Methods for constructing age-specific knowledge graphs from multi-source heterogeneous data on infant and toddler care services

Abstract

This invention relates to the field of infant and toddler care technology and discloses a method for constructing an age-specific knowledge graph from multi-source heterogeneous data of infant and toddler care services. The method collects multi-source heterogeneous data streams from care services, including real-time health indicators, continuous behavioral trajectories, and dynamic environmental parameters. After outlier removal and data alignment, a time-aligned data set is formed. Then, based on the critical periods of infant and toddler neurodevelopment, this data set is divided into multiple developmental stage data groups. For each developmental stage data group, an attention mechanism encoder is used to extract entity features and relational paths, constructing a developmental stage knowledge subgraph. Finally, a graph attention network is used to perform cross-stage association learning on the developmental stage knowledge subgraphs, generating a fused age-specific knowledge graph. This method achieves accurate developmental stage segmentation and deep semantic association of care data, effectively supporting personalized monitoring and scientific intervention in the infant and toddler growth process.

Description

Technical Field

[0001] This invention relates to the field of infant and toddler care technology, specifically a method for constructing age-specific knowledge graphs from multi-source heterogeneous data on infant and toddler care services. Background Technology

[0002] Current data management for infant and toddler care services primarily employs a discrete recording method based on age grouping. Existing technologies for collecting infant and toddler data are mostly limited to a single dimension, failing to achieve simultaneous acquisition of multi-source data such as health, behavior, and environment. Data processing methods are simplistic and lack the ability to deeply mine temporal features. Developmental stages are typically divided using fixed monthly intervals, failing to consider the differences in critical periods of neurodevelopment. Knowledge construction is conducted in isolation, and the correlations between data from different age groups are not effectively utilized. Existing methods need to address key technical issues such as multi-source data fusion, accurate developmental stage segmentation, and cross-stage knowledge association.

[0003] Traditional childcare data management systems suffer from significant shortcomings in data integration and analysis depth. Multi-source data collection suffers from asynchronous timing and inconsistent formats, hindering effective correlation analysis. Rigid developmental stage classification standards fail to adapt to individual developmental differences and critical period characteristics. Knowledge extraction methods are limited to surface features, failing to uncover deep-seated correlations within the data. Subgraph construction occurs independently, lacking a holistic consideration of developmental continuity. Knowledge fusion methods are simplistic, failing to establish cross-stage evolutionary relationship models. Existing technologies necessitate the development of a fully intelligent processing solution encompassing data collection and knowledge fusion. Summary of the Invention

[0004] The purpose of this invention is to provide a method for constructing age-specific knowledge graphs from multi-source heterogeneous data on infant and toddler care services, in order to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, this invention provides a method for constructing an age-specific knowledge graph of multi-source heterogeneous data related to infant and toddler care services, the method comprising:

[0006] Collect multi-source heterogeneous data streams from infant and toddler care services. The multi-source heterogeneous data streams include real-time health indicator data, continuous behavioral trajectory data, and dynamic environmental parameter data. Perform outlier removal and data alignment processing on the multi-source heterogeneous data streams to generate a time-series aligned data set.

[0007] Based on the critical periods of infant and toddler neurodevelopment, the time-aligned dataset is divided into multiple developmental stage data groups according to the critical developmental periods;

[0008] For each developmental stage data set, an attention mechanism encoder is used to extract entity features and relational paths to construct a developmental stage knowledge subgraph.

[0009] A graph attention network is used to perform cross-stage association learning on knowledge subgraphs of each developmental stage, generating a fused age-specific knowledge graph.

[0010] Preferably, the method further includes:

[0011] The entity richness and relation diversity of the age-specific knowledge graph are measured, and the graph structure strength index is calculated based on the entity richness and relation diversity.

[0012] Based on the graph structure strength index, an adaptive quality threshold is set to detect isolated nodes in age-specific knowledge graphs and label suspicious data nodes.

[0013] When there are suspicious data nodes, extract the spatiotemporal features and context embedding vector of each suspicious data node, and determine whether the suspicious data node constitutes a real defect based on the spatiotemporal features and context embedding vector. When it is confirmed to be a real defect, calculate the basic map quality score based on the number of real defect nodes.

[0014] Topological attributes are extracted from real defect nodes, and the topological attributes are used as input features. Density clustering algorithm is used to group the real defect nodes, and the presence of systematic anomalies is detected based on the grouping results.

[0015] When a systematic anomaly is identified, a calibration weight is calculated based on the degree of propagation of the systematic anomaly. The calibration weight is then used to correct the quality score of the basic knowledge graph and to intelligently adjust the optimization interval of the knowledge graph.

[0016] Preferably, when performing outlier removal and data alignment on multi-source heterogeneous data streams, the following steps are included:

[0017] Set a sliding time window to segment the data stream and calculate the statistical outlier value of the data in each window;

[0018] When the fluctuation range of a data point exceeds the outlier threshold, the data point is marked as an outlier and replaced by interpolation.

[0019] A dynamic time warping algorithm is used to align data sequences with different sampling frequencies to a unified time axis, generating a time-aligned data set.

[0020] Preferably, when dividing developmental stage data groups based on infant neurodevelopmental milestones, the data includes:

[0021] Obtain a mapping table between infants' and toddlers' ages and developmental abilities, and determine the age boundaries of each critical developmental period based on the mapping table;

[0022] The time-aligned dataset is divided into continuous data segments according to age boundaries, with each data segment corresponding to a developmental stage data group.

[0023] Developmental labels were added to each developmental stage data set, including the motor skills period, language explosion period, and socio-emotional period.

[0024] Preferably, when using an attention mechanism encoder to extract entity features and relationship paths, the following are included:

[0025] The developmental stage data set is input into the multi-head self-attention layer to calculate the association weights between entities;

[0026] Filter important relationship paths based on association weights, and retain relationships with weights higher than the path threshold;

[0027] A graph encoder is used to map entities and relations to low-dimensional vectors, constructing a knowledge subgraph for developmental stages.

[0028] Preferably, when using graph attention networks for cross-stage association learning, it includes:

[0029] Input the knowledge subgraphs of each developmental stage into the graph attention network and calculate the cross-stage attention coefficients between nodes;

[0030] The node representation is updated by aggregating the feature information of adjacent nodes through graph convolution operations.

[0031] The updated node representations are then subjected to graph pooling to generate a fused age-specific knowledge graph.

[0032] Preferably, when measuring entity richness and relation diversity, the following are included:

[0033] The number of different entity types in the age-specific knowledge graph is counted, and the entity type entropy value is calculated as the entity richness.

[0034] Analyze the uniformity of the distribution of relation paths and calculate the relation distribution divergence as relation diversity;

[0035] The graph structure strength index is obtained by weighted summation of entity richness and relation diversity.

[0036] Preferably, when performing isolated node detection on an age-specific knowledge graph, the following are included:

[0037] Calculate the number of neighboring nodes and connection density for each node;

[0038] When the number of neighboring nodes of a node is lower than the isolation threshold or the connection density is lower than the density threshold, the node is marked as a suspicious data node.

[0039] Record the location identifier and attribute summary of suspicious data nodes.

[0040] Preferably, when extracting the spatiotemporal features and context embedding vectors of suspicious data nodes, the process includes:

[0041] Extracting temporal trend features and spatial distribution features from the historical change sequence of nodes;

[0042] A graph neural network is used to generate a context embedding vector for a node, the context embedding vector containing local topological information of the node;

[0043] The time trend features, spatial distribution features, and context embedding vectors are concatenated into a feature vector, which is then input into the defect classifier to determine whether it is a real defect.

[0044] Preferably, when using a density clustering algorithm to group real defect nodes, the following steps are included:

[0045] Set the neighborhood radius and minimum number of samples for density clustering;

[0046] Calculate the feature distance matrix between real defect nodes, and perform density clustering based on the distance matrix;

[0047] When a large cluster of nodes spanning multiple developmental stages appears in the clustering results, a systematic anomaly is identified.

[0048] When calculating calibration weights based on the degree of propagation of systemic anomalies, the following are included:

[0049] The subgraph extent and node proportions of the impact of systemic anomalies are measured;

[0050] The anomaly impact factor is calculated based on the subgraph range and node ratio, and then mapped to the calibration weight interval.

[0051] The quality score of the basic knowledge graph is adjusted using calibration weights, and the optimization interval of the knowledge graph is recalculated based on the adjusted score.

[0052] Compared with the prior art, the beneficial effects of the present invention are:

[0053] Based on the critical periods of infant and toddler neurodevelopment, the time-aligned dataset is divided into multiple developmental stage data groups. Critical periods of neurodevelopment are scientifically defined according to brain development patterns and cognitive characteristics, reflecting sensitive periods for the development of specific abilities. Stage division considers key developmental windows such as the sensorimotor stage and preoperational stage, ensuring that each stage has unique developmental characteristics. Data grouping is dynamically divided according to critical period time nodes to accommodate individual developmental differences among infants and toddlers. Each developmental stage data group contains a complete record of health indicators, behavioral patterns, and environmental interactions within that stage. For each developmental stage data group, an attention mechanism encoder is used to extract entity features and relationship paths. The attention mechanism encoder analyzes key information in the data, automatically focusing on features important for developmental assessment. Entity feature extraction identifies core elements such as health parameters, behavioral characteristics, and environmental factors. Relationship path mining reveals the interaction mechanisms between entities, such as the correlation between sleep quality and cognitive development. Attention weights are dynamically adjusted to accurately capture the differences in feature importance across different developmental stages. The constructed developmental stage knowledge subgraph fully represents the developmental patterns and internal connections of that stage.

[0054] Graph attention networks are used to perform cross-stage association learning on knowledge subgraphs for each developmental stage. The graph attention network treats each developmental stage's knowledge subgraph as a node, establishing connections between stages. The network aggregates information from adjacent stages through a message-passing mechanism, learning the transmission patterns of developmental features. Attention coefficients calculate the importance of inter-stage connections, identifying key developmental transition paths. Cross-stage association learning captures the continuity and stage-specific leaps in developmental trajectories. The learning process considers dynamic factors such as developmental speed, ability accumulation, and key turning points. A fused age-appropriate knowledge graph is generated, integrating the subgraphs of each stage and their relationships. The fused graph maintains the integrity of features within each stage while reflecting the evolutionary patterns between stages. Graph nodes represent developmental entities and stages, and edges represent feature relationships and stage associations. The hierarchical structure design supports multi-granularity queries from micro-features to macro-development. The dynamic update mechanism of the graph adapts to the continuous input of new data and pattern evolution. Through the synergistic effect of critical period segmentation, attention feature extraction, and graph network fusion, a systematic construction of infant and toddler developmental knowledge is achieved. Stage-based processing accurately reflects developmental patterns, the attention mechanism improves feature extraction accuracy, and graph network fusion ensures the integrity of the knowledge system. This approach provides comprehensive knowledge support for personalized childcare services. Attached Figure Description

[0055] Figure 1 This is a schematic diagram illustrating the working principle of the method for constructing an age-specific knowledge graph of multi-source heterogeneous data for infant and toddler care services as described in this invention.

[0056] Figure 2 A flowchart for outlier removal and data alignment processing of multi-source heterogeneous data streams;

[0057] Figure 3 A flowchart for extracting entity features and relation paths for an attention mechanism encoder;

[0058] Figure 4 This is a distribution analysis diagram of suspicious and defective nodes;

[0059] Figure 5 A comparative chart of systematic abnormality indicators in knowledge graphs for infants and young children at different developmental stages. Detailed Implementation

[0060] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0061] Please see Figure 1 This invention provides a method for constructing age-specific knowledge graphs from multi-source heterogeneous data on infant and toddler care services. This method integrates multi-source data streams and performs intelligent analysis based on developmental stages to achieve structured representation of knowledge. The overall implementation scheme includes four main steps: data acquisition and preprocessing, developmental stage segmentation, knowledge subgraph construction, and cross-stage fusion. Specific implementation methods are as follows:

[0062] This method is achieved by collecting multi-source heterogeneous data streams from infant and toddler care services. These multi-source heterogeneous data streams include real-time health indicator data, continuous behavioral trajectory data, and dynamic environmental parameter data. Outlier removal and data alignment are performed on the multi-source heterogeneous data streams to generate a time-aligned data set. Based on the critical periods of infant and toddler neurodevelopment, the time-aligned data set is divided into multiple developmental stage data groups according to the critical periods. For each developmental stage data group, an attention mechanism encoder is used to extract entity features and relational paths to construct a developmental stage knowledge subgraph. A graph attention network is used to perform cross-stage association learning on the knowledge subgraphs of each developmental stage to generate a fused age-specific knowledge graph.

[0063] Example 1: See Figure 2In outlier removal and data alignment, a sliding time window is used to segment the data stream, and the statistical outlier value of the data within each window is calculated. When the fluctuation range of a data point exceeds the outlier threshold, the data point is marked as an outlier and interpolated for replacement. A dynamic time warping algorithm is used to align data sequences with different sampling frequencies to a unified time axis, generating a time-aligned dataset. When dividing the developmental stage data groups based on infant neurodevelopmental milestones, a mapping table between infant age and developmental ability is obtained, and the age boundaries of each critical developmental period are determined according to the mapping table. The time-aligned dataset is divided into continuous data segments according to age boundaries, with each segment corresponding to a developmental stage data group. Developmental labels are added to each developmental stage data group, including the motor skills period, language explosion period, and socio-emotional period.

[0064] In practice, a sliding time window is used to segment the data stream, calculating the statistical outlier value within each window. When the fluctuation of a data point exceeds the outlier threshold, the data point is marked as an outlier and replaced by interpolation. A dynamic time warping algorithm is used to align data sequences with different sampling frequencies to a unified time axis, generating a time-aligned dataset. When dividing developmental stage data groups based on infant neurodevelopmental milestones, a mapping table between infant age and developmental ability is obtained, and the age boundaries of each critical developmental period are determined based on the mapping table. The time-aligned dataset is then divided into continuous data segments according to age boundaries, with each segment corresponding to a developmental stage data group. Developmental labels are added to each developmental stage data group, based on the matching of statistical characteristics of the data within the developmental stage data group with preset rules. For example, when the variance of displacement sensor data in a dataset increases significantly and remains high, it is labeled as the "motor skills stage"; when the spectral complexity of infant babbling in audio data exceeds a threshold, it is labeled as the "language burst stage"; when the duration of facial gaze and the frequency of interactive responses both increase in visual sensor data, it is labeled as the "social-emotional stage". Developmental labels include the motor skills stage, the language burst stage, and the social-emotional stage.

[0065] In practice, the size of the sliding time window is adaptively adjusted according to the data sampling frequency to ensure that each window contains enough data points for statistical outlier calculation. The calculation of statistical outliers is based on the distribution characteristics of the data within the window, for example, using the mean μ and standard deviation σ to define the outlier threshold. The formula is expressed as: if data point x satisfies... ,but Marked as outliers, where This represents the arithmetic mean of the data within the sliding time window. This represents the standard deviation of the data within the sliding time window. This represents the predefined outlier threshold multiplier. The size of the sliding time window is dynamically determined based on the shortest sampling period of the data source. For example, for heart rate data at 1-minute intervals, the window size can be set to 10 minutes; for activity data at 5-minute intervals, the window size would be set to 50 minutes. The value of the outlier threshold multiplier k is related to the data type. For health indicator data that conforms to a normal distribution (such as body temperature), k is usually set to 2 to 3; for behavioral data with impulsive peaks (such as crying decibels), k can be appropriately relaxed to 4 to 5 to retain valid behavioral events. When outliers are identified, linear interpolation or spline interpolation methods are used to replace them to maintain data continuity. The dynamic time warping algorithm aligns asynchronously acquired data to a common time axis by calculating the minimum path distance between different sequences, generating a time-aligned dataset.

[0066] In some embodiments, the step size of the sliding time window can be set to an overlapping mode to reduce boundary effects; the outlier threshold multiplier k is dynamically adjusted according to the data type, for example, a stricter threshold is used for health indicator data. In some embodiments, the distance metric of the dynamic time warping algorithm can be Euclidean distance or Manhattan distance to adapt to the alignment requirements of different data characteristics. Optionally, the interpolation replacement method can be selected to predict based on the trend of neighboring data points to enhance the smoothness of the data. Optionally, the output format of the time-series aligned dataset includes timestamp indexes and normalized values ​​to facilitate subsequent processing.

[0067] It is understandable that the mapping table between infants' and toddlers' ages and developmental abilities is derived from authoritative developmental standards, such as the World Health Organization's growth curves or developmental milestone guidelines; age boundaries are defined according to the critical period theory to ensure that each developmental stage data set covers a specific neurodevelopmental window. It is also understandable that the assignment of developmental labels is based on typical performance of developmental milestones: the motor skills period corresponds to the large muscle group coordination stage, the language explosion period corresponds to the rapid vocabulary growth stage, and the socio-emotional period corresponds to the development of emotion recognition and interaction abilities. The divided developmental stage data sets are stored as structured datasets, accompanied by metadata describing the developmental labels and age ranges.

[0068] Example 2: See Figure 3When extracting entity features and relationship paths using an attention mechanism encoder, the developmental stage data sets are input into a multi-head self-attention layer to calculate the association weights between entities. Important relationship paths are selected based on these weights, retaining relationships with weights higher than a path threshold. A graph encoder maps entities and relationships to low-dimensional vectors, constructing a developmental stage knowledge subgraph. When using a graph attention network for cross-stage association learning, the knowledge subgraphs of each developmental stage are input into the graph attention network to calculate cross-stage attention coefficients between nodes. Feature information from adjacent nodes is aggregated through graph convolution operations to update node representations. The updated node representations are then subjected to graph pooling to generate a fused age-specific knowledge graph.

[0069] In its implementation, the attention mechanism encoder operates by inputting the developmental stage data set into a multi-head self-attention layer, calculating the association weights between entities. These weights are calculated using a dot product attention mechanism of query vectors, key vectors, and value vectors. Important relationship paths are filtered based on these association weights, retaining relationships with weights higher than a path threshold. This path threshold is set according to the quantile distribution of relationship weights in the developmental stage data set. A graph encoder maps the filtered entities and relationships into low-dimensional vector representations. The graph encoder employs a translation-based model or a neural network encoder to construct a developmental stage knowledge subgraph containing entity and relationship vectors.

[0070] In practical implementation, the calculation process of the multi-head self-attention layer involves linearly projecting the input entity sequence into multiple subspaces, and independently calculating the attention score in each subspace. The formula for calculating the attention score is as follows:

[0071]

[0072] in: This represents the output of the h-th attention head. The projection matrix of the query vector. The projection matrix of the key vector. This indicates the transpose operation. The projection matrix of the value vector. This represents the dimension of the key vector. The path threshold is set using a dynamic adjustment strategy, such as taking the upper quartile of all relation path weights. The graph encoder's mapping process learns distributed representations of entities and relations, ensuring that semantically similar entities are close in distance within the vector space.

[0073] In practical implementation, when using a graph attention network for cross-stage association learning, knowledge subgraphs of each developmental stage are input into the graph attention network. The graph attention network calculates cross-stage attention coefficients between nodes, which measure the association strength between nodes in knowledge subgraphs of different developmental stages. Feature information from neighboring nodes is aggregated through graph convolution operations to update node representations. Graph convolution operations perform feature propagation and transformation within the node's neighborhood. The updated node representations are then subjected to graph pooling operations. Graph pooling reduces the graph size through downsampling while preserving important topological features, generating a fused age-specific knowledge graph.

[0074] In some embodiments, the number of attention heads can be set to 8 or 16 to capture relational patterns in different subspaces. In some embodiments, dynamic adjustment of the path threshold can introduce a sliding window mechanism, updating it based on the weight distribution of the most recent batch of data. Optionally, the graph encoder can be trained using negative sampling techniques to optimize model parameters by maximizing the scores of positive samples and minimizing the scores of negative samples. Optionally, the graph pooling operation can choose to use self-attention pooling or global pooling methods to adapt to knowledge subgraphs of different sizes.

[0075] It is understandable that calculating cross-stage attention coefficients requires considering the semantic similarity and temporal relevance between nodes, providing a quantitative basis for cross-stage connections. Similarly, it is understandable that the number of layers in a graph convolution operation determines the neighborhood range of feature aggregation; multi-layer graph convolution can capture a wider range of contextual information. The age-specific knowledge graph after graph pooling integrates the association information of multiple developmental stages, forming a unified representation of infant developmental trajectories. The fused age-specific knowledge graph supports cross-stage developmental pattern analysis and prediction tasks.

[0076] Example 3: Measure the entity richness and relation diversity of an age-specific knowledge graph, and calculate the graph structure strength index based on these metrics. Set an adaptive quality threshold based on the graph structure strength index, perform isolated node detection on the age-specific knowledge graph, and label suspicious data nodes. Count the number of different entity types in the age-specific knowledge graph and calculate the entity type entropy as the entity richness. Analyze the uniformity of relation path distribution and calculate the relation distribution divergence as the relation diversity. Summate the entity richness and relation diversity with weights to obtain the graph structure strength index.

[0077] In a specific implementation, a graph structure strength index is calculated based on entity richness and relation diversity. An adaptive quality threshold is then set based on this index to detect isolated nodes and label suspicious data nodes in the age-graded knowledge graph. Specifically, entity richness is measured by counting the number of different entity types in the age-graded knowledge graph and calculating entity type entropy values, which reflect the uniformity and diversity of entity type distribution. Relation diversity is measured by analyzing the uniformity of relation path distribution and calculating relation distribution divergence, which characterizes the balance of different relation types appearing in the graph. Entity richness and relation diversity are fused using a weighted summation operation to generate the graph structure strength index, which is used to evaluate the overall structural integrity of the age-graded knowledge graph.

[0078] In practical implementation, the calculation of entity type entropy is based on information entropy theory, and the formula is expressed as:

[0079]

[0080] in: Represents the entropy value of the entity type. This represents the total number of entity types in the age-based knowledge graph. Indicates the first The frequency of entity types appearing in age-specific knowledge graphs. The divergence of relation distribution is calculated using the Gini impurity metric; a larger divergence value indicates a more uneven distribution of relation types. Graph structure strength index. The calculation formula is: ,in: This represents entity richness (i.e., entity type entropy value). This represents the diversity of relationships (i.e., the divergence of relationship distribution). and These represent the weight coefficients corresponding to entity richness and relation diversity, respectively.

[0081] In some embodiments, entity type entropy can be calculated separately for knowledge subgraphs at specific developmental stages to obtain a more granular structural assessment. In some embodiments, the calculation of relation distribution divergence can incorporate the semantic depth of relation paths as an adjustment factor to more accurately reflect the actual diversity of relations. Optionally, weighting coefficients... and The value can be preset based on domain knowledge; for example, it can be assigned a value when more attention is paid to entity diversity. Larger values. Optionally, the calculation of the spectral structure strength index can incorporate normalization to ensure that the index value falls within a fixed range for easier comparison.

[0082] It is understandable that when setting an adaptive quality threshold based on the graph structure strength index, the threshold and the graph structure strength index are negatively correlated; a lower graph structure strength index corresponds to a higher quality threshold for more stringent isolated node detection. The isolated node detection process traverses each node in the age-specific knowledge graph, calculating the number of neighboring nodes and the connection density. When the number of neighboring nodes is below the isolated node threshold or the connection density is below the density threshold, the node is labeled as a suspicious data node. The labeling information for suspicious data nodes includes the node identifier, its developmental stage, and a summary of its structural features.

[0083] Example 4: When detecting isolated nodes in an age-specific knowledge graph, the number of neighboring nodes and connection density of each node are calculated. When the number of neighboring nodes of a node is lower than the isolation threshold or the connection density is lower than the density threshold, the node is marked as a suspicious data node. The location identifier and attribute summary of the suspicious data node are recorded. When extracting the spatiotemporal features and context embedding vectors of suspicious data nodes, time trend features and spatial distribution features are extracted from the node's historical change sequence. A graph neural network is used to generate the node's context embedding vector, which contains the node's local topological information. The time trend features, spatial distribution features, and context embedding vector are concatenated into a feature vector, which is then input into a defect classifier to determine whether it is a real defect.

[0084] In practice, the isolated node detection process is achieved by calculating the number of neighboring nodes and the connection density of each node in the age-graded knowledge graph. When the number of neighboring nodes of a node is lower than the isolation threshold or the connection density is lower than the density threshold, the age-graded knowledge graph system marks the node as a suspicious data node and records the location identifier and attribute summary of the suspicious data node. Specifically, the calculation of the number of neighboring nodes is based on a traversal and statistical analysis of the adjacency matrix of the age-graded knowledge graph, and the calculation of the connection density is based on the ratio of the number of edges in the node's local subgraph to the maximum possible number of edges, expressed by the formula:

[0085]

[0086] in: Indicates connection density. This represents the actual number of edges in the local subgraph centered at the current node. This represents the total number of nodes contained in the local subgraph. The isolation threshold and density threshold are preset based on the global average connectivity characteristics of the age-specific knowledge graph; for example, the isolation threshold is set to 50% of the average degree of the age-specific knowledge graph, and the density threshold is set to 60% of the global density of the age-specific knowledge graph. Referring to Table 1, the attribute summary of suspicious data nodes includes node type, degree centrality, and the label of the developmental stage to which they belong.

[0087] Table 1: Summary of Attributes of Suspicious Data Nodes

[0088] ;

[0089] In practice, when extracting the spatiotemporal features and context embedding vectors of suspicious data nodes, the spatiotemporal features are derived from the node's historical change sequence. The time trend features are calculated using a sliding window to determine the mean rate of change of node attributes, and the spatial distribution features are obtained by measuring the variance of the node's position coordinates in the age-specific knowledge graph. The context embedding vector is generated using a graph neural network. This network samples multi-hop neighbors centered on the suspicious data node, aggregates local topological information through a message-passing mechanism, and outputs a fixed-dimensional vector representation. The time trend features, spatial distribution features, and context embedding vector are concatenated into a comprehensive feature vector, which is input into a pre-trained defect classifier to determine whether it is a genuine defect. The defect classifier uses a support vector machine or neural network model, and its output is a binary classification probability.

[0090] In some embodiments, the calculation of the number of neighboring nodes can incorporate edge weights, giving greater importance to strong connections. In some embodiments, the generation of the context embedding vector can employ a hierarchical attention mechanism to dynamically adjust the contribution weights of neighbors with different hop counts. Optionally, the calculation of time trend features can incorporate a seasonal decomposition method to separate long-term trends from short-term fluctuations. Optionally, the training data for the defect classifier is derived from historically labeled real defect node samples, and cross-validation is used to optimize the model parameters.

[0091] It is understandable that location identifiers employ unique encoding rules to ensure the traceability of suspicious data nodes within the age-specific knowledge graph. It is also understandable that attribute summaries are recorded in a key-value pair structure for ease of subsequent querying and analysis. The extraction of spatiotemporal features relies on the version history logs of the age-specific knowledge graph, and the generation of context embedding vectors depends on the embedding layer parameters of the graph neural network. The judgment results of the defect classifier are used to trigger the quality calibration process of the age-specific knowledge graph.

[0092] See Figure 4In the analysis of suspicious nodes and actual defective nodes in the age-specific knowledge graph of infant and toddler care services, the data distribution shows significant differences across developmental stages. From a data perspective, the total number of nodes is 50 in the socio-emotional stage, 44 in the language explosion stage, and 56 in the motor skills stage; the number of suspicious nodes is 18, 13, and 14 respectively; the number of actual defective nodes is 10 in both the socio-emotional and language explosion stages, and 7 in the motor skills stage. The percentage of suspicious nodes decreases with developmental stage, with the highest percentage in the socio-emotional stage, followed by the language explosion stage, and the lowest in the motor skills stage. This distribution characteristic stems from the isolated node detection and defect assessment conducted on multi-source heterogeneous data from different developmental stages (social-emotional stage, language explosion stage, and motor skills stage) during the construction of the age-specific knowledge graph. By calculating the number of neighboring nodes and connection density to identify suspicious nodes, and then combining spatiotemporal features and contextual embedding vectors to determine the real defects, the differences in the number and proportion of nodes at each stage are finally presented. This provides data support for the quality calibration and optimization of knowledge graphs in infant and toddler care services, and also reflects the heterogeneity and defect distribution patterns of data in different critical periods of neurodevelopment.

[0093] Example 5: When using density clustering to group real defective nodes, the neighborhood radius and minimum sample number parameters for density clustering are set; the feature distance matrix between real defective nodes is calculated, and density clustering is performed based on the distance matrix; when a large-scale node group spanning developmental stages appears in the clustering results, a systematic anomaly is identified. When calculating the calibration weights based on the propagation degree of the systematic anomaly, the subgraph range and node proportion affected by the systematic anomaly are measured; the anomaly impact factor is calculated based on the subgraph range and node proportion, and the anomaly impact factor is mapped to the calibration weight interval; the calibration weights are used to weight and adjust the basic graph quality score, and the optimization interval of the knowledge graph is recalculated based on the adjusted score.

[0094] In practical implementation, density clustering requires presetting a neighborhood radius parameter ε and a minimum sample number parameter η. The neighborhood radius parameter ε defines the distance threshold at which a node is considered a neighbor, and the minimum sample number parameter η specifies the minimum number of neighboring nodes required for a core node. The feature distance matrix between real defective nodes is calculated based on their topological attributes, including node degree, clustering coefficient, and centrality index. The feature distance matrix uses Euclidean distance to measure the difference in feature vectors between nodes. Density clustering is then performed based on the feature distance matrix, dividing real defective nodes into multiple node clusters. When a large-scale node cluster spanning multiple developmental stages appears in the clustering results, a systematic anomaly is identified. "Spanning multiple developmental stages" means that the real defective nodes in the node cluster originate from multiple data sets at different developmental stages, and "large-scale node cluster" means that the number of real defective nodes in the node cluster exceeds a pre-set size threshold.

[0095] In practice, the propagation degree of systematic anomalies is quantified by measuring the subgraph extent and node proportion affected by the systematic anomaly. The subgraph extent refers to the number and size of connected subgraphs affected by the systematic anomaly, and the node proportion refers to the percentage of affected nodes relative to the total number of nodes in the age-specific knowledge graph. The anomaly impact factor is calculated based on the subgraph extent and node proportion, expressed by the following formula:

[0096]

[0097] in: Indicates abnormal influencing factors. This represents the total number of connected subgraphs affected by systematic anomalies. Indicates the first The number of nodes in the affected connected subgraph. This represents the total number of nodes in the age-specific knowledge graph. This represents the total number of nodes affected by the systemic anomaly. The anomaly impact factor is mapped to a calibration weight interval, set to [0.5, 1.0]. The mapping function uses linear scaling, and the formula is as follows: ,in This represents the calculated calibration weights. The calibration weights are used to weight and adjust the basic spectrum quality score; the adjusted score is calculated using the following formula: ,in This indicates the adjusted spectrum quality score. This represents the quality score of the basic knowledge graph. The optimization interval of the knowledge graph is recalculated based on the adjusted score. The optimization interval is negatively correlated with the adjusted score; the lower the adjusted score, the shorter the optimization interval should be.

[0098] In some embodiments, the parameter ε of the density clustering algorithm can be determined by analyzing the distance distribution between real defect nodes, for example, by selecting the inflection point value of the distance distribution. In some embodiments, the calculation of the anomaly impact factor can incorporate the weight of the subgraph depth, assigning higher importance to subgraphs containing core entities. Optionally, the mapping interval of the calibration weight can be adjusted according to the fault tolerance requirements of the actual application scenario, for example, set to [0.3, 0.9]. Optionally, the calculation of the optimization interval can use a piecewise function, setting a fixed inspection cycle for different scoring intervals.

[0099] It is understandable that large-scale node clusters spanning developmental stages indicate that defect patterns may originate from common biases in data acquisition or processing, rather than isolated random errors. It is also understandable that the propagation of systematic anomalies reflects the severity of the problem, and higher anomaly impact factors necessitate more timely knowledge graph optimization interventions. The introduction of calibration weights enables the graph quality score to dynamically reflect structural risks, and the intelligent adjustment of optimization intervals achieves efficient allocation of maintenance resources.

[0100] See Figure 5 This study presents three key indicators related to systematic anomalies in knowledge graphs in infants and toddlers at different developmental stages (motor skills stage (0-6 months), language explosion stage (7-12 months), and socio-emotional stage (13-24 months)): average number of affected connected subgraphs, average percentage of affected nodes (%), and average anomaly impact factor (×100). Regarding the average number of affected connected subgraphs, the figure is 3.9 in the motor skills stage, 4.1 in the language explosion stage, and 4.3 in the socio-emotional stage, showing a gradual increase with developmental stage, indicating that the number of connected subgraphs affected by systematic anomalies gradually increases during infant and toddler growth. The average percentage of affected nodes (%) is 13.4 in the motor skills stage, 14.3 in the language explosion stage, and reaches 15.2 in the socio-emotional stage, also showing an upward trend, reflecting that the proportion of nodes affected by systematic anomalies in the total number of nodes continuously increases with developmental stage. The average abnormality impact factor (×100) was 46.2 during the motor skills period, rose to 48.1 during the language explosion period, and slightly decreased to 47.1 during the socio-emotional period. Overall, it was at a high level, indicating that the impact of systemic abnormalities on knowledge graphs was significant during each critical developmental period of infants and young children, with the highest abnormality impact factor during the language explosion period.

[0101] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for constructing an age-specific knowledge graph of multi-source heterogeneous data on infant and toddler care services, characterized in that, The method includes: Collect multi-source heterogeneous data streams from infant and toddler care services. The multi-source heterogeneous data streams include real-time health indicator data, continuous behavioral trajectory data, and dynamic environmental parameter data. Perform outlier removal and data alignment processing on the multi-source heterogeneous data streams to generate a time-series aligned data set. Based on the critical periods of infant and toddler neurodevelopment, the time-aligned dataset is divided into multiple developmental stage data groups according to the critical developmental periods; For each developmental stage data set, an attention mechanism encoder is used to extract entity features and relational paths to construct a developmental stage knowledge subgraph. A graph attention network is used to perform cross-stage association learning on knowledge subgraphs of each developmental stage to generate a fused age-specific knowledge graph. The method further includes: The entity richness and relation diversity of the age-specific knowledge graph are measured, and the graph structure strength index is calculated based on the entity richness and relation diversity. Based on the graph structure strength index, an adaptive quality threshold is set to detect isolated nodes in age-specific knowledge graphs and label suspicious data nodes. When there are suspicious data nodes, extract the spatiotemporal features and context embedding vector of each suspicious data node, and determine whether the suspicious data node constitutes a real defect based on the spatiotemporal features and context embedding vector. When it is confirmed to be a real defect, calculate the basic map quality score based on the number of real defect nodes. Topological attributes are extracted from real defect nodes, and the topological attributes are used as input features. Density clustering algorithm is used to group the real defect nodes, and the presence of systematic anomalies is detected based on the grouping results. When a systematic anomaly is identified, a calibration weight is calculated based on the degree of propagation of the systematic anomaly. The calibration weight is then used to correct the quality score of the basic knowledge graph and to intelligently adjust the optimization interval of the knowledge graph. When extracting the spatiotemporal features and context embedding vectors of suspicious data nodes, the following are included: Extracting temporal trend features and spatial distribution features from the historical change sequence of nodes; A graph neural network is used to generate a context embedding vector for a node, the context embedding vector containing local topological information of the node; The time trend features, spatial distribution features, and context embedding vectors are concatenated into a feature vector, which is then input into the defect classifier to determine whether it is a real defect. When using density clustering algorithms to group real defect nodes, the following steps are included: Set the neighborhood radius and minimum number of samples for density clustering; Calculate the feature distance matrix between real defect nodes, and perform density clustering based on the distance matrix; When a large cluster of nodes spanning multiple developmental stages appears in the clustering results, a systematic anomaly is identified. When calculating calibration weights based on the degree of propagation of systemic anomalies, the following are included: The subgraph extent and node proportions of the impact of systemic anomalies are measured; The anomaly impact factor is calculated based on the subgraph range and node ratio, and then mapped to the calibration weight interval. The quality score of the basic knowledge graph is adjusted using calibration weights, and the optimization interval of the knowledge graph is recalculated based on the adjusted score.

2. The method for constructing an age-specific knowledge graph of multi-source heterogeneous data on infant and toddler care services according to claim 1, characterized in that, When performing outlier removal and data alignment on multi-source heterogeneous data streams, the following steps are included: Set a sliding time window to segment the data stream and calculate the statistical outlier value of the data in each window; When the fluctuation range of a data point exceeds the outlier threshold, the data point is marked as an outlier and replaced by interpolation. A dynamic time warping algorithm is used to align data sequences with different sampling frequencies to a unified time axis, generating a time-aligned data set.

3. The method for constructing an age-specific knowledge graph of multi-source heterogeneous data on infant and toddler care services according to claim 1, characterized in that, When dividing developmental stage data groups based on infant and toddler neurodevelopmental milestones, the following are included: Obtain a mapping table between infants' and toddlers' ages and developmental abilities, and determine the age boundaries of each critical developmental period based on the mapping table; The time-aligned dataset is divided into continuous data segments according to age boundaries, with each data segment corresponding to a developmental stage data group. Developmental labels were added to each developmental stage data set, including the motor skills period, language explosion period, and socio-emotional period.

4. The method for constructing an age-specific knowledge graph of multi-source heterogeneous data on infant and toddler care services according to claim 1, characterized in that, When using an attention-based encoder to extract entity features and relational paths, the following are included: The developmental stage data set is input into the multi-head self-attention layer to calculate the association weights between entities; Filter important relationship paths based on association weights, and retain relationships with weights higher than the path threshold; A graph encoder is used to map entities and relations to low-dimensional vectors, constructing a knowledge subgraph for developmental stages.

5. The method for constructing an age-specific knowledge graph of multi-source heterogeneous data on infant and toddler care services according to claim 4, characterized in that, When using graph attention networks for cross-stage association learning, including: Input the knowledge subgraphs of each developmental stage into the graph attention network and calculate the cross-stage attention coefficients between nodes; The node representation is updated by aggregating the feature information of adjacent nodes through graph convolution operations. The updated node representations are then subjected to graph pooling to generate a fused age-specific knowledge graph.

6. The method for constructing an age-specific knowledge graph of multi-source heterogeneous data on infant and toddler care services according to claim 1, characterized in that, When measuring entity richness and relation diversity, the following are included: The number of different entity types in the age-specific knowledge graph is counted, and the entity type entropy value is calculated as the entity richness. Analyze the uniformity of the distribution of relation paths and calculate the relation distribution divergence as relation diversity; The graph structure strength index is obtained by weighted summation of entity richness and relation diversity.

7. The method for constructing an age-specific knowledge graph of multi-source heterogeneous data on infant and toddler care services according to claim 6, characterized in that, When performing isolated node detection on age-specific knowledge graphs, the following is included: Calculate the number of neighboring nodes and connection density for each node; When the number of neighboring nodes of a node is lower than the isolation threshold or the connection density is lower than the density threshold, the node is marked as a suspicious data node. Record the location identifier and attribute summary of suspicious data nodes.

Images