preset relaxation strategy, in order to explore whether potential semantically relevant bidding information can be matched under conditions that do not fully satisfy the original constraints.

[0072] The first round of rule relaxation specifically includes: relaxing the matching constraints of the subscription rules in a single step, according to a preset step size. This preset step size is used to control the degree of relaxation in a single round, ensuring the smoothness and controllability of the relaxation process. In one embodiment, the rule relaxation includes lowering the similarity threshold for keyword matching, that is, reducing the cosine similarity threshold for semantic expansion determination from the original value by one step unit. In the word vector space, semantic expansion is performed by calculating the cosine similarity between the n-dimensional word vectors of the subscription keywords and the n-dimensional word vectors of the candidate matching words. The formula for calculating the cosine similarity is: the dot product of the word vectors of the subscription keywords and the word vectors of the candidate matching words divided by the product of their magnitudes. The closer the value is to 1, the more semantically similar the two words are.

[0073] For example, if a user subscribes to the keyword "nitrogen generation system", and the cosine similarity between its word vector and the word vector of "nitrogen generator" is 0.82, which exceeds the preset semantic matching threshold, then "nitrogen generator" is included in the semantic expansion word set; if the cosine similarity is lower than the threshold, it is not expanded.

[0074] Based on the results of the first round of rule relaxation, the system adjusts the rules of each matching processor in the streaming matching engine, incorporating the semantic extension word set into the matching word library of the keyword matching processor, so that the subsequent matching process can retain the original keyword matching while adding the matching judgment of semantic extension words.

[0075] Similarly, streaming matching analysis is re-executed to obtain a candidate set for the first round of relaxed bidding. A three-dimensional vector transformation and non-dominated solution set search are then performed on this candidate set to determine the first round of relaxed recommendation sequence and to perform an empty set check.

[0076] If the first round of relaxed recommendation sequence is a non-empty set, then the first round of relaxed recommendation sequence is directly used as the final recommendation sequence output; if the first round of relaxed recommendation sequence is still an empty set, then it is determined that the relaxation of the rules in the current round is not sufficient to recall any bidding information, and a multi-round relaxation process is triggered.

[0077] In the multi-round relaxation process, the above rule relaxation operation is executed iteratively. That is, in each round, the similarity threshold is further reduced by a preset step size, the semantic expansion word set is redefined, the rule parameters of each matching processor in the streaming matching engine are updated, streaming matching analysis, three-dimensional vector transformation and non-dominated solution set search are re-executed, a new round of relaxed recommendation sequence is generated, and empty set judgment is performed.

[0078] This process is repeated until the generated relaxed recommendation sequence is a non-empty set, at which point the iteration terminates and the final non-empty recommendation sequence is output as the recommendation result.

[0079] Through the aforementioned round-by-round relaxation mechanism, while ensuring a smooth and controllable relaxation process, it is possible to effectively recall bidding information that is semantically related to the user's subscription intent but was not precisely matched due to differences in expression.

[0080] S3: For the recommended sequence, a distributed scheduling framework is deployed to perform timed and real-time push management; Furthermore, for the recommended sequence, a distributed scheduling framework is deployed to perform timed and real-time push management. Step S3 of this application includes: Based on the set push time period, frequency mode, and on / off status, a distributed scheduling framework based on the recommendation sequence is deployed; according to the distributed scheduling framework, timed push tasks and real-time push tasks are executed and sent to the user terminal via the message interface.

[0081] In this embodiment, the generated recommendation sequence is managed through a distributed scheduling framework for scheduled and real-time push operations. Specifically, the push configuration parameters preset by the user are first obtained, including at least the push period, frequency mode, and on / off status. The push period defines the time interval for daily push operations; the frequency mode determines the time interval and triggering rules for push tasks, and includes at least three selectable modes: daily scheduled push mode; real-time push mode, which periodically checks and pushes newly matched bidding information according to a preset short interval; and a custom time period push mode, where the user receives pushes at a custom time interval within the push period; the on / off status indicates the user's willingness to enable or disable the push function. When the on / off status is off, the system does not perform any push operations.

[0082] After obtaining the aforementioned push configuration parameters, a distributed scheduling framework based on recommendation sequences is deployed based on these parameters. Specifically, push task trigger events are generated according to the push time period and frequency pattern. The task dispatcher is responsible for assigning these trigger events to available worker nodes for execution, thereby achieving high-availability scheduling and load balancing of push tasks, completing the deployment of the distributed scheduling framework, and executing scheduled and real-time push tasks accordingly.

[0083] Subsequently, the push message is sent to the user terminal via a message interface, which includes at least one or more combinations of WeChat template message interface, SMS interface, and in-application push interface to support multi-channel message delivery. Through the aforementioned distributed scheduling mechanism, the timeliness and reliability of push tasks can be guaranteed even in scenarios with large-scale concurrent user pushes.

[0084] The bidding information recommendation method based on chain of responsibility matching provided in this application has the following technical effects: 1. A dynamic adaptive responsibility chain streaming matching approach is adopted. Multi-dimensional features of subscribed rules are extracted to calculate condition density, and a matching responsibility chain is formed by greedily sorting based on the processor's historical filtering rate. This chain is automatically rearranged after rule changes. High-filtering-efficiency processors are prioritized, significantly reducing invalid matching calculations and improving the streaming engine's throughput. Dynamic optimization of the matching execution order significantly improves the efficiency of streaming matching processing of bidding information.

[0085] 2. A three-dimensional vector multi-objective non-dominated ranking recommendation is adopted. A three-dimensional feature vector is constructed based on relevance, quality, and information entropy diversity, and non-dominated solution set selection is performed in conjunction with implicit user interactions. This balances the accuracy, quality, and content richness of the recommendation results, avoiding bias from a single metric.

[0086] 3. A counterfactual semantic multi-round relaxation and completion mechanism: When no matching results are found, the semantic expansion module is activated, gradually relaxing the retrieval constraints based on word vector cosine similarity, and iterating until recommended content is generated. This solves the problem of no push notifications and information omissions caused by overly strict user subscription rules.

[0087] Example 2, based on the same inventive concept as the bidding information recommendation method based on chain of responsibility matching in the previous examples, such as... Figure 2 As shown, this application provides a bidding information recommendation system based on chain of responsibility matching, wherein the system includes: Engine Configuration Unit 11: When the subscription rules are entered into the database, multi-dimensional feature extraction is performed to construct a subscription tag system. The matching processors in the streaming matching engine are chained and the rules are initialized to obtain the streaming matching engine after the responsibility chain sequence is configured. If the subscription rule is modified and the condition density changes, the responsibility chain is reordered. The condition density is positively correlated with the number of keywords, the number of regional restrictions, the number of type restrictions and the number of excluded words in the subscription rule. Recommendation Analysis Unit 12: By passing the bidding information through the streaming matching engine, performing streaming matching analysis based on the matching processor, obtaining the bidding candidate set, and performing three-dimensional vector transformation and non-dominated solution set search based on multi-objective interaction, a recommendation sequence is generated; Push management unit 13: For the recommended sequence, it performs timed and real-time push management by deploying a distributed scheduling framework; Specifically, by determining whether the recommended sequence is empty, if it is empty, the semantic extension matching processor is activated, and the rules are relaxed in the word vector space of the bidding domain based on counterfactual inference. The rules are relaxed round by round until a non-empty recommended sequence is generated, which replaces the recommended sequence.

[0088] Furthermore, the engine configuration unit 11 is used to perform the following steps: extracting multi-dimensional features from the subscription rules, wherein the multi-dimensional features include keyword density, keyword quantity, geographical coverage breadth, and type coverage range; calculating the historical filtering rate of each matching processor based on historical matching logs, and weighting and correcting the historical filtering rate according to the condition density of the subscription rules to obtain the expected filtering rate, wherein the expected filtering rate is the degree of information filtering by each matching processor for the subscription rules; and performing chain sorting and rule initialization of the matching processors based on the greedy principle according to the expected filtering rate, determining the responsibility chain sequence, and synchronizing the responsibility chain sequence to the streaming matching engine.

[0089] The keyword density is calculated as follows: (number of keywords + number of excluded words) / preset keyword limit; the geographical coverage is calculated as: (number of selected provinces + number of selected cities) / total number of provinces; the type coverage is calculated as: number of selected announcement types / total number of announcement types.

[0090] Furthermore, the engine configuration unit 11 is used to perform the following steps: if the subscription rule is modified, determine whether the condition density has changed and determine the condition density variable; when the condition density variable is greater than a preset threshold, generate a rearrangement instruction, reorder the responsibility chain of the matching processor, generate a new responsibility chain sequence and synchronize it to the streaming matching engine.

[0091] Furthermore, the recommendation analysis unit 12 is used to perform the following steps: for the bidding candidate set, each candidate bidding information is represented as a three-dimensional vector as a candidate three-dimensional vector, wherein the three-dimensional vector is composed of a relevance target score, a quality target score, and a diversity target score; retrieve the historical interaction data of the bidding candidate set to mine implicit feedback information; and perform a non-dominated solution set search based on multi-objective interaction according to the implicit feedback information and the candidate three-dimensional vector to generate a recommendation sequence.

[0092] Furthermore, the recommendation analysis unit 12 is used to perform the following steps: setting a first preference weight vector based on multiple objectives; traversing each three-dimensional vector in the candidate three-dimensional vectors, and determining the effective three-dimensional vectors through non-dominated sorting, wherein if the first three-dimensional vector is not inferior to the second vector on any objective and is superior to the second three-dimensional vector on at least one objective, the first three-dimensional vector is regarded as the dominant vector; according to the first preference weight vector, the effective three-dimensional vectors are weighted and summed to obtain a first recommendation sequence, wherein the recommendation sequence contains multiple mapping groups represented as preference weight vector-recommendation sequence.

[0093] Furthermore, the system is also used to perform the following steps: if the bidding candidate set is empty, generate a semantic expansion matching instruction; according to the semantic expansion matching instruction, activate the semantic expansion matching processor, perform a round of rule relaxation processing based on counterfactual inference in the word vector space of the bidding domain, adjust the rules of the matching processor in the streaming matching engine, and determine a round of relaxed bidding candidate set, wherein a single rule relaxation constraint is performed with a preset step size; perform three-dimensional vector transformation and non-dominated solution set search on the round of relaxed bidding candidate set to determine a round of relaxed recommendation sequence; determine whether the round of relaxed recommendation sequence is empty, and if so, perform multiple rounds of relaxation processing until the relaxed recommendation sequence is non-empty.

[0094] Among them, the diversity target score is the information entropy of the label distribution in the bidding candidate set; in the word vector space, semantic expansion is performed by calculating the cosine similarity between the n-dimensional word vectors based on the subscription keywords and the n-dimensional word vectors of the candidate matching words.

[0095] Furthermore, the push management unit 13 is used to perform the following steps: deploy a distributed scheduling framework based on a recommendation sequence according to the set push time period, frequency mode and on / off status; execute timed push tasks and real-time push tasks according to the distributed scheduling framework, and send them to the user terminal via the message interface.

[0096] Through the foregoing detailed description of the tender information recommendation method based on chain of responsibility matching, those skilled in the art can clearly understand the tender information recommendation method based on chain of responsibility matching in this embodiment. As for the apparatus disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and relevant parts can be referred to the method section description.

[0097] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A bidding information recommendation method based on chain of responsibility matching, characterized in that, The method includes: When the subscription rules are entered into the database, multi-dimensional feature extraction is performed to construct a subscription tag system. The matching processors in the streaming matching engine are chained and the rules are initialized to obtain the streaming matching engine after the responsibility chain sequence is configured. If the subscription rule is modified and the condition density changes, the responsibility chain is reordered. The condition density is positively correlated with the number of keywords, the number of regional restrictions, the number of type restrictions, and the number of excluded words in the subscription rule. By passing the bidding information through the streaming matching engine and performing streaming matching analysis based on the matching processor, a bidding candidate set is obtained, and a three-dimensional vector transformation and non-dominated solution set search based on multi-objective interaction are performed to generate a recommendation sequence. For the recommended sequence, a distributed scheduling framework is deployed to perform timed and real-time push management; Specifically, by determining whether the recommended sequence is empty, if it is empty, the semantic extension matching processor is activated, and the rules are relaxed in the word vector space of the bidding domain based on counterfactual inference. The rules are relaxed round by round until a non-empty recommended sequence is generated, which replaces the recommended sequence.

2. The bidding information recommendation method based on chain of responsibility matching as described in claim 1, characterized in that, The streaming matching engine, after obtaining the chain of responsibility sequence configuration, includes: Extract multi-dimensional features from the subscription rules, wherein the multi-dimensional features include keyword density, keyword quantity, geographical coverage breadth, and type coverage range; Based on the historical matching logs, the historical filtering rate of each matching processor is statistically analyzed. According to the condition density of the subscription rules, the historical filtering rate is weighted and corrected to obtain the expected filtering rate, wherein the expected filtering rate is the degree of information filtering by each matching processor for the subscription rules. Based on the expected filtering rate, the matching processor is subjected to chain sorting and rule initialization based on the greedy principle to determine the responsibility chain sequence, and the responsibility chain sequence is synchronized to the streaming matching engine.

3. The bidding information recommendation method based on chain of responsibility matching as described in claim 2, characterized in that, The keyword density is calculated as: (number of keywords + number of excluded words) / preset keyword limit; The calculation of the geographical coverage is: (number of selected provinces + number of selected cities) / total number of provinces; The calculation of the coverage of the type is: number of selected announcement types / total number of announcement types.

4. The bidding information recommendation method based on chain of responsibility matching as described in claim 3, characterized in that, If the subscription rules are modified, determine whether the condition density has changed, and determine the condition density variable; When the condition density variable is greater than a preset threshold, a rearrangement instruction is generated to reorder the responsibility chain of the matching processor, generate a new responsibility chain sequence, and synchronize it to the streaming matching engine.

5. The bidding information recommendation method based on chain of responsibility matching as described in claim 1, characterized in that, Perform 3D vector transformation and non-dominated solution set search based on multi-objective interaction to generate recommendation sequences, including: For the candidate tender set, each candidate tender information is represented as a three-dimensional vector, which serves as the candidate three-dimensional vector. The three-dimensional vector consists of a relevance target score, a quality target score, and a diversity target score. Retrieve historical interaction data from the bidding candidate set to uncover implicit feedback information; Based on the implicit feedback information and the candidate 3D vectors, a non-dominated solution set search based on multi-objective interaction is performed to generate a recommendation sequence.

6. The bidding information recommendation method based on chain of responsibility matching as described in claim 5, characterized in that, Perform a non-dominated solution set search based on multi-objective interactions to generate a recommendation sequence, including: Define a first preference weight vector based on multiple objectives; Traverse each of the candidate three-dimensional vectors and determine the effective three-dimensional vectors through non-dominated sorting. If the first three-dimensional vector is not inferior to the second three-dimensional vector on each target and is superior to the second three-dimensional vector on at least one target, the first three-dimensional vector is regarded as the dominant vector. Based on the first preference weight vector, the effective three-dimensional vectors are weighted and summed to form the first recommendation sequence, wherein the recommendation sequence contains multiple mapping groups represented as preference weight vector-recommendation sequence.

7. The bidding information recommendation method based on chain of responsibility matching as described in claim 6, characterized in that, If the tender candidate set is empty, generate a semantic extension matching instruction; According to the semantic expansion matching instruction, the semantic expansion matching processor is activated, and a round of rule relaxation processing based on counterfactual inference is performed in the word vector space of the bidding field. The matching processor in the streaming matching engine is adjusted to determine a round of relaxed bidding candidate set, wherein a single rule relaxation constraint is performed with a preset step size. Perform three-dimensional vector transformation and non-dominated solution set search on the candidate set of the first round of relaxed bidding to determine the recommended sequence of the first round of relaxed bidding. Determine whether the first round of relaxed recommendation sequence is an empty set. If it is, perform multiple rounds of relaxed processing until the relaxed recommendation sequence is a non-empty set.

8. The bidding information recommendation method based on chain of responsibility matching as described in claim 7, characterized in that, The diversity target score is the information entropy of the label distribution in the tender candidate set; In the aforementioned word vector space, semantic expansion is performed by calculating the cosine similarity between the n-dimensional word vectors based on the subscribed keywords and the n-dimensional word vectors of the candidate matching words.

9. The bidding information recommendation method based on chain of responsibility matching as described in claim 1, characterized in that, For the recommended sequence, a distributed scheduling framework is deployed to perform timed and real-time push management, including: Deploy a distributed scheduling framework based on recommendation sequences according to the set push time period, frequency mode, and on / off status; Based on the distributed scheduling framework, scheduled push tasks and real-time push tasks are executed and sent to the user terminal via the message interface.

10. A bidding information recommendation system based on chain of responsibility matching, characterized in that, The system is used to perform the tender information recommendation method based on chain of responsibility matching as described in any one of claims 1-9, the system comprising: Engine Configuration Unit: When subscription rules are entered into the database, multi-dimensional feature extraction is performed to construct a subscription tag system. Chain sorting and rule initialization are performed on each matching processor in the streaming matching engine to obtain the streaming matching engine after the responsibility chain sequence is configured. If the subscription rule is modified and the condition density changes, the responsibility chain is reordered. The condition density is positively correlated with the number of keywords, the number of regional restrictions, the number of type restrictions, and the number of excluded words in the subscription rule. Recommendation Analysis Unit: By passing the bidding information through the streaming matching engine, streaming matching analysis based on the matching processor is performed to obtain the bidding candidate set, and three-dimensional vector transformation and non-dominated solution set search based on multi-objective interaction are performed to generate the recommendation sequence; Push Management Unit: For the recommended sequence, it performs timed and real-time push management by deploying a distributed scheduling framework; Specifically, by determining whether the recommended sequence is empty, if it is empty, the semantic extension matching processor is activated, and the rules are relaxed in the word vector space of the bidding domain based on counterfactual inference. The rules are relaxed round by round until a non-empty recommended sequence is generated, which replaces the recommended sequence.

Images