Data operation system and method based on knowledge graph

Abstract

The invention relates to the technical field of artificial intelligence and big data analysis, in particular to a data operation system and method based on a knowledge graph, and the method comprises the steps: extracting an entity and semantic relationship from multi-source business data, and constructing a dynamic evolvable initial knowledge graph; node features are aggregated, and multi-dimensional situation state vectors are generated in combination with gating loop unit modeling behavior path dependence; through a structure-semantic coupling attribution scoring mechanism, a statistical information gain and semantic similarity are fused to identify a core driving factor, and a causal regression model of the factor and an operation target is established; dynamically adjusting the edge weight and the structure of the atlas in real time, and triggering a new path discovery mechanism to continuously optimize the atlas; and according to a quantitative business target, reversely extracting a high-confidence influence path from the atlas, and generating a personalized strategy combination through intervention simulation and multi-target Pareto optimization, thereby realizing intelligent recommendation and decision closed loop driven by an operation target. According to the invention, higher-precision operation situation awareness and strategy generation are realized.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence and big data analysis technology, and specifically to a data operation system and method based on a knowledge graph. Background Art

[0002] With the deepening of digital transformation, enterprises urgently need to build an operational support system with semantic understanding, dynamic modeling and intelligent decision-making capabilities when facing complex and changing market environments and user behaviors; traditional rule-driven data analysis and recommendation methods can no longer meet the current comprehensive needs such as multi-source heterogeneous data fusion, multi-objective operation strategy coordination, and dynamic user response feedback processing; knowledge graphs, with their advantages in entity relationship modeling, semantic reasoning and explainable modeling, provide a new technical path for realizing data asset linkage, context modeling and strategy optimization in complex business scenarios.

[0003] The Chinese invention patent with announcement number CN113535972B discloses a knowledge graph link prediction model method and device that integrates contextual semantics, including an input sampling subgraph. In the input, each training data is an r-subgraph composed of n entities, and n refers to the number of entities in the r-subgraph. According to the sampling strategy of breadth-first search, a series of relationship sequences can be generated, that is, an r-subgraph is a relationship path containing n entities. The subgraph can contain three aspects of semantic information: structural information in the knowledge graph, type attribute meta-information and ontology information of the subgraph (details are written in the attribute information extraction part), and contextual natural semantic knowledge information contained in entities and relationships. It has a reasonable structure and uses LSTM and GCN to jointly capture the contextual semantic representation of entities in the knowledge graph, improve the embedding representation capability of entities, and apply it to link prediction tasks to improve model performance.

[0004] In recent years, with the development of technologies such as graph neural networks, causal reasoning models, and feedback reinforcement learning, how to combine knowledge graphs to build an intelligent operation system that is driven by operational goals and has state modeling and self-evolution capabilities has become an important research direction in the field of data operations. Summary of the Invention

[0005] The purpose of this invention is to address the problems existing in the background technology and propose a data operation system and method based on knowledge graph.

[0006] The technical solution of the present invention is a data operation method based on knowledge graph, which includes the following specific implementation steps:

[0007] S1. Divide the operation segments based on the time interval and contextual semantic similarity of the user behavior sequence, extract semantic entities, filter potential relationship triplets with semantic attention, and construct the initial knowledge graph with multi-granularity subgraph division;

[0008] S2. Based on the relationship-aware graph convolutional network, we aggregate the node relationship features within the subgraph, model the time evolution sequence of the behavior path through the gated recurrent unit, and integrate mean pooling and goal-oriented attention mechanism to unify the heterogeneous state vector space, generating a situation state vector with context structure and behavior dependency.

[0009] S3. Build a structural-semantic coupling attribution scoring method that integrates statistical information gain and semantic similarity to identify the key drivers of operational objectives and establish a causal regression model between the factors and the objectives.

[0010] S4. Dynamically adjust the credibility of triples by comparing behavioral paths, use the prediction error of key performance indicators to trigger abnormal path detection and decomposition to generate new structure candidates, and use the effect of strategic intervention as feedback signal to integrate the time decay factor to continuously optimize the graph edge weight;

[0011] S5. Based on the current operational goals, reversely extract high-confidence paths from the knowledge graph, combine strategy intervention simulation with multi-objective Pareto optimization, and generate personalized strategy combinations.

[0012] Preferably, the initial knowledge graph construction process is as follows:

[0013] Generate operational segments with business semantics based on the time interval of user behavior sequences and the similarity of contextual semantics;

[0014] The bidirectional long short-term network (BiLSTM) and conditional random field (CRF) model are used to extract semantic entities in operation segments, filter potential entity relationships based on semantic attention, and generate a set of relationship triples.

[0015] The fragment-level triples are integrated to construct a heterogeneous graph, and the semantic subgraphs are automatically divided through semantic clustering to generate the initial knowledge graph.

[0016] Preferably, the process of generating the situation state vector is:

[0017] Applying a relation-aware graph convolutional network at the semantic subgraph level to aggregate features of neighboring nodes under different relation types, and retaining the original node information through self-connection to generate an updated semantic state representation;

[0018] Based on the node embedding sequence, the gated recurrent unit is used to model the temporal dependency of the behavior path, generating a sequence of path state vectors reflecting the historical evolution trajectory.

[0019] By fusing the overall trend of mean pooling and goal-oriented attention pooling operations on the path state vector sequence, the heterogeneous state space is unified and the situation state vector is constructed.

[0020] Preferably, the causal regression model construction process is:

[0021] Identify potential causal paths based on knowledge graphs and behavioral paths, construct user attribution graphs based on statistical or semantic rules, and locate the driving relationship between core behavioral nodes and observable behavioral nodes;

[0022] Through the attribution scoring mechanism that couples the statistical correlation of information gain with the semantic similarity structure, key driving factors are sorted and screened, and a multi-dimensional quantitative evaluation of the factor influence is conducted;

[0023] Construct a structure-preserving causal regression model, establish a predictive relationship between core driving factors and target indicators, quantify the linear weights of single factors and the interaction effects between factors, and support reverse reasoning calculations based on target values.

[0024] Preferably, the optimization process of the graph edge weight is as follows:

[0025] Based on the consistency difference between the user behavior path and the graph deduction path, the credibility score of the triple relationship is dynamically updated through positive and negative feedback learning rates;

[0026] Using the prediction error of key performance indicators (KPIs) as a trigger condition, the algorithm integrates path embedding, semantic decoding, and context consistency verification to dynamically discover and screen potential high-impact paths, supplement important structures not modeled in the graph, and perform structural optimization and semantic enhancement.

[0027] The strategy intervention effect is used as an external feedback signal, and the strategy gain coefficient and time decay factor are combined to dynamically adjust the graph edge weight. The strategy intervention effect is used as a feedback signal and the time decay factor is integrated to continuously optimize the graph edge weight.

[0028] The preferred implementation process for dynamically discovering and screening potential high-impact paths is as follows:

[0029] Anomaly perception is defined based on the difference between the actual key performance indicator (KPI) value and the graph prediction value. If the difference exceeds the set threshold, the graph structure reconstruction is triggered.

[0030] The user behavior path is encoded into a vector representation through a path autoencoder, and its cosine similarity deviation with the central path vector is calculated. The abnormality probability is output and the abnormality degree of the path structure is quantified to identify potential new semantic paths.

[0031] The graph structure of paths with abnormal probability higher than the set threshold is decomposed to screen potential high-impact paths.

[0032] Preferably, the decomposition process of graph structure decomposition for paths with abnormal probability higher than a set threshold is:

[0033] Slice the highly abnormal paths and slide the windows to form several entity pairs with semantic associations.

[0034] Using the graph embedding-based relationship prediction model TransE, the relationship type of each pair of entities is calculated to obtain new structural candidate triples;

[0035] A contextual behavior sequence consistency test is introduced to count the high-frequency co-occurrence of triples in typical paths to ensure their contextual stability. The semantic similarity between triples and the main path is calculated using a path embedding model to determine the consistency of structural evolution. The temporal synergy between their occurrence frequency and key operational indicators is also evaluated.

[0036] All triples that pass the semantic scoring and context consistency check are added to the graph structure candidate pool.

[0037] Preferably, the personalized strategy combination generation process is as follows:

[0038] Define a quantitative objective function to calculate the comprehensive contribution of the strategy combination to the operational goal, mine a set of high-confidence paths related to the target node from the graph through inverse reasoning, and output the target path set to ensure that the confidence of the path impact exceeds the preset threshold;

[0039] Extract strategic intervention factors from the target path set, optimize the strategy combination effect evaluation by simulating the target indicator change values ​​under different context clusters and calculating the average expected effect;

[0040] The multi-objective Pareto frontier optimization method is used to generate a strategy combination by weighted summation of strategy factors, and the comprehensive expected effect is maximized through resource cost constraints;

[0041] Based on the generated optimal strategy combination, combined with user status, behavioral intention and context labels, the correlation score between users and strategies is calculated through graph personalized reasoning path matching, and the most relevant strategy is finally selected.

[0042] The technical solution of the present invention is a data operation system based on a knowledge graph, which is used to execute the data operation method based on a knowledge graph as claimed in the claim, comprising:

[0043] The operational context modeling and knowledge graph construction module is used to extract entities, attributes, and semantic relationships from multi-source business data, build an initial graph structure that includes the operational logic context, and form a three-layer nested semantic graph structure by integrating rule knowledge, behavior logs, and context scenarios to achieve structured cognitive expression under the semantics of operational goals.

[0044] The state vector modeling module is used to construct dynamic state representations of each entity based on the knowledge graph. It extracts the state transition characteristics of entities during operation based on state evolution tensor modeling and time window slice analysis, and maps them into multidimensional state vectors for subsequent driving factor identification and strategy simulation input.

[0045] The driver identification and causal indicator modeling module is used to identify key driver factors that affect operational goals from the graph based on path perception and causal graph reasoning algorithms, construct an indicator correlation matrix and response delay model between factors and indicators, and provide intervention dimensions and measurement basis for operational strategies;

[0046] The multi-source semantic feedback fusion and graph update module is used to receive user responses, operational intervention results, and external context changes in real time. Through the semantic signal decoder and graph edge weight estimation model, it dynamically adjusts the graph structure and edge attributes to achieve continuous evolution and semantic enhancement and update of the graph.

[0047] The goal-driven policy optimization and recommendation generation module is used to build objective functions and perform policy simulation, resource constraint optimization, and personalized recommendations based on the current business goals, context status, and the path set inferred from the graph, generating an explainable intelligent policy combination.

[0048] Compared with the prior art, the above technical solution of the present invention has the following beneficial technical effects:

[0049] This paper designs a data operations system and method based on a knowledge graph. By building a dynamically evolving knowledge graph and a causal-driven intelligent decision-making closed loop, it effectively solves three core problems in traditional operations systems: static knowledge representation is difficult to adapt to business changes, attribution analysis is one-sided, and strategy recommendations lack interpretability. Specifically:

[0050] (1) Improve the accuracy of operational situation awareness: Based on the integration of R-GCN and GRU context structure and behavior path dependency, a dynamic state vector is generated to reduce the error in business state representation and significantly enhance the real-time perception capability of scenarios such as user conversion and churn warning;

[0051] (2) Strengthening causal attribution explainability: Identifying key drivers through a structural-semantic coupling scoring mechanism (information gain + semantic similarity) to improve the accuracy and traceability of operational interventions;

[0052] (3) Dynamically optimize strategic resources: Combine target inverse reasoning with Pareto frontier optimization to generate personalized strategic combinations, thereby improving GMV (Gross Merchandise Volume) conversion efficiency under resource constraints while reducing ineffective marketing costs.

[0053] (4) Ensure the system's ability to continuously evolve: Dynamically adjust the graph structure and edge weights based on multi-source feedback (user behavior, intervention results, external changes), improve the knowledge base's adaptability to new business scenarios, and effectively support long-term operational decisions. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 This is a system architecture diagram of a knowledge graph-based data operation system proposed by the present invention;

[0055] Figure 2 This is a flow chart of a data operation method based on knowledge graph proposed in the present invention. DETAILED DESCRIPTION

[0056] Example 1, as Figure 1 As shown, the present invention proposes a data operation system based on knowledge graph, including: operation context modeling and knowledge graph construction module, situation state vector modeling module, driving factor identification and causal indicator modeling module, multi-source semantic feedback fusion and graph update module and goal-driven strategy optimization and recommendation generation module.

[0057] The operational context modeling and knowledge graph construction module is used to extract entities, attributes, and semantic relationships from multi-source business data, build an initial graph structure that includes the operational logic context, and form a three-layer nested semantic graph structure by integrating rule knowledge, behavior logs, and context scenarios to achieve structured cognitive expression under the semantics of operational goals.

[0058] The state vector modeling module is used to construct dynamic state representations of each entity based on the knowledge graph. It extracts the state transition characteristics of entities during operation based on state evolution tensor modeling and time window slice analysis, and maps them into multidimensional state vectors for subsequent driving factor identification and strategy simulation input.

[0059] The driver identification and causal indicator modeling module is used to identify key driver factors that affect operational goals from the graph based on path perception and causal graph reasoning algorithms, construct an indicator correlation matrix and response delay model between factors and indicators, and provide intervention dimensions and measurement basis for operational strategies;

[0060] The multi-source semantic feedback fusion and graph update module is used to receive user responses, operational intervention results, and external context changes in real time. Through the semantic signal decoder and graph edge weight estimation model, it dynamically adjusts the graph structure and edge attributes to achieve continuous evolution and semantic enhancement and update of the graph.

[0061] The goal-driven strategy optimization and recommendation generation module is used to build objective functions and perform strategy simulation, resource constraint optimization, and personalized recommendations based on the current business goals, context status, and the path set inferred from the graph, generating an explainable intelligent strategy combination and feeding it back to the operation system.

[0062] Example 2, as Figure 2 As shown, the present invention proposes a data operation method based on a knowledge graph, which is applied to a data operation system based on a knowledge graph proposed in Example 1. The specific implementation steps are as follows:

[0063] S1. The operational context modeling and knowledge graph construction module abstracts structural expressions with stable business semantics from raw operational data. By introducing context-awareness, fragment-driven, implicit causal analysis, and multi-granularity graph construction mechanisms, it establishes an initial knowledge graph structure with evolvable characteristics. The specific implementation steps are as follows:

[0064] S11. For typical business goals (including but not limited to user conversion improvement, customer churn warning, content reach efficiency, etc.), first perform process semantic fragmentation modeling on the original data stream;

[0065] That is: user behavior sequence B u ={b1,...,b i ,...,b n},b i =(a i ,t i ,c i ) is the input, combined with the time interval δ t With contextual semantic similarity sim, the behavior is aggregated into an operation segment S under the conditions of temporal continuity and semantic proximity j :

[0066] S j ={b k |δ t (b k ,b k+1 )<θ t ∧sim(c k ,c k+1 )>θ s};

[0067] Among them, B u represents the complete behavior sequence of user u; b i Indicates the i-th behavior record tuple; a i Indicates behavioral actions, including but not limited to click, jump, search, and payment; i Indicates the timestamp of the behavior; c iIndicates the context information when the behavior occurs, including but not limited to page tags, content keywords, and business modules; S j represents the jth constructed operational semantic process fragment; δ t (b k ,b k+1 ) indicates behavior b k with b k+1 The time interval function of θ t Indicates the time continuity threshold, which is an adjustable parameter. In this embodiment, it is set to 10 to 15 minutes; sim(c k ,c k+1 ) represents the semantic similarity between behavior contexts; θ s represents the context semantic similarity threshold;

[0068] For example, if a user browses multiple product pages in a short period of time and stays in the comment section for a long time, this process can be identified as a "comparison decision segment";

[0069] S12. In each operation segment S j In the deep context modeling network (this embodiment uses a bidirectional long short-term network BiLSTM + conditional random field CRF) to extract the dynamic entity set in the operational context:

[0070] E j ={e1,...,e i ,...,e m},e i ∈S j ;

[0071] Among them, E j represents the set of semantic entities extracted from the jth process segment; e i Represents the i-th semantic entity, including but not limited to "user type", "page topic", and "content module";

[0072] Combined with the semantic attention drift analysis model to identify potential relationships between behaviors:

[0073] α i,j =softmax(W1h i +W2h j +b);

[0074] Among them, α i,j Represents entity e i With e j The semantic relationship attention score between i and h jrepresents the semantic representation vector of the entity in the context, which is encoded by the pre-trained model BERT; W1 and W2 represent trainable weight matrices; b represents the bias term; softmax(·) represents the softmax activation function;

[0075] If α i,j >λ, filter potential semantic relations and output preliminary relation triple candidates: T j ={(e i ,r ij ,e j )|α ij >λ};

[0076] Among them, λ represents the attention threshold, which controls whether the two entities are considered to have an implicit relationship; T j represents the set of triples generated in the jth segment; r ij Represents entity e i With e j The type of semantic relationship between them (including but not limited to "jump", "interest transfer", "interruption");

[0077] S13, segment level triples T j Map to the knowledge graph structure and construct a heterogeneous graph G = (V, E, R);

[0078] V: node set, including but not limited to users, behaviors, content, and events;

[0079] E: edge set, generated by triple mapping;

[0080] R: a set of relationship types;

[0081] To enhance the graph’s sensitivity to behavioral changes and its ability to represent context, a multi-granularity subgraph partitioning mechanism is introduced:

[0082] Among them, G k represents the kth semantic subgraph, i.e., a local business behavior scenario; K represents the number of subgraphs, which is automatically determined based on semantic clustering; (V k ,E k ) represents the subgraph G k The set of entities and edges within .

[0083] S2, the situation state vector modeling module uses a graph-driven state representation learning method, combining context structure, behavior path dependency, and relational semantic embedding to achieve high-dimensional feature expression of the state vector. Its specific implementation steps are as follows:

[0084] S21, based on the operation knowledge graph G = (V, E, R) constructed in step S1, subgraph G k As a unit, for each entity node vi ∈V to aggregate contextual structural features and adopt the relational graph convolutional network (R-GCN) method based on relational perception for modeling:

[0085]

[0086] in, Represents node v i The embedding vector at the lth layer of graph convolution reflects its semantic state in the graph structure; Represents the updated node representation after the l+1th layer of graph convolution, aggregating to a wider neighborhood; Represents the node v i The set of neighbor nodes connected under the relationship type r; R represents the set of all defined semantic relationship types in the graph, including but not limited to "click-jump" and "stay-jump"; Represents the trainable weight matrix of relation r in the l-th layer of graph convolution, which is used to extract the features of the relation; Represents the weight matrix of the self-connection, retaining its original information and enhancing stability; c i,r represents the normalization coefficient; ReLU(·) represents the ReLU activation function;

[0087] S22. After obtaining the structural representation of each node in the graph, further model the time dependency of the behavior path and set the path P for a certain operation process. u ={v1→v2→…→v T}, construct the path representation:

[0088] s t =GRU(h t ,s t-1 );

[0089] Based on this: we can obtain the path-level state vector sequence {s1,s2,…,s T};

[0090] Among them, h t represents the node embedding at path position t, the output h from step S21 i ;s t-1 represents the state latent vector at the previous moment, that is, the behavioral evolution trajectory up to t-1; GRU(·) represents the gated recurrent unit, which is used to process long dependency information in the sequence and compress historical behavior patterns;

[0091] S23, considering the heterogeneity of the state vector distribution generated by different subgraphs and different behavior paths, further unify the state vector space, introduce multi-scale normalization and semantic compression mechanism, and construct the final state vector z u :z u =MeanPool({s t})+γ·Attention({s t},q);

[0092] Among them, {s t} represents the state representation vector sequence at each moment in a behavior path; MeanPool(·) represents the mean pooling operation, which calculates the mean of all state vectors in the sequence to obtain the overall behavior trend; q represents the query vector, which can be the semantic representation of the operation goal (including but not limited to "conversion tendency" and "silence warning"), used to guide attention focus; γ represents the adjustable coefficient, which controls the fusion ratio of mean pooling and attention pooling, and is set to 0.2~0.5 in this embodiment; Attention(·) represents the attention pooling operation, which selects the state that is most critical to the operation goal based on the query vector q.

[0093] S3. The driving factor identification and causal indicator modeling module identifies the behavioral factors that have a decisive impact on key operational goals (including but not limited to registration conversion, paid retention, and user churn) based on user behavior graphs, operational status vectors, and historical indicator data, and establishes an explanatory mapping relationship between these factors and the target indicators. The specific implementation process is as follows:

[0094] S31. Introduce an automatic attribution graph construction method based on behavioral path causal extraction, identify potential driving factor paths from knowledge graphs and behavioral paths, and construct a causal attribution graph:

[0095] D u ={(v i ,r,v j )|v i ∈V core ,r∈R causal ,v j ∈V obs ,P ij ∈P causal};

[0096] Among them, D u represents the attribution graph path set of user u, which is the set of behavioral paths that are inferred to have causal logic in the graph; v i Indicates the core behavior node, representing the target behavior (including but not limited to ordering and unsubscribing); v j Indicates observable behavior nodes (including but not limited to click, stay, browse); r indicates v j to v iPath relationship type between core Represents the set of all candidate core behavior nodes, representing the target behavior space, defined by business operation goals, including but not limited to "complete registration" and "open push"; V obs Represents the set of all observable behavior nodes, representing possible influencing behaviors, and comes from behavioral data such as logs, click streams, and embedded points; R causal represents the set of edge types that are considered to have causal meaning; P ij Indicates that from v j to v i The path between causal Represents a set of paths that are considered causally possible by statistical or semantic rules;

[0097] S32. After the causal attribution graph is constructed, not all candidate factors have a significant impact on the operational objectives. Therefore, a structural-semantic coupling attribution scoring mechanism is introduced to simultaneously perform attribution strength modeling on the two dimensions of statistical correlation and semantic similarity:

[0098] Define the attribution score function: α k =IG(X k ,Y)+τ·sim(z u ,e k );

[0099] Among them, α k represents the attribution intensity score of the kth driving factor, which is used to measure its impact on the target indicator; X k represents the kth candidate driving factor; Y represents the target operating indicator, including but not limited to user payment conversion rate, retention rate, and awakening rate; IG(X k ,Y) represents information gain, which measures the factor X k The amount of information enhancement on Y; τ represents the balance coefficient, which controls the weight between statistical information and structural semantics; sim(·,·) represents the similarity function, which is used to calculate the closeness between the state vector and the factor embedding;

[0100] Then all factors are adjusted by α k Sort and select the first N as the core driving factor set {X1,...,X k ,...,X N};

[0101] Based on this, we score each candidate factor on both the "statistical significance" and "reasonableness in the current semantic context" dimensions, thus avoiding redundancy issues such as "a factor was significant historically but is currently irrelevant";

[0102] S33, based on the above selected core driving factor set {X1,...,X k,...,X N A structure-preserving causal regression model (SPCR) is used to establish a prediction model for factors and target indicators, and supports reverse reasoning of indicator target values:

[0103]

[0104] in, Represents the operational target indicator predicted by the model; β k Represents factor X k The linear regression coefficient, that is, the direct influence weight of a single factor; φ i,j Represents the factor X i and X j The interaction coefficient of , that is, the joint effect of factors; ε represents the model error term, that is, the unobservable or noise part, which meets the assumptions of the conventional regression model.

[0105] S4, Multi-source semantic feedback fusion and graph update module builds a multi-source semantic feedback fusion framework. By integrating user behavior evolution, indicator performance fluctuations, and operational intervention effects, it drives real-time updates to the knowledge graph structure and edge weight adjustments, thereby continuously improving the graph's ability to fit and predict real operational situations. The specific implementation process is as follows:

[0106] S41. Based on the consistency between user behavior sequences and the deduction paths in the graph, a dynamic relationship credibility adjustment framework is constructed:

[0107] Each triple relationship (v i ,r,v j ) is defined as:

[0108]

[0109] Where ΔP match Indicates the consistency enhancement between the triple relationship and the user's actual behavior path in the recent behavior sequence; ΔP div It represents the deviation measure of the triple relationship being replaced or bypassed in the current behavior path; η1 represents the positive feedback learning rate, which controls the degree of influence of the user matching behavior on the credibility enhancement; η2 represents the negative feedback learning rate, which controls the degree of influence of the user deviation behavior on the credibility weakening; represents the credibility score of relation r at time t; represents the updated relationship credibility after the current round of feedback fusion;

[0110] It should be noted that ΔP match With ΔP divDerived from the user behavior sequence database through behavioral path comparison; ΔP match From "whether the relationship still frequently appears in the user's real behavior path", ΔP div Indicates whether the user begins to detour or deviate from the original path, for example, from "browse → save → order" to "browse → exit";

[0111] For example, in the past, the system identified "view promotion page → click details → purchase" as a high-weight path through the graph. However, the frequency of "click details" on this path has recently decreased and is no longer a valid conversion. After the system detects the deviation, it dynamically reduces the credibility of this triple using the above formula to reduce its impact on subsequent reasoning.

[0112] S42. Build a new path discovery method that uses key indicator prediction errors as trigger conditions. This method introduces key performance indicator (KPI) prediction errors into the structure update logic, and uses path compression and semantic decoding techniques to avoid redundant additions, improve structure discovery efficiency, and compensate for important paths in the graph that are not modeled but may affect the target. Specifically:

[0113] A1. Define abnormality perception of indicators: ΔY = Y real -Y pred ;

[0114] Among them, ΔY represents the indicator prediction error, which means that the graph cannot accurately model the KPI under the current structure; Y real Indicates the actual indicator value obtained from the business database at the current moment; Y pred Represents the theoretical prediction value calculated by the knowledge graph and prediction model;

[0115] If |ΔY| exceeds the set threshold, structural reconstruction is triggered;

[0116] A2. Construct a path set P based on the user's most recent M behavior sequences = {P1,...,P i ,…,P M}, using the path autoencoder E path Learn its structural representation;

[0117] Calculate each path P i and the mean path vector μ P Deviation:

[0118] p anomaly (P i )=1-cos(E path (P i ),μ P );

[0119] Among them, p anomaly (P i) represents the path P i The degree of structural abnormality of E is used to identify potential new semantic paths; path (P i ) represents the embedding vector generated by the path self-encoder for the path; P i represents the user's i-th behavior path, i.e., a specific sequence of operations; cos(·,·) represents the cosine similarity function; μ P Represents the mean of the user path vector in the current time window (center path vector);

[0120] A3. Decompose the graph structure of the paths whose abnormal probability is higher than the set threshold. The decomposition process is as follows:

[0121] B1. Set the high abnormal path P i =[a1→…→a i →…a k ] Slice-type window sliding is performed to form several entity pairs with semantic associations;

[0122] For example, the path [search → browse → favorite → exit] can be parsed as: (search, r 1 , browse), (browse, r 2 ,Collection),(Collection,r 3 , jump out);

[0123] B2. Using the graph embedding-based relationship prediction model (TransE), for each pair of entities v s 、v t Calculate the most likely relationship type r * , obtain a new structural candidate triple (v s ,r * ,v t );

[0124] B3. Introduce contextual behavior sequence consistency testing to count the high-frequency co-occurrence of triples in typical paths to ensure contextual stability. Calculate the semantic similarity between triples and the main path using a path embedding model to determine the consistency of structural evolution. Evaluate the temporal synergy between their frequency of occurrence and key operational indicators (including but not limited to conversion rate and retention rate).

[0125] B4. All triples that pass the semantic scoring and context consistency check are added to the graph structure candidate pool;

[0126] S43. Use the strategy intervention behavior as external feedback input, participate in the graph edge weight adjustment, and conduct feedback linkage between the graph and the strategy:

[0127] in, represents the updated edge weight after policy feedback; represents the graph edge weight of the triple relationship r at time t; η3 represents the strategy gain coefficient, which controls the positive impact of strategy effectiveness on the graph edge weight; η4 represents the time decay coefficient, which controls the natural aging speed of the edge weight; τ represents the time decay rate constant, which reflects the failure rate of old knowledge; δ effect Indicates the change in target business indicators after a strategy is launched; e -τt represents the edge weight decay factor, which automatically reduces the weight of unupdated knowledge over time;

[0128] For example, the system recommends pushing a "limited-time discount pop-up" to non-converted users. One week after the strategy was launched, the system detected that the order rate of the push path increased by 4.1%. The edge weight of the push-related path is positively strengthened, and the strategy effectiveness is automatically recorded. If a strategy has no positive effect for a long time, its related edge weight will naturally decay over time.

[0129] S5. The goal-driven strategy optimization and recommendation generation module builds a strategy optimization and recommendation generation framework driven by operational goals. This framework extracts influencing factor paths from the graph structure, performs performance inversion evaluation and intervention simulation, and ultimately achieves personalized strategy combination reasoning and precise recommendations for specific business goals (including but not limited to conversion rate improvement, user retention, and GMV (Gross Merchandise Volume) growth). The specific implementation process is as follows:

[0130] S51. Define a quantitative objective function based on the operational goals of the current business (including but not limited to improving user retention rate and average customer spending). In the knowledge graph, based on the inverse reasoning mechanism, mine all key paths and node subgraphs that may have a positive or negative impact on the goal. Specifically:

[0131] (1) Construct the objective function: L goal =ω1·ΔM kpi +ω2·ΔU act -ω3·C intv ;

[0132] Among them, L goal Represents the target optimization function value, that is, the comprehensive contribution of the strategy combination to the operation target; ΔM kpi Indicates the improvement of key performance indicators (KPIs), including but not limited to GMV, retention rate, conversion rate, etc., in percentage or absolute value difference; ΔU act Indicates the positive change in key user behaviors (including but not limited to click count, stay time, repurchase rate, etc.); C intvThe resource cost consumed by strategy execution (including but not limited to operating manpower, preferential costs, and marketing budget), expressed in monetary units or converted cost values; ω1, ω2, and ω3 are weighting coefficients, representing the company's current focus on different goals, obtained through expert evaluation, A / B testing, or historical data training;

[0133] (2) Based on the knowledge graph G = (V, E), let the target node be v g , traverse the graph backward to find the set of paths that have direct or indirect dependencies with the target:

[0134] P goal ={P i |P i →v g ∧conf(P i )≥ψ};

[0135] Among them, P goal Represents a set of nodes with the target node v g The set of relevant high-confidence paths; P i represents the i-th path, which is composed of a set of triples in the graph; v g Indicates the target node, corresponding to a business goal (such as completing a purchase); conf(P i ) represents the path P i The confidence level of the impact on the target node is derived from the comprehensive calculation of graph edge weights and semantic consistency; ψ represents the minimum confidence threshold for path validity judgment, which is used to eliminate low-quality paths;

[0136] S52, from the target path set P goal Extract the relationship nodes related to the strategy nodes (including but not limited to: coupon issuance, wording changes) to form the intervention factor set F intv , and further evaluate its effects in different contextual scenarios through the intervention simulation module to optimize the strategy combination;

[0137] The simulation modeling is as follows:

[0138] Among them, SimEff(f i ,C j ) represents the strategy factor f i In context cluster C j The simulation effect score under f i represents the i-th strategic factor (including but not limited to issuing coupons, modifying wording, and adjusting time periods), extracted from the graph node; C j represents the jth context cluster (including but not limited to high-activity and high-value old users), obtained through behavioral clustering; s represents the simulation situation, from the context cluster C j The user state vector sampled from ; In the simulation scenario s, the factor f is applied i The post-predicted change value of the target indicator; It represents the average of the simulation results of multiple user samples within the context cluster, that is, the expected strategy effect;

[0139] S53. Based on the simulated scoring matrix, a multi-objective Pareto frontier optimization method is introduced to generate strategy combinations: The constraints are: C total ≤B;

[0140] Among them, S score represents the total score of the current candidate strategy combination, that is, the comprehensive expected effect of the strategy combination on the target; k represents the number of strategy factors being evaluated; β i represents the weighted coefficient of the i-th strategic factor, reflecting its impact on the target in historical data; C * Indicates the key context tags identified in the current system (including but not limited to high-bounce users); C total represents the overall execution resource cost of the candidate strategy combination; B represents the acceptable total budget, which is set or dynamically adjusted by the business team;

[0141] S54: Based on the generated optimal strategy combination, combined with the current user status, behavioral intention, and contextual tags, personalized reasoning path matching and strategy implementation are performed on the graph:

[0142]

[0143] in, RelScore(u,s) represents the personalized optimal recommendation strategy generated for user u; RelScore(u,s) represents the relevance score between user u and strategy S, which measures the semantic path similarity or historical response similarity between the user and the strategy; argmax(·) represents the selection of the most relevant strategy in the candidate set; S best Indicates the set of policy candidates recommended by the current system.

[0144] The embodiments of the present invention are described in detail above with reference to the accompanying drawings, but the present invention is not limited thereto. Various changes can be made within the scope of knowledge possessed by those skilled in the art without departing from the spirit of the present invention.

Claims

1. A data operation method based on knowledge graph, characterized in that: The specific implementation steps include the following: S1. Divide the operation segments based on the time interval and contextual semantic similarity of the user behavior sequence, extract semantic entities, filter potential relationship triplets with semantic attention, and construct the initial knowledge graph with multi-granularity subgraph division; S2. Based on the relationship-aware graph convolutional network, we aggregate the node relationship features within the subgraph, model the time evolution sequence of the behavior path through the gated recurrent unit, and integrate mean pooling and goal-oriented attention mechanism to unify the heterogeneous state vector space, generating a situation state vector with context structure and behavior dependency. S3. Build a structural-semantic coupling attribution scoring method that integrates statistical information gain and semantic similarity to identify the key drivers of operational objectives and establish a causal regression model between the factors and the objectives. S4. Dynamically adjust the credibility of triples by comparing behavioral paths, use the prediction error of key performance indicators to trigger abnormal path detection and decomposition to generate new structure candidates, and use the effect of strategic intervention as feedback signal to integrate the time decay factor to continuously optimize the graph edge weight; S5. Based on the current operational goals, reversely extract high-confidence paths from the knowledge graph, combine strategy intervention simulation with multi-objective Pareto optimization, and generate personalized strategy combinations.

2. A data operation method based on knowledge graph according to claim 1, characterized in that: The initial knowledge graph construction process is as follows: Generate operational segments with business semantics based on the time interval of user behavior sequences and the similarity of contextual semantics; The bidirectional long short-term network (BiLSTM) and conditional random field (CRF) model are used to extract semantic entities in operation segments, filter potential entity relationships based on semantic attention, and generate a set of relationship triples. The fragment-level triples are integrated to construct a heterogeneous graph, and the semantic subgraphs are automatically divided through semantic clustering to generate the initial knowledge graph.

3. A data operation method based on knowledge graph according to claim 2, characterized in that: The generation process of the situation state vector is: Applying a relation-aware graph convolutional network at the semantic subgraph level to aggregate features of neighboring nodes under different relation types, and retaining the original node information through self-connection to generate an updated semantic state representation; Based on the node embedding sequence, the gated recurrent unit is used to model the temporal dependency of the behavior path, generating a sequence of path state vectors reflecting the historical evolution trajectory. By fusing the overall trend of mean pooling and goal-oriented attention pooling operations on the path state vector sequence, the heterogeneous state space is unified and the situation state vector is constructed.

4. A data operation method based on knowledge graph according to claim 3, characterized in that: The process of building a causal regression model is as follows: Identify potential causal paths based on knowledge graphs and behavioral paths, construct user attribution graphs based on statistical or semantic rules, and locate the driving relationship between core behavioral nodes and observable behavioral nodes; Through the attribution scoring mechanism that couples the statistical correlation of information gain with the semantic similarity structure, key driving factors are sorted and screened, and a multi-dimensional quantitative evaluation of the factor influence is conducted; Construct a structure-preserving causal regression model, establish a predictive relationship between core driving factors and target indicators, quantify the linear weights of single factors and the interaction effects between factors, and support reverse reasoning calculations based on target values.

5. A data operation method based on knowledge graph according to claim 4, characterized in that: The optimization process of graph edge weight is as follows: Based on the consistency difference between the user behavior path and the graph deduction path, the credibility score of the triple relationship is dynamically updated through positive and negative feedback learning rates; Using the prediction error of key performance indicators (KPIs) as a trigger condition, the algorithm integrates path embedding, semantic decoding, and context consistency verification to dynamically discover and screen potential high-impact paths, supplement important structures not modeled in the graph, and perform structural optimization and semantic enhancement. The strategy intervention effect is used as an external feedback signal, and the strategy gain coefficient and time decay factor are combined to dynamically adjust the graph edge weight. The strategy intervention effect is used as a feedback signal and the time decay factor is integrated to continuously optimize the graph edge weight.

6. A data operation method based on knowledge graph according to claim 5, characterized in that: The implementation process of dynamically discovering and screening potential high-impact paths is as follows: Anomaly perception is defined based on the difference between the actual key performance indicator (KPI) value and the graph prediction value. If the difference exceeds the set threshold, the graph structure reconstruction is triggered. The user behavior path is encoded into a vector representation through a path autoencoder, and its cosine similarity deviation with the central path vector is calculated. The abnormality probability is output and the abnormality degree of the path structure is quantified to identify potential new semantic paths. The graph structure of paths with abnormal probability higher than the set threshold is decomposed to screen potential high-impact paths.

7. The data operation method based on knowledge graph according to claim 6 is characterized in that: The decomposition process of graph structure decomposition for paths with abnormal probability higher than the set threshold is as follows: Slice the highly abnormal paths and slide the windows to form several entity pairs with semantic associations. Using the graph embedding-based relationship prediction model TransE, the relationship type of each pair of entities is calculated to obtain new structural candidate triples; A contextual behavior sequence consistency test is introduced to count the high-frequency co-occurrence of triples in typical paths to ensure their contextual stability. The semantic similarity between triples and the main path is calculated using a path embedding model to determine the consistency of structural evolution. The temporal synergy between their occurrence frequency and key operational indicators is also evaluated. All triples that pass the semantic scoring and context consistency check are added to the graph structure candidate pool.

8. The data operation method based on knowledge graph according to claim 7 is characterized in that: The process of generating a personalized strategy combination is as follows: Define a quantitative objective function to calculate the comprehensive contribution of the strategy combination to the operational goal, mine a set of high-confidence paths related to the target node from the graph through inverse reasoning, and output the target path set to ensure that the confidence of the path impact exceeds the preset threshold; Extract strategic intervention factors from the target path set, optimize the strategy combination effect evaluation by simulating the target indicator change values ​​under different context clusters and calculating the average expected effect; The multi-objective Pareto frontier optimization method is used to generate a strategy combination by weighted summation of strategy factors, and the comprehensive expected effect is maximized through resource cost constraints; Based on the generated optimal strategy combination, combined with user status, behavioral intention and context labels, the correlation score between users and strategies is calculated through graph personalized reasoning path matching, and the most relevant strategy is finally selected.

9. A data operation system based on a knowledge graph, which is used to execute a data operation method based on a knowledge graph according to any one of claims 1 to 8, characterized in that: include: The operational context modeling and knowledge graph construction module is used to extract entities, attributes, and semantic relationships from multi-source business data, build an initial graph structure that includes the operational logic context, and form a three-layer nested semantic graph structure by integrating rule knowledge, behavior logs, and context scenarios to achieve structured cognitive expression under the semantics of operational goals. The state vector modeling module is used to construct dynamic state representations of each entity based on the knowledge graph. It extracts the state transition characteristics of entities during operation based on state evolution tensor modeling and time window slice analysis, and maps them into multidimensional state vectors for subsequent driving factor identification and strategy simulation input. The driver identification and causal indicator modeling module is used to identify key driver factors that affect operational goals from the graph based on path perception and causal graph reasoning algorithms, construct an indicator correlation matrix and response delay model between factors and indicators, and provide intervention dimensions and measurement basis for operational strategies; The multi-source semantic feedback fusion and graph update module is used to receive user responses, operational intervention results, and external context changes in real time. Through the semantic signal decoder and graph edge weight estimation model, it dynamically adjusts the graph structure and edge attributes to achieve continuous evolution and semantic enhancement and update of the graph. The goal-driven policy optimization and recommendation generation module is used to build objective functions and perform policy simulation, resource constraint optimization, and personalized recommendations based on the current business goals, context status, and the path set inferred from the graph, generating an explainable intelligent policy combination.

Images