Content recommendation method based on cross-device behavior pattern analysis and dynamic portraits

Abstract

The invention discloses a content recommendation method based on cross-device behavior pattern analysis and dynamic portraits, and relates to the field of content recommendation. The invention aims to solve the defects of a traditional recommendation system in a personalized and cross-device scene. By introducing a cross-device behavior pattern analysis technology and combining device fingerprints (51Degrees or a custom rule set) and user agent character string analysis, behavior characteristics on different devices (mobile phones, tablet computers and computers) are comprehensively captured, the limitation of single device data is broken through, and a comprehensive user portrait is constructed. In the aspect of content processing, a multi-topic model (LDA and LSA) and an adaptive feature extraction algorithm are combined, potential semantics are dynamically mined from a high-dimensional text, and a multi-dimensional article portrait is formed. Furthermore, through cross-dimensional similarity calculation (a cosine similarity and collaborative filtering hybrid model), user historical behaviors and equipment preferences are integrated, a dynamic feedback mechanism is introduced, a recommendation strategy is optimized in real time, and recommendation accuracy and diversity are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of content recommendation, and in particular to a content recommendation method based on cross-device behavior pattern analysis and dynamic profiling. Background Art

[0002] With the rapid development of Internet technology, the amount and diversity of digital content has increased dramatically. How to effectively recommend appropriate content to users has become a key issue. Traditional recommendation systems mainly rely on keyword matching and basic user preference analysis to provide search results. These technologies rely on complex algorithms and large-scale data indexing to quickly extract relevant content from massive information.

[0003] With the continuous advancement of artificial intelligence and machine learning technologies, the market demand for more intelligent and personalized recommendation systems is growing, including the use of advanced data analysis technology to deeply understand user behavior and preferences, and provide more customized recommendation results based on this information. Data is generally managed by a content management system (CMS, a software application used to manage and publish web content). In addition, based on the trend of users searching on multiple devices, the demand for technologies that can provide a consistent and high-quality search experience across devices is also growing.

[0004] Although traditional recommendation systems are very efficient in processing large amounts of data, they still have the following limitations in personalization and deep understanding of user needs: (1) Lack of in-depth understanding of user behavior: The keyword matching method analyzes the keywords in the user's search query and then searches for documents or items containing these keywords in the content library. This method is simple and direct, but it often fails to understand the user's deep intentions and context, and lacks sufficient understanding and analysis of the user's implicit needs, such as browsing habits and device usage preferences. It fails to fully utilize the behavioral data generated by users on multiple devices, and thus cannot fully capture the user's behavior patterns and preferences; (2) Insufficient personalized recommendations: Collaborative filtering technology makes recommendations by analyzing the similarities between users or items. User-based collaborative filtering recommends items by finding other users with similar preferences to the target user, while item-based collaborative filtering recommends new items that are similar to items that the user has previously liked. Content-based recommendation methods recommend new items that are similar to items that the user has previously liked by analyzing the attributes and content (text, images, videos, etc.) of the items. It usually involves feature extraction and similarity calculation. These recommendation systems attempt to achieve personalized recommendations by tracking the user's search history, but this method is often limited to the user's historical behavior and fails to fully consider the dynamic changes in user interests. It is usually based on simple statistics or matching algorithms and cannot deeply explore the user's potential needs and interests; (3) Lack of cross-device recommendation capabilities: The hybrid recommendation system combines the above methods and aims to improve the accuracy and coverage of recommendations by integrating different recommendation technologies. However, as users increasingly search and browse on multiple devices, traditional recommendation systems often fail to provide a consistent and high-quality recommendation experience across different devices. They fail to fully utilize cross-device behavior data to optimize recommendation algorithms, making it impossible to achieve personalized recommendations across devices. Compared with the existing technology, the present invention can not only provide more intelligent, accurate and personalized recommendation services, but also solve the problems of insufficient personalization and shallow understanding of user needs in traditional recommendation systems, bringing users a new and high-quality content recommendation experience. In addition, the present invention pays special attention to the analysis of cross-device behavior patterns, effectively improving the consistency and coherence of recommendation experience between different devices, and meeting the actual needs of modern users for multi-device use. Summary of the invention

[0005] In view of the deficiencies in the prior art, the present invention provides a content recommendation method based on cross-device behavior pattern analysis and dynamic profiling, aiming to solve the limitations of traditional recommendation systems in personalization and in-depth understanding of user needs, especially in terms of user behavior understanding, personalized recommendations, and cross-device consistency. The present invention is implemented through the following technical solutions, and the specific implementation steps are as follows: S1: Cross-device behavior pattern analysis. Collect user behavior data on different devices, including search history, click behavior, page dwell time, browsing depth and speed, etc., and identify and record the type of device used by the user and its corresponding operating system and browser information. De-duplicate, filter outliers and standardize the collected user behavior data to ensure data accuracy and consistency; S2: Item feature extraction. Using natural language processing technology, the TF-IDF algorithm is used to extract keywords from the text content; the topic models LDA (Latent Dirichlet Allocation) and LSA (Latent Semantic Analysis) are used to perform topic analysis on the text content to identify the main topics of the article; machine learning methods are used to evaluate the sentiment of the text, including positive, negative or neutral. Information on various items is obtained from the content management system, features are extracted from the metadata of the items, and metadata of the items is collected, including category labels, author information, release time, update time, content length, etc. A labeling system is constructed based on the categories and attributes of the items, and each item is labeled accordingly for subsequent classification and retrieval. The text information of the items is preprocessed, including removing special characters, stop words, etc., and performing word segmentation to ensure the cleanliness and structuring of the text data; S3: User portrait construction and update. Integrate all features obtained from content analysis and attribute analysis to form a complete item feature vector. Construct item portraits based on feature vectors, including basic information, themes, and emotions of the items. Based on the analysis results of user behavior data, create multi-dimensional user portraits that include user device preferences, content interest levels, information consumption patterns, and search activity time patterns. Use long short-term memory networks (LSTM) to analyze users' web browsing depth and speed, and infer users' attention to content and reading habits. Use the K-means clustering algorithm to comprehensively identify and extract users' personalized preferences. When an item is edited or a new comment is added, update the item's feature vector accordingly; design a dynamic learning and optimization mechanism to update user portraits in real time to ensure the accuracy and timeliness of item portraits to adapt to dynamic changes in user preferences and behaviors; S4: Personalized recommendation algorithm. Based on detailed user portraits and item portraits, the similarity between the two is calculated to select items that are highly consistent with user preferences as the preliminary candidate set. Collaborative filtering technology is used to predict the scores that users may give to each item in the candidate set, and the candidate content is prioritized according to the predicted scores to provide users with recommended content that best suits their interests and preferences. In order to improve the diversity and novelty of recommended content, the system will eliminate items in the candidate pool that are repeated or whose similarity exceeds the preset threshold, effectively avoiding the over-concentration of recommendation results. In addition, through in-depth analysis of candidate content and user historical behavior patterns, the candidate list is further streamlined to exclude options that do not match user interests. The system also has dynamic response capabilities, and can instantly adjust the presentation order of recommended content based on the user's real-time interactive data to ensure that the recommendation results are closely aligned with the user's instantaneous demand changes. Finally, the system designs a closed-loop feedback mechanism that automatically adjusts the key parameters of the recommendation algorithm by continuously collecting and analyzing the user's actual response to the recommended content, so as to continuously iterate and optimize the performance of the recommendation system and improve the user experience; S5: Evaluation and optimization. A multi-dimensional performance indicator system is used for comprehensive evaluation, and the key indicators include accuracy, recall, and coverage. Accuracy measures the proportion of items in the recommendation list that are actually of interest to users, reflecting the accuracy of the recommendation; recall focuses on the proportion of items of interest that the system can successfully recommend to users among all potential items of interest, reflecting the comprehensiveness of the recommendation; and coverage evaluates the proportion of the range of items that the recommendation system can reach relative to the entire item library, emphasizing the breadth of the system. Based on the comprehensive evaluation results of the above indicators, the team will tune and iterate the recommendation algorithm model in a targeted manner. Specific measures may involve fine-tuning of algorithm parameters, feature engineering optimization, model architecture improvement, etc. The goal is to significantly improve the accuracy of recommendations and user experience, while enhancing the overall performance of the system.

[0006] Preferably, in said S1, it also includes: Cross-device data integration collects user behavior data on different devices, including smartphones, tablets, and desktop computers, and uses this data to obtain comprehensive online behavior patterns of users. These data cover the user's search history, click behavior, page dwell time, browsing depth and speed, etc., so as to build a comprehensive view of user behavior. Device preference identification can identify the user's device preferences by analyzing the user's usage frequency and habits on different devices, which is crucial for providing consistency and personalized experience across devices. Using deep learning technologies such as long short-term memory networks (LSTM), real-time monitoring and analysis of dynamic changes in user behavior can be used to adjust recommendation strategies in a timely manner to adapt to changes in user behavior.

[0007] Preferably, in said S2, further comprising: The TF-IDF algorithm is used to extract keywords from the text, and the topic model (LDA / LSA) is used for topic analysis to identify the main topics of the article, and the sentiment tendency of the text is evaluated through machine learning methods. Enhancing the understanding of the content of the item helps to match user interests more accurately. Category labels, author information, release time and other content are extracted from the metadata of the item, and a label system is built according to the category and attributes of the item to facilitate classification and retrieval. All features are integrated to form a complete item feature vector, which supports the construction of a detailed item portrait.

[0008] Preferably, in said S3, further comprising: Multi-device behavior characteristics, based on the frequency and habits of users on different devices, clarify their device preferences. Evaluate the user's interest in various types of content through statistical analysis methods and divide the interest levels; analyze the user's browsing speed, dwell time and reading completeness to determine the information consumption pattern; record the user's activity level at different time periods and summarize the time pattern of search activities. Used to continuously monitor the depth and speed changes of users browsing web pages. Once a significant change is found (from fast browsing to deep reading), adjust the user's inference of content attention and reading habits in a timely manner. Real-time adaptability ensures that the recommended content always fits the user's latest interests. Use regularly to group users with similar behavioral characteristics. When the user's behavior pattern changes, resulting in a change in the cluster group to which it belongs, update its personalized preferences accordingly. A real-time monitoring mechanism is designed to immediately trigger the update of the user portrait when the user has new interests or changes in behavior patterns.

[0009] Preferably, in said S4, further comprising: Cross-device consistency: In the recommendation algorithm, considering that users may switch between multiple devices, the system can instantly adjust the presentation order of recommended content based on the user's real-time interaction data to ensure that users can obtain recommendation results that closely match their instantaneous demand changes regardless of the device they are on. The cosine similarity algorithm is used to quantify the similarity between the two, and items that highly match user preferences are selected as candidate sets. Collaborative filtering technology is used to predict the ratings that users may give to candidate content, and the ranking is based on the ratings to ensure the quality and relevance of the recommended content. A mechanism for identifying and eliminating duplicate and highly similar content is implemented to improve the diversity and novelty of recommended content and avoid over-concentration of recommended results.

[0010] Compared with the prior art, the present invention discloses a content recommendation method based on cross-device behavior pattern analysis and dynamic portrait. The present invention has the following beneficial effects: 1. Enhanced understanding of user behavior: Through cross-device behavior pattern analysis, we can more deeply understand the user's deep intentions and context, including browsing habits and device usage preferences, so as to provide recommendations that better meet user needs; 2. Improve the quality of personalized recommendations: The construction of dynamic user portraits and the application of personalized recommendation algorithms can capture the dynamic changes of user interests and provide more accurate personalized recommendations, surpassing the reliance of traditional recommendation systems on user historical behavior; 3. Improve cross-device recommendation experience: It can use cross-device behavior data to optimize the recommendation algorithm, provide a consistent and high-quality recommendation experience across different devices, promote data integration and collaboration across multiple devices, and solve the shortcomings of traditional recommendation systems in cross-device recommendation capabilities; 4. Optimize resource allocation and content matching: Through more accurate user behavior analysis and recommendation, suitable content can be recommended to users more effectively, resource allocation can be optimized, and the efficiency and effectiveness of content matching can be improved; 5. Enhance user experience and satisfaction: By providing more customized and accurate recommendations, the present invention can enhance user experience and satisfaction, thereby potentially increasing user stickiness and the commercial value of the platform. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments are briefly introduced below. It should be understood that the following drawings only illustrate certain embodiments of the present invention and should not be regarded as limiting the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative work: Figure 1 : Data collection and feature extraction flow chart, showing the complete process from user behavior data collection to item feature extraction; Figure 2 : The portrait construction and recommendation flowchart shows the complete process from user portrait construction and updating to personalized recommendation algorithm. DETAILED DESCRIPTION

[0012] Step 1: Cross-device behavior pattern analysis S11 embeds specific event triggers through front-end code to capture users' search history information, click behavior, page dwell time, etc. in real time. Use JavaScript or SDK in mobile applications, combined with device fingerprint (51Degrees or custom rule set) and user agent string parsing (User-Agent) to accurately identify device type, operating system and browser information. Data obtained from different sources (website logs, API interfaces, database queries) needs to be uniformly formatted to ensure data consistency and integrity; S12 uses hash tables or other deduplication algorithms to deduplicate the collected user search history, click behavior and other data. It uses rule-based or machine learning filtering models to filter out obviously abnormal or invalid data, extremely short residence time, unreasonable high-frequency searches, etc. It uses data standardization tools and algorithms to convert user behavior data into a unified format and metric standard for subsequent analysis and processing; S13 efficiently obtains multiple information resources of items from the content management system through standardized data interfaces, covering various media forms such as text, images, and videos. At the same time, the system automatically collects metadata of items, including category tags, author details, first publication time, last update time, content length and other key attributes. With the help of text extraction technology, the system can accurately parse and extract the core text content of items, titles, descriptions, tags and other basic information, providing high-quality data support for subsequent content processing, classification, indexing and recommendation. In the preprocessing stage of the text information of the items, S14 uses regular expressions and professional text processing libraries to perform a series of sophisticated data cleaning operations, including removing special characters, filtering stop words, etc., and then performs efficient word segmentation to ensure the purity and structured level of the text data. At the same time, using data cleaning and conversion technology, the metadata of the items is comprehensively sorted and standardized to ensure the consistency of the data format, the reliability of the data quality, and the availability of the data content.

[0013] Step 2: Item feature extraction In the text content processing and analysis module of S21, the system first uses the TF-IDF (Term Frequency-Inverse Document Frequency) algorithm in natural language processing technology to segment the text, count the word frequency, and calculate the inverse document frequency to determine the keywords in the text. Next, the topic models LDA (Latent Dirichlet Allocation) and LSA (Latent Semantic Analysis) are used to train the preprocessed text data. Through the learning process of the model, the potential topic distribution of the text is obtained, and then the main topic areas of the article are identified. In addition, the sentiment analysis model in the field of machine learning is applied, and the text data is input into the model to evaluate its sentiment tendency and judge whether the sentiment color of the text content is positive, negative or neutral. The TF-IDF value is composed of two parts: term frequency (Term Frequency, TF) and inverse document frequency (Inverse Document Frequency, IDF); (1) Term frequency (TF): Term frequency refers to the number of times a word appears in a document. In order to prevent certain high-frequency words from dominating, a standardized method is used to calculate the TF value: TF(t,d) = number of times word t appears in document d / total number of words in document d (2) Inverse Document Frequency (IDF): Inverse Document Frequency is used to measure the importance of a word. It is inversely proportional to the number of documents containing the word. That is, the fewer documents contain the word, the more important the word is. The calculation formula of IDF is as follows:

[0014] Where N is the total number of documents, nt is the number of documents containing word tt; (3) TF-IDF value: The final TF-IDF value can be calculated using the following formula:

[0015] S22 extracts features such as category tags, author information, publishing time, content length, etc. from the metadata of items, and analyzes the author information, including the style of their past works and audience feedback; it builds a label system based on the categories and attributes of items, and labels each item accordingly to facilitate subsequent classification and retrieval; S23 aggregates and integrates keywords, themes, emotions, metadata and other features, and uses vector splicing or feature fusion technology to form a complete feature vector of the item. Based on the integrated feature vector, an item portrait is constructed, which includes the basic information, theme, emotion and other features of the item; S24 monitors the editing status of items and the addition of new comments in real time. When there are changes, it re-executes some of the above steps to update the characteristics and portraits of the items.

[0016] Step 3: User portrait construction and update S31 uses statistical analysis methods to evaluate users' preferences for various devices, and uses machine learning algorithms to accurately evaluate users' interest in different content, analyze users' browsing speed and dwell time, and determine information consumption patterns; S32 uses a long short-term memory network (LSTM) to analyze the user's browsing depth and speed and predict changes in the user's attention to content. The LSTM model can continuously learn and monitor the evolution of user browsing behavior. Once a significant change in user behavior is detected, it will immediately adjust the prediction of the user's content attention and reading habits to ensure that the recommended content is consistent with the user's latest interests. The TensorFlow or PyTorch framework can be used during the training process. Among them, the structure of the LSTM unit consists of three core components: the forget gate, the input gate, and the output gate. The specific implementation is as follows: A. Forget gate: responsible for deciding which information to discard from the cell state. Its calculation formula is:

[0017] Among them, Wf is the weight matrix of the forget gate, bf is the bias term, σ represents the sigmoid activation function, ht−1 represents the hidden state at the previous moment, and xt is the input data at the current moment; B. Input gate: Determines what new information should be added to the cell state. This process is divided into two steps: First, the input gate is used to determine which information needs to be updated, and secondly, new candidate cell states are generated. The calculation formulas are: ,

[0018] Among them, Wi and Wc are the weight matrices of the input gate and candidate cell state, bi and bc are the corresponding bias terms, and tanh represents the hyperbolic tangent activation function; Cell state update: Combine the results of the forget gate and the input gate to update the cell state. The calculation formula is:

[0019] C. Output gate: determines what information is output from the cell state. The calculation formula is:

[0020] The final hidden state ht is obtained by activating the cell state Ct through tanh and then multiplying it by the output gate:

[0021] Among them, Wo is the weight matrix of the output gate, and bo is the bias term; S33 regularly uses K-means clustering analysis to classify users with similar behavioral characteristics into corresponding groups. The clustering task can be completed with the help of the scikit-learn library. When the user's behavior pattern changes, resulting in a change in the cluster group to which he belongs, the system will promptly update his personalized preference settings to adapt to the dynamic changes in user interests; S34 In the user behavior dynamic modeling and user portrait update module, the system inputs the user's browsing behavior data, including page dwell time, browsing depth, browsing speed, etc., as the input sequence XtXt into the long short-term memory network (LSTM) model. By training the LSTM model, it can effectively capture the dynamic change pattern of user behavior over time. Based on the prediction results of the model, the system can update the user's digital portrait in a timely manner to ensure the timeliness and accuracy of the user portrait; S35 sets up listeners to monitor users' real-time interactive behaviors. Once new interests or behavioral pattern changes are detected, the user profile update is triggered immediately. Asynchronous communication is achieved through the WebSocket protocol or message queue (RabbitMQ). An automated feedback loop is designed to fine-tune the parameters in the user profile based on the user's immediate response to ensure that it always reflects the latest preferences and behaviors.

[0022] Step 4: Personalized recommendation algorithm In order to improve the accuracy and relevance of personalized recommendations, S41 uses user portrait and item portrait vectorization technology. For user portraits, by comprehensively analyzing multi-dimensional information such as user historical behavior, preferences and social attributes, vectorization processing is used to simplify the data structure, laying the foundation for subsequent similarity calculations. On this basis, the cosine similarity algorithm is used to measure the similarity between user portrait vectors and item portrait vectors, aiming to quantify the correlation between the two through mathematical methods and provide users with more personalized recommendation content. The relevant formula is as follows: (1) User portraits and item portraits are converted into vector form For user portrait, assuming there are n features, represented as U1, U2, ..., Un, the user portrait vector U can be expressed as: U = [U1, U2, ..., Un] For the object image, assuming there are m features, represented as I1, I2, ..., In, the object image vector I can be expressed as: I = [I1, I2, ..., In] Cosine similarity, for the user portrait vector U and the item portrait vector I, the cosine similarity cosine(U,I)cosine(U,I) is defined as:

[0023] Where: UI represents the dot product of vector U and vector I. ||U|| and ||I|| represent the modulus lengths of vector U and vector I respectively.

[0024] S42: For the items in the candidate pool, obtain the rating data of other similar users on these items. Use a matrix decomposition-based or model-based collaborative filtering algorithm to predict the target user's rating of the candidate items; and preliminarily sort the candidate content according to the predicted ratings; S43 uses hashing algorithms or other fast comparison methods to remove duplicate content from the candidate pool. It analyzes the semantic relevance of candidate content and the user's past behavior, uses natural language processing technology to extract key features for comparison, and filters out content that is too similar or has low relevance to the user's past behavior; S44 monitors the real-time interactive behavior of users in real time, and dynamically adjusts the ranking of candidate content using an online learning algorithm or a real-time update mechanism according to the real-time behavior of users; S45 adjusts the recommendation strategy in a timely manner based on the analysis of user behavior changes and personal preferences, increases the recommendation weight of new areas of interest, or adjusts the recommendation ratio of different types of content. It collects explicit feedback from users on recommended content, analyzes feedback data, and uses optimization algorithms such as gradient descent to adjust the parameters of the recommendation algorithm.

[0025] Step 5: Evaluate and optimize S51 uses three key indicators, namely, accuracy, recall and coverage, to evaluate the effectiveness of the recommendation system. Accuracy refers to the proportion of items in the recommendation list that actually match the user's interests (i.e., are truly relevant), reflecting the accuracy of the recommendation results. Recall refers to the proportion of truly relevant items in the recommendation list to the total number of items related to the user's interests. This indicator is mainly used to measure the system's ability to discover the user's potential interests. The coverage indicator focuses on the breadth of the recommendation system, that is, the proportion of the range of items that the system can recommend to users to the total project library, which is directly related to the diversity of recommended content and the potential for exploring unknown areas; During the performance optimization and evaluation of the recommendation system, S52 first collects multi-dimensional feedback information from users on the recommended content, including explicit feedback (ratings and comments given by users) and implicit feedback (click behavior, browsing time, etc.). At the same time, the output results of the recommendation system are recorded in detail, including key information such as the list of items recommended each time and the time of recommendation; Based on the data collected above, the system calculates core evaluation indicators such as accuracy, recall, and coverage through the following methods to comprehensively measure the effectiveness of the recommendation system: A. Accuracy: It is defined as the ratio of the number of truly relevant items in the recommendation list to the total number of items in the recommendation list, that is, accuracy = number of relevant items recommended / total number of items recommended; B. Recall rate: It is defined as the ratio of the number of truly relevant items in the recommendation list to the total number of all actually relevant items, that is, recall rate = number of recommended relevant items / total number of relevant items C. Coverage rate: It is defined as the ratio of the number of different types of items that the recommendation system can recommend to the total number of all different types of items in the system, that is: coverage rate = number of recommended types of items / total number of types of items.

[0026] S53 performs grid search or random search on the key parameters of the recommendation algorithm (learning rate, regularization coefficient) to find the optimal configuration. Constantly try to add new features or improve the quality of existing features to improve the recommendation effect. Explore more complex model structures - deep neural networks to further improve recommendation accuracy; S54 designs an integrated solution for content-based recommendation systems based on the architecture and business processes of existing recommendation systems. Ensure that the new system can seamlessly connect with the existing system and share data and functions. Develop necessary interfaces and APIs to achieve data exchange and function calls between the new system and the existing system. Test the new system to ensure system stability and accuracy. Optimize and adjust based on test results to improve system performance and user experience.

Claims

1. A content recommendation method based on cross-device behavior pattern analysis and dynamic profiling, characterized in that: The method comprises the following steps: S1: Cross-device behavior pattern analysis, collecting user behavior data, and identifying and recording the type of device used by the user and its corresponding operating system and browser information; S2: Item feature extraction: extract keywords from text content using natural language processing technology, perform topic analysis using topic models, evaluate the sentiment of text, and extract features from item metadata; S3: User portrait construction and update: Based on the analysis results of user behavior data, a multi-dimensional user portrait is created. The long short-term memory network is used to analyze the user's web browsing depth and speed, and the K-means clustering algorithm is used to identify and extract the user's personalized preferences. S4: Personalized recommendation algorithm, which is built based on user portraits and item portraits. It calculates the similarity between the two to select items that are highly consistent with user preferences as the initial candidate set. It uses collaborative filtering technology to predict the ratings that users may give to each item in the candidate set, and prioritizes the candidate content based on the predicted ratings. S5: Evaluation and optimization, using a multi-dimensional performance indicator system for comprehensive evaluation. Key indicators include accuracy, recall, and coverage. Based on the comprehensive evaluation results of the above indicators, the recommendation algorithm model is tuned and iterated.

2. The method according to claim 1, characterized in that The S1 further comprises: Integrate data across devices to obtain comprehensive online behavior patterns of users; Device preference identification: analyzing the user's usage frequency and habits on different devices and identifying the user's device preferences.

3. The method according to claim 1, characterized in that The S2 further includes: Use the TF-IDF algorithm to extract keywords from the text, use the topic model LDA / LSA to perform topic analysis, identify the main topics of the article, and evaluate the sentiment tendency of the text; Extract metadata of items and build a labeling system to facilitate classification and retrieval.

4. The method according to claim 1, characterized in that: The S3 further includes: Identify users’ device preferences, assess users’ interest in various content categories, determine information consumption patterns, and summarize the time patterns of search activities; Use LSTM to analyze the user's browsing depth and speed, and infer the user's attention to content and reading habits; Regular cluster analysis is performed to group users with similar behavioral characteristics, and when users’ behavioral patterns change, their personalized preferences are updated.

5. The method according to claim 1, characterized in that The S4 further comprises: Quantify the similarity between the user portrait vector and the item portrait vector, predict the target user's rating of the candidate items, and preliminarily sort the candidate content; Eliminate duplicate or overly similar content from the candidate pool to increase the diversity and novelty of recommended content; Dynamically adjust the ranking of candidate content, adjust the recommendation strategy in a timely manner, increase the recommendation weight of new areas of interest, collect clear feedback from users, and adjust the recommendation algorithm parameters.

6. The method according to claim 1, characterized in that The S5 further includes: Comprehensive evaluation is carried out using multi-dimensional indicators such as accuracy, recall, and coverage; Targetedly tune and iterate the recommendation algorithm model, including but not limited to fine-tuning of algorithm parameters, feature engineering optimization, and model architecture improvement, to significantly improve the accuracy of recommendations and user experience, and enhance the overall performance of the system.

Images