An artificial intelligence-based personalized information precise pushing system and method

Abstract

The application discloses an individualized information accurate pushing system and method based on artificial intelligence, and relates to the technical field of individualized recommendation.The application generates a context label with confidence by fusing user multi-platform behavior data and external space-time environment information and using an improved space-time density clustering algorithm; a causal directed acyclic graph is constructed by using a causal forest algorithm, heterogeneous causal effects are quantitatively analyzed, a user potential intention is reversed, and a standardized intention reverse probability distribution is output; a Transformer intention state transition model which is causally constrained by the causal directed acyclic graph is trained in combination with a historical intention and a behavior sequence, multi-step probability deduction is performed, and an intention evolution path is generated; information is retrieved from a dynamic knowledge graph according to the predicted path, a push script of a matching context is generated by using NLP technology, and the best push time is calculated in combination with a user position trajectory, so that the accurate touch of individualized information is realized.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of personalized recommendation, and particularly relates to an intelligent personalized information accurate pushing system and method. BACKGROUND

[0002] With the rapid development of information technology, personalized information pushing has become a key means to improve user experience and business value. However, the existing personalized pushing technology generally has defects, which seriously restricts its accuracy and effectiveness.

[0003] In the aspect of user portrait construction, the existing technology mostly relies on single platform or single type of data, and it is difficult to form a comprehensive and three-dimensional user view. This data island leads to a one-sided and static user portrait, which cannot capture the dynamic demand changes of users in different scenarios. In the aspect of intent recognition and analysis, the existing mainstream technology is mostly based on correlation analysis, that is, the correlation between user behavior and item click rate, purchase rate and the like is counted to infer user interest, but the deep causal logic behind the behavior cannot be revealed. The ambiguity of causal relationship leads to low accuracy of intent recognition, and the core motivation and potential demand of users cannot be accurately captured due to the interference of irrelevant behaviors. In the aspect of future behavior prediction, the existing technology mostly stays in the prediction of current or short-term interest of users, and lacks the dynamic deduction ability of future behavior trend, so that the pushing time selection completely depends on simple rules or thresholds, and the evolution path and best triggering time of user intent cannot be predicted, resulting in that the pushing content is ignored when the user does not have demand, or the pushing content appears when the user's demand has passed, which affects the conversion effect of pushing and user experience. SUMMARY

[0004] The application aims to provide an intelligent personalized information accurate pushing system and method to solve the problems in the prior art.

[0005] To achieve the above-mentioned purpose, the application provides the following technical scheme.

[0006] In a first aspect, the application provides an intelligent personalized information accurate pushing method based on artificial intelligence, comprising:

[0007] Real-time collection of user behavior data on e-commerce, social and health and fitness platforms, acquisition of external spatio-temporal environment data and user location trajectory information;

[0008] Fusion of the collected data, generation of a context label with confidence by using an improved spatio-temporal density clustering algorithm, definition of a specific scene in which the user is currently located, and feature engineering by taking the generated context label and user historical behavior sequence as input;

[0009] The heterogeneous causal effects among the behavior data, the context label and the potential intention are quantitatively analyzed by using a causal forest algorithm, and a causal directed acyclic graph is constructed; the potential intention of the user in the current context is reversed by using the graph for counterfactual reasoning, and a standardized intention reverse probability distribution is output;

[0010] By using the constructed causal directed acyclic graph, the intention state transition model based on the Transformer architecture is trained in combination with the historical intention sequence and the behavior sequence of the user;

[0011] Based on the intention reverse 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 predicted time windows;

[0012] According to the predicted intention evolution path, relevant information is retrieved from the existing dynamic knowledge graph, and NLP technology based on template and retrieval enhancement is used to create a push script matched with the current context; in combination with the user location trajectory information, the best push time is calculated.

[0013] In combination with the first aspect, in a first implementation manner of the first aspect of the application, the real-time collection of the behavior data of the user on the e-commerce, social and health and sports platforms, the acquisition of the external space-time environment data and the user location trajectory information, include:

[0014] The behavior data of the user on the e-commerce platform, including commodity browsing, adding to cart, purchasing and evaluating; the behavior data of the user on the social platform, including content publishing, liking, commenting, collecting and following; and the behavior data of the user on the health and sports platform, including exercise recording are collected in real time;

[0015] The external space-time environment data, including time stamp, weather condition and holiday information, are acquired in real time; and the geographical location, moving path, moving speed, residence point and interest point of the user are acquired.

[0016] In combination with the first aspect, in a second implementation manner of the first aspect of the application, the collected data is fused, and an improved space-time density clustering algorithm is used to generate a context label with a confidence degree, and a specific scene in which the user is currently located is defined, including:

[0017] The collected behavior data, external space-time environment data and user location trajectory information are fused to construct a high-dimensional space-time feature vector; the behavior data is quantified into behavior types and behavior intensities through a preset rule, and the above features are normalized to form a unified time series feature data set;

[0018] An improved DBSCAN algorithm based on density peak value and time decay factor is used for clustering analysis of the time series feature data set, a time decay factor and a behavior intensity weight are introduced, and the influence of the data points in clustering is calculated.

[0019] For each cluster formed after clustering analysis, aggregate the core features of all data points in the cluster to generate a preset context label; based on the compactness of the cluster, the consistency of the time continuity and behavior intensity of the data points in the cluster, calculate the confidence of each context label through a weighted summation algorithm;

[0020] Take the generated context label with confidence as the definition of the specific scene where the user is currently located, continuously receive new user data, and trigger reevaluation of the current scene; when the matching degree of the new data point with the current context label is lower than a preset threshold, or a new high-confidence context label appears, update the current scene definition of the user.

[0021] In combination with the first aspect, in a third implementation manner of the first aspect of the application, the generated context label and the user historical behavior sequence are taken as inputs for feature engineering, including:

[0022] The generated context label with confidence is mapped into a high-dimensional sparse vector through one-hot encoding; each behavior event in the user historical behavior sequence is constructed into a structured behavior event vector according to the behavior type, behavior intensity and timestamp;

[0023] The context label vector at the current moment and the historical behavior sequence vector of the user within a preset backtracking time window are spliced in time sequence, forming a mixed feature sequence that fuses the current context and historical behavior;

[0024] The mixed feature sequence is input into a pre-trained feature embedding model to map each vector into a unified continuous vector space, generating an embedding feature matrix that can capture the semantic association between features.

[0025] In combination with the first aspect, in a fourth implementation manner of the first aspect of the application, the causal forest algorithm is used to quantitatively analyze the heterogeneous causal effects between behavior data, context labels and potential intentions, and a causal directed acyclic graph is constructed, including:

[0026] The behavior features and context features in the generated embedding feature matrix are taken as processing variables, and the potential intentions are taken as result variables; through the causal forest algorithm, the causal effect strength of each processing variable on each potential intention result under different context and behavior combinations is analyzed, the key causal factors that have the greatest impact on a specific intention are identified, and the heterogeneity of the causal effects is revealed;

[0027] 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 behavior data and context labels as parent nodes.

[0028] In a fifth implementation form of the first aspect, the using the graph to perform counterfactual reasoning to reverse the potential intention of the user in the current situation and output a standardized intention reverse probability distribution comprises:

[0029] Taking the defined current user situation label as a reference situation, locking the specific behavior node of the current user in the constructed causal directed acyclic graph;

[0030] Performing counterfactual intervention in the causal directed acyclic graph, obtaining the probability distribution change of the potential intention node by removing the specific behavior of the current user; calculating the difference of intention probability before and after intervention to quantify the causal contribution of the behavior to the intention;

[0031] Comprehensively considering the comprehensive effect of the current situation label and all historical behaviors in the causal directed acyclic graph, reversing the probability distribution of the unobserved potential intention of the user in the current situation by a Bayesian inference algorithm, normalizing it into a standardized intention reverse probability distribution, and displaying several most possible intentions and probabilities of the user in the current situation.

[0032] In a sixth implementation form of the first aspect, the using the constructed causal directed acyclic graph to train an intention state transition model based on a Transformer architecture in combination with the historical intention sequence and the behavior sequence of the user comprises:

[0033] Defining the intention reverse probability distribution of the user in a historical time period as an intention state, defining the specific behavior of the user in the time period as a trigger behavior, and constructing an intention state transition sequence for supervised learning, wherein each training sample contains an intention state at time t, a trigger behavior, and an intention state at next time t+1;

[0034] Designing an intention state transition model based on a Transformer architecture, the input sequence of the model being an intention state sequence and a corresponding trigger behavior sequence, and the output sequence being a subsequent intention state prediction sequence; using the constructed causal directed acyclic graph as external knowledge constraints and integrating it into the training process of the model, encoding the legal transition relationship defined in the causal directed acyclic graph into a causal mask to act on the attention layer of the model to ensure that the model learns and follows the state transition path defined in the graph and consistent with the causal logic; using the subsequent intention state in the historical real data as a label to supervise the training of the model and optimize the accuracy of the prediction of future intention states.

[0035] In a seventh implementation form of the first aspect, the method further includes: performing multi-step probabilistic deduction based on the intention inversion probability distribution and the intention state transition model to generate a plurality of intention evolution paths each including an intention node, a transition probability and a predicted occurrence time window.

[0036] The intention inversion probability distribution generated at the current time is taken as an initial state of the multi-step probabilistic deduction, representing probabilities of all possible intentions of the user in the current context; the current deduction state is input into the intention state transition model to predict probabilities that the intention state of the user will transition to all other possible intentions in a future time step, generating a transition probability distribution; the predicted transition probability distribution is taken as a new state for next iteration, and the transition probability from the current intention node to the next intention node is recorded at each state transition; and a predicted occurrence time window is estimated according to a preset life cycle of the intention.

[0037] After the deduction of all steps is completed, a complete set of intention evolution paths is generated based on the recorded transition information each time, each path being a sequence including an initial intention node, subsequent intention nodes generated by deduction, and a transition probability and a predicted occurrence time window associated between adjacent nodes in the path, and the method outputs a plurality of intention evolution paths with different probabilities.

[0038] In an eighth implementation form of the first aspect, the method further includes: retrieving relevant information from an existing dynamic knowledge graph according to the predicted intention evolution paths, and creating a push script matched with the current context by using NLP technology generated based on a template and retrieval enhancement; and calculating an optimal push timing in combination with user location trajectory information.

[0039] For each generated intention evolution path, a target intention node in the path is taken as a core query condition to perform retrieval in the dynamic knowledge graph, to obtain entities, attributes and relationships directly associated with the intention, forming a candidate information set;

[0040] A hybrid NLP technology is adopted to filter, from a predefined script template library, a basic template with the highest semantic and style matching degree according to the context label and the intention evolution path of the current user, to take the retrieved candidate information set as enhanced information, and to use a generative model to intelligently fill the enhanced information into a corresponding placeholder in the basic template to generate a final personalized push script.

[0041] The movement mode, residence point and interest point of the user are analyzed, the spatio-temporal accessibility of the user to the geographical position associated with the push content within a predicted time window is calculated, the accessibility scores under each path are comprehensively compared with the coincidence degree of the time window, the time point with the highest accessibility and the optimal matching degree with the time window is selected, the push time prediction window of the intention is activated, and the final basis for executing the precise push decision is obtained.

[0042] In a second aspect, the present application provides an artificial intelligence-based personalized information precise push system, comprising:

[0043] A multi-source data perception and fusion module includes a data acquisition unit and a data fusion and contextualization unit. The data acquisition unit is responsible for real-time acquisition of raw data streams from three dimensions of user platform interaction, external environment and user location trajectory. The data fusion and contextualization unit preprocesses the collected multi-source heterogeneous data, and aggregates discrete data points into context labels with confidence through an improved spatio-temporal clustering algorithm.

[0044] A causal intention inference 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, identifies key factors driving user intentions, and builds a causal directed acyclic graph based on this. The intention inversion unit uses the built causal graph to perform counterfactual reasoning to infer several potential intentions and probabilities that the user is most likely to have under the given current context and behavior, and outputs a standardized intention inversion probability distribution. The intention state transition modeling unit trains an intention state transition model based on the Transformer architecture in combination with historical intention sequences and behavior sequences, learns the transition rules between intention states that conform to causal logic using the causal graph as external knowledge constraints, and is used to predict future intentions.

[0045] An intention evolution path deduction module includes a multi-step probability deduction unit that takes the current intention inversion probability distribution as a starting point, iteratively calls the intention state transition model, and generates complete intention evolution paths containing future intention nodes, transition probabilities between nodes and predicted time windows.

[0046] A personalized content generation module includes a knowledge graph retrieval unit and a hybrid NLP generation unit. The knowledge graph retrieval unit uses the target intention in the intention evolution path as a query condition to retrieve related entities, attributes and relationships in a dynamic knowledge graph. The hybrid NLP generation unit combines the retrieved information and the context label of the current user, 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 script.

[0047] The intelligent timing decision-making and execution module includes a spatiotemporal accessibility analysis unit and a push timing activation and execution unit. The spatiotemporal accessibility analysis unit combines the user's location trajectory information and the geographic location associated with the push content to calculate the spatiotemporal accessibility of the user to the relevant location within the predicted time window of the intent evolution path. The push timing activation and execution unit comprehensively compares the accessibility scores and the overlap of the time windows under each intent evolution path, 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.

[0048] Compared with the prior art, the beneficial effects of the present invention are:

[0049] 1. This invention integrates multi-dimensional data such as user behavior, spatiotemporal environment, and location trajectory, and generates contextual labels with confidence through an improved clustering algorithm, thereby achieving an accurate definition of the specific scenario in which the user is located.

[0050] 2. This invention introduces the causal forest algorithm to quantitatively analyze the heterogeneous causal effects between behavior, context and intention, constructs a causal directed acyclic graph, and accurately inverts the user's potential intention in the current context through counterfactual reasoning.

[0051] 3. This invention combines causal graphs and user history sequences to train a Transformer-based intent state transition model, and generates an intent evolution path containing intent nodes, transition probabilities, and time windows through multi-step probabilistic deduction. Attached Figure Description

[0052] Fig. 1 This is a schematic diagram illustrating the steps of a personalized information delivery method based on artificial intelligence according to the present invention.

[0053] Fig. 2 This is a flowchart of a personalized information delivery method based on artificial intelligence according to the present invention. Detailed Implementation

[0054] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0055] Example: Figs. 1-2 As shown, the present invention provides a technical solution.

[0056] like Fig. 1A schematic diagram illustrating the steps of a personalized information delivery method based on artificial intelligence is provided in this invention. The method includes:

[0057] Step S100: Collect user behavior data in real time on e-commerce, social and health and fitness platforms, obtain external spatiotemporal environment data and user location trajectory information; fuse the collected data, use an improved spatiotemporal density clustering algorithm to generate context labels with confidence, and define the specific scene in which the user is currently located; use the generated context labels and the user's historical behavior sequence as input to perform feature engineering;

[0058] Specifically, it collects users' browsing, adding to cart, purchasing, and reviewing behaviors on e-commerce platforms in real time; it collects users' content posting, liking, commenting, collecting, and following behaviors on social media platforms; and it collects users' exercise records on health and fitness platforms.

[0059] Real-time acquisition of external spatiotemporal environmental data, including timestamps, weather conditions, and holiday information; acquisition of user's geographic location, movement path, movement speed, frequent locations, and points of interest.

[0060] The collected behavioral data, external spatiotemporal environment data, and user location trajectory information are fused from multiple sources to construct a high-dimensional spatiotemporal feature vector; the behavioral data is quantified into behavior type and behavior intensity according to preset rules, and the above features are normalized to form a unified time series feature dataset;

[0061] An improved DBSCAN algorithm based on density peak and time decay factor is used to perform cluster analysis on time series feature datasets. The time decay factor and behavior intensity weight are introduced to calculate the influence of data points in clustering.

[0062] For each cluster formed after clustering analysis, the core features of all data points within the cluster are aggregated to generate a preset context label; based on the cluster density, the temporal continuity of data points within the cluster, and the consistency of behavioral intensity, the confidence level of each context label is calculated using a weighted summation algorithm.

[0063] The generated context labels with confidence levels are used as the definition of the user's current specific scenario. New user data is continuously received, triggering a re-evaluation of the current scenario. When the matching degree between new data points and the current context labels is lower than a preset threshold, or when new high-confidence context labels appear, the definition of the user's current scenario is updated.

[0064] The generated contextual labels with confidence scores are mapped into high-dimensional sparse vectors through one-hot encoding; each behavioral event in the user's historical behavior sequence is constructed into a structured behavioral event vector based on the behavior type, behavior intensity, and timestamp.

[0065] The current context label vector and the user's historical behavior sequence vector within the preset backtracking time window are concatenated in chronological order to form a hybrid feature sequence that integrates the current context and historical behavior.

[0066] The mixed feature sequence is input into a pre-trained feature embedding model, which maps each vector to a unified continuous vector space to generate an embedding feature matrix that can capture the semantic relationships between features.

[0067] In one 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 their shopping cart. Subsequently, user A posted a status update on a social media platform about going to a suburban park over the weekend and received likes. Simultaneously, a health and fitness platform recorded user A's GPS trajectory from home at 15:00, with a fixed location at longitude 116.4°, latitude 39.9°, a movement speed of 15 km / h, and a point of interest at a suburban park at longitude 116.5°, latitude 40.1°. This behavioral data, external data, and location trajectory were fused into a high-dimensional spatiotemporal feature vector. The behavioral data was quantified as a shopping interest intensity of 0.8 and a social sharing intensity of 0.6. After normalization, a time-series dataset was formed. An improved DBSCAN algorithm was used for clustering, with a time decay factor α = 0.5 and a behavioral intensity weight β = 0. 7. Analysis revealed that the data points were densely clustered and temporally continuous within the weekend outdoor leisure cluster, specifically between 14:00 and 15:30. This cluster aggregated core features such as e-commerce shopping, social sharing, and movement to suburban parks, generating the contextual label "weekend outdoor leisure." A confidence score of 0.85 was calculated using a weighted average of density (0.85), temporal continuity (0.9), and behavioral consistency (0.8). This label was mapped to a vector [0, 0, 1, 0] using one-hot encoding. The preset label set was [working from home, commuting to exercise, weekend outdoor leisure, weekday leisure]. The user's historical behavior sequence, including the vector of last Saturday's park browsing guide [0.2, 0.1, 0.7, 0.1], was concatenated with the current contextual vector and input into a pre-trained embedding model. This generated an embedding feature matrix that captures the semantic association between outdoor shopping and social sharing, providing feature input for subsequent intent inversion.

[0068] Step S200: Using the causal forest algorithm, the heterogeneous causal effects between behavioral data, contextual labels, and latent intentions are quantitatively analyzed to construct a causal directed acyclic graph; the graph is used to perform counterfactual reasoning to invert the user's latent intentions in the current context and output a standardized probability distribution of intention inversion.

[0069] Specifically, behavioral and contextual features in the generated embedded feature matrix are used as processing variables, and latent intentions are used as outcome variables. The causal forest algorithm is used to analyze the causal effect strength of each processing variable on the outcome of each latent intention under different combinations of context and behavior, identify the key causal factors that have the greatest impact on a specific intention, and reveal the heterogeneity of causal effects.

[0070] Based on the heterogeneous causal effects analyzed by the causal forest algorithm, a causal directed acyclic graph is constructed with latent intentions as the core nodes and behavioral data and contextual labels as the parent nodes.

[0071] 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;

[0072] Counterfactual intervention is performed in a causal directed acyclic graph by removing the current user's specific behavior to obtain the change in the probability distribution of potential intent nodes; the difference in intent probability before and after the intervention is calculated to quantify the causal contribution of the behavior to the intent.

[0073] Taking into account the combined effect of the current context label and all historical behaviors in the causal directed acyclic graph, the probability distribution of the user's unobserved potential intentions in the current context is inverted using the Bayesian inference algorithm. This distribution is then normalized into a standardized intention inversion probability distribution, showing the most likely intentions and probabilities that the user will generate in the current context.

[0074] In one specific embodiment, user A's e-commerce add-to-cart behavior feature value (0.8), social sharing behavior feature value (0.6), and weekend outdoor leisure context label are used as processing variables, while potential intentions such as purchasing camping equipment, planning a weekend trip, and checking the weather are used as outcome variables. These are input into a causal forest algorithm for analysis. The algorithm identifies that in the weekend outdoor leisure context, e-commerce add-to-cart behavior has the strongest causal effect on purchasing camping equipment, with a heterogeneity effect value of 0.75, while social sharing behavior has the second strongest causal effect on planning a weekend trip, at 0.60. Based on this, a node is constructed with purchasing camping equipment as the core node, and e-commerce add-to-cart behavior as its parent node. A causal directed acyclic graph (DAG) of the weekend outdoor leisure scenario tag was constructed. Based on the current scenario, the e-commerce add-to-cart behavior node of user A was identified in the graph. Counterfactual intervention was performed to remove this behavior, and the probability of purchasing camping equipment decreased from the baseline probability of 0.7 to 0.3. The causal contribution of this behavior to the intention was quantified as 0.4. Combining the combined effect of the current scenario and all historical behaviors, the standardized probability distribution of user A's potential intention was derived through Bayesian inference: purchasing camping equipment 0.65, planning a weekend trip 0.25, checking the weather 0.10. It was determined that his strongest potential intention at present is to purchase camping equipment.

[0075] Step S300: Using the constructed causal directed acyclic graph, combined with the user's historical intent sequence and behavior sequence, train an intent state transition model based on the Transformer architecture.

[0076] Specifically, the probability distribution of user intent inversion within a historical time period is defined as the intent state, and the specific behavior generated by the user within that time period is defined as the triggering behavior. An intent state transition sequence for supervised learning is constructed, wherein each training sample contains the intent state at time t, the triggering behavior, and the intent state at the next time t+1.

[0077] Design an intent state transition model based on the Transformer architecture. The input sequence of the model is the intent state sequence and the corresponding trigger behavior sequence, and the output sequence is the subsequent intent state prediction sequence. A constructed causal directed acyclic graph (DAG) is used as an external knowledge constraint and integrated into the model's training process. When the model makes state transition predictions, the legal transition relationships defined in the DAG are encoded as causal masks and applied to the model's attention layer to ensure that the model learns and follows the causal logic-compliant state transition paths defined in the graph. Subsequent intent states from historical real data are used as labels to supervise the model's training and optimize its accuracy in predicting future intent states.

[0078] In one specific embodiment, user A's intent state at 14:30 on October 28th is defined as a probability distribution: purchasing camping equipment (0.65), planning a weekend trip (0.25), and checking the weather (0.10). Their act of adding a tent to their cart is defined as the triggering behavior. Combining this with their intent state from the previous Saturday, October 21st (planning a weekend trip (0.8) and the triggering behavior of posting a park guide), a supervised learning sample sequence is constructed. An intent state transition model containing a two-layer Transformer encoder is designed. The input is the sequence [(October 21st intent state, posting park guide behavior), (October 28th intent state, adding tent to cart behavior)], and the model output is the probability distribution for the next... The model predicts the intent 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 "buying camping equipment" to "planning a weekend trip" is forcibly set to zero to ensure that the model does not learn transition paths that violate causal logic. Finally, the model is trained under supervision using historical real data, such as the real intent after October 21st being "buying camping equipment". By minimizing the cross-entropy loss between the predicted distribution and the real label, the model is optimized to accurately predict the probability that user A's intent state will transition to "buying camping equipment" (0.85), "planning a weekend trip" (0.10), and "checking the weather" (0.05) at the next moment after adding a tent.

[0079] Step S400: Based on the intent inversion probability distribution and intent state transition model, perform multi-step probability deduction to generate multiple intent evolution paths that include intent nodes, transition probabilities, and expected occurrence time windows;

[0080] Specifically, the intent inversion probability distribution generated at the current moment is used as the initial state for multi-step probability deduction, representing the probability of all possible intentions of the user in the current context; the current deduction state is input into the intent state transition model to predict the probability that the user's intent state will transition to all other possible intentions within a future time step, generating a transition probability distribution; the predicted transition probability distribution is used as the new state for the next iteration, and at each state transition, the transition probability from the current intent node to the next intent node is recorded, and an expected occurrence time window is estimated based on the preset lifecycle of the intent;

[0081] After completing all steps of deduction, a complete set of intent evolution paths is generated based on the recorded information of each transition. Each path is a sequence, including the initial intent node, the subsequent intent nodes generated through deduction, the corresponding transition probabilities and expected occurrence time windows between adjacent nodes in the path, and multiple intent evolution paths with different probabilities are output.

[0082] In one specific embodiment, the initial state T0 is taken as the probability distribution of user A's current intent. The projection period is set to 3 days, with each step being 1 day. The T0 state is input into the intent state transition model. The model predicts that on the first day T1, the user's intent state will transition to purchasing camping equipment (0.75), planning a weekend trip (0.15), and checking the weather (0.10). The transition probability from the initial intent to purchasing camping equipment is recorded as 0.75. Based on the preset 1-2 day lifespan of this intent, its time window is estimated as [the current day, the next day]. The T1 state is used as a new input for further projection. The model predicts that on the second day T2, the intent state will transition to purchasing camping equipment (0.40) and planning a weekend trip (0.5). 5. Check the weather (0.05), record the transition probability from purchasing camping equipment to planning a weekend trip as 0.40, and estimate the time window as [the next day, the third day] based on the 1-3 day lifecycle of the intention to plan a weekend trip. After completing all steps of deduction, two high-probability intention evolution paths are finally generated. Path 1 is purchasing camping equipment T0 - purchasing camping equipment T1 - planning a weekend trip T2, including transition probabilities of 0.75 and 0.40 and time windows of [the same day, the next day] and [the next day, the third day]. Path 2 is purchasing camping equipment T0 - planning a weekend trip T1 - planning a weekend trip T2, including transition probabilities of 0.25 and 0.55 and time windows of [the same day, the next day] and [the next day, the third day].

[0083] Step S500: Based on the predicted intent evolution path, retrieve relevant information from the existing dynamic knowledge graph, and use template-based and retrieval-enhanced NLP technology to create push copy that matches the current context; combine user location trajectory information to calculate the optimal push timing.

[0084] 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 the entities, attributes and relationships directly related to the intent, forming a candidate information set;

[0085] Using hybrid NLP technology, based on the current user's contextual tags and intent evolution path, the system selects the basic template with the highest semantic and style matching degree from a predefined copywriting template library. The retrieved candidate information set is used as enhancement information, and a generative model is used to intelligently fill it into the corresponding placeholders in the basic template to generate the final personalized push copy.

[0086] Analyze users' movement patterns, frequent locations, and points of interest to calculate the spatiotemporal accessibility of the user's geographic location associated with the push content within the expected time window. By comprehensively comparing the accessibility scores of each path with the overlap of the time window, select the moment with the highest accessibility and the best match with the time window to activate the push timing prediction window for that intent, which serves as the final basis for executing precise push decisions.

[0087] In one specific embodiment, following the generated intent evolution path, with the target intent node of the path, "planning a weekend outing," as the core, a dynamic knowledge graph is searched to obtain related entities such as a suburban park, sunny weather, and barbecue rental points, forming a candidate information set. Based on user A's current weekend outdoor leisure context tags, a basic template is selected from the copywriting template library. The retrieved suburban park, sunny weather, and barbecue rental points are used as augmenting information and intelligently filled in through a generative model to generate the final personalized push copy. Analyzing user A's GPS trajectory reveals that their usual location is at home, their point of interest is the suburban park, and the expected time window is [the next day, the third day]. The calculation shows that the spatiotemporal accessibility score of the user leaving home for the suburban park at 10:00 the next day is the highest, at 0.95, and the overlap with the time window is optimal. Therefore, 10:00 the next day is selected as the best push time, activating the push time prediction window for this intent, and preparing to execute precise push.

[0088] like Fig. 2 The flowchart of a personalized information delivery method based on artificial intelligence is shown. This invention provides a personalized information delivery method based on artificial intelligence, comprising:

[0089] Through multi-source data acquisition, user behavior data on e-commerce, social, and health and fitness platforms is captured in real time, while external spatiotemporal environmental data and user location trajectory information are acquired simultaneously. The collected multi-source heterogeneous data is fused to construct a high-dimensional spatiotemporal feature vector, which is then processed using the DBSCAN algorithm that incorporates a time decay factor and behavior intensity weights to identify the specific scenario in which the user is currently located and calculate a confidence score for each context label. The matching degree between new data and the current context is continuously evaluated, and recalculation is triggered when there is a mismatch. 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. The two are concatenated in chronological order and input into a pre-trained feature embedding model to generate an embedding feature matrix that can capture deep semantic relationships between features.

[0090] By using behavioral and contextual features as processing variables and latent intentions as outcome variables, the causal forest algorithm is used to quantify the heterogeneous causal effects between them, and a causal directed acyclic graph is constructed. Counterfactual reasoning is performed to inversely determine the latent intentions of users that are not directly observed in the current context, and finally a standardized intention probability distribution is output.

[0091] A Transformer-based intent state transition model is trained using historical data. The training process incorporates the aforementioned causal graph as a constraint to ensure that the transition path learned by the model conforms to causal logic. Based on the current intent state and the triggering behavior, the model accurately predicts the intent state at the next moment. Based on the trained model and the currently inverted intent, multi-step probabilistic inference is performed. Using the output of the previous moment as input, the model recursively predicts the changes in multiple future time steps, generating multiple intent evolution paths that include intent nodes, transition probabilities, and expected occurrence time windows, thus depicting multiple possible evolution trajectories of the user's intent in the future.

[0092] Based on the predicted intent path, relevant information is retrieved from the dynamic knowledge graph. Using template-based and retrieval-enhanced NLP technology, personalized push copy that highly matches the current context is automatically created. By comprehensively analyzing the user's real-time location trajectory, movement pattern, and the time window of intent occurrence, the optimal push timing with the best spatiotemporal accessibility is calculated. At the moment of calculation, the generated personalized information is accurately pushed to the user.

[0093] 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 implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A method for precise personalized information delivery based on artificial intelligence, characterized in that, include: Real-time collection of user behavior data on e-commerce, social and health and fitness platforms, and acquisition of external spatiotemporal environmental 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 scores to define the specific scenario in which the user is currently located. The generated context labels and the user's historical behavior sequence are used as inputs for feature engineering. The causal forest algorithm is used to quantify the heterogeneous causal effects among behavioral data, contextual labels, and latent intentions, and to construct a causal directed acyclic graph. This graph is then used for counterfactual reasoning to invert the user's latent intention in the current context, and to output a standardized probability distribution of the inverted intention. By utilizing the constructed causal directed acyclic graph and combining the user's historical intent sequence and behavior sequence, an intent state transition model based on the Transformer architecture is trained. Based on the intent inversion probability distribution and intent state transition model, multi-step probability deduction is performed to generate multiple intent evolution paths that include intent 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. Using template-based and retrieval-enhanced NLP technology, push notification copy matching the current context is created. Combined with user location trajectory information, the optimal push timing is calculated.

2. The method for precise personalized information delivery based on artificial intelligence according to claim 1, characterized in that, The real-time collection of user behavior data on e-commerce, social, and health and fitness platforms, and the acquisition of external spatiotemporal environmental data and user location trajectory information, include: Real-time data collection of user browsing, adding to cart, purchasing, and review behavior on e-commerce platforms; data collection of user content posting, liking, commenting, saving, and following behavior on social media platforms; data collection of user exercise records on health and fitness platforms; Real-time acquisition of external spatiotemporal environmental data, including timestamps, weather conditions, and holiday information; acquisition of user's geographic location, movement path, movement speed, frequent locations, and points of interest.

3. The personalized information delivery method 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 contextual labels with confidence levels, defining 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 fused from multiple sources to construct a high-dimensional spatiotemporal feature vector; the behavioral data is quantified into behavior type and behavior intensity according to 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 datasets. The time decay factor and behavior intensity weight are introduced to calculate the influence of data points in clustering. For each cluster formed after clustering analysis, the core features of all data points within the cluster are aggregated to generate a preset context label; based on the cluster density, the temporal continuity of data points within the cluster, and the consistency of behavioral intensity, the confidence level of each context label is calculated using a weighted summation algorithm. The generated context labels with confidence levels are used as the definition of the user's current specific scenario. New user data is continuously received, triggering a re-evaluation of the current scenario. When the matching degree between new data points and the current context labels is lower than a preset threshold, or when new high-confidence context labels appear, the definition of the user's current scenario is updated.

4. The method for precise personalized information delivery based on artificial intelligence according to claim 1, characterized in that, The step of using the generated contextual labels and user historical behavior sequences as input for feature engineering includes: The generated contextual labels with confidence scores are mapped into high-dimensional sparse vectors through one-hot encoding; each behavioral event in the user's historical behavior sequence is constructed into a structured behavioral 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 backtracking time window are concatenated in chronological order to form a hybrid feature sequence that integrates the current context and historical behavior. The mixed feature sequence is input into a pre-trained feature embedding model, which maps each vector to a unified continuous vector space to generate an embedding feature matrix that can capture the semantic relationships between features.

5. The method for precise personalized information delivery based on artificial intelligence according to claim 1, characterized in that, The aforementioned method employs a causal forest algorithm to quantify and analyze the heterogeneous causal effects among behavioral data, contextual labels, and latent intentions, constructing a causal directed acyclic graph, including: Using behavioral and contextual features in the generated embedded feature matrix as processing variables and latent intentions as outcome variables, the causal forest algorithm is used to analyze the causal effect strength of each processing variable on the outcome of each latent intention under different combinations of context and behavior, 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 latent intentions as the core nodes and behavioral data and contextual labels as the parent nodes.

6. The method for precise personalized information delivery based on artificial intelligence according to claim 1, characterized in that, The process of using this graph for counterfactual reasoning to invert the user's potential intention in the current context and output 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; Counterfactual intervention is performed in a causal directed acyclic graph by removing the current user's specific behavior to obtain the change in the probability distribution of potential intent nodes; the difference in intent probability before and after the intervention is calculated to quantify the causal contribution of the behavior to the intent. Taking into account the combined effect of the current context label and all historical behaviors in the causal directed acyclic graph, the probability distribution of the user's unobserved potential intentions in the current context is inverted using the Bayesian inference algorithm. This distribution is then normalized into a standardized intention inversion probability distribution, showing the most likely intentions and probabilities that the user will generate in the current context.

7. The method for precise personalized information delivery based on artificial intelligence according to claim 1, characterized in that, The process involves using the constructed causal directed acyclic graph, combined with the user's historical intent sequence and behavior sequence, to train an intent state transition model based on the Transformer architecture, including: The probability distribution of user intent inversion within a historical time period is defined as the intent state, and the specific behavior generated by the user within that time period is defined as the triggering behavior. An intent state transition sequence for supervised learning is constructed, where each training sample contains the intent state at time t, the triggering behavior, and the intent state at the next time t+1. Design an intent state transition model based on the Transformer architecture. The input sequence of the model is the intent state sequence and the corresponding trigger behavior sequence, and the output sequence is the subsequent intent state prediction sequence. A constructed causal directed acyclic graph (DAG) is used as an external knowledge constraint and integrated into the model's training process. When the model makes state transition predictions, the legal transition relationships defined in the DAG are encoded as causal masks and applied to the model's attention layer to ensure that the model learns and follows the causal logic-compliant state transition paths defined in the graph. Subsequent intent states from historical real data are used as labels to supervise the model's training and optimize its accuracy in predicting future intent states.

8. The method for precise personalized information delivery based on artificial intelligence according to claim 1, characterized in that, The process, based on the intent inversion probability distribution and intent state transition model, involves multi-step probabilistic deduction to generate multiple intent evolution paths, each containing an intent node, transition probability, and expected occurrence time window. The current intention inversion probability distribution is used as the initial state for multi-step probability deduction, representing the probability of all possible intentions of the user in the current context. The current deduction 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 a future 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 lifecycle of the intention. After completing all steps of deduction, a complete set of intent evolution paths is generated based on the recorded information of each transition. Each path is a sequence, including the initial intent node, the subsequent intent nodes generated through deduction, the corresponding transition probabilities and expected occurrence time windows between adjacent nodes in the path, and multiple intent evolution paths with different probabilities are output.

9. The method for precise personalized information delivery based on artificial intelligence according to claim 1, characterized in that, The process involves retrieving relevant information from an existing dynamic knowledge graph based on the predicted intent evolution path, and using template-based and retrieval-enhanced NLP technology to create push notification copy that matches the current context. By combining user location trajectory information, the optimal timing for push notifications is calculated, including: For each generated intent evolution path, the target intent node in the path is used as the core query condition. The dynamic knowledge graph is searched to obtain the 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 system selects the basic template with the highest semantic and style matching degree from a predefined copywriting template library. The retrieved candidate information set is used as enhancement information, and a generative model is used to intelligently fill it into the corresponding placeholders in the basic template to generate the final personalized push copy. Analyze users' movement patterns, frequent locations, and points of interest to calculate the spatiotemporal accessibility of the user's geographic location associated with the push content within the expected time window. By comprehensively comparing the accessibility scores of each path with the overlap of the time window, select the moment with the highest accessibility and the best match with the time window to activate the push timing prediction window for that intent, which serves 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 as described in any one of claims 1-9, characterized in that, include: Multi-source data perception and fusion 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 preprocesses the collected multi-source heterogeneous data and aggregates discrete data points into contextual labels with confidence through an improved spatiotemporal clustering algorithm. The causal intent inference module includes a causal effect analysis unit, an intent inversion unit, and an intent state transition modeling unit. The causal effect analysis unit employs a causal forest algorithm to quantify the heterogeneous causal effects between behavior, context, and intent, identifying key factors driving user intent and constructing a causal directed acyclic graph based on this. The intent inversion unit utilizes the constructed causal graph and, through counterfactual reasoning, inverts the most likely potential intents and probabilities of the user given the current context and behavior, outputting a standardized intent inversion probability distribution. The intent state transition modeling unit combines historical intent sequences and behavior sequences to train an intent state transition model based on a Transformer architecture. Using the causal graph as external knowledge constraint, it learns the causal logic-compliant transition patterns between intent states to predict future intents. Intent Evolution Path Deduction Module: Includes a multi-step probability deduction unit. This unit starts with the current intent inversion probability distribution and generates multiple complete intent evolution paths by iteratively calling the intent state transition model, which include future intent nodes, the transition probabilities between nodes, and the expected occurrence time windows. Personalized content generation 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 the query condition to retrieve relevant entities, attributes and relationships in the dynamic knowledge graph; the hybrid NLP generation unit combines the retrieved information and 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. The intelligent timing decision-making and execution module includes a spatiotemporal accessibility analysis unit and a push timing activation and execution unit. The spatiotemporal accessibility analysis unit combines the user's location trajectory information and the geographic location associated with the push content to calculate the spatiotemporal accessibility of the user to the relevant location within the predicted time window of the intent evolution path. The push timing activation and execution unit comprehensively compares the accessibility scores and the overlap of the time windows under each intent evolution path, 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