SEO keyword intelligent analysis system based on edge AI chip

Abstract

The application relates to the technical field of information retrieval, in particular to an SEO keyword intelligent analysis system based on an edge AI chip, which comprises three core modules: a data acquisition module that acquires user behaviors, webpage contents and device environment data in real time; an edge computing module that is internally provided with an AI chip, processes data in a local device, generates user preference features, and generates optimized queries in combination with search keywords; and a search result processing module that sends the optimized queries to a search engine and processes the results individually, wherein the application sinks AI computing power to terminal devices, constructs a closed loop system of real-time acquisition, local intelligent processing and personalized result enhancement, and solves the long-term contradiction between personalized service and user privacy protection. Sensitive data is processed locally without being uploaded to the cloud, thereby eliminating the risk of privacy leakage from the source; the edge AI chip realizes real-time search individualization with extremely low delay and high precision, and significantly improves the intelligent level of the search engine and the user satisfaction.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of information retrieval, in particular to an SEO keyword intelligent analysis system based on edge AI chips. BACKGROUND

[0002] With the explosive growth of Internet information, search engines have become the core portal for users to access information. Search engine optimization (SEO) strategies and personalized search results are crucial for improving user experience. Current personalized search services are mainly implemented based on cloud servers. The typical approach is: the cloud server continuously collects a large amount of user search history, click behavior, geographic location, etc. data, builds user portraits through complex algorithms, and optimizes queries or reorders search results using the portrait when the user initiates a new search.

[0003] However, this centralized processing mode based on the cloud has several inherent defects:

[0004] First, high risk of privacy leakage: sensitive behavior data and preference models of users are stored in cloud servers, facing the risk of unauthorized access, leakage or abuse, making users hesitant to use personalized functions.

[0005] Second, network delay exists: each user interaction data needs to be uploaded to the cloud for processing and feedback, and in poor network conditions, it will bring perceptible delay, affecting the smoothness and real-time of search experience.

[0006] Third, personalized feedback lags: cloud model updates are often not real-time, unable to learn from the user's just-ended browsing behavior and adjust the strategy for the next search immediately, resulting in inaccurate and timely personalized recommendations.

[0007] In addition, some existing local browser plugins can record user behavior, but lack strong local computing power, cannot perform complex real-time artificial intelligence analysis and modeling, and have relatively single functions, making it difficult to achieve deep and truly real-time search personalization.

[0008] In view of this, an SEO keyword intelligent analysis system based on edge AI chips is proposed. SUMMARY

[0009] The purpose of the present application is to provide an SEO keyword intelligent analysis system based on edge AI chips to provide highly personalized search experience and fundamentally protect user privacy and reduce delay.

[0010] To solve the above technical problems, the present application provides an SEO keyword intelligent analysis system based on edge AI chips, comprising:

[0011] a data collection module configured to collect behavior data, webpage content data, and device environment data in real time during user interaction with a search engine;

[0012] an edge computing module including an AI chip, connected to the data collection module, configured to perform real-time processing and analysis of the collected data on a local device, generate user preference features, and generate an optimized search query based on a user input search keyword and the user preference features;

[0013] a search result processing module connected to the edge computing module, configured to send the optimized search query to a search engine and perform personalized processing of search results returned by the search engine.

[0014] As a further improvement of the technical solution, the data collection module includes:

[0015] a user behavior collection unit configured to capture user click operations, page dwell time, scrolling operations, and mouse tracks on a search results page through a browser plug-in;

[0016] a webpage content collection unit configured to obtain title text, body content, and meta tag information of a user access page through a client-side document object model parsing technology;

[0017] an environment data collection unit configured to obtain latitude and longitude coordinates and system time stamps of a user device through a device application program interface, and obtain search trend data of a region where the user is located through a network interface.

[0018] As a further improvement of the technical solution, the edge computing module includes:

[0019] a data preprocessing unit configured to perform cleaning, normalization, and feature extraction operations on collected raw data;

[0020] a user preference modeling unit configured to run based on an AI chip, construct a user preference model through real-time incremental learning of an online learning algorithm, and output a user preference feature vector through the user preference model, wherein the user preference model is a multi-layer neural network model;

[0021] a query optimization unit configured to perform fusion calculation on a user input search keyword and a user preference feature vector when detecting a user submission of a search request, and output an optimized search query containing user preference information.

[0022] As a further improvement of the technical solution, the neural network model implemented by the user preference modeling unit implements user preference learning through the following ways:

[0023] extracting associated features of page content and search queries in user historical click records;

[0024] analyzing the distribution pattern of the user's stay time on different types of web pages;

[0025] learning the user's preference degree for the website domain name or content source with historical access frequency higher than a preset threshold;

[0026] Each dimension of the user preference feature vector corresponds to a specific preference category.

[0027] As a further improvement of the technical solution, the query optimization unit generates an optimized search query by the following steps:

[0028] performing lexical analysis and semantic analysis on the search keywords input by the user;

[0029] calculating the semantic similarity between the search keywords input by the user and each preference dimension in the user preference feature vector, and selecting the preference dimension with the highest similarity score;

[0030] combining the semantic label corresponding to the preference dimension with the original search keywords to generate an expanded query;

[0031] Generating multiple candidate optimized queries and selecting them based on the user preference model.

[0032] As a further improvement of the technical solution, the search result processing module includes:

[0033] A query agent unit is configured to intercept the original query request sent by the user to the search engine, and after obtaining the user's authorization, replace the original query request with an optimized search query and send it to the search engine;

[0034] A result reordering unit is configured to calculate the relevance of the search results returned by the search engine based on the user preference features;

[0035] A visual enhancement unit is configured to highlight the search results that meet the user's preferences in the user interface.

[0036] As a further improvement of the technical solution, the result reordering unit processes the search results by the following way:

[0037] Calculate the matching score of each search result with the user preference features;

[0038] According to the matching score, the search result list is reordered;

[0039] Retain the integrity of the original search results, and only adjust the display order of the search results.

[0040] As a further improvement of the technical solution, it also includes:

[0041] The cloud edge collaborative updating module is connected with the edge computing module, and is used for realizing model parameter synchronization of a user preference model and knowledge base updating between the edge computing module and a cloud server.

[0042] The cloud edge collaborative updating module periodically receives the model parameters of the anonymized user preference model from the edge computing module, and transmits the model parameters to the cloud server in an encrypted manner.

[0043] The cloud edge collaborative updating module also receives the updated model parameters and global knowledge data issued from the cloud server, and delivers the model parameters and the global knowledge data to the edge computing module for use.

[0044] The model parameters include neural network weights and bias terms.

[0045] Compared with the prior art, the present application has the following beneficial effects:

[0046] 1. In the SEO keyword intelligent analysis system based on the edge AI chip, the AI computing capability is sunk to the user terminal device (edge side), and a closed-loop system integrating real-time data acquisition, local intelligent processing and personalized result enhancement is constructed, so that the long-term contradiction between "personalized service" and "user privacy protection" is fundamentally solved.

[0047] 2. In the SEO keyword intelligent analysis system based on the edge AI chip, the system completes the processing of all sensitive data and user preference modeling on the local device, without uploading the original behavior data to the cloud, so as to eliminate the risk of privacy leakage at the source, and at the same time, by virtue of the powerful computing power of the edge AI chip, the real-time search personalization with extremely low delay and high precision is realized, and the intelligent level of the search engine and the user satisfaction are significantly improved. BRIEF DESCRIPTION OF DRAWINGS

[0048] Figure 1 The figure is a whole system block diagram of the present application. DETAILED DESCRIPTION

[0049] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application. EMBODIMENT

[0050] Currently, the search engine personalization service of the prior art completely relies on a cloud server to process user data, and a typical practice is as follows: the cloud server continuously collects a large amount of search history, click behavior, geographic location and other data of a large number of users, constructs a user portrait through a complex algorithm, and when a user initiates a new search, the portrait is used to optimize the query or reorder the search results, but this practice also has problems such as high privacy leakage risk, network delay and personalized feedback lag;

[0051] In view of this, with reference to Figure 1 The purpose of the present application is to provide an SEO keyword intelligent analysis system based on an edge AI chip, which provides a three-layer architecture distributed solution, the core of which is to sink the computing power to the user terminal device, including:

[0052] A data collection module is configured to collect behavior data, webpage content data and device environment data generated in the process of user interaction with a search engine in real time; for example:

[0053] On a Windows / mac OS personal computer, the implementation is achieved through a Chrome browser extension and a local background daemon (Daemon), the browser extension is responsible for capturing user behavior in the browser, and the daemon is responsible for calling operating system API to obtain device environment data, after the data is collected, the collected data is serialized into Protocol Buffers format through an inter-process communication (IPC) mechanism, and written into a shared memory area, so as to be quickly read by an edge computing module;

[0054] An edge computing module is connected to the data collection module, and is configured to perform real-time processing and analysis on the collected data on the local device, generate user preference features, and generate an optimized search query according to the user input search keyword and the user preference features; for example:

[0055] A C++ dynamic link library compiled locally is used to directly call the underlying interface provided by the hardware vendor to drive the NPU (neural network processor). The library contains a monitoring thread that continuously polls the above-mentioned shared memory area, and once new data is detected to be written, the NPU driver is immediately called to load the data into the memory of the NPU for model inference calculation, and the whole process is completed locally without the need for network connection;

[0056] A search result processing module is connected to the edge computing module, and is configured to send the optimized search query to the search engine and perform personalized processing on the search results returned by the search engine; for example:

[0057] The local proxy server and the browser extension are cooperated to achieve the local proxy server is written in Go language, and a local port (such as 8080 port) is listened to, the browser extension modifies the proxy settings (or configures the PAC script) of the browser, and redirects all requests to the preset search engine domain name (such as www.google.com, www.bing.com) to the local port (such as 127.0.0.1:8080), the local proxy server communicates with the edge computing module after intercepting the search request, obtains the optimized query word, replaces the parameters in the original request, and then forwards the request to the real search engine. Subsequently, it parses and personalizes the page content returned by the search engine, and then returns it to the browser for rendering and display;

[0058] Through the above technical means, a complete and efficient localized personalized search closed loop is constructed, all original data processing related to user privacy is completed inside the user equipment, high-speed communication is realized through shared memory and local loop network, the risk of data leakage is eliminated from the source, the user privacy is effectively protected, at the same time, the special computing capacity of the local NPU is utilized, the delay of personalized processing is reduced, the real-time search personalization is realized without awareness, and the search efficiency and satisfaction of the user are improved. Embodiments

[0059] In order to construct an accurate user preference model, it is necessary to comprehensively, accurately and low-power collect multi-dimensional original data on the local device, including how the user interacts with the search results, what content they view and the environment they are in, while ensuring that the collection process does not seriously affect the device performance and the normal operation of the browser. Therefore, the data collection module includes:

[0060] The user behavior collection unit captures the user's click operation, page dwell time, scrolling operation and mouse trajectory on the search result page through the browser plug-in; specifically:

[0061] The unit is realized through a browser extension, the extension uses a page navigation complete event listener to detect the timing of the search result page loading completion, then injects a content script into the page through the script injection (chrome.scripting.execute Script) method, the injected script collects user behavior by listening to page click (records the X Path path and timestamp of the event.target element), mouse movement (samples mouse coordinates at a frequency of 10 times per second), scrolling (records the value of performance.now() when the page is loaded and the user leaves the page respectively and calculates the difference), and other events, all collected data is sent to the background script of the extension for temporary storage and arrangement;

[0062] The webpage content collection unit obtains the title text, the body content and the meta tag information of the page accessed by the user through the client document object model analysis technology; specifically:

[0063] The unit is also integrated in the browser extension. When the user clicks on the search result and navigates to the target webpage, the extension detects the completion of the loading of the target webpage and injects a special content script into the target webpage. The script obtains the page title (by calling document.title) and the meta description information (by document.query Selector('meta[name="description"]')?.content). For the body content, the script uses a special content extraction library (a copy of the lightweight and modified Readability.js library) to automatically analyze the page structure, identify and extract the core article content.

[0064] The environmental data collection unit obtains the latitude and longitude coordinates of the user's device and the system timestamp through the device application programming interface, and obtains the search trend data of the area where the user is located through the network interface; specifically:

[0065] The implementation of the unit is related to the device platform. On the mobile platform, the latitude and longitude coordinates are obtained by calling the location service interface (requesting GPS or network location update); the system timestamp is obtained through the system interface. The regional search trend data is obtained by initiating a network request to the application programming interface provided by the search engine to obtain popular keyword data; on the PC platform, the location information may need to be manually input by the user or roughly located through the IP address;

[0066] Through the above technical means, a complete and workable local data collection scheme is provided. By fully utilizing the capabilities of browser extensions and operating system interfaces, multi-dimensional and fine-grained data necessary for building a user preference model are accurately captured. All operations are completed locally, laying a solid data foundation for subsequent real-time analysis, while firmly protecting the privacy of user data;

[0067] In embodiment 3, the collected raw data is unstructured and disordered, and cannot be directly used for machine learning models. At the same time, on terminal devices with limited computing resources, memory and power consumption, efficient data cleaning, conversion and running of a complex user preference model are required to complete real-time inference and query optimization, which is a great technical challenge. Therefore, the edge computing module includes:

[0068] The data preprocessing unit is used for cleaning, normalizing and feature extraction of the collected raw data; specifically:

[0069] After reading the raw data from the shared memory, the unit performs the following standardization process:

[0070] Data cleaning: remove useless characters in the text using regular expression matching; according to the business rules, invalid data is removed.

[0071] Data normalization: for numerical features, the minimum-maximum normalization method is used to linearly transform them to a fixed interval.

[0072] Feature extraction: for text content, a pre-trained lightweight neural network model is used to convert text strings into fixed-length feature vectors, which well represent the semantic information of the text.

[0073] The user preference modeling unit is based on an AI chip and uses an online learning algorithm to incrementally learn in real time to build a user preference model, and outputs a user preference feature vector through the user preference model. The user preference model is a multi-layer neural network model. Specifically:

[0074] This unit is a multi-layer perceptron model, and its structure is defined as: input layer (768 dimensions), first fully connected layer (256 neurons, using Re LU activation function), Dropout layer (dropout rate is 0.5), output layer (10 neurons, using Softmax activation function). The model is built using a mobile machine learning framework and is deployed on an NPU for inference through a dedicated hardware acceleration mechanism. The model is trained using an online learning mode, using momentum stochastic gradient descent as the optimizer, with a learning rate of 0.01 and a momentum parameter of 0.9. The loss function uses categorical cross-entropy, and the model weights are updated once every N (e.g. N=100) new sample batches, enabling continuous incremental learning.

[0075] The query optimization unit, upon detecting that the user has submitted a search request, fuses and calculates the user input search keywords with the user preference feature vector, and outputs an optimized search query containing user preference information. Specifically:

[0076] The unit is a Python script running on the CPU of the device, and exchanges data with the preference modeling unit in the NPU through shared memory. When a user submits a search request, the following process is immediately started: first, use a segmentation library (such as jieba.lcut(query)) to segment and tag the parts of speech of the original keywords input by the user, and keep the nouns (n) and verbs (v) as the core stems; then, read the latest output of the user preference feature vector (such as 10 dimensions) from the shared memory; finally, generate a set of candidate optimized search queries through a predefined fusion function (for example, concatenate the word vectors of the core stems with the preference feature vector, and then input them into a small linear classifier), and select the one with the highest confidence as the final output;

[0077] Through the above technical means, using dedicated AI hardware acceleration, model lightweight technology (quantization), and efficient online learning algorithm, the complex neural network model is successfully run on resource-constrained terminal devices in a low-power and high-performance manner, which realizes large-scale data processing and real-time model inference that can only be done on the cloud in the past, and provides core technical support for truly real-time and adaptive search personalization.

[0078] In embodiment 4, since simple keyword matching cannot deeply understand the real interests and preferences of users, how to make the machine learning model automatically learn the deep and semantic preference patterns from the user's behavior, so as to generate a digital portrait (feature vector) that can accurately represent the user's interests, is a core problem to be solved. Therefore, the neural network model implemented by the user preference modeling unit learns user preferences in the following ways:

[0079] Extract the association features of page content and search queries in the user's historical click records;

[0080] Analyze the user's stay time distribution patterns on different types of web pages;

[0081] Learn the user's preference degree for website domain names or content sources with a historical access frequency higher than a preset threshold;

[0082] Each dimension of the user preference feature vector corresponds to a specific preference category. For example, the neural network model learns in the following multiple mechanisms, specifically:

[0083] Correlation feature learning: During the model training phase, the system automatically constructs training sample pairs. For example, take (user search term, user final clicked page title) as a positive sample pair. The model adopts a double tower structure (Siamese Network), one tower processes search terms, and the other tower processes page titles and abstracts. Training is performed by maximizing the cosine similarity of the two vectors in the latent space in the positive sample pair, so that the model learns to map semantically similar searches and clicks to similar vector spaces. For example, when a user searches for "apple" multiple times and clicks on a page about "iPhone", the model will automatically learn the strong semantic correlation between "apple" and "iPhone", rather than the correlation with "fruit";

[0084] Temporal distribution pattern learning: The model inputs the user's stay time t on each category page into the fully connected layer as an important continuous numerical feature. The model will automatically learn the weight of this feature through training. For example, the model will find that there is a strong positive correlation between long stay on technical document type pages and high output scores of the user's "programming" preference dimension, so that this feature is given more attention during prediction;

[0085] Source preference learning: During data preprocessing, the system will parse the URL string into its top-level domain name (for example, extract github.com from https: / / github.com / path / to / repo). This domain name string will be input into an embedding layer of the model as a category feature. The embedding layer maps each domain name to a low-dimensional, dense floating-point number vector (for example, 8-dimensional). During training, the model adjusts the parameters of the embedding layer and the subsequent network, so that when the input feature contains a vector representing github.com, the value corresponding to the "technology" or "programming" preference dimension in the final output vector will be significantly increased;

[0086] Preference vector output: The output layer of this neural network model is precisely configured with 10 neurons, each corresponding to a pre-defined preference category (for example: neuron 0: technology, neuron 1: finance, neuron 2: sports,..., neuron 9: short video). The activation value (after Softmax conversion) of each neuron represents the model's prediction of the user's instantaneous preference intensity for that category. These 10 output values together form a 10-dimensional user preference feature vector, which is a highly condensed and interpretable digital portrait of user interest;

[0087] By the above technical means, the model can surpass the surface keyword matching, and mine the abstract and deep interest preferences from the user's behavior sequence. The generated preference feature vector is a precise and efficient semantic representation, which provides a strong and reliable basis for subsequent query optimization and search result reordering, and greatly improves the accuracy of personalization.

[0088] In embodiment 5, since the search keywords input by the user are often short and ambiguous. How to convert the internal abstract and numerical user preference vector into a specific and effective query word that the search engine can understand and process, so as to significantly improve the relevance of the search results at the first query, is the goal of the present embodiment. Therefore, the query optimization unit generates an optimized search query by the following steps:

[0089] Step one, lexical analysis and semantic analysis: lexical analysis and semantic analysis are performed on the search keywords input by the user; specifically: using the locally integrated jieba word segmentation library to process the original keywords input by the user; calling the function jieba.lcut_for_search(query) to perform search engine mode word segmentation, and obtaining a keyword list. Then, call jieba.posseg.lcut(query) to perform part-of-speech tagging, and keep the words with core word properties such as nouns (n) and verbs (v) as core stems representing the user's search intent;

[0090] Step two, preference dimension matching: calculate the semantic similarity between the search keywords input by the user and each preference dimension in the user preference feature vector, and select the preference dimension with the highest similarity score; specifically: convert one or more core stems obtained in the first step and 10 preference category labels (such as "technology", "programming", "sports") into word vectors. Here, a pre-trained 300-dimensional Word2Vec word vector model downloaded locally can be used. Calculate the cosine similarity between each core stem vector and each category label vector, and the calculation formula is: Then, sort all similarity scores and select the preference category label with the highest score (for example, input "Python", and the similarity with the "programming" label is the highest);

[0091] Step three, query expansion: combine the semantic label corresponding to the preference dimension with the original search keyword to generate an expanded query; specifically: using a rule-based strategy, combine the semantic label corresponding to the matched preference dimension with the original search keyword to generate a new, more information-rich expanded query. The combination strategy includes but is not limited to: direct addition (such as "Python" + "programming" = "Python programming"), addition of site restrictions (such as "Python" + "site:github.com" = "Python site:github.com"), or addition of file types (such as "Python" + "filetype:pdf");

[0092] Step four, ranking selection: generate multiple candidate optimization queries and select based on the user preference model, specifically: the system generates 3-5 candidate optimization queries according to the preset rules. Then, each candidate query is simulated as a "virtual" search request and input into the local user preference model for rapid inference. The model outputs a prediction value representing the probability of the user clicking on the search results returned by the candidate query. Finally, the candidate query with the highest predicted click probability is selected as the final optimization search query to be sent to the search engine;

[0093] Through the above steps, the user's short-term search intent (keyword) and long-term interest preference (preference vector) can be effectively fused at the semantic level, generating a more informative and intent-specific search query that can guide the search engine to return a more relevant and user-desired result list from the source.

[0094] Embodiment 6, considering how to seamlessly take over the search process without violating the user's existing search habits and ensuring the user's informed consent, send the optimized query to the search engine, and perform personalized enhancement processing on the returned results, finally present to the user in an intuitive form, therefore, the search result processing module includes:

[0095] The query agent unit is configured to intercept the original query request sent by the user to the search engine, and after obtaining the user's authorization, replace the original query request with the optimized search query and send it to the search engine; specifically:

[0096] The unit is implemented as a browser extension. The extension uses the chrome.declarativeNetRequest API (Manifest V3) to intercept network requests. The extension declares the search engine URL patterns to be intercepted (such as *: / / www.google.com / search*) in the pre-rules list. When the request is intercepted, the extension's background script triggers a popup page to clearly explain the "smart search optimization" function to the user and request authorization. Only after the user clicks "Agree" to authorize, the extension will use the chrome.declarativeNetRequest.updateDynamicRules API to dynamically modify the request URL, replacing the original query parameters with optimized query parameters, and then releasing the request;

[0097] The result reordering unit is used to calculate the relevance of search results returned by the search engine based on user preference features. Specifically, this unit is integrated into the local proxy server. After the proxy server receives the HTML response returned by the search engine, it uses the goquery library (similar to jQuery) in Go language to parse the HTML document. By calling goquery.NewDocumentFromResponse(resp) and doc.Find("div.g").Each(...), it traverses and extracts each search result item (div.g is the container selector of the Google search result page). For each result item, further extract its title (h3 tag text), URL (a tag href attribute), and abstract (div.st tag text);

[0098] The visualization enhancement unit is used to highlight the search results that meet the user's preferences in the user interface. Specifically:

[0099] The unit then works with the result reordering unit. For each search result item extracted, the title and abstract text are sent to the local user preference model for relevance scoring. Then, according to the scoring results, all search result items are sorted in descending order. Subsequently, using goquery to manipulate DOM, first clear the content of the original search result container #rso, then insert the result items in the new order. At the same time, add a custom CSS class, such as highlight-result, to the div.g elements of the top 3 result items with the highest relevance, and insert inline CSS styles in the returned HTML document section: <style>.highlight-result { border-left: 3px solid #4CAF50; padding-left: 10px;}< / style> , thus achieving green highlight display of the preferred results;

[0100] Through the above technical means, the whole process of query optimization and result enhancement is automatically and non-intrusively completed under the premise of fully respecting the user's right to know and choice; finally, the most valuable information is presented to the user in a more intuitive and efficient way, greatly improving the user's information acquisition efficiency and search satisfaction.

[0101] In view of the fact that the integrity and diversity of the results must be ensured when the search results are individually reordered, and the "information cocoon" is avoided due to excessive individualization, so that the user misses the new information that may be important but the model temporarily considers irrelevant, therefore, the result reordering unit processes the search results in the following way:

[0102] Calculate the matching degree score of each search result with the user preference feature; specifically:

[0103] For each original search result item returned by the search engine (i.e. each div.g element), concatenate its title text and abstract text into a complete string S; input the string S into the local user preference model to perform preprocessing and NPU inference process, and the model outputs a 10-dimensional preference vector , which represents the content attributes of the search result itself; the matching degree score is obtained by calculating the dot product (Dot Product) of the vector and the current user preference vector , and the calculation formula is: This dot product value reflects the overall fit between the search result content and the user's personal interest, and the higher the value, the stronger the relevance;

[0104] According to the matching degree score, the search result list is reordered; specifically:

[0105] The system creates an empty list to store the tuple (search_result_dom_element, ). Traverse all search result items, calculate the of each and store it in the list. Then, use the built-in sorting algorithm (such as Go's sort.Slice function) to sort the list in descending order according to ;

[0106] The integrity of the original search results is preserved, and only the display order of the search results is adjusted; specifically: in the final output HTML, the original DOM elements (without any content modification) of each search result item are rendered to the page in strict accordance with the sorted order. Do not delete any result, do not modify any internal HTML content of any result (title, link, abstract remain the same as returned by the search engine), only change their arrangement order;

[0107] Through the above steps, while providing powerful personalized ranking functions, the integrity and neutrality of the search engine's returned results are strictly guaranteed. All the information seen by the user is unaltered, and the system only prioritizes according to its personal preferences, effectively avoiding the negative effects of the "information cocoon", balancing the needs of personalization and information diversity.

[0108] Embodiment 8, considering that the local models distributed on millions of terminal devices are isolated and cannot learn evolution from global data. How to enable these models to absorb collective wisdom and continuously optimize performance while firmly preventing any raw data from leaving the user device and protecting privacy is a key challenge. Therefore, the system further comprises:

[0109] A cloud-edge collaborative updating module connected to the edge computing module, for realizing model parameter synchronization and knowledge base updating of the user preference model between the edge computing module and the cloud server;

[0110] The cloud-edge collaborative updating module regularly receives the model parameters of the anonymized user preference model from the edge computing module, and transmits the model parameters to the cloud server in an encrypted manner; for example:

[0111] Edge-side uploading: the updating service on the device triggers once a week as a Cron Job, which calls related interfaces such as TFLite's Interpreter::GetTensor() to extract the parameters of all trainable layers in the current local user preference model, including the weight matrix W and the bias vector b; then, using the EVP_EncryptInit_ex series of functions of the OpenSSL library, the AES-256-GCM encryption algorithm is used to encrypt the parameters (W, b) using a device-unique key generated by a hardware security module (such as TPM) or secure enclave at the time of device manufacturing or first activation. The encrypted data block is uploaded to a specific endpoint (such as https: / / api.example.com / model-updates) of the cloud server through an HTTPS POST request (with TLS certificate verification), and the request header never contains any information that can directly identify the user's identity (such as IMEI, username);

[0112] Cloud processing: the cloud server runs the Federated Averaging (FedAvg) algorithm. The server waits until it collects a sufficient number (e.g., 10,000 devices) of encrypted model updates, and then decrypts them in a secure environment. The aggregation process is: first, the weight matrix W of all devices is averaged, and then the bias vector b is averaged. The sum is then divided by the total number of devices N to obtain the averaged global weight The same operation is performed on the bias vector b. This new and fuses the collective knowledge of numerous user devices;

[0113] The cloud-edge collaborative update module also receives the updated model parameters and global knowledge data issued from the cloud server and delivers them to the edge computing module for use; for example:

[0114] Cloud-side issuance: the cloud aggregates the global parameters and into a file, uses the cloud's private key to digitally sign the file to ensure its integrity and trusted source, and then publishes the signed file to a content distribution network (CDN); the update service on the device regularly (e.g., daily) queries a fixed URL on the CDN to check if there is a new version of the model file (by comparing the version number or file hash value);

[0115] Edge-side update: after the device downloads the new global model file, it first uses the pre-built cloud public key in the system to verify the digital signature; after verification, the device calls the TFLite Interpreter::SetTensor() or Interpreter::CopyToBuffer() interface to overwrite the corresponding parameters in the local model with the new global parameters and , thus completing a model update;

[0116] Through the above technical means, a secure and privacy-protected distributed machine learning ecosystem is established. Under the premise of absolutely ensuring that the original data of any single user never leaves the device, the sharing of global user knowledge and the collaborative evolution of model performance are realized. This makes the local model of each device continuously and securely become more "intelligent" and accurate, achieving the dual goals of privacy protection and model performance improvement.

[0117] The above shows and describes the basic principles, main features and advantages of the present application. Those skilled in the art should understand that the present application is not limited to the above embodiments, and the above embodiments and descriptions in the specification are only preferred examples of the present application and are not intended to limit the present application. Without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application, and these changes and improvements all fall within the scope of the claimed present application. The scope of protection of the present application is defined by the appended claims and their equivalents.

Claims

1. An SEO keyword intelligent analysis system based on an edge AI chip, characterized in that, The application relates to a search engine optimization method and system based on edge computing and cloud computing. The application comprises the following steps: A data acquisition module is used to collect behavior data, webpage content data and device environment data generated in the process of user interaction with a search engine in real time; An edge computing module is connected with the data acquisition module and comprises an AI chip, which is used to process and analyze the collected data in real time on a local device, generate user preference features, and generate an optimized search query according to a search keyword input by a user and the user preference features; The application comprises the following steps: a C++ dynamic link library is compiled locally to call a bottom interface provided by a hardware manufacturer to drive an NPU; the dynamic link library comprises a monitoring thread which is used to continuously poll a shared memory area and detect new data written into the shared memory area, and the NPU driver is immediately called to load the data into the memory of the NPU for model inference calculation when new data is detected; A search result processing module is connected with the edge computing module and is used to send the optimized search query to a search engine and perform personalized processing on search results returned by the search engine; The application comprises the following steps: a local proxy server and a browser extension are used to work cooperatively, the local proxy server is programmed by using a Go language and listens to a local port, and the browser extension modifies the proxy settings of a browser to replace parameters in an original request with optimized query words when the browser sends a search request to a preset search engine domain name; the local proxy server communicates with the edge computing module to obtain the optimized query words, and then forwards the request to a real search engine; A cloud-edge collaborative updating module is connected with the edge computing module and is used to synchronize model parameters of a user preference model and update a knowledge base between the edge computing module and a cloud server; The cloud-edge collaborative updating module receives model parameters of an anonymized user preference model from the edge computing module at regular time intervals, and transmits the model parameters to the cloud server in an encrypted manner; The cloud-edge collaborative updating module also receives updated model parameters and global knowledge data from the cloud server and delivers the model parameters and the global knowledge data to the edge computing module for use; 2.The edge-AI chip-based SEO keyword intelligent analysis system according to claim 1, characterized in that, The model parameters comprise neural network weights and bias terms. The data acquisition module comprises the following units: A user behavior acquisition unit is used to capture a click operation, a page staying time, a scroll operation and a mouse track of a user on a search result page through a browser plug-in; A webpage content acquisition unit is used to obtain title text, body content and meta tag information of a webpage accessed by a user through a client document object model parsing technology; 3.The edge-AI chip-based SEO keyword intelligent analysis system according to claim 1, characterized in that, An environment data acquisition unit is used to obtain latitude and longitude coordinates and a system timestamp of a user device through a device application program interface and obtain search trend data of a region where the user is located through a network interface. The edge computing module comprises the following units: A data preprocessing unit is used to perform cleaning, normalization and feature extraction operations on collected original data; A user preference modeling unit is based on an AI chip and is used to construct a user preference model through real-time incremental learning of an online learning algorithm and output a user preference feature vector through the user preference model, wherein the user preference model is a multilayer neural network model. The query optimization unit fuses and calculates the search keywords input by the user and the user preference feature vector to output an optimized search query containing user preference information when detecting that the user submits a search request. 4.The edge-AI-chip-based SEO keyword intelligent analysis system according to claim 3, characterized in that: The neural network model implemented by the user preference modeling unit learns user preferences in the following ways: Extracting the association features of page content and search queries in user historical click records; Analyzing the dwell time distribution patterns of users on different types of web pages; Learning the preference degree of users for website domains or content sources with a historical access frequency higher than a preset threshold; Each dimension of the user preference feature vector corresponds to a specific preference category. 5.The edge-AI chip based SEO keyword intelligent analysis system according to claim 3, characterized in that, The query optimization unit generates an optimized search query in the following steps: Performing lexical analysis and semantic analysis on the search keywords input by the user; Calculating the semantic similarity between the search keywords input by the user and each preference dimension in the user preference feature vector, and selecting the preference dimension with the highest similarity score; Combining the semantic label corresponding to the preference dimension with the original search keywords to generate an expanded query; Generating multiple candidate optimized queries and selecting them based on the user preference model. 6.The edge-AI-chip-based SEO keyword intelligent analysis system according to claim 1, characterized in that: The search result processing module includes: The query agent unit intercepts the original query request sent by the user to the search engine and, after obtaining user authorization, replaces the original query request with an optimized search query and sends it to the search engine; The result reordering unit calculates the relevance of the search results returned by the search engine based on the user preference features; The visual enhancement unit highlights the search results that meet the user's preferences in the user interface. 7.The edge-AI-chip-based SEO keyword intelligent analysis system according to claim 6, characterized in that: The result reordering unit processes the search results in the following ways: Calculating the matching degree score of each search result with the user preference features; Reordering the search result list according to the matching degree score; Preserving the integrity of the original search results and only adjusting the display order of the search results.

Images