CN-121980080-A - Intelligent recommendation method based on scene configuration

CN121980080ACN 121980080 ACN121980080 ACN 121980080ACN-121980080-A

Abstract

The invention discloses an intelligent recommendation method, system and equipment based on cooperation of a rule engine and a large model. The method comprises the steps of analyzing scene configuration input by a user into feature vectors and searching in a rule model library, executing double-path decision according to confidence weight of a search result, directly recommending standard packages if a high confidence rule is hit, triggering a dynamic variation mechanism if the high confidence rule is missed, and generating a temporary scheme based on atomic commodity constraint by using a large language model. The system further performs weight optimization on the stock rules in response to the adoption or corrective feedback of the user, or registers the temporary solution as a new standard suite, and generates new rules through feature interval generalization. According to the invention, through the organic combination of deterministic search and generative reasoning, the recommendation problem of a complex long-tail scene is effectively solved, the self-adaptive growth and iteration of a knowledge base are realized by utilizing a feedback closed loop, and the coverage rate and accuracy of a recommendation system are obviously improved.

Inventors

XU KAIHUA
GUO XIAOCHUAN
XIE CHEN
WU BIN
XIA YINGJIE

Assignees

博康智能信息技术有限公司

Dates

Publication Date: 20260505
Application Date: 20260116

Claims (10)

1. The intelligent recommendation method based on scene configuration is characterized by comprising the following steps of: acquiring scene configuration data input by a user, wherein the scene configuration data comprises feature vectors describing the dimension of the scene of the user; Searching and matching in a preset rule model base based on the scene configuration data, and judging whether a standard suit meeting preset conditions exists or not; if the search matching is successful, selecting a corresponding standard suit as a recommendation basis; if the search matching fails or the matching result does not reach the preset standard, generating a temporary suit combination meeting the scene configuration data based on an atomic commodity library by using a large language model; Generating and displaying a recommended scheme according to the standard suit or the temporary suit combination; acquiring adoption feedback data of a user aiming at the recommended scheme; And when the user adopts the temporary set combination, executing a rule curing process, converting the temporary set combination into a new standard set, and establishing a mapping relation between the scene configuration data and the new standard set to generate a new rule and storing the new rule into the rule model library.
2. The method of claim 1, wherein a plurality of mapping rules are stored in the rule model library, each mapping rule comprising a conditional function, a target package index, and a confidence weight; the searching and matching in a preset rule model base based on the scene configuration data comprises the following steps: traversing the rule model library, and judging whether the scene configuration data meets constraint conditions or not by utilizing the condition function to obtain a candidate rule set; And calculating the confidence coefficient weight of each rule in the candidate rule set, and obtaining the maximum confidence coefficient weight.
3. The method according to claim 2, wherein if the search fails or the matching result does not reach the preset criteria, specifically comprising: Judging whether the candidate rule set is empty or not, or whether the maximum confidence weight is lower than a preset judging threshold value or not; If yes, judging to trigger a dynamic mutation mechanism, and executing the step of generating temporary package combination meeting the scene configuration data based on an atomic commodity library by using a large language model.
4. The method of claim 3, wherein the generating a temporary package combination that satisfies the scenario configuration data based on an atomic commodity library using a large language model comprises: Retrieving commodity metadata and hard constraint rules in the atomic commodity library; Constructing a structured prompt word containing the scene configuration data, the hard constraint rule and the task target; And calling the large language model to perform reasoning, and generating a temporary commodity list which meets the hard constraint rule and adapts to the scene configuration data as the temporary suit combination.
5. The method of claim 1, wherein the generating a recommendation from the standard suite or the temporary suite combination comprises: acquiring a commodity list and price information contained in a target set, wherein the target set is the standard set or the temporary set combination; constructing a generated prompting word by combining the scene configuration data, the target suit and the constraint of a preset output format; And generating a recommendation scheme containing a commodity list, a total price and an adaptive interpretation document based on the generated prompt word by using the large language model.
6. The method of claim 2, wherein updating the rule model library in response to the adoption feedback data further comprises: if the recommendation scheme is generated based on the standard suit and the user feedback is directly adopted, carrying out numerical updating on the confidence coefficient weight of the mapping rule corresponding to the standard suit according to the preset learning rate parameter so as to improve the priority of the mapping rule.
7. The method according to claim 1, wherein the performing a rule curing procedure converts the temporary set combination into a new standard set, in particular comprising: Assigning a unique suit code to the temporary suit combination and registering the unique suit code in a standard suit library; Extracting key feature dimensions in the scene configuration data triggering the current recommendation; And constructing a new logic judgment function based on the key feature dimension, associating the logic judgment function with the unique set code, and generating the new rule.
8. The method of claim 7, wherein constructing a new logical decision function based on the key feature dimension comprises: Performing interval generalization processing on the numerical type features in the scene configuration data; determining effective coverage areas of the features according to specific numerical values of the numerical type features and preset generalized tolerance coefficients; The logical judgment function is constructed so that the logical judgment function can cover a feature area centered on the scene configuration data.
9. The method of claim 8, wherein the rule curing process further comprises: If the user performs commodity adding and deleting correction on the recommended scheme before adopting, judging whether the corrected final suit combination exists in a standard suit library; And if not, taking the final suite combination as the temporary suite combination, and executing the rule curing process.
10. An electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor implements the method of any one of claims 1 to 9 when the program is executed by the processor.

Description

Intelligent recommendation method based on scene configuration Technical Field The invention relates to the technical field of computer data processing and artificial intelligence, in particular to an intelligent recommendation method based on scene configuration. Background Along with the deep integration of industrial internet and electronic commerce, an intelligent recommendation system aiming at complex service scenes (such as industrial selection, home customization, security configuration and the like) has become a key tool for improving service efficiency. Such scenarios typically involve a large number of parameter configurations, stringent constraints, and diversified user requirements, placing extremely high demands on accuracy, flexibility, and response speed of the recommendation system. Traditional intelligent recommendation systems rely primarily on predefined expert rule bases or decision tree models for logical reasoning. The deterministic logic-based architecture is a static and closed knowledge representation in nature, although it has high accuracy and interpretability in handling standard, high frequency business scenarios. With the expansion of business boundaries and the exponential growth of long-tailed demand, it becomes extremely impractical to attempt to exhaust all possible scenarios and commodity combining rules. When configuration parameters input by a user exceed the coverage range of preset rules or belong to undefined long tail scenes, the traditional rule-based system cannot effectively infer, only a default scheme with a null result or extremely poor generality can be returned, the application range of the system is limited, and the maintenance cost is increased explosively along with the increase of the number of rules. On the other hand, although the Large Language Model (LLM) has shown strong semantic understanding and generating capability in recent years, a new idea is provided for solving the long tail problem, and the direct application of the Large Language Model (LLM) to strict business configuration recommendation still has significant defects. The general large model lacks accurate grasp of physical properties and hard constraints (such as physical compatibility, voltage matching, stock state and the like) of atomic commodity in a specific vertical field, and is extremely easy to generate a reasonable-looking but actually incapable of being executed in a landing manner 'illusion' recommended scheme. In addition, the high computational effort costs and relatively high response delays of large models also make it difficult to completely replace conventional systems to handle all high frequency, simple query requests. More critical is that the prior art has faults in handling user feedback and knowledge precipitation. When a user manually modifies or creatively combines the recommendation results for a particular scenario, these precious interaction experiences are often stored only as discrete histories and cannot be automatically converted into reusable general rules at the bottom of the system. This means that the system cannot realize "solidification" and "generalization" of knowledge from actual business interactions, so that the system still needs to repeatedly consume computing power to recalculate when facing similar new scenes, and cannot realize adaptive growth and spiral increase of the recommendation capability of the system. Disclosure of Invention The technical problem to be solved by the invention is that the existing recommendation system based on rules is difficult to cover long-tail complex scenes, and the generated recommendation which simply depends on a large model is lack of certainty and is difficult to precipitate system knowledge. In view of the above, the invention provides an intelligent recommendation method and electronic equipment based on scene configuration, and knowledge self-evolution of a recommendation system is realized by constructing a 'search + generation' dual-path mechanism and introducing a rule solidification flow based on feedback. The first aspect of the invention provides an intelligent recommendation method based on scene configuration. The method mainly comprises the steps of firstly obtaining scene configuration data input by a user, wherein the scene configuration data comprise feature vectors describing the scene dimension of the user, carrying out search matching in a preset rule model library based on the scene configuration data, judging whether standard packages meeting preset conditions exist or not, selecting corresponding standard packages as recommendation bases if the search matching is successful, generating temporary package combinations meeting the scene configuration data based on an atomic commodity library by using a large language model if the search matching fails or a matching result does not reach the preset standard, generating and displaying a recommendation scheme according to the standard pac