Personalized information accurate pushing system and method based on artificial intelligence

Abstract

The invention discloses a personalized information accurate pushing system and method based on artificial intelligence, and relates to the technical field of personalized recommendation, and the method comprises the steps: fusing user multi-platform behavior data and external space-time environment information, and generating a situation label with confidence through employing an improved space-time density clustering algorithm; constructing a causal directed acyclic graph by adopting a causal forest algorithm, quantitatively analyzing a heterogeneity causal effect, inverting a potential intention of the user, and outputting standardized intention inversion probability distribution; in combination with a historical intention and a behavior sequence, training a Transform intention state transition model constrained by causality of a causality directed acyclic graph, and performing multi-step probability deduction to generate an intention evolution path; information is retrieved from the dynamic knowledge graph according to the prediction path, a pushing copywriting matched with the situation is generated through the NLP technology, the optimal pushing opportunity is calculated in combination with the position track of the user, and accurate reaching of personalized information is achieved.

Description

Technical Field

[0001] The present invention relates to the field of personalized recommendation technology, and in particular to a personalized information precision push system and method based on artificial intelligence. Background Art

[0002] With the rapid development of information technology, personalized information push has become a key means to enhance user experience and commercial value. However, existing personalized push technologies generally have shortcomings, which seriously restrict their accuracy and effectiveness.

[0003] When it comes to constructing user profiles, existing technologies mostly rely on a single platform or a single type of data, making it difficult to form a comprehensive, three-dimensional view of users. This data silo results in a one-sided and static user profile, failing to capture the dynamic changes in user needs in different scenarios. At the level of intent recognition and analysis, existing mainstream technologies are mostly based on correlation analysis, which infers user interests by statistically correlating user behavior with item click-through rates, purchase rates, and other factors. However, this fails to reveal the underlying causal logic behind the behavior. The ambiguity of causal relationships results in low accuracy in intent recognition, and the technology is easily interfered with by irrelevant behaviors, failing to accurately capture users' core motivations and potential needs. When it comes to predicting future behavior, existing technologies mostly remain at the stage of predicting users' current or short-term interests, lacking the ability to dynamically deduce future behavioral trends. This results in the selection of push timing relying entirely on simple rules or thresholds, making it impossible to predict the evolutionary path of user intent and the optimal triggering time. This results in push content being ignored when the user has not yet generated demand, or appearing only after the user's demand has become outdated, impacting the push conversion effect and user experience. Summary of the Invention

[0004] The purpose of the present invention is to provide a personalized information precision push system and method based on artificial intelligence to solve the problems raised in the prior art.

[0005] To achieve the above object, the present invention provides the following technical solutions: In a first aspect, the present invention provides a method for accurately pushing personalized information based on artificial intelligence, comprising: Collect user behavior data on e-commerce, social networking, and health and fitness platforms in real time, and obtain external spatiotemporal environment data and user location trajectory information; The collected data is fused and an improved spatiotemporal density clustering algorithm is used to generate context labels with confidence levels to define the specific scenario the user is currently in. The generated context labels and the user's historical behavior sequence are used as input for feature engineering. A causal forest algorithm is used to quantitatively analyze the heterogeneous causal effects between behavioral data, context labels, and potential intent, constructing a causal directed acyclic graph. This graph is then used to perform counterfactual reasoning, inverting the user's potential intent in the current context and outputting a standardized probability distribution for the inverted intent. Using the constructed causal directed acyclic graph and combining the user's historical intention sequence and behavior sequence, we train an intention state transition model based on the Transformer architecture. Based on the intention inversion probability distribution and intention state transition model, multi-step probability deduction is performed to generate multiple intention evolution paths containing intention nodes, transition probabilities, and expected occurrence time windows; Based on the predicted intent evolution path, relevant information is retrieved from the existing dynamic knowledge graph, and NLP technology based on template and retrieval enhancement generation is used to create push copy that matches the current context; combined with user location trajectory information, the optimal push timing is calculated.

[0006] In conjunction with the first aspect, in a first implementation of the first aspect of this application, the real-time collection of user behavior data on e-commerce, social networking, and health and fitness platforms, and the acquisition of external spatiotemporal environment data and user location trajectory information, include: Real-time collection of users' product browsing, add-to-cart, purchase, and evaluation behaviors on e-commerce platforms; collection of users' content publishing, likes, comments, favorites, and follow-up behaviors on social platforms; and collection of users' exercise records on health and fitness platforms. Acquire external spatiotemporal environmental data in real time, including timestamps, weather conditions, and holiday information; obtain the user's geographic location, movement path, movement speed, permanent locations, and points of interest.

[0007] In combination with the first aspect, in a second implementation of the first aspect of the present application, the collected data is fused, and an improved spatiotemporal density clustering algorithm is used to generate a context label with a confidence level to define the specific scenario in which the user is currently located, including: The collected behavioral data, external spatiotemporal environment data, and user location trajectory information are integrated into multi-source heterogeneous data to construct a high-dimensional spatiotemporal feature vector. The behavioral data is quantified into behavior type and behavior intensity using preset rules, and the above features are normalized to form a unified time series feature dataset. An improved DBSCAN algorithm based on density peak and time decay factor is used to perform cluster analysis on time series feature data sets. Time decay factor and behavior intensity weight are introduced to calculate the influence of data points in the cluster. For each cluster formed after cluster analysis, the core features of all data points within the cluster are aggregated to generate a preset context label. Based on the compactness of the cluster, the temporal continuity of the data points within the cluster, and the consistency of the behavioral intensity, the confidence level of each context label is calculated using a weighted summation algorithm. The generated context label with confidence is used as the definition of the specific scenario the user is currently in. New user data is continuously received to trigger a re-evaluation of the current scenario. When the matching degree between the new data point and the current context label is lower than the preset threshold, or a new high-confidence context label appears, the user's current scenario definition is updated.

[0008] In combination with the first aspect, in a third implementation of the first aspect of the present application, the feature engineering is performed using the generated context labels and the user's historical behavior sequence as input, including: The generated context labels with confidence are mapped into high-dimensional sparse vectors through one-hot encoding. Each behavior event in the user's historical behavior sequence is constructed into a structured behavior event vector based on the behavior type, behavior intensity, and timestamp. The current context label vector and the user's historical behavior sequence vector within the preset retrospective time window are concatenated in chronological order to form a mixed feature sequence that combines the current context and historical behavior. The mixed feature sequence is input into a pre-trained feature embedding model, which maps each vector into a unified continuous vector space and generates an embedding feature matrix that can capture the semantic associations between features.

[0009] In combination with the first aspect, in a fourth implementation of the first aspect of the present application, the causal forest algorithm is used to quantitatively analyze the heterogeneous causal effects between behavioral data, context labels, and potential intentions, and to construct a causal directed acyclic graph, including: The behavioral and situational features in the generated embedded feature matrix are used as treatment variables, and the potential intention is used as the outcome variable. The causal forest algorithm is used to analyze the causal effect strength of each treatment variable on each potential intention outcome under different situational and behavioral combinations, identify the key causal factors that have the greatest impact on specific intentions, and reveal the heterogeneity of causal effects. Based on the heterogeneous causal effects analyzed by the causal forest algorithm, a causal directed acyclic graph is constructed with potential intentions as core nodes and behavioral data and situational labels as parent nodes.

[0010] In conjunction with the first aspect, in a fifth implementation of the first aspect of the present application, using the graph to perform counterfactual reasoning, inverting the user's potential intention in the current situation, and outputting a standardized intention inversion probability distribution includes: Using the defined current user context label as the baseline context, the specific behavior node of the current user is locked in the constructed causal directed acyclic graph; Perform a counterfactual intervention in a causal directed acyclic graph. By removing the current user's specific behavior, the probability distribution of potential intention nodes is changed. The difference in intention probability before and after the intervention is calculated to quantify the causal contribution of the behavior to the intention. Taking into account the combined effects of the current context label and all historical behaviors in the causal directed acyclic graph, the Bayesian inference algorithm is used to invert the probability distribution of the user's unobserved potential intentions in the current context, normalize it into a standardized intention inversion probability distribution, and display the several intentions and probabilities that the user is most likely to have in the current context.

[0011] In conjunction with the first aspect, in a sixth implementation of the first aspect of the present application, the constructed causal directed acyclic graph is used to combine the user's historical intention sequence and behavior sequence to train an intention state transition model based on the Transformer architecture, including: The inverse probability distribution of the user's intention in a historical time period is defined as the intention state, and the specific behavior generated by the user in this time period is defined as the triggering behavior. The intention state transition sequence for supervised learning is constructed, where each training sample contains the intention state at time t, the triggering behavior, and the intention state at the next time t+1; An intention state transition model based on the Transformer architecture is designed. The input sequence of the model is the intention state sequence and the corresponding trigger behavior sequence, and the output sequence is the subsequent intention state prediction sequence. The constructed causal directed acyclic graph is used as an external knowledge constraint and integrated into the model training process. When the model makes state transition predictions, the legal transition relationships defined in the causal directed acyclic graph are encoded as causal masks, which act on the model's attention layer to ensure that the model learns and follows the state transition path defined in the graph that conforms to the causal logic. The subsequent intention states in historical real data are used as labels to supervise the model training and optimize its accuracy in predicting future intention states.

[0012] In combination with the first aspect, in a seventh implementation of the first aspect of the present application, the multi-step probability deduction is performed based on the intention inversion probability distribution and the intention state transition model to generate multiple intention evolution paths including intention nodes, transition probabilities, and expected occurrence time windows, including: The intention inversion probability distribution generated at the current moment is used as the initial state of the multi-step probability deduction to represent the probabilities of all possible user intentions in the current context. The current deduced state is input into the intention state transition model to predict the probability that the user's intention state will transition to all other possible intentions within the next time step, generating a transition probability distribution. The predicted transition probability distribution is used as the new state for the next iteration. At each state transition, the transition probability from the current intention node to the next intention node is recorded, and an expected occurrence time window is estimated based on the preset life cycle of the intention. After completing the deduction of all steps, a complete set of intention evolution paths is generated based on each recorded transfer information. Each path is a sequence, including the initial intention node, the subsequent intention nodes generated by deduction, and the corresponding transfer probabilities and expected occurrence time windows associated with adjacent nodes in the path. Multiple intention evolution paths with different probabilities are output.

[0013] In conjunction with the first aspect, in an eighth implementation of the first aspect of the present application, the method of retrieving relevant information from an existing dynamic knowledge graph based on the predicted intent evolution path, and using NLP technology based on template and retrieval enhancement generation to create a push copy that matches the current context; and calculating the optimal push timing in combination with user location trajectory information, including: For each generated intent evolution path, the target intent node in the path is used as the core query condition to search in the dynamic knowledge graph to obtain entities, attributes, and relationships directly related to the intent, forming a candidate information set; Using hybrid NLP technology, based on the current user's contextual tags and intent evolution path, the predefined copy template library is screened for the base template with the highest semantic and stylistic match. The retrieved candidate information set is used as enhanced information and intelligently filled into the corresponding placeholders of the base template using a generative model to generate the final personalized push copy. Analyze the user's mobility patterns, frequent stops, and points of interest, and calculate the spatiotemporal accessibility of the user's geographic location associated with the pushed content within the expected time window. Comprehensively compare the overlap between the accessibility scores of each path and the time window, select the moment with the highest accessibility and the best match with the time window, and activate the push timing prediction window for the intention as the final basis for executing precise push decisions.

[0014] In a second aspect, the present invention provides a personalized information precision push system based on artificial intelligence, comprising: Multi-source data perception and fusion module: This module includes a data acquisition unit and a data fusion and contextualization unit. The data acquisition unit is responsible for collecting raw data streams in real time from three dimensions: user-platform interaction, external environment, and user location trajectory. The data fusion and contextualization unit pre-processes the collected multi-source heterogeneous data and aggregates discrete data points into contextual labels with confidence levels using an improved spatiotemporal clustering algorithm. Causal Intention Inference Module: This module includes a causal effect analysis unit, an intention inversion unit, and an intention state transition modeling unit. The causal effect analysis unit uses a causal forest algorithm to quantitatively analyze the heterogeneous causal effects between behaviors, contexts, and intentions, identify the key factors driving user intentions, and construct a causal directed acyclic graph based on this. The intention inversion unit uses the constructed causal graph to perform counterfactual reasoning. Given the current context and behavior, it inverts the user's most likely potential intentions and probabilities, and outputs a standardized intention inversion probability distribution. The intention state transition modeling unit combines historical intention sequences and behavior sequences to train an intention state transition model based on the Transformer architecture. Using the causal graph as an external knowledge constraint, it learns the transition rules between intention states that conform to causal logic for predicting future intentions. Intention evolution path deduction module: This module includes a multi-step probability deduction unit. This unit uses the current intention inversion probability distribution as the starting point and iteratively calls the intention state transition model to generate multiple complete intention evolution paths that include future intention nodes, transition probabilities between nodes, and expected occurrence time windows. Personalized content generation module: This module includes a knowledge graph retrieval unit and a hybrid NLP generation unit. The knowledge graph retrieval unit uses the target intent in the intent evolution path as a query condition to retrieve relevant entities, attributes, and relationships in the dynamic knowledge graph. The hybrid NLP generation unit combines the retrieved information with the current user's contextual tags, selects the most matching basic template from the template library, and uses a generative model to intelligently fill in the retrieved information to generate the final personalized push copy. Intelligent timing decision and execution module: includes a spatiotemporal reachability analysis unit and a push timing activation and execution unit; the spatiotemporal reachability analysis unit combines the user's location trajectory information and the geographic location associated with the pushed content to calculate the spatiotemporal reachability of the user to the relevant location within the expected time window of the intention evolution path; the push timing activation and execution unit comprehensively compares the reachability scores under each intention evolution path and the overlap of the time window, selects the optimal path and the timing with the highest matching degree, activates the final push timing prediction window, and triggers the system to execute the push operation.

[0015] Compared with the prior art, the present invention has the following beneficial effects: 1. The present invention integrates multi-dimensional data of user behavior, spatiotemporal environment and location trajectory, and generates context labels with confidence through an improved clustering algorithm to achieve accurate definition of the specific scenario in which the user is located.

[0016] 2. This invention introduces a causal forest algorithm to quantitatively analyze the heterogeneous causal effects between behaviors, situations, and intentions, construct a causal directed acyclic graph, and accurately invert the user's potential intentions in the current situation through counterfactual reasoning.

[0017] 3. The present invention combines causal graphs with user history sequences to train a Transformer-based intention state transition model, and generates an intention evolution path containing intention nodes, transition probabilities, and time windows through multi-step probability deduction. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 This is a schematic diagram of the steps of a personalized information accurate push method based on artificial intelligence of the present invention; Figure 2 This is a program flow chart of a method for accurately pushing personalized information based on artificial intelligence in the present invention. DETAILED DESCRIPTION

[0019] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0020] Example: Figure 1-Figure 2 As shown, the present invention provides a technical solution. like Figure 1 As shown in a schematic diagram of the steps of a personalized information precision push method based on artificial intelligence, the present invention provides a personalized information precision push method based on artificial intelligence, comprising: Step S100: Collect user behavior data on e-commerce, social networking, and health and fitness platforms in real time to obtain external spatiotemporal environment data and user location trajectory information; fuse the collected data and use an improved spatiotemporal density clustering algorithm to generate context labels with confidence levels to define the specific scenario the user is currently in; use the generated context labels and the user's historical behavior sequence as input to perform feature engineering; Specifically, it collects users' product browsing, adding to cart, purchasing, and review behaviors on e-commerce platforms in real time; collects users' content publishing, liking, commenting, collecting, and following behaviors on social platforms; and collects users' exercise records on health and fitness platforms; Acquire external spatiotemporal environmental data in real time, including timestamps, weather conditions, and holiday information; obtain the user's geographic location, movement path, movement speed, permanent locations, and points of interest.

[0021] The collected behavioral data, external spatiotemporal environment data, and user location trajectory information are integrated into multi-source heterogeneous data to construct a high-dimensional spatiotemporal feature vector. The behavioral data is quantified into behavior type and behavior intensity using preset rules, and the above features are normalized to form a unified time series feature dataset. An improved DBSCAN algorithm based on density peak and time decay factor is used to perform cluster analysis on time series feature data sets. Time decay factor and behavior intensity weight are introduced to calculate the influence of data points in the cluster. For each cluster formed after cluster analysis, the core features of all data points within the cluster are aggregated to generate a preset context label. Based on the compactness of the cluster, the temporal continuity of the data points within the cluster, and the consistency of the behavioral intensity, the confidence level of each context label is calculated using a weighted summation algorithm. The generated context label with confidence is used as the definition of the specific scenario the user is currently in. New user data is continuously received to trigger a re-evaluation of the current scenario. When the matching degree between the new data point and the current context label is lower than the preset threshold, or a new high-confidence context label appears, the user's current scenario definition is updated.

[0022] The generated context labels with confidence are mapped into high-dimensional sparse vectors through one-hot encoding. Each behavior event in the user's historical behavior sequence is constructed into a structured behavior event vector based on the behavior type, behavior intensity, and timestamp. The current context label vector and the user's historical behavior sequence vector within the preset retrospective time window are concatenated in chronological order to form a mixed feature sequence that combines the current context and historical behavior. The mixed feature sequence is input into a pre-trained feature embedding model, which maps each vector into a unified continuous vector space and generates an embedding feature matrix that can capture the semantic associations between features.

[0023] In a specific embodiment, user A browsed camping tents and portable barbecue grills on an e-commerce platform at 14:30 on October 28, 2023 and added them to the shopping cart. He then posted a dynamic about his weekend trip to the country park on a social platform and received likes. At the same time, the health and fitness platform recorded his GPS trajectory starting from home at 15:00, with a permanent point at longitude 116.4 and latitude 39.9, a moving speed of 15km / h, and a point of interest at the country park at longitude 116.5 and latitude 40.1. These behavioral data, external data, and location trajectories are fused into a high-dimensional spatiotemporal feature vector, where the behavioral data is quantified as a shopping interest intensity of 0.8 and a social sharing intensity of 0.6, and normalized to form a time series data set. The improved DBSCAN algorithm is used for clustering, where the time decay factor ɑ=0.5 and the behavioral intensity weight β=0 .7. Analysis found that the data points are densely packed and temporally continuous in the weekend outdoor leisure cluster, between 2:00 PM and 3:30 PM. This cluster aggregates core features such as e-commerce shopping, social sharing, and movement to suburban parks. This generates a contextual label, "weekend outdoor leisure," and calculates a confidence score of 0.85 using a weighted calculation of 0.85 based on a density of 0.85, temporal continuity of 0.9, and behavioral consistency of 0.8. This label is mapped to a vector [0, 0, 1, 0] using one-hot encoding. The preset label set is [working from home, commuting and fitness, weekend outdoor leisure, weekday leisure]. The user's historical behavior sequence, "browsing a park guide last Saturday" (0.2, 0.1, 0.7, 0.1), is concatenated with the current contextual vector and input into the pre-trained embedding model to generate an embedding feature matrix that captures the semantic association between outdoor shopping and social sharing, providing feature input for subsequent intent inversion.

[0024] Step S200: Using a causal forest algorithm, the heterogeneous causal effects between behavioral data, context labels, and potential intent are quantitatively analyzed to construct a causal directed acyclic graph. This graph is used to perform counterfactual reasoning, invert the user's potential intent in the current context, and output a standardized probability distribution of intent inversion. Specifically, the behavioral and situational features in the generated embedded feature matrix are used as treatment variables, and the potential intention is used as the outcome variable. Using the causal forest algorithm, we analyze the causal effect strength of each treatment variable on each potential intention outcome under different situational and behavioral combinations, identify the key causal factors that have the greatest impact on specific intentions, and reveal the heterogeneity of causal effects. Based on the heterogeneous causal effects analyzed by the causal forest algorithm, a causal directed acyclic graph is constructed with potential intentions as core nodes and behavioral data and situational labels as parent nodes.

[0025] Using the defined current user context label as the baseline context, the specific behavior node of the current user is locked in the constructed causal directed acyclic graph; Perform a counterfactual intervention in a causal directed acyclic graph. By removing the current user's specific behavior, the probability distribution of potential intention nodes is changed. The difference in intention probability before and after the intervention is calculated to quantify the causal contribution of the behavior to the intention. Taking into account the combined effects of the current context label and all historical behaviors in the causal directed acyclic graph, the Bayesian inference algorithm is used to invert the probability distribution of the user's unobserved potential intentions in the current context, normalize it into a standardized intention inversion probability distribution, and display the several intentions and probabilities that the user is most likely to have in the current context.

[0026] In a specific embodiment, user A's e-commerce purchase behavior characteristic value of 0.8, social sharing behavior characteristic value of 0.6, and weekend outdoor leisure context label are used as processing variables, and potential intentions such as purchasing camping equipment, planning a weekend trip, and checking the weather are used as outcome variables. These are input into the causal forest algorithm for analysis. The algorithm identifies that in the weekend outdoor leisure context, e-commerce purchase behavior has the highest causal effect strength on the purchase of camping equipment, with a heterogeneous effect value of 0.75, while social sharing behavior has the second strongest causal effect strength on the plan of weekend trip, which is 0.60. Based on this, a core node is constructed with the purchase of camping equipment as its core node, and its parent node is the e-commerce purchase behavior. and a causal directed acyclic graph with weekend outdoor leisure context labels; taking the current context as the benchmark, locking the e-commerce add-to-cart behavior node of user A in the graph, performing counterfactual intervention, removing this behavior, and inverting the probability of purchasing camping equipment from the baseline probability of 0.7 to 0.3, quantifying the causal contribution of this behavior to the intention as 0.4; combining the combined effects of the current context and all historical behaviors, using Bayesian reasoning to invert the standardized probability distribution of user A's potential intention is: purchasing camping equipment 0.65, planning a weekend trip 0.25, and checking the weather 0.10, clarifying that his current strongest potential intention is to purchase camping equipment.

[0027] Step S300: Using the constructed causal directed acyclic graph, combined with the user's historical intention sequence and behavior sequence, train an intention state transition model based on the Transformer architecture; Specifically, the inverse probability distribution of the user's intention in a historical time period is defined as the intention state, and the specific behavior generated by the user in this time period is defined as the triggering behavior. The intention state transition sequence for supervised learning is constructed, where each training sample contains the intention state at time t, the triggering behavior, and the intention state at the next time t+1; An intention state transition model based on the Transformer architecture is designed. The input sequence of the model is the intention state sequence and the corresponding trigger behavior sequence, and the output sequence is the subsequent intention state prediction sequence. The constructed causal directed acyclic graph is used as an external knowledge constraint and integrated into the model training process. When the model makes state transition predictions, the legal transition relationships defined in the causal directed acyclic graph are encoded as causal masks, which act on the model's attention layer to ensure that the model learns and follows the state transition path defined in the graph that conforms to the causal logic. The subsequent intention states in historical real data are used as labels to supervise the model training and optimize its accuracy in predicting future intention states.

[0028] In a specific embodiment, the intention state of user A at 14:30 on October 28 is defined as a probability distribution, with purchase of camping equipment 0.65, plan for weekend trip 0.25, and query of weather 0.10. The behavior of purchasing a tent is defined as a triggering behavior. Combined with the intention state of planning a weekend trip 0.8 on October 21 last Saturday and the triggering behavior of publishing a park guide, a supervised learning sample sequence is constructed; an intention state transfer model containing a two-layer Transformer encoder is designed, with the input being the sequence [(intention state on October 21, publish park guide behavior), (intention state on October 28, purchase of tent behavior)], and the model output being the next Prediction of the intention state at each moment; during training, the constructed causal directed acyclic graph is encoded as a causal mask. For example, when calculating attention, the transition probability weight from "purchasing camping equipment" to "planning a weekend trip" is forced to zero to ensure that the model does not learn a transition path that violates the causal logic; finally, the model is supervised and trained with historical real data, such as the subsequent real intention of purchasing camping equipment on October 21. By minimizing the cross-entropy loss between the predicted distribution and the real label, the model is optimized so that it can accurately predict the probability of user A's intention state shifting to purchasing camping equipment 0.85, planning a weekend trip 0.10, and checking the weather 0.05 at the next moment after purchasing a tent.

[0029] Step S400: Based on the intention inversion probability distribution and the intention state transition model, multi-step probability deduction is performed to generate multiple intention evolution paths containing intention nodes, transition probabilities, and expected occurrence time windows; Specifically, the intention inversion probability distribution generated at the current moment is used as the initial state of the multi-step probability deduction to represent the probability of all possible intentions of the user in the current situation; the current deduction state is input into the intention state transition model to predict the probability that the user's intention state will transfer to all other possible intentions within the next time step, and a transition probability distribution is generated; the predicted transition probability distribution is used as the new state of the next iteration. At each state transition, the transition probability from the current intention node to the next intention node is recorded, and an expected occurrence time window is estimated based on the preset life cycle of the intention; After completing the deduction of all steps, a complete set of intention evolution paths is generated based on each recorded transfer information. Each path is a sequence, including the initial intention node, the subsequent intention nodes generated by deduction, and the corresponding transfer probabilities and expected occurrence time windows associated with adjacent nodes in the path. Multiple intention evolution paths with different probabilities are output.

[0030] In a specific embodiment, the inversion probability distribution of user A's intention at the current moment is used as the initial state T0, and the next three days are set as the deduction cycle, with each step being 1 day; the T0 state is input into the intention state transition model, and the model predicts that on the first day T1, the user's intention state will transfer to purchasing camping equipment 0.75, planning a weekend trip 0.15 and checking the weather 0.10, and the transition probability from the initial intention of purchasing camping equipment to itself is recorded as 0.75. According to the preset 1-2 day life cycle of the intention, its time window is estimated to be [this day, the next day]; the T1 state is used as a new input for re-deduction, and the model predicts that on the second day T2, the intention state will transfer to purchasing camping equipment 0.40, planning a weekend trip 0.5 5. Check the weather at 0.05, record the transition probability from purchasing camping equipment to planning a weekend trip as 0.40, and estimate the time window to be [next day, third day] based on the 1-3 day life cycle of the intention to plan a weekend trip. After completing all the steps, two high-probability intention evolution paths are generated. Path one is purchasing camping equipment T0-purchasing camping equipment T1-planning a weekend trip T2, with transition probabilities of 0.75 and 0.40 and time windows of [same day, next day] and [next day, third day]. Path two is purchasing camping equipment T0-planning a weekend trip T1-planning a weekend trip T2, with transition probabilities of 0.25 and 0.55 and time windows of [same day, next day] and [next day, third day].

[0031] Step S500: Based on the predicted intent evolution path, relevant information is retrieved from the existing dynamic knowledge graph, and NLP technology based on template and retrieval enhancement generation is used to create push copy that matches the current context; combined with user location trajectory information, the optimal push timing is calculated.

[0032] Specifically, for each generated intent evolution path, the target intent node in the path is used as the core query condition to search in the dynamic knowledge graph to obtain entities, attributes, and relationships directly related to the intent, forming a candidate information set; Using hybrid NLP technology, based on the current user's contextual tags and intent evolution path, the predefined copy template library is screened for the base template with the highest semantic and stylistic match. The retrieved candidate information set is used as enhanced information and intelligently filled into the corresponding placeholders of the base template using a generative model to generate the final personalized push copy. Analyze the user's mobility patterns, frequent stops, and points of interest, and calculate the spatiotemporal accessibility of the user's geographic location associated with the pushed content within the expected time window. Comprehensively compare the overlap between the accessibility scores of each path and the time window, select the moment with the highest accessibility and the best match with the time window, and activate the push timing prediction window for the intention as the final basis for executing precise push decisions.

[0033] In a specific embodiment, the generated intent evolution path is taken over, with the target intent node of the path, planning a weekend trip, as the core, and the dynamic knowledge graph is searched to obtain related entities such as a suburban park, sunny weather, barbecue grill rental points, etc., to form a candidate information set; based on the contextual label of user A's current weekend outdoor leisure, a basic template is screened from the copy template library, and the retrieved suburban park, sunny weather, and barbecue grill rental points are used as enhanced information, and intelligently filled in through a generative model to generate the final personalized push copy; the GPS trajectory of user A is analyzed, and it is found that his permanent place is home, his point of interest is the suburban park, and the expected occurrence time window is [next day, third day]. It is calculated that the spatiotemporal accessibility score of the user leaving home for the suburban park at 10:00 the next day is the highest, which is 0.95, and the overlap with the time window is the best. Therefore, 10:00 the next day is selected as the best push time, and the push time prediction window of the intent is activated to prepare for precise push.

[0034] like Figure 2 As shown in the program flow chart of a personalized information precision push method based on artificial intelligence, the present invention provides a personalized information precision push method based on artificial intelligence, comprising: Through multi-source data collection, user behavior data on e-commerce, social networking, and health and fitness platforms is captured in real time, and external spatiotemporal environmental data and user location trajectory information are simultaneously obtained. The collected multi-source heterogeneous data is fused and constructed into a high-dimensional spatiotemporal feature vector. The DBSCAN algorithm, which introduces a time decay factor and behavior intensity weights, is used for processing to identify the user's current specific scenario and calculate a confidence level for each context label. The match between new data and the current context is continuously evaluated, and recalculation is triggered when a mismatch occurs. Feature engineering is performed, and the generated context labels are converted into vectors through one-hot encoding. The user's historical behavior sequence is constructed into a structured behavior event vector. After the two are spliced ​​in chronological order, they are input into a pre-trained feature embedding model to generate an embedded feature matrix that captures the deep semantic associations between features. Using behavioral and situational features as processing variables and potential intent as the outcome variable, the causal forest algorithm is used to quantify the heterogeneous causal effects between them and construct a causal directed acyclic graph. Counterfactual reasoning is performed to invert the user's potential intent that is not directly observed in the current context, ultimately outputting a standardized probability distribution of intent. A Transformer-based intent state transition model is trained using historical data. The model's training process incorporates the aforementioned causal graph as a constraint to ensure that the transition paths learned by the model conform to causal logic. Based on the current intent state and triggering behavior, the model accurately predicts the intent state at the next moment. Based on the trained model and the currently inverted intent, a multi-step probabilistic deduction is performed, using the previous moment's output as input. Changes over multiple time steps in the future are recursively predicted, generating multiple intent evolution paths containing intent nodes, transition probabilities, and expected occurrence time windows, depicting the various possible future evolution trajectories of the user's intent. Based on the predicted intent path, relevant information is retrieved from the dynamic knowledge graph, and NLP technology based on template and retrieval enhancement generation is used to automatically create personalized push copy that is highly matched with the current context. The user's real-time location trajectory, movement pattern and time window of intent are comprehensively analyzed to calculate the best push timing with optimal spatiotemporal accessibility. At the moment of calculation, the generated personalized information is accurately pushed to the user.

[0035] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above and that the invention can be embodied in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered in all respects as illustrative and non-restrictive, and the scope of the invention is defined by the appended claims, not the foregoing description, and all variations within the meaning and range of equivalents of the claims are intended to be included therein. Any reference sign in a claim should not be construed as limiting the claim to which it relates.

Claims

1. A personalized information accurate push method based on artificial intelligence, characterized in that: include: Collect user behavior data on e-commerce, social networking, and health and fitness platforms in real time, and obtain external spatiotemporal environment data and user location trajectory information; The collected data is fused and an improved spatiotemporal density clustering algorithm is used to generate context labels with confidence levels to define the specific scenario the user is currently in. The generated context labels and the user's historical behavior sequence are used as input for feature engineering. A causal forest algorithm is used to quantitatively analyze the heterogeneous causal effects between behavioral data, context labels, and potential intent, constructing a causal directed acyclic graph. This graph is then used to perform counterfactual reasoning, inverting the user's potential intent in the current context and outputting a standardized probability distribution for the inverted intent. Using the constructed causal directed acyclic graph and combining the user's historical intention sequence and behavior sequence, we train an intention state transition model based on the Transformer architecture. Based on the intention inversion probability distribution and intention state transition model, multi-step probability deduction is performed to generate multiple intention evolution paths containing intention nodes, transition probabilities, and expected occurrence time windows; Based on the predicted intent evolution path, relevant information is retrieved from the existing dynamic knowledge graph, and NLP technology based on template and retrieval enhancement generation is used to create push copy that matches the current context; combined with user location trajectory information, the optimal push timing is calculated.

2. The method for accurately pushing personalized information based on artificial intelligence according to claim 1, characterized in that: The real-time collection of user behavior data on e-commerce, social networking, and health and fitness platforms, and acquisition of external spatiotemporal environment data and user location trajectory information, includes: Real-time collection of users' product browsing, add-to-cart, purchase, and evaluation behaviors on e-commerce platforms; collection of users' content publishing, likes, comments, favorites, and follow-up behaviors on social platforms; and collection of users' exercise records on health and fitness platforms. Acquire external spatiotemporal environmental data in real time, including timestamps, weather conditions, and holiday information; obtain the user's geographic location, movement path, movement speed, permanent locations, and points of interest.

3. The method for accurately pushing personalized information based on artificial intelligence according to claim 1, characterized in that: The collected data is fused and an improved spatiotemporal density clustering algorithm is used to generate a context label with confidence, defining the specific scenario the user is currently in, including: The collected behavioral data, external spatiotemporal environment data, and user location trajectory information are integrated into multi-source heterogeneous data to construct a high-dimensional spatiotemporal feature vector. The behavioral data is quantified into behavior type and behavior intensity using preset rules, and the above features are normalized to form a unified time series feature dataset. An improved DBSCAN algorithm based on density peak and time decay factor is used to perform cluster analysis on time series feature data sets. Time decay factor and behavior intensity weight are introduced to calculate the influence of data points in the cluster. For each cluster formed after cluster analysis, the core features of all data points within the cluster are aggregated to generate a preset context label. Based on the compactness of the cluster, the temporal continuity of the data points within the cluster, and the consistency of the behavioral intensity, the confidence level of each context label is calculated using a weighted summation algorithm. The generated context label with confidence is used as the definition of the specific scenario the user is currently in. New user data is continuously received to trigger a re-evaluation of the current scenario. When the matching degree between the new data point and the current context label is lower than the preset threshold, or a new high-confidence context label appears, the user's current scenario definition is updated.

4. The method for accurately pushing personalized information based on artificial intelligence according to claim 1, characterized in that: The generated context labels and user historical behavior sequences are used as input to perform feature engineering, including: The generated context labels with confidence are mapped into high-dimensional sparse vectors through one-hot encoding. Each behavior event in the user's historical behavior sequence is constructed into a structured behavior event vector based on the behavior type, behavior intensity, and timestamp. The current context label vector and the user's historical behavior sequence vector within the preset retrospective time window are concatenated in chronological order to form a mixed feature sequence that combines the current context and historical behavior. The mixed feature sequence is input into a pre-trained feature embedding model, which maps each vector into a unified continuous vector space and generates an embedding feature matrix that can capture the semantic associations between features.

5. The method for accurately pushing personalized information based on artificial intelligence according to claim 1, characterized in that: The causal forest algorithm is used to quantitatively analyze the heterogeneous causal effects between behavioral data, context labels, and potential intentions, and to construct a causal directed acyclic graph, including: The behavioral and situational features in the generated embedded feature matrix are used as treatment variables, and the potential intention is used as the outcome variable. The causal forest algorithm is used to analyze the causal effect strength of each treatment variable on each potential intention outcome under different situational and behavioral combinations, identify the key causal factors that have the greatest impact on specific intentions, and reveal the heterogeneity of causal effects. Based on the heterogeneous causal effects analyzed by the causal forest algorithm, a causal directed acyclic graph is constructed with potential intentions as core nodes and behavioral data and situational labels as parent nodes.

6. The method for accurately pushing personalized information based on artificial intelligence according to claim 1, characterized in that: The method of using the graph to perform counterfactual reasoning, inverting the user's potential intention in the current situation, and outputting a standardized probability distribution of intention inversion includes: Using the defined current user context label as the baseline context, the specific behavior node of the current user is locked in the constructed causal directed acyclic graph; Perform a counterfactual intervention in a causal directed acyclic graph. By removing the current user's specific behavior, the probability distribution of potential intention nodes is changed. The difference in intention probability before and after the intervention is calculated to quantify the causal contribution of the behavior to the intention. Taking into account the combined effects of the current context label and all historical behaviors in the causal directed acyclic graph, the Bayesian inference algorithm is used to invert the probability distribution of the user's unobserved potential intentions in the current context, normalize it into a standardized intention inversion probability distribution, and display the several intentions and probabilities that the user is most likely to have in the current context.

7. The method for accurately pushing personalized information based on artificial intelligence according to claim 1, characterized in that: The constructed causal directed acyclic graph is combined with the user's historical intention sequence and behavior sequence to train an intention state transition model based on the Transformer architecture, including: The inverse probability distribution of the user's intention in a historical time period is defined as the intention state, and the specific behavior generated by the user in this time period is defined as the triggering behavior. The intention state transition sequence for supervised learning is constructed, where each training sample contains the intention state at time t, the triggering behavior, and the intention state at the next time t+1; An intention state transition model based on the Transformer architecture is designed. The input sequence of the model is the intention state sequence and the corresponding trigger behavior sequence, and the output sequence is the subsequent intention state prediction sequence. The constructed causal directed acyclic graph is used as an external knowledge constraint and integrated into the model training process. When the model makes state transition predictions, the legal transition relationships defined in the causal directed acyclic graph are encoded as causal masks, which act on the model's attention layer to ensure that the model learns and follows the state transition path defined in the graph that conforms to the causal logic. The subsequent intention states in historical real data are used as labels to supervise the model training and optimize its accuracy in predicting future intention states.

8. The method for accurately pushing personalized information based on artificial intelligence according to claim 1, characterized in that: Based on the intention inversion probability distribution and intention state transition model, multi-step probability deduction is performed to generate multiple intention evolution paths containing intention nodes, transition probabilities and expected occurrence time windows, including: The intention inversion probability distribution generated at the current moment is used as the initial state of the multi-step probability deduction to represent the probabilities of all possible user intentions in the current context. The current deduced state is input into the intention state transition model to predict the probability that the user's intention state will transition to all other possible intentions within the next time step, generating a transition probability distribution. The predicted transition probability distribution is used as the new state for the next iteration. At each state transition, the transition probability from the current intention node to the next intention node is recorded, and an expected occurrence time window is estimated based on the preset life cycle of the intention. After completing the deduction of all steps, a complete set of intention evolution paths is generated based on each recorded transfer information. Each path is a sequence, including the initial intention node, the subsequent intention nodes generated by deduction, and the corresponding transfer probabilities and expected occurrence time windows associated with adjacent nodes in the path. Multiple intention evolution paths with different probabilities are output.

9. The method for accurately pushing personalized information based on artificial intelligence according to claim 1, characterized in that: According to the predicted intent evolution path, relevant information is retrieved from the existing dynamic knowledge graph, and NLP technology based on template and retrieval enhancement generation is used to create push copy that matches the current context; Combined with user location information, the optimal push timing is calculated, including: For each generated intent evolution path, the target intent node in the path is used as the core query condition to search in the dynamic knowledge graph to obtain entities, attributes, and relationships directly related to the intent, forming a candidate information set; Using hybrid NLP technology, based on the current user's contextual tags and intent evolution path, the predefined copy template library is screened for the base template with the highest semantic and stylistic match. The retrieved candidate information set is used as enhanced information and intelligently filled into the corresponding placeholders of the base template using a generative model to generate the final personalized push copy. Analyze the user's mobility patterns, frequent stops, and points of interest, and calculate the spatiotemporal accessibility of the user's geographic location associated with the pushed content within the expected time window. Comprehensively compare the overlap between the accessibility scores of each path and the time window, select the moment with the highest accessibility and the best match with the time window, and activate the push timing prediction window for the intention as the final basis for executing precise push decisions.

10. A personalized information precision push system based on artificial intelligence, using the personalized information precision push method based on artificial intelligence according to any one of claims 1 to 9, characterized in that: include: Multi-source data perception and fusion module: This module includes a data acquisition unit and a data fusion and contextualization unit. The data acquisition unit is responsible for collecting raw data streams in real time from three dimensions: user-platform interaction, external environment, and user location trajectory. The data fusion and contextualization unit pre-processes the collected multi-source heterogeneous data and aggregates discrete data points into contextual labels with confidence levels using an improved spatiotemporal clustering algorithm. Causal Intention Inference Module: This module includes a causal effect analysis unit, an intention inversion unit, and an intention state transition modeling unit. The causal effect analysis unit uses a causal forest algorithm to quantitatively analyze the heterogeneous causal effects between behaviors, contexts, and intentions, identify the key factors driving user intentions, and construct a causal directed acyclic graph based on this. The intention inversion unit uses the constructed causal graph to perform counterfactual reasoning. Given the current context and behavior, it inverts the user's most likely potential intentions and probabilities, and outputs a standardized intention inversion probability distribution. The intention state transition modeling unit combines historical intention sequences and behavior sequences to train an intention state transition model based on the Transformer architecture. Using the causal graph as an external knowledge constraint, it learns the transition rules between intention states that conform to causal logic for predicting future intentions. Intention evolution path deduction module: This module includes a multi-step probability deduction unit. This unit uses the current intention inversion probability distribution as the starting point and iteratively calls the intention state transition model to generate multiple complete intention evolution paths that include future intention nodes, transition probabilities between nodes, and expected occurrence time windows. Personalized content generation module: This module includes a knowledge graph retrieval unit and a hybrid NLP generation unit. The knowledge graph retrieval unit uses the target intent in the intent evolution path as a query condition to retrieve relevant entities, attributes, and relationships in the dynamic knowledge graph. The hybrid NLP generation unit combines the retrieved information with the current user's contextual tags, selects the most matching basic template from the template library, and uses a generative model to intelligently fill in the retrieved information to generate the final personalized push copy. Intelligent timing decision and execution module: includes a spatiotemporal reachability analysis unit and a push timing activation and execution unit; the spatiotemporal reachability analysis unit combines the user's location trajectory information and the geographic location associated with the pushed content to calculate the spatiotemporal reachability of the user to the relevant location within the expected time window of the intention evolution path; the push timing activation and execution unit comprehensively compares the reachability scores under each intention evolution path and the overlap of the time window, selects the optimal path and the timing with the highest matching degree, activates the final push timing prediction window, and triggers the system to execute the push operation.

Images