Content data active exploration method and system based on cognitive map and traceability reasoning

Abstract

The present application relates to the technical field of data processing, and is a content data active exploration method and system based on cognitive graph and trace inference. The method comprises: encoding a multi-modal input signal of target content data into a trigger vector, and matching the trigger vector with a content cognitive ontology graph to anchor a target attribution node; automatically generating competitive hypotheses of each target attribution node in the content cognitive ontology graph, and constructing an exploration path; calculating the utility value according to the prior probability and the execution cost of each exploration path, and iteratively executing the exploration path in the order of utility value from large to small, updating the posterior probability of each competitive hypothesis in real time, and outputting an attribution diagnosis conclusion according to the posterior probability of each competitive hypothesis in response to the satisfaction of a preset termination condition. The present application can actively understand and verify the causality of content data, and improve the attribution accuracy of content data.

Description

Technical Field

[0001] This invention relates to the field of data processing technology. More specifically, this invention relates to a method and system for proactively exploring content data based on cognitive mapping and abductive reasoning. Background Technology

[0002] With the explosive growth of generative artificial intelligence and digital media technologies, internet content data is characterized by massive volume, multimodality, and high-frequency updates. Content operation has moved from simple traffic distribution to a stage of refined intelligent decision-making. As a key link in content ecosystem governance and operational optimization, how to monitor the operational status of massive amounts of content data in real time and accurately pinpoint the causes of anomalies from complex and ever-changing indicator fluctuations, such as declining traffic and reduced interaction rates, has become a core requirement for improving the decision-making efficiency and risk control capabilities of content platforms. This is especially true when dealing with terabytes of unstructured data, where a data processing method capable of deeply understanding the business context and possessing attribution analysis capabilities is urgently needed.

[0003] Existing content data processing or monitoring technologies typically employ keyword filtering based on preset rules, natural language query retrieval, or clustering analysis methods based on vector similarity. Specifically, these methods require operators to predefine explicit query conditions or configure fixed DSL query statements; when the system receives instructions or triggers a threshold, it uses a full-text search engine or models such as CLIP / BERT to calculate embedding vectors for clustering, thereby outputting hot topic clusters or indicator trend charts to assist manual data analysis.

[0004] However, since the aforementioned methods are essentially passive response-based retrieval and clustering mechanisms, they heavily rely on explicit user input and lack the ability to proactively understand content data and verify causality. Furthermore, existing technologies typically cluster based solely on semantic similarity, resulting in a single and rigid attribution path. This leads to a lack of falsifiability in the attribution analysis process, making it highly susceptible to confirmation bias due to noisy data, severely hindering the timeliness and accuracy of decision-making. Summary of the Invention

[0005] To address the technical problem that existing technologies lack the ability to proactively understand and verify causal relationships of content data, resulting in poor accuracy of content data attribution and timeliness of decision-making, this invention discloses a method and system for proactively exploring content data based on cognitive mapping and abductive reasoning.

[0006] In a first aspect, this invention discloses a method for proactively exploring content data based on cognitive graphs and abductive reasoning. The method includes: receiving multimodal input signals about target content data; performing feature encoding on the multimodal input signals to obtain trigger vectors, and performing feature matching between the trigger vectors and a preset content cognitive ontology graph to anchor at least one target attribution node; starting from the target attribution node, performing bidirectional traversal in the content cognitive ontology graph to generate competing hypotheses for each target attribution node, wherein the competing hypotheses include supporting evidence nodes that support the target attribution node and refuting evidence nodes that refute the target attribution node; mapping the evidence nodes in each competing hypothesis to calling instructions for corresponding data probes to construct at least one exploration path; calculating utility values ​​based on the prior probabilities and execution costs of each exploration path, and iteratively executing the exploration paths in descending order of utility values, updating the posterior probabilities of each competing hypothesis in real time, and determining whether a preset termination condition is met; and outputting an attribution diagnosis conclusion based on the posterior probabilities of each competing hypothesis in response to the satisfaction of the preset termination condition.

[0007] Furthermore, the multimodal input signal includes at least one of text description, event information, and indicator trends.

[0008] Furthermore, the content cognition ontology graph includes multiple cognitive nodes and directed logical edges connecting the cognitive nodes; the directed logical edges include at least a first type of logical edge representing a support relationship and a second type of logical edge representing a refutation relationship.

[0009] Furthermore, a bidirectional traversal is performed in the content cognition ontology graph to generate a set of competing hypotheses, including: starting from any target attribution node, performing a forward traversal along the first type of logical edge to obtain supporting evidence nodes, and performing a reverse traversal along the second type of logical edge to obtain rejecting evidence nodes. The supporting evidence nodes and the rejecting evidence nodes constitute the competing hypotheses of the target attribution node.

[0010] Furthermore, the calculation of utility value based on the prior probability and execution cost of each exploration path includes: the utility value is positively correlated with the prior probability of the target attribution node and negatively correlated with the execution cost parameter; wherein the prior probability is determined based on the similarity between the trigger feature vector and the target attribution node, and the execution cost parameter is calculated based on the query latency time of the data probe and the number of data scan rows.

[0011] Furthermore, the data probe includes at least one of the following: a snapshot probe, used to calculate static indicator values ​​within a specified time window for structured historical data; a streaming probe, used to construct a sliding time window to monitor abnormal fluctuations in the data stream for a real-time message queue; and a graph traversal probe, used to perform correlation analysis along the relationship links between creators, content, or user entities for a relational graph database.

[0012] Further, the real-time update of the posterior probability of each competing hypothesis includes: calculating the likelihood term of each competing hypothesis in the current round, wherein the likelihood term is used to characterize the probability of observing the evidence data obtained by the exploration path under the condition that the competing hypothesis is true; taking the posterior probability obtained in the previous round as the prior probability of the current round, multiplying the prior probability with the likelihood term using Bayesian rules and normalizing it to obtain the posterior probability of each competing hypothesis in the current round; wherein, for the first iteration, its prior probability is determined based on the feature similarity between the trigger feature vector and the target attribution node.

[0013] Further, the calculation of the likelihood term for each competing hypothesis in the current round includes: obtaining evidence data returned by the exploration path executed in the current round; if the evidence data is a supporting evidence node, mapping the confidence level of the evidence data to a positive likelihood value using the Sigmoid function; if the evidence data is a rejecting evidence node, mapping the confidence level of the evidence data to a negative coefficient using the Sigmoid function, and recording the difference between 1 and the negative coefficient as the negative likelihood value; in response to the competing hypothesis corresponding to the executed exploration path, the likelihood term is the product of the positive likelihood value and the negative likelihood value, otherwise, the likelihood term is a preset neutral likelihood value.

[0014] Furthermore, the preset termination condition includes at least one of the following: the posterior probability of any competing hypothesis exceeds a preset confidence threshold; the prior probability or utility value of the remaining unverified competing hypotheses is lower than a preset minimum necessary threshold; and the cumulative execution cost of the executed data probes exceeds a preset resource quota.

[0015] Secondly, the present invention discloses a content data proactive exploration system based on cognitive graphs and abductive reasoning, comprising: a memory for storing a computer program; and a processor for executing the computer program to implement the steps of the content data proactive exploration method based on cognitive graphs and abductive reasoning as described above.

[0016] The beneficial effects of this invention are as follows: By receiving multimodal input signals about target content data and anchoring them to a pre-defined content cognitive ontology graph, the system generates competing hypotheses containing supporting and rejecting evidence nodes through bidirectional traversal of the graph. Then, based on prior probabilities and execution costs, it calculates utility values ​​to dynamically schedule data probes. Using Bayesian rules, it updates posterior probabilities in real time until termination conditions are met. This enables proactive exploration from fuzzy business signals to precise attribution conclusions. It can proactively parse multidimensional contextual information through multimodal input signals. Through the dialectical verification logic of competing hypotheses and intelligent scheduling based on utility values, it effectively solves the problems of existing technologies relying on passive retrieval and lacking causal verification. This significantly improves the automation level, attribution accuracy, and decision-making timeliness of content data anomaly diagnosis.

[0017] Competing hypotheses can be automatically constructed based on the content cognitive ontology graph, ensuring that each attribution hypothesis is accompanied by a rigorous verification chain, avoiding confirmation bias caused by one-way reasoning, and improving the interpretability and accuracy of attribution analysis.

[0018] During the iterative execution of the investigation path, the posterior probabilities of each competing hypothesis can be dynamically adjusted as evidence data is acquired, ensuring that the attribution conclusion is based on the mathematical convergence result of accumulated evidence, rather than subjective conjecture. At the same time, the calculation of positive and negative likelihood values ​​not only quantifies the strength of the evidence, but also gives the exclusionary evidence the ability to be vetoed. That is, the probability is significantly reduced by the negative likelihood value, which improves the accuracy and anti-interference ability of the attribution diagnosis. Attached Figure Description

[0019] Figure 1 This is a flowchart of a method for proactively exploring content data based on cognitive mapping and abductive reasoning, according to an embodiment of the present invention. Figure 2 This is a schematic diagram of a content data proactive exploration system based on cognitive graphs and abductive reasoning, according to an embodiment of the present invention. Detailed Implementation

[0020] 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, not all, of the embodiments of the present invention. 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.

[0021] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0022] This invention provides a method for proactively exploring content data based on cognitive mapping and abductive reasoning. This method aims to address the problem that existing data aggregation methods passively rely on keyword retrieval, lacking proactive understanding of target content data and causal verification capabilities, thus leading to inaccurate attribution results.

[0023] For example, in a short video operation scenario, operators receive a vague monitoring alert: "The interaction volume of camping videos has declined recently." This monitoring alert can be regarded as target content data. Traditional processing methods require manual inspection of each indicator, which is time-consuming, labor-intensive, and difficult to determine the root cause. The technical solution of this application can proactively analyze the target content data, anchor it to specific attribution hypotheses such as "theme fatigue," and automatically dispatch data probes to verify "click-through rate decay slope > 0.1 / day" and "new creator influx ratio > 20%," ultimately providing a precise diagnostic conclusion and obtaining accurate attribution results, achieving a technological leap from passive aggregation to proactive investigation.

[0024] like Figure 1 The diagram shown is a flowchart of a content data proactive exploration method based on cognitive graphs and abductive reasoning according to an embodiment of the present invention. The specific implementation of the present invention will be described in detail below.

[0025] S1, receive a multimodal input signal about target content data; perform feature encoding on the multimodal input signal to obtain a trigger vector, and perform feature matching on the trigger vector with a preset content cognitive ontology graph to anchor at least one target attribution node.

[0026] In one embodiment, the target content data can be monitoring alerts received by operations personnel, and the cause of the target content data will be diagnosed subsequently; the multimodal input signal includes at least one of text description, event information and indicator trends, which can comprehensively reflect the status of the target content data.

[0027] For example, in a short video operation scenario, the target content data is "the interaction volume of recent camping videos has declined". The text description can be a natural language query entered by the operator, such as "the interaction volume has dropped abnormally". The event information can be a key time in the marketing calendar, such as "Double Eleven promotion". The indicator trend can be the change trend of key performance indicators over a period of time, such as "the downward trend of video completion rate". By integrating information from multiple sources, it is possible to understand complex contexts that cannot be reflected by a single data source.

[0028] Furthermore, after obtaining the multimodal input signals, feature encoding is performed on the multimodal input signals. This process involves deep learning models, using the BERT model to encode the text descriptions, a Long Short-Term Memory (LSTM) network to encode the indicator trends, and a Multilayer Perceptron (MLP) to encode the event information. The results of the feature encoding are then fused into a fixed-dimensional trigger vector.

[0029] The preset content cognition ontology graph is a directed heterogeneous graph, including cognitive nodes and directed logical edges connecting cognitive nodes; each cognitive node corresponds to a preset attribution cause, and the directed logical edges are used to reflect the evidentiary relationship of the cognitive node; the directed logical edges include at least a first type of logical edge representing a support relationship and a second type of logical edge representing a refutation relationship.

[0030] As an example, the cognitive nodes include subject matter fatigue, style homogenization, emotional imbalance, and platform throttling sensitivity. The first type of logical edge is defined as "requires," indicating evidence that will inevitably appear if the attribution hypothesis is true. The second type of logical edge is defined as "refuted by," indicating that if this evidence appears, the attribution hypothesis is likely false. This design simulates the dialectical thinking of human experts, both confirming and falsifying, providing a basis for constructing competing hypotheses. The content cognitive ontology graph also includes multiple evidence nodes. Taking the cognitive node "subject matter fatigue" as an example, "Subject matter fatigue → requires → click-through rate decay slope > 0.1 / day" indicates that a click-through rate decay slope > 0.1 / day is a supporting evidence node for subject matter fatigue; "Subject matter fatigue → refuted by → new creator influx ratio > 20%" indicates that a new creator influx ratio > 20% is a repulsive evidence node for subject matter fatigue. Specifically, a click-through rate decay slope > 0.1 / day and a new creator influx ratio > 20% correspond to evidence nodes in the content cognitive ontology graph.

[0031] To accurately map fuzzy multimodal input signals to specific business concepts, the trigger vector is matched with a preset content cognitive ontology graph to anchor at least one target attribution node. This process is achieved by calculating the similarity between the trigger vector and the embedding vectors of each cognitive node in the content cognitive ontology graph, where the embedding vectors of each cognitive node in the content cognitive ontology graph can be obtained by a node embedding algorithm.

[0032] The process of anchoring at least one target attribution node satisfies the following relation:

[0033]

[0034] in, Representing cognitive nodes Match score; Indicates the trigger vector; Cognitive nodes in the content cognition ontology graph The embedding vector; This represents the function for calculating cosine similarity. This represents the set of all cognitive nodes in the content cognitive ontology graph; This represents the final set of target attribution nodes; The number of nodes to be selected is 3, for example, meaning there are 3 target attribution nodes.

[0035] By calculating cosine similarity, the correlation between multimodal input signals and various potential attribution causes is quantified, thereby quickly selecting the most likely attribution causes (such as "subject fatigue" and "style homogenization") as the starting point for subsequent investigation, avoiding the waste of resources caused by blind investigation.

[0036] Thus, through multimodal fusion and ontology graph anchoring, this step achieves the mapping from vague business signals to precise cognitive concepts, establishing a clear objective for subsequent proactive exploration.

[0037] S2, starting from the target attribution node, performs a bidirectional traversal in the content cognition ontology graph to generate competing hypotheses for each target attribution node. The competing hypotheses include supporting evidence nodes that support the target attribution node and rejecting evidence nodes that refute the target attribution node.

[0038] In one embodiment, performing a bidirectional traversal in the content cognitive ontology graph to generate a set of competing hypotheses includes: starting from any target attribution node, performing a forward traversal along a first type of logical edge to obtain supporting evidence nodes, and performing a reverse traversal along a second type of logical edge to obtain rejecting evidence nodes, wherein the supporting evidence nodes and the rejecting evidence nodes constitute competing hypotheses of the target attribution node.

[0039] Understandably, for the anchored "theme fatigue" node, we find "click-through rate decay slope > 0.1 / day" as supporting evidence along the first type of logical edge, and "new creator influx ratio > 20%" as excluding evidence along the second type of logical edge. These two evidence nodes together constitute a competing hypothesis about "theme fatigue".

[0040] For the anchored "style homogenization" node, we find "creator's historical style similarity > 0.6" as supporting evidence along the first type of logical edge, and "increase in loyal fan interaction rate > 0.1" as excluding evidence along the second type of logical edge. These two evidence nodes together constitute a competing hypothesis about "style homogenization".

[0041] Thus, for each target attribution node, a corresponding competing hypothesis can be automatically constructed, which is used to calculate the likelihood term of the target attribution node in subsequent steps.

[0042] S3 maps the evidence nodes in each competing hypothesis to the corresponding data probe invocation instructions, constructing at least one exploration path.

[0043] In one embodiment, for each competing hypothesis, it is necessary to invoke the data probe corresponding to the evidence node to obtain evidence data. The data probe includes at least one of the following: a snapshot probe, used to calculate static indicator values ​​within a specified time window for structured historical data; a streaming probe, used to construct a sliding time window to listen for abnormal fluctuations in the data stream for a real-time message queue; and a graph traversal probe, used to perform correlation analysis along the relationship links between creators, content, or user entities for a relational graph database.

[0044] It should be noted that the data probes corresponding to each evidence node are pre-set and can be stored in a preset content cognition ontology graph. That is, in addition to cognition nodes, directed logical edges connecting the cognition nodes and evidence nodes, the content cognition ontology graph also includes data probes corresponding to the evidence nodes. The data probes are used to obtain the evidence data of the corresponding evidence nodes.

[0045] For example, the competing hypothesis of "content fatigue" includes two evidence nodes: "click-through rate decay slope > 0.1 / day" and "new creator influx ratio > 20%". For "click-through rate decay slope > 0.1 / day", it is mapped to calling the snapshot probe Probe(CTR_slope), which connects to the data warehouse to calculate the click-through rate trend over the past 7 days. For the node "new creator influx ratio > 20%", it is mapped to calling the streaming probe Probe(Proportion of newcreators influx), which listens for real-time data in the Flink streaming engine.

[0046] In another example, for the competing hypothesis of "style homogenization," it includes two evidence nodes: "creator historical style similarity > 0.6" and "increase in fan engagement rate > 0.1." For "creator historical style similarity > 0.6," it is mapped to calling the graph traversal probe Probe("tracing back the creator's historical content from the current video"). This probe connects to a relational graph database, performs multi-hop sampling along the relationship link of "video-creator-historical video," and calculates the cosine similarity variance of the historical content feature vector. For "increase in fan engagement rate > 0.1," it is mapped to calling the snapshot probe Probe(core_fan_interaction_rate). This probe connects to a structured data warehouse and calculates the ratio of likes and comments from users who have followed the video for more than a certain threshold (such as 90 days) under the current video.

[0047] Thus, through bidirectional traversal of the graph and probe mapping, this step automatically generates an executable exploration path containing verification logic from both positive and negative perspectives, ensuring the comprehensiveness and operability of the attribution analysis.

[0048] S4 calculates the utility value based on the prior probability and execution cost of each exploration path, iteratively executes the exploration paths in descending order of utility value, updates the posterior probability of each competing hypothesis in real time, and determines whether the preset termination condition is met.

[0049] In one embodiment, the utility value is used to measure the cost-effectiveness of executing a certain exploration path and to guide the scheduling order of exploration paths. The calculation of the utility value based on the prior probability and execution cost of each exploration path includes: the utility value is positively correlated with the prior probability of the target attribution node and negatively correlated with the execution cost parameter; wherein the prior probability is determined based on the similarity between the trigger feature vector and the target attribution node, and the execution cost parameter is calculated based on a weighted average of the query latency time of the data probe and the number of data rows scanned.

[0050] No. Utility value of each exploration path Satisfying the relation:

[0051] in, Indicates the first The exploration paths correspond to competing hypotheses The prior probability; Indicates the estimated query time for the data probe; This indicates the estimated number of rows of data the data probe needs to scan; and These are time weighting coefficients and scan weighting coefficients, respectively. The unit of the time weighting coefficient is the reciprocal of the estimated query time, used to eliminate the dimension of the estimated query time. The unit of the scan weighting coefficient is the reciprocal of the estimated number of data rows, used to eliminate the dimension of the estimated number of data rows. For example, It can be set to 0.6, with units of , It can be set to 0.4, with the unit being... , Indicates the calculation of the first The total cost of all data probes in the exploration path.

[0052] in, In the relation, the molecule The competing hypotheses calculated in step S1 The higher the prior probability of the matching score for the corresponding target attribution node, the more suspicious the corresponding target attribution node is, and the greater its utility value; the denominator is the th... The total cost of all data probes in the exploration path refers to the cost of executing the first probe. The total time and computational resources required for each investigation path are considered the sum of all possible costs. The lower the total cost, the greater the utility value. Iteratively executing investigation paths in descending order of utility value prioritizes those with high suspicion and minimal verification costs. This scheduling strategy allows for the fastest possible approach to the truth within limited time and computational constraints, significantly reducing decision-making delays.

[0053] In one embodiment, during the iterative execution of the exploration path, the real-time update of the posterior probability of each competing hypothesis includes: calculating the likelihood term of each competing hypothesis in the current round, wherein the likelihood term is used to characterize the probability of observing the evidence data obtained by the exploration path under the condition that the competing hypothesis is true; using the posterior probability obtained in the previous round as the prior probability of the current round, multiplying the prior probability with the likelihood term using Bayesian rules and normalizing it to obtain the posterior probability of each competing hypothesis in the current round; wherein, for the first iteration, its prior probability is determined based on the feature similarity between the trigger feature vector and the target attribution node.

[0054] The update of the posterior probability satisfies the following relation:

[0055] in, Indicates the updated competing hypothesis The posterior probability; The set of competing hypotheses comprising all competing hypotheses; and They represent competing hypotheses. and competing hypotheses The likelihood term; and These represent the competing hypotheses before the update. and competing hypotheses The prior probability.

[0056] In one embodiment, calculating the likelihood term for each competing hypothesis in the current round includes: obtaining evidence data returned by the exploration path executed in the current round; if the evidence data is a supporting evidence node, mapping the confidence level of the evidence data to a positive likelihood value using the Sigmoid function; if the evidence data is a rejecting evidence node, mapping the confidence level of the evidence data to a negative coefficient using the Sigmoid function, and recording the difference between 1 and the negative coefficient as the negative likelihood value; in response to the competing hypothesis corresponding to the executed exploration path, the likelihood term is the product of the positive likelihood value and the negative likelihood value, otherwise, the likelihood term is a preset neutral likelihood value.

[0057] In response to the competing hypothesis corresponding to the executed exploration path, the likelihood term satisfies the following relation: ; in, For competing hypotheses The likelihood term, This indicates the supporting evidence set. This indicates a set of excluded evidence; Use the Sigmoid activation function; The confidence level of the returned evidence data is calculated by weighting statistical significance and data quality. The adjustment factor is 1.0, for example.

[0058] To facilitate understanding, the calculation of the likelihood term for each competing hypothesis in the current round is explained exemplarily. Assume that in the current round, exploration path A is scheduled and executed based on utility value; exploration path A corresponds to the competing hypothesis of "subject fatigue," which involves calling the snapshot probe Probe(CTR_slope) and the streaming probe Probe(Proportion of new creatorsinflux). Probe(CTR_slope) corresponds to the supporting evidence node "click-through rate decay slope > 0.1 / day," and Probe(Proportion of new creators influx) corresponds to the rejecting evidence node "new creator influx ratio > 20%." In addition to the competing hypothesis of "subject fatigue," the current set of competing hypotheses also includes the competing hypotheses of "style homogenization" and "platform throttling sensitivity."

[0059] In the current round, the evidence returned by the snapshot probe (CTR_slope) is as follows: the click-through rate decay slope over the past 7 days is 0.15 / day, which is greater than 0.1 / day, supporting the validity of the evidence, and its confidence level is [not specified]. The evidence returned by the flow cytometry probe (Proportion of new creators influx) is as follows: the proportion of new creators influx is 21%, which is greater than 20%, thus validating the evidence, and its confidence level is [not specified]. The positive likelihood value is then... This indicates that the click-through rate data strongly supports the hypothesis of subject matter fatigue; the inverse likelihood value is... This indicates that the influx of new creators highly likely refutes the hypothesis of subject matter fatigue (i.e., it negates the hypothesis); ultimately, the likelihood term of the competing hypothesis of "subject matter fatigue" is the product of the positive and negative likelihood values, which is... Since the "style homogenization" hypothesis and the "platform throttling sensitivity" hypothesis do not include the two data probes in exploration path A (i.e., they have no direct logical connection with the click-through rate slope and the influx of new creators), they are directly assigned a neutral likelihood value of 0.5. The likelihood terms of the competing hypotheses of "style homogenization" and "platform throttling sensitivity" are both 0.5.

[0060] In summary, due to the veto effect of exclusionary evidence, the likelihood of the "topic fatigue" competing hypothesis is only 0.045, far below the neutral value, and its posterior probability will decrease significantly, thus being excluded from the competition. Meanwhile, the evidence data returned in the current round neither supports nor excludes the competing hypotheses of "style homogenization" and "platform throttling sensitivity," remaining neutral and not introducing additional information bias. Therefore, the likelihood of the competing hypotheses of "style homogenization" and "platform throttling sensitivity" is directly set to 0.5. In subsequent posterior probability updates, since the likelihood of the "topic fatigue" competing hypothesis is lower than that of the "style homogenization" and "platform throttling sensitivity" competing hypotheses, the posterior probability of the "topic fatigue" competing hypothesis will decrease relatively, thus achieving differentiated verification of the competing hypotheses in a single exploration.

[0061] S5, in response to the satisfaction of the preset termination condition, outputs the attribution diagnosis conclusion based on the posterior probability of each competing hypothesis.

[0062] In one embodiment, in order to stop unnecessary probes in a timely manner, the preset termination condition includes at least one of the following: the posterior probability of any competing hypothesis exceeds a preset confidence threshold; the prior probability or utility value of the remaining unverified competing hypotheses is lower than a preset minimum necessary threshold; and the cumulative execution cost of the executed data probes exceeds a preset resource quota.

[0063] For example, the confidence threshold can be set to 0.85, the minimum necessary threshold can be set to 0.1, and the preset resource quota is the maximum resource allowed to be consumed when calling the data probe. For example, the maximum allowed query time is 5 seconds, and the maximum number of data rows allowed to be scanned is 2000 rows. Once the termination condition is met, the competing hypothesis with the highest posterior probability is immediately output as the diagnostic conclusion, and the target attribution node corresponding to the competing hypothesis is the cause of the target content data.

[0064] Thus, through dynamic scheduling based on utility values, the introduction of posterior probability updates, and an intelligent termination mechanism, efficient, accurate, and logically rigorous data exploration and attribution are achieved, accurately pinpointing the causes of the target content data.

[0065] like Figure 2 The diagram shown is a system schematic of a content data proactive exploration system based on cognitive graphs and abductive reasoning according to an embodiment of the present invention. Furthermore, the present invention also provides a content data proactive exploration system based on cognitive graphs and abductive reasoning, which is integrated into a user terminal device. From a hardware architecture perspective, the system includes: a memory for storing computer program code that executes the above-described algorithm, and locally cached decision information set data; and a processor for executing the program in the memory. The processor is configured to run the steps described above, such as anchoring target attribution nodes, generating competing hypotheses, executing exploration paths, and updating posterior probabilities, thereby achieving a technological leap from passive aggregation to proactive exploration.

[0066] The system in this embodiment also includes other components well known to those skilled in the art, such as communication interfaces. Their settings and functions are known in the art, and therefore will not be described in detail here.

[0067] In this invention, the aforementioned memory can be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0068] While this specification has shown and described numerous embodiments of the invention, it will be apparent to those skilled in the art that such embodiments are provided by way of example only. Many modifications, alterations, and alternatives will occur to those skilled in the art without departing from the spirit and essence of the invention. It should be understood that various alternatives to the embodiments of the invention described herein may be employed in the practice of this invention.

Claims

1. A method for proactively exploring content data based on cognitive mapping and abductive reasoning, characterized in that: The detection method includes: Receive multimodal input signals about target content data, wherein the multimodal input signals include at least one of text description, event information, and indicator trends; The multimodal input signal is feature-encoded to obtain a trigger vector, and the trigger vector is matched with a preset content cognition ontology graph to anchor at least one target attribution node. The BERT model is used to encode the text description, the long short-term memory network is used to encode the indicator trend, and the multilayer perceptron is used to encode the event information. The results of the feature encoding are fused into a fixed-dimensional trigger vector. Starting from the target attribution node, a bidirectional traversal is performed in the content cognition ontology graph to generate competing hypotheses for each target attribution node. The competing hypotheses include supporting evidence nodes that support the target attribution node and refuting evidence nodes that refute the target attribution node. Map the evidence nodes in each competing hypothesis to the corresponding data probe invocation instructions to construct at least one exploration path; The utility value is calculated based on the prior probability and execution cost of each exploration path, and the exploration paths are iteratively executed in descending order of utility value. The posterior probability of each competing hypothesis is updated in real time, and it is determined whether a preset termination condition is met. The calculation of the utility value based on the prior probability and execution cost of each exploration path includes: the utility value is positively correlated with the prior probability of the target attribution node and negatively correlated with the execution cost parameter; wherein the prior probability is determined based on the similarity between the trigger vector and the target attribution node, and the execution cost parameter is calculated by weighting the query latency time of the data probe and the number of data scan rows. The real-time update of the posterior probability of each competing hypothesis includes: calculating the likelihood term of each competing hypothesis in the current round, wherein the likelihood term is used to characterize the probability of observing the evidence data obtained by the exploration path under the condition that the competing hypothesis is true; taking the posterior probability obtained in the previous round as the prior probability of the current round, multiplying the prior probability with the likelihood term using Bayesian rules and normalizing it to obtain the posterior probability of each competing hypothesis in the current round; wherein, for the first iteration, its prior probability is determined based on the feature similarity between the trigger vector and the target attribution node; In response to the satisfaction of the preset termination condition, the attribution diagnosis conclusion is output based on the posterior probability of each competing hypothesis.

2. The method for proactively exploring content data based on cognitive mapping and abductive reasoning according to claim 1, characterized in that, The content cognition ontology graph includes multiple cognitive nodes and directed logical edges connecting the cognitive nodes; A directed logical edge includes at least a first-class logical edge representing a support relation and a second-class logical edge representing a refutation relation.

3. The method for proactively exploring content data based on cognitive mapping and abductive reasoning according to claim 2, characterized in that, Perform a bidirectional traversal of the content-cognitive ontology graph to generate a set of competing hypotheses, including: Starting from any target attribution node, a forward traversal is performed along the first type of logical edge to obtain supporting evidence nodes, and a reverse traversal is performed along the second type of logical edge to obtain rejecting evidence nodes. The supporting evidence nodes and the rejecting evidence nodes constitute the competing hypothesis of the target attribution node.

4. The method for proactively exploring content data based on cognitive mapping and abductive reasoning according to claim 1, characterized in that, The data probe includes at least one of the following: Snapshot probes are used to calculate static metric values ​​within a specified time window for historical data in structured storage. Streaming probes are used to build sliding time windows for real-time message queues to listen for abnormal fluctuations in the data stream; Graph traversal probes are used to perform association analysis along the relationship links between creator, content, or user entities in relational graph databases.

5. The method for proactively exploring content data based on cognitive mapping and abductive reasoning according to claim 1, characterized in that, The calculation of the likelihood terms for each competing hypothesis in the current round includes: Obtain the evidence data returned by the exploration path executed in the current round; If the evidence data is a supporting evidence node, the confidence level of the evidence data is mapped to a positive likelihood value using the Sigmoid function; if the evidence data is a rejecting evidence node, the confidence level of the evidence data is mapped to a reverse coefficient using the Sigmoid function, and the difference between 1 and the reverse coefficient is recorded as the reverse likelihood value. In response to competing hypotheses corresponding to the exploration paths performed, the likelihood term is the product of the positive and negative likelihood values; otherwise, the likelihood term is the pre-defined neutral likelihood value.

6. The method for proactively exploring content data based on cognitive mapping and abductive reasoning according to claim 1, characterized in that, The preset termination condition includes at least one of the following: The posterior probability of any competing hypothesis exceeds the preset confidence threshold; The prior probabilities or utility values ​​of the remaining unverified competing hypotheses are all below the pre-set minimum necessary threshold; and, The cumulative execution cost of the executed data probes has exceeded the preset resource quota.

7. A content data proactive exploration system based on cognitive mapping and abductive reasoning, characterized in that, include: Memory, used to store computer programs; A processor, configured to implement the steps of the content data active exploration method based on cognitive graphs and abductive reasoning as described in any one of claims 1 to 6 when executing the computer program.

Images