An information intelligent recommendation method driven by user intention recognition

Abstract

The application discloses a kind of user intention recognition drive's information intelligent recommendation method, specifically includes: reading the page stay time in user session, click times, scroll proportion and query word semantic difference, write into user behavior record set;User behavior record set is vectorized, and interaction depth vector is generated;Information entropy is calculated to interaction depth vector, and information entropy change rate is generated, and stage label determination is carried out;Read stage label and theme keyword, extract relevant information, and construct semantic graph structure;In cognitive stage, diversity screening is executed to generate cognitive candidate set;In research and judgment stage, information density and viewpoint polarity are calculated to generate research and judgment candidate set;Scoring is carried out to candidate set, and recommendation information sequence is generated by sorting. The application accurately captures user demand, provides personalized, strong real-time recommendation.

Description

Technical Field

[0001] This invention relates to the field of intelligent recommendation technology, and in particular to an intelligent information recommendation method driven by user intent recognition. Background Technology

[0002] With the rapid development of information technology, internet data is experiencing explosive growth, especially in areas such as news, social media, and e-commerce platforms. Faced with massive amounts of information, users often find it difficult to quickly filter out content that interests them. Traditional information recommendation systems are mostly based on static user profiles and collaborative filtering algorithms, recommending content through statistical analysis of historical data and user interests. However, these methods have many limitations.

[0003] Existing recommender systems generally rely on static user interest models, which can only reflect users' long-term preferences and cannot effectively capture users' immediate needs. This results in poor timeliness of recommendation results and a degraded user experience. Furthermore, traditional systems lack sufficient perception and differentiation of changes in user needs at different cognitive stages, failing to provide corresponding recommendations based on users' varying needs at the cognitive stage (e.g., understanding background information) and the judgment stage (e.g., in-depth analysis, decision support). This lack of dynamic adaptability prevents personalized and precise information delivery.

[0004] To address these issues, recent research has focused on dynamic user intent recognition and its driving force for intelligent recommendation methods. These methods aim to dynamically determine a user's current need stage based on real-time behavioral data and interaction patterns, thereby providing more personalized recommendations that better match their current requirements. However, existing dynamic recommendation methods still face the following challenges: First, how to accurately capture and analyze changes in user intent; second, how to combine semantic analysis with multi-dimensional data such as information density and opinion polarity for precise filtering during the recommendation process; and finally, how to handle the diverse needs of users at different cognitive stages, ensuring that recommended content broadly covers user interests while providing sufficiently in-depth analytical information.

[0005] Therefore, this invention proposes an intelligent recommendation method based on user intent recognition. By dynamically identifying the user's demand stages and combining a semantic topology filtering mechanism, it overcomes the limitations of traditional recommendation systems and improves the accuracy and real-time performance of recommendations. Summary of the Invention

[0006] One objective of this invention is to propose an intelligent information recommendation method driven by user intent recognition. This invention introduces a dynamic intent recognition algorithm and a bimodal semantic topology filtering mechanism to perform in-depth analysis of users' real-time behavior data and interaction trajectories, perform information entropy calculation and intent stage determination, and construct a personalized recommendation information filtering process. It has the advantages of accurate recommendation results, strong real-time performance, and high personalization adaptability.

[0007] A user intent recognition-driven intelligent information recommendation method according to an embodiment of the present invention includes the following steps:

[0008] Read the page dwell time, topic click count, page scroll ratio, and semantic difference of consecutive search query terms from the user session, and write them into the user behavior record set;

[0009] The user behavior record set is vectorized to generate an interaction depth vector containing components such as dwell time, click frequency, scroll ratio, and query semantic refinement.

[0010] Perform information entropy calculation on the interaction depth vector, calculate the information entropy change rate, and input the information entropy change rate into the hidden Markov state decision operation to generate cognitive stage labels or judgment stage labels.

[0011] Read cognitive stage tags or judgment stage tags, user topic keywords, extract information sets that match topic keywords, and construct a semantic graph structure with semantic similarity connections.

[0012] When the stage label is a cognitive stage label, the maximum coverage diversity filtering operation is performed in the semantic graph structure to generate a cognitive candidate set;

[0013] When the stage label is the assessment stage label, the information density value and the opinion polarity distribution value are calculated in the semantic graph structure to generate the assessment candidate set;

[0014] Scoring is performed on the cognitive candidate set and the judgment candidate set, and the results are sorted to generate a sequence of recommendation information.

[0015] Optionally, writing to the user behavior record set specifically includes:

[0016] Read the user terminal's page access records during the session, extract the page access start time value and page access end time value, perform time difference calculation to generate the page dwell time value, and write the page dwell time value into the user behavior record set;

[0017] Read the page click records of the user terminal during the session, perform topic classification processing on the click records according to the topic keyword matching results, perform click count statistics calculation under the same topic category to generate topic click count value, and write the topic click count value into the user behavior record set;

[0018] Read the scroll position record of the user terminal during page browsing, extract the maximum scroll position value and the total page length value, and perform proportional calculation to generate the page scroll ratio value;

[0019] Read the text of two consecutive search queries from the user terminal, perform word segmentation on the two search query texts to generate a word set, perform semantic vector mapping on the word set to generate a query semantic vector, perform vector difference calculation on the query semantic vectors of adjacent time positions to generate a query semantic difference value, and write the page scroll ratio value and the query semantic difference value into the user behavior record set.

[0020] Optionally, generating the interaction depth vector, which includes dwell time, click frequency, scroll ratio, and query semantic refinement components, includes the following steps:

[0021] Read the page dwell time, topic click count, page scroll ratio, and query semantic difference values ​​from the user behavior record set, perform feature sorting operations according to the unified time index order, generate a behavior feature vector at each time position, and write the behavior feature vector into the behavior feature sequence.

[0022] Component normalization is performed on the behavioral feature sequence to generate dwell time component, click frequency component, scroll ratio component, and query semantic refinement component. Vector combination operation is performed on each component in a fixed position order to generate an interaction depth vector. The interaction depth vector is written into the interaction depth vector sequence in time index order.

[0023] Optionally, the calculation of information entropy, specifically including the calculation of the rate of change of information entropy, includes:

[0024] Read the interaction depth vector sequence, extract the dwell time component value, click frequency component value, scroll ratio component value and query semantic refinement component value at each time index position, write the four component values ​​into the component calculation unit, and perform component summation operation on the four component values ​​to generate the component sum value.

[0025] Perform a component sum division operation on the dwell time component, click frequency component, scroll ratio component, and query semantic refinement component to generate dwell time probability value, click frequency probability value, scroll ratio probability value, and query semantic probability value, and write the four probability values ​​into the probability set;

[0026] Read each probability value from the probability set, perform a logarithmic operation on each probability value to generate a logarithmic probability value, perform a component product operation on the probability value and the corresponding logarithmic probability value to generate an entropy component value, perform a sum operation on the entropy component values ​​to generate the information entropy value at the current time index position, and write the information entropy value into the information entropy sequence in chronological order.

[0027] Optionally, generating cognitive stage labels or judgment stage labels includes the following steps:

[0028] Read the information entropy sequence, extract the first information entropy value and the second information entropy value at adjacent time index positions, perform difference calculation on the first information entropy value and the second information entropy value to generate the information entropy change rate value, and write the information entropy change rate value into the change rate sequence according to the time index order.

[0029] Read the rate of change sequence and the interaction depth vector sequence, extract the information entropy rate of change value and the corresponding interaction depth vector at each time index position, perform vector combination operation on the dwell time component value, click frequency component value, scroll ratio component value and query semantic refinement component value in the interaction depth vector to generate the observation vector, and write the observation vector and information entropy rate of change value into the state determination input sequence.

[0030] Read the state determination input sequence, perform hidden Markov state calculation operation at each time index position, perform emission probability calculation on the observation vector and the information entropy change rate value to generate state probability value, perform maximum probability selection operation on the state probability value to generate cognitive stage label value or judgment stage label value, and write the stage label value into the stage label sequence in time index order.

[0031] Optionally, the semantic graph structure for constructing semantic similarity connections specifically includes:

[0032] Read the stage tag sequence and the topic keywords in the user's current session, extract the stage tag value and the corresponding topic keyword at each time index position, and perform word segmentation on the topic keywords to generate a keyword set;

[0033] For each keyword in the keyword set, perform semantic vector mapping to generate a semantic vector. Combine the generated semantic vector with the stage label value to construct a semantic relationship matrix between stage labels and keywords.

[0034] The elements in the semantic relation matrix are weighted, and the weights are calculated based on the semantic similarity between the stage label values ​​and the keywords. The weighted semantic relation matrix is ​​then written into the semantic graph construction module.

[0035] Clustering is performed on the nodes in the semantic graph to generate several topic clusters. Each topic cluster is then matched with the user's historical behavior data to generate a set of matched information.

[0036] Based on the matched information set, a candidate information set that meets the needs of the current stage is selected.

[0037] Optionally, performing the maximum coverage diversity filtering operation includes the following steps:

[0038] Read all nodes in the semantic graph and their corresponding semantic similarity values. Calculate and generate a set of similarity values ​​for each node based on the semantic relationship between each node and the topic keywords. For each node in the set of similarity values, perform a sorting operation, arranging the nodes in descending order of semantic similarity. Select the top N nodes with the highest semantic similarity values ​​as candidate nodes and write these candidate nodes into the candidate node set.

[0039] For each node in the candidate node set, perform diversity calculation. First, extract the semantic feature vector of each candidate node, perform a dot product with the semantic feature vector of each other node in the candidate set to generate a similarity value, and sum all the similarity values ​​of each node to calculate the cumulative similarity value of each node.

[0040] The cumulative similarity value of each node is sorted in descending order. The node with the smallest cumulative similarity value is added to the current candidate node set. The similarity generation process continues until the candidate node set reaches the preset size or meets the set diversity requirements.

[0041] The selected set of candidate nodes is output as the cognitive candidate set.

[0042] Optionally, the step of calculating the information density value and the opinion polarity distribution value in the semantic graph structure to generate the candidate set for judgment specifically includes:

[0043] Read all nodes and related semantic feature vectors in the semantic graph structure, analyze the semantic feature vectors of the nodes and the text content associated with the nodes, extract the number of entities in the text, the proportion of data charts and the complexity of sentences, and generate the semantic information density value of each node.

[0044] Calculate the opinion polarity distribution value for each node, perform sentiment analysis on the relevant text of the node, extract the proportion of positive, negative and neutral opinions, and combine the semantic features of the node to calculate the opinion polarity distribution value of the generated node.

[0045] The semantic information density value and opinion polarity distribution value of all nodes are combined and calculated to generate a comprehensive score value for each node. The score value is obtained by weighting the information density and opinion polarity distribution value according to a predetermined weight.

[0046] Sort the nodes according to their comprehensive score values ​​from highest to lowest, select the top N nodes with the highest scores, and generate a candidate set for evaluation.

[0047] Optionally, the step of performing scoring calculations on the candidate set and sorting them to generate a recommendation information sequence includes the following steps:

[0048] Read each node and its associated comprehensive score from the cognitive candidate set and the judgment candidate set. Based on the preset scoring calculation formula, perform a weighted scoring calculation on each node. The weight is determined according to the ratio between the semantic information density value and the opinion polarity distribution value of the node.

[0049] Sorting is performed on the ratings of all nodes in the cognitive candidate set and the judgment candidate set. The nodes are arranged in descending order of rating, and the top N nodes with the highest scores are selected as the final set of recommendation information.

[0050] Perform deduplication on each node in the recommendation information set to ensure that each recommendation information appears only once. After removing duplicate nodes, re-sort the remaining nodes.

[0051] Based on the sorting results, the nodes in the recommended information set are output in order to form a recommended information sequence, which is then passed to the user display module for presentation.

[0052] The beneficial effects of this invention are:

[0053] (1) By using a dual-modal semantic topology filtering mechanism, the breadth of information coverage is achieved in the cognition stage, and the accuracy and depth of recommended content are ensured in the judgment stage, thus avoiding the limitation of traditional recommendation systems to a single algorithm and improving the user experience.

[0054] (2) The recommendation strategy is dynamically adjusted according to the user's real-time needs. The recommendation is filtered by combining multiple dimensions such as semantic similarity, information density and opinion polarity to generate personalized recommendation content, which significantly improves the accuracy of the recommendation and user satisfaction.

[0055] (3) Dynamically analyze user behavior data and interaction trajectories, adjust recommendation strategies through information entropy change rate and hidden Markov model, avoid the defects of static model, enable recommendation system to effectively respond to changes in user needs, and thus provide more accurate recommendation results. Attached Figure Description

[0056] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0057] Figure 1 This is a flowchart of a user intent recognition-driven intelligent information recommendation method proposed in this invention;

[0058] Figure 2 This is a schematic diagram of a dynamic intent recognition model for a user intent recognition-driven intelligent information recommendation method proposed in this invention.

[0059] Figure 3This is a schematic diagram of the bimodal semantic topology filtering mechanism of the user intent recognition-driven intelligent information recommendation method proposed in this invention. Detailed Implementation

[0060] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0061] refer to Figures 1-3 A user intent recognition-driven intelligent information recommendation method includes the following steps:

[0062] Read the page dwell time, topic click count, page scroll ratio, and semantic difference of consecutive search query terms from the user session, and write them into the user behavior record set;

[0063] The user behavior record set is vectorized to generate an interaction depth vector containing components such as dwell time, click frequency, scroll ratio, and query semantic refinement.

[0064] Perform information entropy calculation on the interaction depth vector, calculate the information entropy change rate, and input the information entropy change rate into the hidden Markov state decision operation to generate cognitive stage labels or judgment stage labels.

[0065] Read cognitive stage tags or judgment stage tags, user topic keywords, extract information sets that match topic keywords, and construct a semantic graph structure with semantic similarity connections.

[0066] When the stage label is a cognitive stage label, the maximum coverage diversity filtering operation is performed in the semantic graph structure to generate a cognitive candidate set;

[0067] When the stage label is the assessment stage label, the information density value and the opinion polarity distribution value are calculated in the semantic graph structure to generate the assessment candidate set;

[0068] Scoring is performed on the cognitive candidate set and the judgment candidate set, and the results are sorted to generate a sequence of recommendation information.

[0069] In this embodiment, writing to the user behavior record set specifically includes:

[0070] Read the user terminal's page access records during the session, extract the page access start time value and page access end time value, perform time difference calculation to generate the page dwell time value, and write the page dwell time value into the user behavior record set;

[0071] Read the page click records of the user terminal during the session, perform topic classification processing on the click records according to the topic keyword matching results, perform click count statistics calculation under the same topic category to generate topic click count value, and write the topic click count value into the user behavior record set;

[0072] Read the scroll position record of the user terminal during page browsing, extract the maximum scroll position value and the total page length value, and perform proportional calculation to generate the page scroll ratio value;

[0073] Read the text of two consecutive search queries from the user terminal, perform word segmentation on the two search query texts to generate a word set, perform semantic vector mapping on the word set to generate a query semantic vector, perform vector difference calculation on the query semantic vectors of adjacent time positions to generate a query semantic difference value, and write the page scroll ratio value and the query semantic difference value into the user behavior record set.

[0074] In this embodiment, the word segmentation process involves reading the text of two consecutive search queries from the user terminal, preprocessing each query text to remove stop words and punctuation marks, then using a word segmentation tool to divide the text into independent word units, generating a word set, and merging the word sets from the two queries.

[0075] In this embodiment, generating an interaction depth vector that includes dwell time, click frequency, scroll ratio, and query semantic refinement components includes the following steps:

[0076] Read the page dwell time, topic click count, page scroll ratio, and query semantic difference values ​​from the user behavior record set, perform feature sorting operations according to the unified time index order, generate a behavior feature vector at each time position, and write the behavior feature vector into the behavior feature sequence.

[0077] Component normalization is performed on the behavioral feature sequence to generate dwell time component, click frequency component, scroll ratio component, and query semantic refinement component. Vector combination operation is performed on each component in a fixed position order to generate an interaction depth vector. The interaction depth vector is written into the interaction depth vector sequence in time index order.

[0078] In this embodiment, performing information entropy calculation, specifically calculating the rate of change of information entropy, includes:

[0079] Read the interaction depth vector sequence, extract the dwell time component value, click frequency component value, scroll ratio component value and query semantic refinement component value at each time index position, write the four component values ​​into the component calculation unit, and perform component summation operation on the four component values ​​to generate the component sum value.

[0080] Perform a component sum division operation on the dwell time component, click frequency component, scroll ratio component, and query semantic refinement component to generate dwell time probability value, click frequency probability value, scroll ratio probability value, and query semantic probability value, and write the four probability values ​​into the probability set;

[0081] Read each probability value from the probability set, perform a logarithmic operation on each probability value to generate a logarithmic probability value, perform a component product operation on the probability value and the corresponding logarithmic probability value to generate an entropy component value, perform a sum operation on the entropy component values ​​to generate the information entropy value at the current time index position, and write the information entropy value into the information entropy sequence in chronological order.

[0082] In this embodiment, generating cognitive stage labels or judgment stage labels includes the following steps:

[0083] Read the information entropy sequence, extract the first and second information entropy values ​​at adjacent time index positions, perform difference calculation on the first and second information entropy values ​​to generate the information entropy change rate value, and write the information entropy change rate value into the change rate sequence in time index order.

[0084] Read the rate of change sequence and the interaction depth vector sequence, extract the information entropy rate of change value and the corresponding interaction depth vector at each time index position, perform vector combination operation on the dwell time component value, click frequency component value, scroll ratio component value and query semantic refinement component value in the interaction depth vector to generate the observation vector, and write the observation vector and information entropy rate of change value into the state determination input sequence.

[0085] Read the state determination input sequence, perform hidden Markov state calculation operation at each time index position, perform emission probability calculation on the observation vector and the information entropy change rate value to generate state probability value, perform maximum probability selection operation on the state probability value to generate cognitive stage label value or judgment stage label value, and write the stage label value into the stage label sequence in time index order.

[0086] In this embodiment, extracting the first information entropy value and the second information entropy value specifically includes:

[0087] Read the information entropy sequence, traverse the information entropy values ​​in the sequence one by one according to the unified time index order, extract the current information entropy value at the current time index position, and determine the information entropy value as the first information entropy value; at the same time, read the information entropy value at the next time index position adjacent to the current time index position, and determine the information entropy value as the second information entropy value.

[0088] In this embodiment, performing the Hidden Markov State Calculation specifically includes:

[0089] Initialize the state set, transition probability matrix, and emission probability matrix;

[0090] Based on the state probability of the previous time step and the observation data of the current time step, the probability distribution of the current state is calculated using the forward algorithm;

[0091] The state with the highest probability is selected as the current intent stage label, and this label is written into the stage label sequence for use by the subsequent filtering module.

[0092] In this embodiment, performing the emission probability calculation specifically includes:

[0093] The observation vector and information entropy change rate are extracted from the state judgment input sequence at the current time index position. The observation vector contains the user's behavioral characteristics in a specific time period (such as page dwell time, click frequency, scroll ratio, etc.), while the information entropy change rate reflects the degree of change in user behavior. Through these observation data, the emission probability under each latent state (such as the cognitive stage or the judgment stage) is calculated.

[0094] Initialize an emission probability matrix, which defines the emission probability of each possible observation in each hidden state. The emission probability represents the likelihood of observing a particular observation in a given hidden state. For example, in the cognitive stage, behavioral features such as page dwell time and number of clicks will correspond to different emission probabilities.

[0095] Using a known emission probability matrix or probability distribution model (such as Gaussian distribution or Bayesian inference), the probability of observing specific behavioral data (such as page dwell time, click frequency, etc.) in each hidden state is calculated. This process determines the probability of observing a certain observation data in a given state by using historical data or a pre-trained model.

[0096] The state probability value at the current moment is updated by combining the emission probability and the state transition probability. The emission probability helps describe the likelihood of observing a specific behavior in a given hidden state, while the state probability reflects the probability that the user is in a certain hidden state at the current moment.

[0097] In this embodiment, constructing the semantic graph structure for semantic similarity connections specifically includes:

[0098] Read the stage tag sequence and the topic keywords in the user's current session, extract the stage tag value and the corresponding topic keyword at each time index position, and perform word segmentation on the topic keywords to generate a keyword set;

[0099] For each keyword in the keyword set, perform semantic vector mapping to generate a semantic vector. Combine the generated semantic vector with the stage label value to construct a semantic relationship matrix between stage labels and keywords.

[0100] The elements in the semantic relation matrix are weighted, and the weights are calculated based on the semantic similarity between the stage label values ​​and the keywords. The weighted semantic relation matrix is ​​then written into the semantic graph construction module.

[0101] Clustering is performed on the nodes in the semantic graph to generate several topic clusters. Each topic cluster is then matched with the user's historical behavior data to generate a set of matched information.

[0102] Based on the matched information set, a candidate information set that meets the needs of the current stage is selected.

[0103] In this embodiment, performing the maximum coverage diversity filtering operation includes the following steps:

[0104] Read all nodes in the semantic graph and their corresponding semantic similarity values. Calculate and generate a set of similarity values ​​for each node based on the semantic relationship between each node and the topic keywords. For each node in the set of similarity values, perform a sorting operation, arranging the nodes in descending order of semantic similarity. Select the top N nodes with the highest semantic similarity values ​​as candidate nodes and write these candidate nodes into the candidate node set.

[0105] For each node in the candidate node set, perform diversity calculation. First, extract the semantic feature vector of each candidate node, perform a dot product with the semantic feature vector of each other node in the candidate set to generate a similarity value, and sum all the similarity values ​​of each node to calculate the cumulative similarity value of each node.

[0106] The cumulative similarity value of each node is sorted in descending order. The node with the smallest cumulative similarity value is added to the current candidate node set. The process of generating similarity values ​​continues until the candidate node set reaches a preset size or meets the set diversity requirements.

[0107] The selected set of candidate nodes is output as the cognitive candidate set.

[0108] In this embodiment, performing diversity calculation specifically includes:

[0109] Extract the semantic feature vector of each node in the candidate node set. The semantic feature vector of each node is calculated based on the content of the node (such as page text, keywords, etc.) through natural language processing or vectorization models (such as Word2Vec, GloVe or BERT). The semantic feature vector represents the semantic information of the node content.

[0110] The semantic similarity between each pair of nodes is calculated by performing cosine similarity calculation on the semantic feature vectors of candidate nodes. The cosine similarity measures the angle between two vectors. The closer the value is to 1, the more similar the two nodes are semantically. The closer the value is to 0, the greater the semantic difference.

[0111] The similarity values ​​between all candidate nodes are summed to generate a cumulative similarity value for each node. This value represents the sum of the semantic similarity between the current node and all other nodes. The smaller the cumulative similarity value, the greater the semantic difference between the node and other nodes, and vice versa.

[0112] By executing a greedy or optimization algorithm, the node whose semantics are most different from those in the current candidate set is selected and added to the final candidate set, thereby ensuring that there is a high degree of semantic diversity among the candidate nodes.

[0113] Based on the calculated cumulative similarity value, the node with the smallest cumulative similarity value is selected as the final candidate node, and the process continues until the candidate node set reaches a preset size or meets the set diversity requirements.

[0114] In this embodiment, the specific steps for generating a candidate set for evaluation by calculating the information density value and the opinion polarity distribution value in the semantic graph structure include:

[0115] Read all nodes and related semantic feature vectors in the semantic graph structure, analyze the semantic feature vectors of the nodes and the text content associated with the nodes, extract the number of entities in the text, the proportion of data charts and the complexity of sentences, and generate the semantic information density value of each node.

[0116] Calculate the opinion polarity distribution value for each node, perform sentiment analysis on the relevant text of the node, extract the proportion of positive, negative and neutral opinions, and combine the semantic features of the node to calculate the opinion polarity distribution value of the generated node.

[0117] The semantic information density value and opinion polarity distribution value of all nodes are combined and calculated to generate a comprehensive score value for each node. The score value is obtained by weighting the information density and opinion polarity distribution value according to a predetermined weight.

[0118] Sort the nodes according to their comprehensive score values ​​from highest to lowest, select the top N nodes with the highest scores, and generate a candidate set for evaluation.

[0119] In this embodiment, calculating the viewpoint polarity distribution value for each node specifically includes:

[0120] After performing text preprocessing to remove stop words and punctuation marks, and then performing word segmentation, the sentiment polarity of each word is analyzed using a sentiment dictionary to obtain positive, negative, and neutral sentiment scores.

[0121] Calculate the proportions of positive, negative, and neutral sentiment in the text to obtain the opinion polarity distribution value for each node, i.e., the distribution of positive, negative, and neutral sentiment.

[0122] In this embodiment, performing score calculations on the candidate set and sorting it to generate a recommendation information sequence includes the following steps:

[0123] Read each node and its associated comprehensive score from the cognitive candidate set and the judgment candidate set. Based on the preset scoring calculation formula, perform a weighted scoring calculation on each node. The weight is determined according to the ratio between the semantic information density value and the opinion polarity distribution value of the node.

[0124] Sorting is performed on the ratings of all nodes in the cognitive candidate set and the judgment candidate set. The nodes are arranged in descending order of rating, and the top N nodes with the highest scores are selected as the final set of recommendation information.

[0125] Perform deduplication on each node in the recommendation information set to ensure that each recommendation information appears only once. After removing duplicate nodes, re-sort the remaining nodes.

[0126] Based on the sorting results, the nodes in the recommended information set are output in order to form a recommended information sequence, which is then passed to the user display module for presentation.

[0127] Example 1:

[0128] This embodiment aims to demonstrate the application scenario of an intelligent information recommendation method based on dynamic intent recognition in a news recommendation platform. On this platform, users continuously browse different news content. The system dynamically identifies the user's intent stage based on real-time behavior and interaction data, thereby pushing news information that better meets the user's needs. In news recommendation platforms, users are exposed to a large amount of news information daily, and the system needs to provide the most relevant and user-demanding news content in a short period of time. Traditional recommendation systems often rely on historical data and static user profiles, which makes it difficult for the system to fully respond to changes in user needs, resulting in recommendation results that cannot match the user's current interests in a timely manner.

[0129] In this scenario, the recommendation method of this invention addresses the shortcomings of traditional recommendation systems. Specifically, the system analyzes real-time user behavior data (such as page dwell time, click count, scroll ratio, etc.) and uses a dynamic intent recognition algorithm to determine whether the user is currently in the cognitive or analytical stage. Users in the cognitive stage require a broader information overview, so the system recommends diverse news content to satisfy their wide range of interests; while users in the analytical stage require more in-depth content, so the system pushes professional content such as data analysis reports and expert commentary. Through a bimodal semantic topology filtering mechanism, the system achieves a balance between the diversity and accuracy of recommended content, thereby providing accurate and diverse news recommendations.

[0130] To verify the feasibility of this invention in practical applications, we conducted a month-long experiment on a news recommendation platform, collecting browsing behavior data from 2000 users and comparing it with the method of this invention using traditional recommendation methods. Experimental results show that the recommendation system using the method of this invention significantly outperforms traditional methods in terms of recommendation accuracy and user satisfaction. Specific data are as follows:

[0131] Table 1: Comparison and Evaluation Table of Recommended Methods

[0132] Traditional recommendation methods 72.5 65.3 10.2 50.1 Recommended method of this invention 88.4 85.1 22.3 83.6

[0133] The comparison shows that the recommendation accuracy of the method of this invention is 88.4%, which is 15.9 percentage points higher than the 72.5% of the traditional recommendation method. This indicates that the dynamic intent recognition model of this invention can accurately capture the user's current need stage (such as the cognitive stage or the judgment stage), thereby significantly improving the relevance and accuracy of the recommended content.

[0134] The user satisfaction rate using the method of this invention is 85.1%, significantly higher than the 65.3% of traditional recommendation methods. This indicates that the present invention can better understand and meet user needs, especially in the process of dynamically changing user intent recognition, the system can provide content that better meets user expectations, thereby improving user satisfaction.

[0135] After using the method of this invention, the click-through rate increased by 22.3%, a significant improvement compared to the 10.2% of traditional recommendation methods. This indicates that the method of this invention, by diversifying the recommended content, avoids overly simplistic recommendations, thereby stimulating user interest and interaction, and improving the system's usability.

[0136] The diversity assessment of the method in this invention is 83.6%, significantly higher than the 50.1% of traditional recommendation methods. This indicates that the bimodal semantic topology filtering mechanism of this invention can ensure the recommendation of broad information during the cognitive stage and more detailed analysis during the judgment stage, thereby significantly improving the diversity of recommended content, avoiding the limitations of information recommendation, and increasing users' overall acceptance of the recommended content.

[0137] Based on the data above, we can see that the present invention outperforms traditional recommendation methods in terms of recommendation accuracy, user satisfaction, click-through rate, and diversity, proving that the present invention has strong competitiveness in practical applications and can better meet users' personalized needs.

[0138] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A user intent recognition-driven intelligent information recommendation method, characterized in that, Includes the following steps: Read the page dwell time, topic click count, page scroll ratio, and semantic difference of consecutive search query terms from the user session, and write them into the user behavior record set; The user behavior record set is vectorized to generate an interaction depth vector containing components such as dwell time, click frequency, scroll ratio, and query semantic refinement. Perform information entropy calculation on the interaction depth vector, calculate the information entropy change rate, and input the information entropy change rate into the hidden Markov state decision operation to generate cognitive stage labels or judgment stage labels. Read cognitive stage tags or judgment stage tags, user topic keywords, extract information sets that match topic keywords, and construct a semantic graph structure with semantic similarity connections. When the stage label is a cognitive stage label, the maximum coverage diversity filtering operation is performed in the semantic graph structure to generate a cognitive candidate set; When the stage label is the assessment stage label, the information density value and the opinion polarity distribution value are calculated in the semantic graph structure to generate the assessment candidate set; Scoring is performed on the cognitive candidate set and the judgment candidate set, and the results are sorted to generate a sequence of recommendation information.

2. The user intent recognition-driven intelligent information recommendation method according to claim 1, characterized in that, The step of reading page dwell time, topic click count, page scroll ratio, and semantic difference of consecutive search query terms from the user session and writing them into the user behavior record set specifically includes: Read the user terminal's page access records during the session, extract the page access start time value and page access end time value, perform time difference calculation to generate the page dwell time value, and write the page dwell time value into the user behavior record set; Read the page click records of the user terminal during the session, perform topic classification processing on the click records according to the topic keyword matching results, perform click count statistics calculation under the same topic category to generate topic click count value, and write the topic click count value into the user behavior record set; Read the scroll position record of the user terminal during page browsing, extract the maximum scroll position value and the total page length value, and perform proportional calculation to generate the page scroll ratio value; Read the text of two consecutive search queries from the user terminal, perform word segmentation on the two search query texts to generate a word set, perform semantic vector mapping on the word set to generate a query semantic vector, perform vector difference calculation on the query semantic vectors of adjacent time positions to generate a query semantic difference value, and write the page scroll ratio value and the query semantic difference value into the user behavior record set.

3. The user intent recognition-driven intelligent information recommendation method according to claim 2, characterized in that, The process of generating an interaction depth vector that includes dwell time, click frequency, scroll ratio, and query semantic refinement components includes the following steps: Read the page dwell time, topic click count, page scroll ratio, and query semantic difference values ​​from the user behavior record set, perform feature sorting operations according to the unified time index order, generate a behavior feature vector at each time position, and write the behavior feature vector into the behavior feature sequence. Component normalization is performed on the behavioral feature sequence to generate dwell time component, click frequency component, scroll ratio component, and query semantic refinement component. Vector combination operation is performed on each component in a fixed position order to generate an interaction depth vector. The interaction depth vector is written into the interaction depth vector sequence in time index order.

4. The user intent recognition-driven intelligent information recommendation method according to claim 3, characterized in that, The calculation of information entropy, specifically including the calculation of the rate of change of information entropy, includes: Read the interaction depth vector sequence, extract the dwell time component value, click frequency component value, scroll ratio component value and query semantic refinement component value at each time index position, write the four component values ​​into the component calculation unit, and perform component summation operation on the four component values ​​to generate the component sum value. Perform a component sum division operation on the dwell time component, click frequency component, scroll ratio component, and query semantic refinement component to generate dwell time probability value, click frequency probability value, scroll ratio probability value, and query semantic probability value, and write the four probability values ​​into the probability set; Read each probability value from the probability set, perform a logarithmic operation on each probability value to generate a logarithmic probability value, perform a component product operation on the probability value and the corresponding logarithmic probability value to generate an entropy component value, perform a sum operation on the entropy component values ​​to generate the information entropy value at the current time index position, and write the information entropy value into the information entropy sequence in chronological order.

5. The user intent recognition-driven intelligent information recommendation method according to claim 4, characterized in that, The generation of cognitive stage labels or judgment stage labels includes the following steps: Read the information entropy sequence, extract the first information entropy value and the second information entropy value at adjacent time index positions, perform difference calculation on the first information entropy value and the second information entropy value to generate the information entropy change rate value, and write the information entropy change rate value into the change rate sequence according to the time index order. Read the rate of change sequence and the interaction depth vector sequence, extract the information entropy rate of change value and the corresponding interaction depth vector at each time index position, perform vector combination operation on the dwell time component value, click frequency component value, scroll ratio component value and query semantic refinement component value in the interaction depth vector to generate the observation vector, and write the observation vector and information entropy rate of change value into the state determination input sequence. Read the state determination input sequence, perform hidden Markov state calculation operation at each time index position, perform emission probability calculation on the observation vector and the information entropy change rate value to generate state probability value, perform maximum probability selection operation on the state probability value to generate cognitive stage label value or judgment stage label value, and write the stage label value into the stage label sequence in time index order.

6. The user intent recognition-driven intelligent information recommendation method according to claim 5, characterized in that, The semantic graph structure for constructing semantic similarity connections specifically includes: Read the stage tag sequence and the topic keywords in the user's current session, extract the stage tag value and the corresponding topic keyword at each time index position, and perform word segmentation on the topic keywords to generate a keyword set; For each keyword in the keyword set, perform semantic vector mapping to generate a semantic vector. Combine the generated semantic vector with the stage label value to construct a semantic relationship matrix between stage labels and keywords. The elements in the semantic relation matrix are weighted, and the weights are calculated based on the semantic similarity between the stage label values ​​and the keywords. The weighted semantic relation matrix is ​​then written into the semantic graph construction module. Clustering is performed on the nodes in the semantic graph to generate several topic clusters. Each topic cluster is then matched with the user's historical behavior data to generate a set of matched information. Based on the matched information set, a candidate information set that meets the needs of the current stage is selected.

7. The user intent recognition-driven intelligent information recommendation method according to claim 6, characterized in that, The process of performing the maximum coverage diversity filtering operation includes the following steps: Read all nodes in the semantic graph and their corresponding semantic similarity values. Calculate and generate a set of similarity values ​​for each node based on the semantic relationship between each node and the topic keywords. For each node in the set of similarity values, perform a sorting operation, arranging the nodes in descending order of semantic similarity. Select the top N nodes with the highest semantic similarity values ​​as candidate nodes and write these candidate nodes into the candidate node set. For each node in the candidate node set, perform diversity calculation. First, extract the semantic feature vector of each candidate node, perform a dot product with the semantic feature vector of each other node in the candidate set to generate a similarity value, and sum all the similarity values ​​of each node to calculate the cumulative similarity value of each node. The cumulative similarity value of each node is sorted in descending order. The node with the smallest cumulative similarity value is added to the current candidate node set. The process of generating similarity values ​​continues until the candidate node set reaches a preset size or meets the set diversity requirements. The selected set of candidate nodes is output as the cognitive candidate set.

8. The user intent recognition-driven intelligent information recommendation method according to claim 7, characterized in that, The specific steps of calculating the information density value and the opinion polarity distribution value in the semantic graph structure to generate the candidate set for judgment include: Read all nodes and related semantic feature vectors in the semantic graph structure, analyze the semantic feature vectors of the nodes and the text content associated with the nodes, extract the number of entities in the text, the proportion of data charts and the complexity of sentences, and generate the semantic information density value of each node. Calculate the opinion polarity distribution value for each node, perform sentiment analysis on the relevant text of the node, extract the proportion of positive, negative and neutral opinions, and combine the semantic features of the node to calculate the opinion polarity distribution value of the generated node. The semantic information density value and opinion polarity distribution value of all nodes are combined and calculated to generate a comprehensive score value for each node. The score value is obtained by weighting the information density and opinion polarity distribution value according to a predetermined weight. Sort the nodes according to their comprehensive score values ​​from highest to lowest, select the top N nodes with the highest scores, and generate a candidate set for evaluation.

9. The user intent recognition-driven intelligent information recommendation method according to claim 8, characterized in that, The step of performing scoring calculations on the cognitive candidate set and the judgment candidate set, and sorting them to generate a recommendation information sequence includes the following steps: Read each node and its associated comprehensive score from the cognitive candidate set and the judgment candidate set. Based on the preset scoring calculation formula, perform a weighted scoring calculation on each node. The weight is determined according to the ratio between the semantic information density value and the opinion polarity distribution value of the node. Sorting is performed on the ratings of all nodes in the cognitive candidate set and the judgment candidate set. The nodes are arranged in descending order of rating, and the top N nodes with the highest scores are selected as the final set of recommendation information. Perform deduplication on each node in the recommendation information set to ensure that each recommendation information appears only once. After removing duplicate nodes, re-sort the remaining nodes. Based on the sorting results, the nodes in the recommended information set are output in order to form a recommended information sequence, which is then passed to the user display module for presentation.

Images