Recommendation method and system based on user product use behavior

Abstract

The invention discloses a recommendation method and system based on a user product use behavior, and belongs to the technical field of deep learning, and the method comprises the steps: providing a software use document; storing the original text content into a database, and storing the original text content into a word vector library in a word vector form at the same time, so that a mapping relationship is established between the database and data in the word vector library; user behaviors are obtained, IO input information of a user is combined with software use logic through burying points, the user behaviors are converted into text information, and multiple pieces of sorted continuous text information are inquired in a word vector library; querying results of the word vector library are spliced with cue words and then questions are asked to the large model; and performing function mapping on the answer result of the large model and the current system software to obtain a function specified by the corresponding answer result, and returning a recommendation result to a page. According to the method, the operation behavior data of the user is collected by implementing a targeted point burying technology, and accurate prediction of the next behavior of the user can be realized in combination with the analysis capability of a large model.

Description

Technical Field

[0001] The present invention relates to the technical field of deep learning, and particularly relates to a recommendation method and system based on user product usage behavior. Background Art

[0002] Most of the existing software behavior recommendation applications rely on information on web pages or the software's own pages to recommend based on users, or infer user intentions based on the content currently copied in the clipboard by the user. There is a lack of coherent thinking in recommending the use of software, making it difficult to guess the intentions of the user's continuous behavior in the software, and the current technology lacks a feedback mechanism and does not have sufficient recognition ability for the recommended results, resulting in the inability of the recommendation ability to grow self - incrementally along with the user's usage time.

[0003] In the prior art, the invention patent with the application number CN201510012093.4 and the name of a method and device for recommending solutions based on user operation behavior includes the following content: 1. User behavior monitoring: This patent describes a method that triggers a second window object by monitoring the user's operation on a first window object; 2. Abnormal window recognition: When the second window object is an abnormal window, this method can determine the type of the abnormal window by extracting the text information in the window; 3. Intelligent recommendation of solutions: According to the type of the abnormal window and the user's operation behavior on the abnormal window, the system will recommend corresponding solutions; 4. Reduction of user operations: This method reduces the steps that users need to manually search for solutions when encountering problems through intelligent recommendation; 5. Enhancement of user experience: Through the user - friendly intelligent recommendation function, it helps users solve problems encountered during the use of the computer, thereby improving the user's usage experience.

[0004] However, the existing page recommendations have the following disadvantages: Lack of flexibility: The current web pages or software are too single - functioned in implementing the next - step recommendation and guessing on the page, and cannot make recommendations for the user's next behavior based on page data, attributes, and user operation behavior. Since different pages may have large differences, the recommendation function very much needs to be manually written, and the implemented recommendation ability is too single.

[0005] Inability to achieve self - growth of the recommendation ability: The previous recommendation pages did not add a feedback mechanism to the recommendation function, and users could not provide feedback after selection. Therefore, the recommendation function cannot achieve self - growth along with the user - recognized behavior.

[0006] Over-reliance on page data itself: In the past, the recommendation function of a page needed to make guesses based on data such as the current page where the page is located or the clipboard data, and did not incorporate data points to capture user behavior. Therefore, it relied heavily on the current page state and current page data, so the recommendation ability was restricted.

[0007] Too weak understanding ability: In the past, the recommended options were directly written by the program and did not change much, lacking the ability to understand what the user wanted to achieve. As a result, it was unable to predict the next possible actions based on user behavior.

[0008] In summary, the problems of the existing technical solutions in the page recommendation function can be summarized as follows: Lack of personalization and dynamic adaptability: The current recommendation systems show obvious limitations in the page recommendation function. They usually rely too much on preset rules and the current page state, without fully considering the personalized needs and behavior patterns of users. This has led to the singularity and lack of flexibility of the recommendation function, and it cannot be dynamically adjusted according to the user's real-time operations and preferences. In addition, due to the lack of an effective feedback mechanism, the recommendation system cannot learn from the user's choices and optimize its recommendation strategy, so it cannot achieve the self-growth of the recommendation ability.

[0009] Relying on page data and ignoring user behavior: Existing recommendation systems often rely too much on the data of the page itself, such as the current page content or clipboard information, while ignoring the importance of capturing user behavior through data point technology. This approach limits the understanding ability of the recommendation system and makes it unable to accurately predict the user's next behavior. Due to the lack of in-depth analysis of user behavior, the recommendation system cannot provide more accurate and forward-looking suggestions, thus affecting the user experience and satisfaction. Summary of the Invention

[0010] The present invention aims to solve the obvious limitations shown by the existing recommendation systems in the recommendation function, and proposes a recommendation method and system based on the user's product usage behavior. By implementing targeted data point technology, collecting the operation behavior data of users, and combining the analysis ability of large models, accurate prediction of the user's next behavior is achieved.

[0011] To achieve the above invention objective, the technical solution of the present invention is as follows: A recommendation method based on the user's product usage behavior, comprising: Providing software usage documentation, where the software usage documentation includes documentation related to software operations and some preset sample user behavior recommendation documents; Storing the original text content of the software usage documentation in a database and simultaneously storing it in a word vector library in the form of word vectors, so as to establish a mapping relationship between the data in the database and the word vector library; Obtain user behavior based on the operations of the IO input device during the user's software usage. Insert data points at important functional nodes in the front-end, back-end, or database operations of the program, and define the specific meaning of the currently triggered function according to the software usage logic. Combine the IO input information with the software usage logic through the obtained user behavior and data point operations, transform the user behavior into text information, and query the word vector library after organizing multiple consecutive text messages to obtain the query result. Use the query result of the word vector library as reference information, and combine it with the prompt words to form a complete question for querying the large model. Start the large model as an independent web service to answer the combined complete question. Through obtaining the answer result of the large model, perform function mapping with the current system software, obtain the function specified by the corresponding answer result, and return the recommended result to the page.

[0012] Furthermore, storing the original text content of the software usage document in the database and simultaneously storing it in the word vector library in the form of word vectors to establish a mapping relationship between the data in the database and the word vector library includes: cutting the plain text content according to the number of words, and sliding forward and backward during each cut to ensure data coherence; splitting the segmented text data into the word vector library through a word vector tool, and generating a unique id for the stored word vector data and returning it; after obtaining the corresponding id, store it together with the original document content data, thereby realizing the mapping storage of vectors and original data.

[0013] Furthermore, in the process of using the query result of the word vector library as reference information and combining it with the prompt words to form a complete question for querying the large model, it is necessary to combine the current software design information, set different prompt words for different modules, pages, and interfaces, and optimize the prompt word information for each different function to make it adapt to different scenarios.

[0014] Furthermore, the process of obtaining the answer result of the large model, performing function mapping with the current system software, and obtaining the function specified by the corresponding answer result includes: through the question-and-answer operation, the natural language model returns its guess of the user's next behavior. Based on multiple possible results inferred by the large model, combined with the current software design operations, obtain the corresponding operation behavior, and return this behavior as the final multiple recommended results to the page.

[0015] Further, it also includes judging whether the user selects and clicks the recommended option on the page, recording whether the user accepts the recommended result, and giving feedback: if the user clicks the recommended option, it is considered that the current recommendation is reasonable and the user agrees with the recommended result. This behavior is regarded as a high-quality recommendation. The user behavior and result are vectorized and stored in the database as reference information for the next recommendation; otherwise, the current recommendation process is regarded as a useless recommendation and abandoned without recording.

[0016] Further, it also includes performing the corresponding option operation after the user clicks the option corresponding to the recommended result on the page, tracking the user behavior, and giving feedback on the execution result: if the execution is not correct, the recommended result is regarded as an abnormal result, and feedback is given on the final result, and the stored reference information is retrieved or changed in a timely manner. This behavior is regarded as a high-quality recommendation, but the recommended result needs to be optimized; if the execution fails, it means that the user agrees with the behavior-guessing recommendation, but there is a logical problem with the recommended result, and feedback is also given on it.

[0017] The present invention also proposes a recommendation system based on user product usage behavior, including: A software usage document, which contains all information related to the software; A word vector storage module, which stores the original text content of the software usage document in the form of word vectors in a word vector library; An original data storage module, which stores the original text content in a database and establishes a mapping relationship with the word vector data; A user behavior acquisition module, which acquires user behavior according to the operations of the IO input device during the user's software usage process; A system buried point recording module, which buries points at important functional nodes of the program front and back ends or database operations, and defines the specific meaning of the currently triggered function according to the software usage logic; A record query module, which combines the captured user behavior and buried point operations, combines the IO input information with the software usage logic, converts the user's behavior into text information, and after sorting out multiple consecutive text information, queries the information in the vector library and obtains the result; A prompt word and query result splicing module, which uses the vector library query result as reference information and combines it with the prompt word to splice into a complete question for asking the large model; A large model answering module, which is started as an independent web service to provide large model question and answer functions; A result generation module, which obtains the answer result of the large model, maps it to the functions of the current system software, obtains the functions specified by the corresponding result, and returns the recommended result to the page.

[0018] Furthermore, the software user guide includes a software operation manual, the software interface architecture and hierarchical relationships, precautions for software operation, and also includes some preset sample user behavior recommendation documents.

[0019] Furthermore, it also includes a recommended result feedback module, which records whether the user accepts the recommended result by judging whether the user selects and clicks on the recommended options on the page and gives feedback.

[0020] Furthermore, it also includes an execution result feedback module, which performs the corresponding option operation after the user clicks on the option corresponding to the recommended result on the page, and gives feedback on the execution result by tracking the user's behavior.

[0021] In summary, the present invention has the following advantages: 1. The recommendation method of the present invention based on the user's product usage behavior incorporates a large model and constructs a large model recommendation feedback function based on a word vector library. It is more flexible in the recommendation method. By implementing targeted data tracking technology, collecting the user's operation behavior data, and combining the analysis ability of the large model, it can achieve accurate prediction of the user's next behavior. 2. The present invention records the user's current multiple consecutive behaviors through the user's software operation behaviors. After combining information such as the large model and software document reference content, it jointly recommends and guesses the user's next behavior, and converts the behavior into click options, quickly simplifying the multi-level operations that the user wants to perform into single-step operations, achieving fast internal software jumps and high-quality speculation of user behaviors. 3. To further improve the performance of the recommendation system, the present invention optimizes the data tracking strategy, defines the specific meaning of the currently triggered function according to the development operation logic of the software function, and at each data tracking point, collects detailed information about the user's behavior by designing rich parameters, ensuring that the captured user behavior data has high relevance and clarity. This will help the large model more accurately understand the user's operation intention and provide more accurate recommendations. 4. The present invention establishes an effective user feedback mechanism, analyzes the user's clicks and executions of the recommended results, continuously optimizes the recommended content, implements a two-way optimization strategy, helps distinguish the performance problems and data accuracy problems of the recommendation system, and thus improves the recommendation results and data quality in a targeted manner, which can improve the accuracy of recommendations and user satisfaction. Brief Description of the Drawings

[0022] Figure 1It is an implementation flow chart of the present invention. DETAILED DESCRIPTION

[0023] In order to explain the present invention more clearly, the present invention is further described below in conjunction with preferred embodiments and drawings. It should be understood by those skilled in the art that the content described below is illustrative rather than restrictive, and should not be used to limit the scope of protection of the present invention.

[0024] Example 1 The present invention provides a recommendation method based on user product usage behavior, such as Figure 1 As shown, the following steps are included: Provide software usage documentation: Software usage documentation contains all information related to the software. To cooperate with the current program operation and operation, the injected documents include software operation manual, software interface architecture and hierarchical relationship, software usage and operation precautions, and other documents related to software operation and use. It also contains some preset user behavior recommendation document samples. Software usage documentation is jointly written and maintained by product, development, and test personnel, and is updated synchronously with the current software development and design.

[0025] Word vector storage: The original text content of the software usage document is stored in the word vector library in the form of word vectors to facilitate subsequent user behavior similar content queries. In this embodiment, the implementation of converting the original text content to the word vector library mainly relies on word embedding technology in natural language processing (NLP), such as Word2Vec, GloVe, and FastText. These technologies learn the vector representation of words by training a large amount of text data, so that semantically similar words are close to each other in the vector space. Taking GloVe as an example, it learns word vectors by counting the co-occurrence information of words in the corpus. GloVe uses a global matrix to capture the co-occurrence relationship between words, taking into account the co-occurrence frequency of words in the entire corpus.

[0026] Storing original data: storing the original text content in the database and establishing a mapping relationship between it and the data in the word vector library. Since the natural language model only supports text content input and does not support vector input, storing the original data allows for more friendly access to the original text content when performing similarity queries later. During implementation, by reading the document or text content, the text content is split according to the document content or the specified data format: the plain text content is cut according to the number of words, and each cut is slid back and forth to ensure data continuity; the text data that is split into segments is stored in the word vector library through the word vector tool, and a unique id is generated for the stored word vector data and returned; after obtaining the corresponding id, it will be stored together with the original document content data, thereby realizing the mapping and preservation of the vector and the original data.

[0027] User behavior acquisition: This includes acquiring the IO input device operations of the user when using the software, and acquiring user behavior by capturing the physical IO device input status and the software application IO input status. For example, IO input device operations such as mouse sliding, mouse hovering, mouse clicking, keyboard input, and shortcut key operations when the user is using the software. Software IO input implementation requires software cooperation. Taking the browser as an example, the browser's built-in event listener can capture input content, mouse click behavior, mouse sliding trajectory, mouse selected data and other information, which will also be recorded as user usage behavior. In addition, different products will run on different devices and software, and all records that can capture relevant user behavior will be recorded by the program.

[0028] System point-of-care recording: Add a behavior capture program to the important functional nodes of the program front-end and back-end or database operations to obtain the user's real operation intention. Since the user's physical operation does not have much practical significance and the user's operation intention cannot be perceived, the present invention defines the specific meaning of the current trigger function based on the development operation logic of the software function at important nodes such as the front-end and back-end or database operations. During implementation, at each point of care, detailed information on user behavior is collected by designing rich parameters, such as which button the user clicks, the page location when clicking, the time the user's mouse stays, etc. These parameters can record the user's behavior more comprehensively, allowing the large model to better understand the user's current behavior. For example, if a user clicks in the upper right corner of a web page, the back end will capture the corresponding request, and combined with the front-end operation, the corresponding behavior will be identified as "viewing personal information", thereby determining what the user's current behavior is doing in the current software. The targeted tracking points set in this step focus on valuable user operations. These operation points are important behaviors when users interact with products, such as clicking buttons, submitting forms, copying text, hovering the mouse, etc., to ensure that the captured user operations are highly relevant and are clear rather than unknown behaviors, so that the recommendation results of the large model can accurately optimize the recommendations around the current software.

[0029] Record Query: Based on the obtained user behaviors and logging operations, transform the user behaviors into text information. After organizing multiple consecutive pieces of text information, query the vector database with this information to obtain the vector database query results, which are used to infer the user's subsequent behaviors. In this step, the IO input information is combined with the software usage logic through logging operations to transform the behaviors into text, thereby obtaining a natural language description of the current user operation behaviors. Organize the user's recent multiple consecutive operation behaviors into consecutive text information, such as: "It is currently 9 pm, the user is on the software's home page, the user copied an IP information, the user clicked on the data panel, and the user selected a data SQL query". Input this piece of text into the vector database to query the relevant information of the documents in the vector and the preset behavior records, so as to obtain the behavior most similar to the user's current behavior and the specific function information corresponding to the current behavior.

[0030] Prompt Word and Query Result Concatenation: Use the query results as reference information and concatenate them with the prompt words to form a complete question. In this step, based on the current user behavior information, obtain the adapted prompt words through the prompt word management tool, and concatenate the prompt words, the user behavior information, and the relevant reference information queried from the vector database together to form a complete question for asking the large model. By combining the design information of the current software, different modules, pages, and interfaces have different prompt words. Optimize the prompt word information for each different function to make it adaptable to different scenarios.

[0031] Large Model Answer: Start the large model as an independent web service to answer the concatenated complete question. In this step, set up an independent process to build the large model service, which is isolated from the current software application. Provide an independent large model question-and-answer service function to facilitate the independent management and deployment of the large model question-and-answer capabilities. For example, ensure that behaviors such as switching parameter information and replacing high-quality models will not affect the normal operation of the main program. The model question-and-answer in cooperation with the prompt words will output a specified result structure for subsequent function transformation.

[0032] Result generation: By obtaining the answer results of the large model, mapping them to the functions of the current system software, obtaining the functions specified by the corresponding results, and returning the recommended results to the page. In this step, through the Q&A operation, the natural language model will return its guess of the user's next behavior. For example, based on the user's current click operation and the reference document information, it is speculated that the user may want to enter a certain page to view the corresponding content. Based on multiple possible results speculated by the large model and combined with the current software design operations, the corresponding operation behaviors are obtained, and these behaviors are returned to the page as the final multiple recommended results. For example, after the user selects an IP address, a recommended list can appear on the page, speculating that the user may want to perform SQL queries, data searches, data parsing, IP jumps, etc. The user can quickly jump to the specified page to complete the specified operation according to the reasoning results, without having to perform multi-level clicks and other operations, and all operations will be completed in one step.

[0033] The present invention incorporates a large model and constructs a large model recommendation feedback function based on a word vector library, which is more flexible in the recommendation method. By implementing targeted data tracking technology, collecting the user's operation behavior data, and combining the analysis ability of the large model, it is possible to accurately predict the user's next behavior.

[0034] Embodiment 2 Based on Embodiment 1, this embodiment proposes a recommendation method based on the user's product usage behavior. The difference from Embodiment 1 is as follows: Furthermore, it also includes recommendation result feedback: By judging whether the user selects and clicks on the page recommendation option, recording whether the user accepts the recommendation result, and feeding back the result record to optimize the database information. Specifically, when the user clicks on the recommendation option, it is considered that the current recommendation is reasonable and the user agrees with the recommendation result. This behavior is regarded as a high-quality recommendation, and the user's behavior and result will be vectorized and stored in the database as reference information for the next recommendation. Conversely, the current recommendation process will be abandoned as a useless recommendation without recording.

[0035] Furthermore, it also includes execution result feedback: After the user clicks on the option and performs the corresponding option operation, such as page jump, content search, generating text content, etc., by tracking the user's behavior, feedback on the execution result is provided to ensure that the user's selection is correctly executed. Specifically, if it is not correctly executed, the recommendation result is regarded as an abnormal result. For example, behaviors such as recommending to jump to a non-existent page require feedback on the final result, and the stored reference information needs to be retrieved or changed in a timely manner. This behavior is regarded as a high-quality recommendation, but the recommendation result needs to be optimized; if the execution fails, it means that the user agrees with the behavior guess recommendation, but there is a logical problem with the recommendation result, and feedback is also required.

[0036] In this embodiment, the feedback content is precisely optimized through user clicks and actual execution result feedback to ensure that the data is more accurate in the reference result. Feedback can be used to confirm whether the recommendation ability is insufficient or there are data errors in the final recommendation result. If the user selects the recommended result, it means that the recommendation result meets expectations. For example, in the background record, the program embedding point will record the user's next operation after selecting the recommendation option. If the user clicks on the jump and immediately cancels the jump behavior, chooses to accept the continuation of the content of the large model and immediately deletes the continuation content, similar to the above operations, the program determines that the recommendation ability of the large model is insufficient, that is, the recommendation direction meets the expected results, but lacks accuracy. If there is a problem with clicking to execute, it means that there are certain data errors in the expected results, and the results need to be optimized. For example, the user selects the large model recommendation, but the subsequent execution result of the recommendation fails or an error is considered to be an error.

[0037] The present invention establishes an effective user feedback mechanism, analyzes users' clicks and executions on recommendation results, continuously optimizes recommended content, and implements a two-way optimization strategy. This helps to distinguish performance issues and data accuracy issues of the recommendation system, thereby improving recommendation results and data quality in a targeted manner, thereby improving recommendation accuracy and user satisfaction.

[0038] Example 3 The present invention provides a recommendation system based on user product usage behavior, including software usage documents, a word vector storage module, an original data storage module, a user usage behavior acquisition module, a system embedding recording module, a record query module, a prompt word and query result splicing module, a large model answer module, and a result generation module.

[0039] Among them, the software usage documentation contains all information related to the software. To cooperate with the current program operation and operation, the injected documents include all documents related to the software operation and use, such as the software operation manual, the software interface architecture and hierarchical relationship, and the software usage and operation precautions. It also contains some preset user behavior recommendation document samples. The software usage documentation is jointly written and maintained by product, development, and test personnel, and is updated synchronously with the current software development and design.

[0040] The word vector storage module stores the original text content in the form of word vectors into the word vector library to facilitate subsequent user behavior similar content queries. In this embodiment, the implementation of text content to word vector library mainly relies on word embedding technology in natural language processing (NLP), such as Word2Vec, GloVe and FastText. These technologies learn the vector representation of words by training a large amount of text data, so that semantically similar words are close to each other in the vector space. Taking GloVe as an example, it learns word vectors by counting the co-occurrence information of words in the corpus. GloVe uses a global matrix to capture the co-occurrence relationship between words, taking into account the co-occurrence frequency of words in the entire corpus. This functional module can flexibly switch to use different word vector technologies according to the actual scenario, converting text content into word vectors for easy query.

[0041] The original data storage module stores the original text content into the database and establishes a one-to-one correspondence with the word vector data. The original data storage module enables the original text content to be obtained more friendly when performing similarity queries later. Specifically, the original data storage module reads the document or text content and splits the file content according to the document content or the specified data format: the plain text content is cut according to the number of words, and each cut will slide back and forth to ensure the continuity of the data; the text data split into segments will be stored in the word vector library through the word vector tool, and a unique id will be generated for the stored word vector data and returned. After obtaining the corresponding id, it will be stored together with the original document content data, thereby realizing the mapping and preservation of the vector and the original data.

[0042] The user behavior acquisition module acquires the IO input device operations such as mouse sliding, mouse hovering, mouse clicking, keyboard input, shortcut key operation, etc. during the user's use of the software, and acquires the user behavior by capturing the physical IO device input status and the software application IO input status. The implementation of software IO input requires the cooperation of software. Taking the browser as an example, the browser's built-in event listener can capture input content, mouse click behavior, mouse sliding trajectory, mouse selection data and other information, which will also be recorded as user usage behavior. In addition, different products will run on different devices and software, and all records that can capture relevant user behavior will be recorded by the program.

[0043] The system logging module captures the real operation intentions of users by adding behavior capture programs at important functional nodes such as the front and back ends of the program or database operations. Specifically, at important nodes such as the front and back ends or database operations, the system logging module defines the specific meaning of the currently triggered function according to the development operation logic of the software function. For example, when a user clicks on the upper right corner of a web page, the back end will capture the corresponding request, and in combination with the front-end operation, this behavior will be identified as "viewing personal information", so as to determine what specific operation the user is performing in the current software.

[0044] The record query module transforms the user's behavior into text information through the captured user behavior and logging operations. After sorting out multiple consecutive text messages, it queries the vector library with this information to obtain the vector library query result, so as to infer the user's next behavior. Specifically, the record query module combines the IO input information with the software usage logic through logging operations, transforms the behavior into text, and obtains the natural language description of the current user operation behavior. Specifically, the user's nearly multiple consecutive operation behaviors are sorted into consecutive text messages, such as: "It is 9 pm now, the user is on the home page of the software, the user copied an IP information, the user clicked on the data panel, and the user selected a data SQL query". This piece of text is poured into the word vector library to query the relevant information of the document in the vector and the preset behavior record, so as to obtain the behavior most similar to the user's current behavior and the specific function information corresponding to the current behavior.

[0045] The prompt word and query result splicing module uses the query result as reference information and splices it with the prompt word to form a complete question. Specifically, the prompt word and query result splicing module obtains the adapted prompt word through the current user behavior information and in combination with the prompt word management tool, and jointly splices the prompt word, the user behavior information, and the relevant reference information queried from the vector library into a complete question for asking the large model. During implementation, by combining the design information of the current software, different prompt words are set for different modules, pages, and interfaces to optimize the prompt word information for each different function and adapt to different scenarios.

[0046] The large model answer module is started as an independent web service to provide the large model question and answer function. By setting up an independent process, a large model service is built, which is isolated from the current software application. An independent large model question and answer service function is provided to facilitate the independent management and deployment of the large model question and answer capabilities. For example, ensuring that behaviors such as switching parameter information and replacing high-quality models will not affect the normal operation of the main program. The model question and answer in cooperation with the prompt word will output a specified result structure for subsequent function transformation.

[0047] The result generation module maps the answer results of the large model to the current system software to obtain the functions specified by the corresponding results. Specifically, through the Q&A operation, the natural language model returns its guess of the user's next behavior. For example, based on the user's current click operation and the reference document information, it is speculated that the user may want to enter a certain page to view the corresponding content. According to the multiple possible results speculated by the large model and combined with the current software design operations, the corresponding operation behaviors are obtained, and these behaviors are returned as the final multiple recommended results to the page. For example, after the user selects an IP address, a recommended list can appear on the page, speculating that the user may want to perform SQL queries, data searches, data parsing, IP jumps, etc. The user can quickly jump to the specified page to complete the specified operation according to the inference results, without having to perform multi-level clicks and other operations, and all operations will be completed in one step.

[0048] Embodiment 4 Based on Embodiment 3, the present invention provides a recommendation system based on user product usage behavior, which is different from Embodiment 3 in that: Furthermore, it also includes a recommended result feedback module. The behavior feedback management module records whether the user accepts the recommended result by judging whether the user selects and clicks on the recommended option, and feeds back the result record to optimize the database information. Specifically, when the user clicks on the recommended option, it is considered that the current recommendation is reasonable and the user agrees with the recommended result. This behavior is regarded as a high-quality recommendation, and the user's behavior and result are vectorized and stored in the database as reference information for the next recommendation; otherwise, the current recommendation process is abandoned as a useless recommendation.

[0049] Even further, it also includes an execution result feedback module. The page jump module performs the corresponding option operations after the user clicks on the option, such as page jumps, content searches, generating text content, etc. By tracking the user's behavior, the execution results are fed back to ensure that the user's selection is correctly executed. Specifically, if it is not correctly executed, the recommended result is regarded as an abnormal result. For example, behaviors such as recommending to jump to a non-existent page require feedback on the final result again, and the stored reference information is retrieved or changed in a timely manner. This behavior is regarded as a high-quality recommendation but the recommended result needs to be optimized; if the execution fails, it means that the user agrees with the behavior guess recommendation, but there is a logical problem with the recommended result, and it is also fed back.

[0050] In addition, as a further optimization, this system can also implement functions such as quick page jumps, selection after page jumps, automatic content input on the page, quick page turning, automatic clicks, automatic text filling, page content searches, etc. within the software, and can also implement operations such as helping to write input text content, quick web page jumps, explanation of selected content, translation of selected content, AI automatic text correction, etc. outside the software.

[0051] The above are only the preferred embodiments of the present invention, and do not impose any form of limitation on the present invention. Any simple modification or equivalent change made to the above embodiments based on the technical essence of the present invention shall fall within the protection scope of the present invention.

Claims

1. A recommendation method based on user product usage behavior, characterized in that: include: Providing software usage documentation, which includes documents related to software operation and some preset user behavior recommendation document samples; The original text content of the software usage document is stored in the database, and at the same time stored in the word vector library in the form of word vectors, so that a mapping relationship is established between the database and the data in the word vector library; Obtain user behavior based on IO input device operations during the use of the software, embed points on important functional nodes of the program front-end and back-end or database operations, and define the specific meaning of the current trigger function based on the software usage logic; By acquiring user behaviors and embedded operations, the IO input information is combined with the software usage logic, the user behaviors are converted into text information, and multiple continuous text information are sorted and queried in the word vector library to obtain the query results; The query results of the word vector library are used as reference information and combined with the prompt words to form a complete question for asking questions to the large model; The large model is launched as an independent web service to answer the complete questions after splicing; By obtaining the answer results of the large model, functional mapping is performed with the current system software, the functions specified by the corresponding answer results are obtained, and the recommended results are returned to the page.

2. A recommendation method based on user product usage behavior as claimed in claim 1, characterized in that: The original text content of the software usage document is stored in the database, and at the same time stored in the word vector library in the form of word vectors, so that a mapping relationship is established between the database and the data in the word vector library, including: cutting the plain text content according to the number of words, sliding back and forth each time cutting to ensure the continuity of the data; the text data split into segments is stored in the word vector library through the word vector tool, and a unique id is generated for the stored word vector data and returned; after obtaining the corresponding id, it will be stored together with the original document content data, thereby realizing the mapping and preservation of the vector and the original data.

3. The recommendation method based on user product usage behavior according to claim 1, characterized in that: The query results of the word vector library are used as reference information and combined with prompt words to form a complete question for asking questions to the large model. It is necessary to combine the current software design information, set different prompt words for different modules, pages, and interfaces, and optimize the prompt word information for each different function to adapt it to different scenarios.

4. A recommendation method based on user product usage behavior as claimed in claim 1, characterized in that: The method of obtaining the answer result of the large model and mapping the function with the current system software to obtain the function specified by the corresponding answer result includes: through the question-and-answer operation, the natural language model returns its guess of the user's next behavior, and based on the multiple possible results inferred by the large model and the current software design operation, the corresponding operation behavior is obtained, and the behavior is returned to the page as the final multiple recommendation results.

5. A recommendation method based on user product usage behavior as claimed in claim 1, characterized in that: It also includes judging whether the user has selected the recommended option on the click page, recording whether the user accepts the recommended result, and providing feedback: if the user clicks the recommended option, it means that the current recommendation is reasonable and agrees with the recommended result. This behavior is regarded as a high-quality recommendation, and the user behavior and result are vectorized as high-quality results and stored in the database for reference information for the next recommendation; Otherwise, the current recommendation process will be discarded as useless recommendation.

6. A recommendation method based on user product usage behavior as claimed in claim 5, characterized in that: It also includes executing the corresponding option operation after the user clicks the option corresponding to the recommended result on the page, and providing feedback on the execution result by tracking the user behavior: if the execution is not correct, the recommendation result will be regarded as an abnormal result, and feedback will be given on the final result. The stored reference information will be promptly retrieved or changed, and the behavior will be regarded as a high-quality recommendation but the recommendation result needs to be optimized; if the execution fails, it means that the user agrees with the behavioral guess recommendation, but there is a logical problem in the recommendation result, and feedback will also be given on it.

7. A recommendation system based on user product usage behavior, characterized in that: include: Software usage documentation, which contains all information related to the software; A word vector storage module, which stores the original text content of the software usage document in the form of word vectors into a word vector library; The original data storage module stores the original text content in the database and establishes a mapping relationship between it and the word vector data; A user behavior acquisition module, wherein the user behavior acquisition module acquires user behavior according to the IO input device operation of the user during the use of the software; A system tracking recording module, which tracks points on important functional nodes of the program front-end and back-end or database operations, and defines the specific meaning of the current trigger function based on the software usage logic; A record query module, which combines IO input information with software usage logic through captured user behaviors and embedded operations, converts user behaviors into text information, and after collating multiple continuous text information, performs vector library query on the information and obtains results; A prompt word and query result splicing module, which uses the vector library query result as reference information and combines the prompt word to splice into a complete question for asking questions to the large model; A large model answering module, which is started as an independent web service and provides a large model question and answering function; The result generation module obtains the answer result of the large model, performs function mapping with the current system software, obtains the function specified by the corresponding result, and returns the recommended result to the page.

8. A recommendation system based on user product usage behavior as claimed in claim 7, characterized in that: The software usage documentation includes the software operation manual, software interface architecture and hierarchical relationship, software usage precautions, and also includes some preset user behavior recommendation document samples.

9. A recommendation system based on user product usage behavior as claimed in claim 7, characterized in that: It also includes a recommendation result feedback module, which determines whether the user selects the recommended option on the click page, records whether the user accepts the recommendation result, and provides feedback.

10. A recommendation system based on user product usage behavior as claimed in claim 7, characterized in that: It also includes an execution result feedback module, which executes the corresponding option operation after the user clicks the option corresponding to the page recommendation result, and provides feedback on the execution result by tracking the user behavior.

Images