Search

CN-121981771-A - Questionnaire matching method based on AI behavior recognition

CN121981771ACN 121981771 ACN121981771 ACN 121981771ACN-121981771-A

Abstract

The invention discloses an investigation questionnaire matching method based on AI behavior recognition, which belongs to the technical field of artificial intelligent behavior analysis and online investigation questionnaire accurate matching, and comprises the following steps of S1, setting semantic requirements; S2, real-time data acquisition and preprocessing, S3, data analysis and demand matching, S4, target questionnaire pop-up, S5, report and suggestion output. By adopting the questionnaire matching method based on the AI behavior recognition, the invention realizes simple rule setting, avoids the earlier-stage portrait cost, and accurately fits the current situation of the user through semantic requirement description, real-time page behavior acquisition and AI large model intention analysis technology, thereby greatly improving the instantaneity and the accuracy of questionnaire matching.

Inventors

  • HUANG PEI
  • LI XIN
  • HE JUN
  • LUO WEI

Assignees

  • 上海众言网络科技有限公司

Dates

Publication Date
20260505
Application Date
20260123

Claims (10)

  1. 1. The questionnaire matching method based on AI behavior recognition is characterized by comprising the following steps: s1, inputting a scene semantic requirement by a user, marking a priority, converting the semantic rule into a system technical rule through a rule generation algorithm, separating words, eliminating stop words, extracting semantic feature vectors by adopting a LoRA lightweight and fine-tuned BERT-base model, calculating semantic feature confidence, and synchronously transmitting after associating priority labels; s2, pre-configuring acquisition rules, capturing user page operation data in real time through event monitoring and DOM structure analysis technology, generating structured behavior data and encrypting and transmitting the structured behavior data after data cutting, outlier rejection, standardization, time sequence feature extraction and differential privacy protection processing; S3, acquiring semantic feature vectors of the S1 and structured behavior data of the S2, combining historical event data and a current DOM structure, obtaining weighted matching degree through feature vector mapping, AI large model intention recognition, dynamic weight optimization and cosine similarity calculation, assisting questionnaire title analysis and an optional user inquiry mechanism, and screening out an optimal matching questionnaire; s4, calculating a dynamic threshold according to the service load, the user activity and the time information, triggering a popup rule according to the user operation state when the weighted matching degree reaches the standard, collecting questionnaire filling state data in real time, and encrypting and transmitting the questionnaire filling state data; S5, decrypting the questionnaire feedback data and the behavior analysis result, counting the core indexes, analyzing the association degree of the behavior and the feedback, predicting the potential demands of the user through XGBoost models, generating personalized reports, and reversely optimizing model parameters by means of a federal learning framework to form a self-learning closed loop.
  2. 2. The AI behavior recognition-based questionnaire matching method according to claim 1, wherein in S1, the scenerized semantic requirement supports single-intent descriptions or multi-intent combination descriptions; The dimension of the hidden layer of the BERT-base model is 768, the fine adjustment round is 10 rounds, and the learning rate is high The semantic feature confidence calculation formula is: ; In the formula, The method comprises the steps of obtaining semantic feature confidence, obtaining the number of intent slots for successful matching, namely the number of elements successfully identified and matched by a system in key elements split in the semantic requirement of a user, obtaining the total number of all key elements obtained after the semantic requirement of the user is split, and obtaining model prediction probability which is the confidence probability of an AI model for processing the semantic requirement on the current semantic requirement identification result; The priority weight calculation formula is: ; In the formula, The method comprises the steps of obtaining a demand priority weight, setting a priority for a user to be the demand priority marked by the user, and obtaining a history matching success rate which is the proportion of the semantic demand successfully matched with the questionnaire in a past scene.
  3. 3. The questionnaire matching method based on AI behavior recognition as claimed in claim 2, wherein in S2, the collection dimension comprises basic behavior data, operation sequence data and environmental characteristic data, the collection frequency is 100 ms/time, an isolated forest algorithm is adopted to eliminate abnormal values, 50 isolated trees with depth limited to 10 are constructed, 200 samples are randomly sampled for each tree, the abnormal value is judged when the abnormal score S is more than or equal to 0.6, and an abnormal score calculation formula is: ; ; And (3) through Z-score standardization and unified data dimension, extracting time sequence features by adopting a sliding window method with a window size of 5s, adopting Laplace noise for differential privacy protection, and finally generating 128-dimensional structured behavior data and transmitting the 128-dimensional structured behavior data through a TLS1.3 protocol.
  4. 4. The AI behavior recognition-based questionnaire matching method of claim 3, wherein the Laplace noise formula is: ; In the formula, The number of original clicks; The privacy protection data obtained after adding noise; For the purpose of a privacy budget, ; To obey the random noise value of the laplace distribution with the position parameter of 0 and the scale parameter of 1.
  5. 5. The AI-behavior-recognition-based questionnaire matching method according to claim 4, wherein in S3, the structured behavior data is mapped into 768-dimensional behavior feature vectors through a full connection layer, a large model is input to recognize the user intention, and a confidence level is output; The dynamic weight optimization strategy updates the behavior characteristic weight once every 500 samples, and the weight updating formula is as follows: ; In the formula, Is the first The cross entropy of the class features loses the gradient, if the gradient is positive, the weight is adjusted positively; Is the first The updated weight of the class behavior feature; Is the first The current weight of the class behavior feature before updating; Is a Sigmoid function; The user-following mechanism is triggered when the intent confidence < 0.7.
  6. 6. The AI-behavior-recognition-based questionnaire matching method according to claim 5, wherein in S3, the matching degree between the semantic feature vector and the behavior feature vector is calculated by cosine similarity, and the weighted matching degree calculation formula is: ; In the formula, The degree of matching is weighted; the priority is the demand priority weight; The number of behavior feature categories for participating in the calculation; Is the first Class behavior is characterized by the current turn Dynamic weights of (2); Is the first Matching degree corresponding to the class behavior characteristics.
  7. 7. The AI-behavior-recognition-based questionnaire matching method of claim 6, wherein in S4, a dynamic threshold is used The calculation rule of (2) is: when the query rate per second is more than 10 ten thousand or the user activity is more than or equal to5 times per month, =0.6; In the early morning 00:00-06:00 or sensitive business scenario, =0.75; In the case of other scenes of the process, =0.7; The popup window occupies 1/3 of the screen height and does not shade the core operation buttons, provides a 'temporary not filling' option and a 'closing' option, wherein the valid period of the 'temporary not filling' option is 7 days, the popup window is delayed for 24 hours under the same scene, and the questionnaire filling state data is encrypted by an AES-256 encryption algorithm.
  8. 8. The AI-behavior-recognition-based questionnaire matching method of claim 7, wherein the user-following mechanism is: And (3) through the page popup single selection popup window, after the intention information supplemented by the user is collected, recalculating the matching degree.
  9. 9. A computer device comprising a processor for coupling with a memory, reading and executing instructions and/or program code in the memory to perform the method of any of claims 1-8.
  10. 10. A computer readable medium, characterized in that the computer readable medium stores computer program code which, when run on a computer, causes the computer to perform the method according to any of claims 1-8.

Description

Questionnaire matching method based on AI behavior recognition Technical Field The invention belongs to the technical field of artificial intelligent behavior analysis and online questionnaire accurate matching, and particularly relates to an questionnaire matching method based on AI behavior recognition. Background In various scenes such as e-commerce, online service, APP operation and the like, the questionnaire is a key tool for collecting user feedback and optimizing product service, and the accuracy of questionnaire matching influences the investigation efficiency and the data value. However, the prior art has the defects that specific rules need to be predefined in advance, the configuration process is complicated, matching deviation is easily caused by incomplete rule setting, user portraits are constructed in advance, the earlier-stage workload is large, data deviation is easily caused, the current real-time situation of the user cannot be accurately judged based on fixed rules or historical data, and timeliness and pertinence of matching are lacking. Thus, a new method is needed. Disclosure of Invention The invention aims to provide an investigation questionnaire matching method based on AI behavior recognition, which realizes simple rule setting, avoids earlier-stage portrait cost, and accurately fits the current scene of a user by semantic requirement description, real-time page behavior acquisition and AI large model intention analysis technology, thereby greatly improving the instantaneity and the accuracy of questionnaire matching. In order to achieve the above purpose, the invention provides an investigation questionnaire matching method based on AI behavior recognition, comprising the following steps: s1, inputting a scene semantic requirement by a user, marking a priority, converting the semantic rule into a system technical rule through a rule generation algorithm, separating words, eliminating stop words, extracting semantic feature vectors by adopting a LoRA lightweight and fine-tuned BERT-base model, calculating semantic feature confidence, and synchronously transmitting after associating priority labels; s2, pre-configuring acquisition rules, capturing user page operation data in real time through event monitoring and DOM structure analysis technology, generating structured behavior data and encrypting and transmitting the structured behavior data after data cutting, outlier rejection, standardization, time sequence feature extraction and differential privacy protection processing; S3, acquiring semantic feature vectors of the S1 and structured behavior data of the S2, combining historical event data and a current DOM structure, obtaining weighted matching degree through feature vector mapping, AI large model intention recognition, dynamic weight optimization and cosine similarity calculation, assisting questionnaire title analysis and an optional user inquiry mechanism, and screening out an optimal matching questionnaire; s4, calculating a dynamic threshold according to the service load, the user activity and the time information, triggering a popup rule according to the user operation state when the weighted matching degree reaches the standard, collecting questionnaire filling state data in real time, and encrypting and transmitting the questionnaire filling state data; S5, decrypting the questionnaire feedback data and the behavior analysis result, counting the core indexes, analyzing the association degree of the behavior and the feedback, predicting the potential demands of the user through XGBoost models, generating personalized reports, and reversely optimizing model parameters by means of a federal learning framework to form a self-learning closed loop. Preferably, in S1, the scenerified semantic requirement supports single-intent description or multi-intent combination description; The dimension of the hidden layer of the BERT-base model is 768, the fine adjustment round is 10 rounds, and the learning rate is high The semantic feature confidence calculation formula is: ; In the formula, The method comprises the steps of obtaining semantic feature confidence, obtaining the number of intent slots for successful matching, namely the number of elements successfully identified and matched by a system in key elements split in the semantic requirement of a user, obtaining the total number of all key elements obtained after the semantic requirement of the user is split, and obtaining model prediction probability which is the confidence probability of an AI model for processing the semantic requirement on the current semantic requirement identification result; The priority weight calculation formula is: ; In the formula, The method comprises the steps of obtaining a demand priority weight, setting a priority for a user to be the demand priority marked by the user, and obtaining a history matching success rate which is the proportion of the semantic demand successfully matched with the