Knowledge graph-driven sales data multi-dimensional analysis and visualization method and system

Abstract

The invention discloses a knowledge graph-driven sales data multi-dimensional analysis and visualization method and system. The method comprises the following steps: constructing a sales domain knowledge graph containing products, customers and transaction entities; on the basis of the knowledge graph, preprocessing the original sales data to obtain preprocessed sales data; according to a knowledge graph and the preprocessed sales data, realizing time sequence adaptive multi-view collaborative graph embedding, and obtaining graph embedding features; fusing the graph embedding feature with the traditional sales data feature to obtain a fused feature; performing multi-dimensional analysis based on the fusion features to obtain an analysis result; performing visual presentation on the analysis result to generate visual data; on the basis of visual data, an interactive exploration interface is provided, deep analysis and insight of data are achieved, by introducing the knowledge graph technology, the complex relation between entities can be effectively captured and utilized, the association rule mining effect is remarkably improved, and the association rule mining efficiency is improved by 77.8% compared with a traditional method.

Description

Technical Field

[0001] The present invention relates to the technical field of sales data multidimensional analysis and visualization methods, and more specifically, to a knowledge graph-driven sales data multidimensional analysis and visualization method and system thereof. Background Art

[0002] With the rapid development of e-commerce and the deepening of digital transformation, sales data analysis has become an important basis for corporate decision-making. Traditional sales data analysis methods mainly include statistical analysis, data mining, and machine learning, which perform well when processing structured data and simple business scenarios. However, with the increase in business complexity and the expansion of data scale, existing technologies face severe challenges in the following aspects:

[0003] 1. Insufficient ability to handle complex relationships: Traditional methods mainly focus on the attribute analysis of a single entity and are unable to effectively handle complex relationships between entities. For example, a well-known e-commerce platform uses traditional association rule mining algorithms to analyze product relationships, but it is unable to fully consider multi-dimensional factors such as product categories, brands, and user preferences, resulting in limited value of the mined association rules.

[0004] 2. Insufficient dynamic feature modeling: Most existing methods assume that data features are static and have difficulty capturing dynamic changes in sales data. For example, a large retail company uses traditional time series models for sales forecasting, but cannot effectively handle the impact of dynamic factors such as holidays and promotions, resulting in unstable forecast accuracy.

[0005] 3. Improper processing of long-tail data: Existing technologies often focus on processing high-frequency data, and do not pay enough attention to the analysis of long-tail data. For example, a luxury e-commerce platform uses conventional classification algorithms to recommend products, but the recommendation effect is poor for customized products with low sales, which affects the sales of long-tail products.

[0006] 4. Difficulty in integrating heterogeneous data from multiple sources: With the diversification of data sources, how to effectively integrate structured, semi-structured and unstructured data has become a major challenge. A large cross-border e-commerce platform tried to use traditional feature engineering methods to integrate data from different sources, but the effect was not good and it was unable to fully utilize the rich data resources.

[0007] 5. Insufficient explainability: With the application of deep learning technology, the explainability of the model has become a prominent issue. A well-known cosmetics company used deep neural networks to predict customer churn. Although the accuracy was high, it was difficult to explain the reasons for the prediction results to business personnel, which affected the practical application of the model.

[0008] 6. Low efficiency of visualization analysis: When faced with large-scale high-dimensional data, traditional visualization technology performs poorly in terms of response speed and interactive experience. When a large home appliance chain tried to visualize and analyze millions of transaction data, the system responded slowly, which seriously affected the analysis efficiency.

[0009] 7. Insufficient system scalability: With the rapid increase in data volume, the existing system architecture design is difficult to meet the analysis needs of high concurrency and large data volumes. A well-known online travel platform often faces the risk of system crashes during business peaks and cannot support real-time large-scale data analysis.

[0010] In the face of these challenges, innovative sales data analysis methods are urgently needed that can comprehensively utilize advanced technologies such as knowledge graphs, deep learning, and high-dimensional data visualization to achieve comprehensive, in-depth, and efficient analysis of sales data. Summary of the invention

[0011] In order to solve the above technical problems, the present invention proposes a knowledge graph-driven sales data multidimensional analysis and visualization method and system.

[0012] The present invention provides a knowledge graph-driven sales data multidimensional analysis and visualization method, comprising the following steps:

[0013] Build a sales domain knowledge graph containing product, customer, and transaction entities;

[0014] Based on the knowledge graph, preprocessing the original sales data to obtain preprocessed sales data;

[0015] According to the knowledge graph and the preprocessed sales data, time-series adaptive multi-view collaborative graph embedding is implemented to obtain graph embedding features;

[0016] Fusion the graph embedding features with traditional sales data features to obtain fusion features;

[0017] Perform multi-dimensional analysis based on the fusion features to obtain analysis results;

[0018] Visually presenting the analysis results to generate visual data;

[0019] Based on the visualized data, an interactive exploration interface is provided to achieve in-depth analysis and insight into the data.

[0020] Specifically, the temporal adaptive multi-view collaborative graph embedding includes the following steps:

[0021] Execute the temporal attention mechanism to obtain the embedding representation of node v at time t

[0022]

[0023] in, represents the embedding representation of node v at time t-1; N(v) represents the set of neighbors of node v; represents the attention weight of node v to neighbor node u at time t; W h ,W a is a learnable weight matrix; a is a learnable vector parameter; σ is the LeakyReLU activation function; || represents a vector concatenation operation;

[0024] Perform multi-view fusion to obtain the fused node representation h v :

[0025]

[0026] in, represents the embedding representation of node v in the i-th view; M represents the number of views, β i represents the weight of the i-th view; MLP represents multi-layer perceptron;

[0027] Perform knowledge distillation and calculate the knowledge distillation loss L KD :

[0028]

[0029] in, and are the outputs of the teacher model and the student model respectively, τ is the temperature parameter, KL is the KL divergence, and σ represents the softmax function;

[0030] Execute the meta-learning anti-long-tail module and update the model parameters θ:

[0031]

[0032] Among them, θ ′ represents the model parameters after the inner loop is updated, α and β are the learning rates of the inner and outer loops respectively, L S and L Q is the loss function of the support set and query set, f θ represents the model function with parameter θ, represents the gradient with respect to θ.

[0033] Specifically, the feature fusion adopts a dynamic cross-modal adaptive feature fusion algorithm, including the following steps:

[0034] Calculate the importance weights of graph embedding features and traditional features:

[0035]

[0036] Among them, φ G and φ T is the dynamic weight generation function, H G is the graph embedding feature matrix, X T is the feature matrix of traditional sales data, sigmoid represents the sigmoid activation function;

[0037] Perform nonlinear transformation on graph embedding features and traditional features:

[0038] H G ′=LeakyReLU(W1·H G +b1)

[0039] X T ′=LeakyReLU(W2·X T +b2)

[0040] Among them, H ′ G represents the transformed graph embedding features, X ′ T represents the transformed traditional features; W1, W2 represent the learnable weight matrices; b1, b2 represent the learnable bias vectors;

[0041] Introduce the mutual information maximization term:

[0042] I(H G ′,X T ′)=E[f(H G ′,X T ′)]-log(E[exp(f(H G ′,X T ″))])

[0043] Among them, f is the discriminator network, X ′ T ′ Yes X ′ T A random permutation of ; E represents expectation;

[0044] Feature fusion based on dynamic weights:

[0045]

[0046] Add residual connections and perform layer normalization:

[0047] F = LayerNorm(F + [H G ∥X T ])

[0048] LayerNorm represents layer normalization operation; ∥ represents vector concatenation operation;

[0049] Optimization goal:

[0050] L=L task +λ1|L MI -λ2|I(H G ′,X T ′)

[0051] Among them, L task is the loss function of the downstream task, L MI is the mutual information maximization loss, and λ1 and λ2 are balancing coefficients.

[0052] Specifically, the multi-dimensional analysis includes sales trend analysis, customer segmentation, product association analysis, customer life cycle value prediction and abnormal transaction detection.

[0053] Specifically, the sales trend analysis adopts the Prophet time series forecasting model:

[0054] y(t)=g(t)+s(t)+h(t)+∈ t

[0055] Among them, g(t) is the trend term, s(t) is the season term, h(t) is the holiday term, ∈ t is the error term.

[0056] Specifically, the customer clustering adopts the improved DBSCAN algorithm and introduces the Mahalanobis distance to measure similarity:

[0057]

[0058] Among them, ρ i represents the local density of sample point i, x i ,x j represents the sample points in the data set, d M (x i ,x j ) represents the sample point x i and x j The Mahalanobis distance between them; σ represents the kernel width parameter, which is used to control the range of density estimation; exp represents the exponential function.

[0059] Specifically, the product association analysis adopts an improved Apriori algorithm and introduces a time decay factor:

[0060]

[0061] Where supp(X) represents the support of item set X, X represents the item set to be analyzed; t represents the time when the transaction occurs; T represents the current time; δ represents the time decay factor, which ranges from (0,1); c t(X) represents the number of occurrences of item set X at time t; N represents the total number of transactions.

[0062] Specifically, the customer lifetime value prediction adopts an RNN model based on the attention mechanism:

[0063]

[0064] Among them, CLTV represents the predicted customer lifetime value; T represents the predicted time span, α t represents the attention weight x at time step t t represents the customer feature vector at time step t; RNN represents the recurrent neural network function.

[0065] Specifically, the visualization uses an improved ForceAtlas2 algorithm to perform force-directed graph layout:

[0066]

[0067] Among them, F r (u,v) represents the repulsive force between nodes u and v; k r represents the repulsion coefficient, deg(u) and deg(v) represent the degrees of nodes u and v respectively; d(u,v) represents the distance between nodes u and v;

[0068] And use the Barnes-Hutt-SNE algorithm for high-dimensional data projection:

[0069]

[0070] Among them, p j|i represents the conditional probability of point i choosing point j as a neighbor in high-dimensional space; x i ,x j ,x k represents a data point in high-dimensional space, σ i represents the local density parameter of point i; ∥·∥ represents the Euclidean distance; exp represents the exponential function.

[0071] The knowledge graph-driven sales data multi-dimensional analysis and visualization system includes:

[0072] The knowledge graph building module is used to build a sales domain knowledge graph containing entities such as products, customers, and transactions;

[0073] A data preprocessing module, used to preprocess the original sales data based on the knowledge graph to obtain preprocessed sales data;

[0074] A graph embedding module, used to implement time-series adaptive multi-view collaborative graph embedding according to the knowledge graph and the preprocessed sales data, and obtain graph embedding features;

[0075] A feature fusion module, used for fusing the graph embedding feature with the traditional sales data feature to obtain a fusion feature;

[0076] A multi-dimensional analysis module, used to perform multi-dimensional analysis based on the fusion features to obtain analysis results;

[0077] A visualization module, used to visualize the analysis results and generate visualization data;

[0078] The interactive exploration module is used to provide an interactive exploration interface based on the visualized data to achieve in-depth analysis and insight into the data.

[0079] The beneficial effects of the present invention are mainly reflected in the following aspects:

[0080] 1. Improve the ability to handle complex relationships: By introducing knowledge graph technology, the present invention can effectively capture and utilize complex relationships between entities, significantly improving the effect of association rule mining, with an average improvement of 3.2, which is 77.8% higher than traditional methods.

[0081] 2. Enhanced dynamic feature modeling capabilities: Using a time-series adaptive multi-view collaborative graph embedding algorithm, the present invention can better capture the dynamic characteristics of sales data. In the sales forecasting task, the MAPE is reduced to 8.2%, which is 47.8% more accurate than the traditional method.

[0082] 3. Improve the long-tail data processing effect: Through the meta-learning anti-long-tail module, the present invention has made significant progress in the analysis of long-tail products, with an F1 score of 0.73, which is 62.2% higher than the traditional method.

[0083] 4. Effective integration of multi-source heterogeneous data: By adopting a dynamic cross-modal adaptive feature fusion algorithm, the present invention can effectively integrate data from different sources and types. In the customer clustering task, the silhouette coefficient reaches 0.68, which is 61.9% higher than the traditional method.

[0084] 5. Improve model interpretability: By combining knowledge graphs and visualization technology, the present invention greatly improves the interpretability of analysis results, enabling decision makers to better understand and apply analysis results.

[0085] 6. Improve the efficiency of visualization analysis: Through improved visualization algorithms and high-performance rendering technology, the present invention reduces the visualization response time to 0.8 seconds, which is 65.2% more efficient than traditional methods.

[0086] 7. Enhanced system scalability: By adopting microservice architecture and containerized deployment strategy, the system scalability of the present invention is significantly improved, and can stably support 12,000 QPS, which is 4 times that of traditional methods.

[0087] In summary, the present invention comprehensively solves the various challenges faced by existing sales data analysis methods by innovatively combining knowledge graphs, deep learning and high-dimensional data visualization technology, providing enterprises with a powerful, efficient and scalable sales data analysis tool, which strongly supports data-driven business decisions. BRIEF DESCRIPTION OF THE DRAWINGS

[0088] Figure 1 This is a flow chart of the knowledge graph-driven sales data multidimensional analysis and visualization method of the present invention.

[0089] Figure 2 This is a schematic diagram of the framework of the knowledge graph-driven sales data multidimensional analysis and visualization system of the present invention. DETAILED DESCRIPTION

[0090] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the specific implementation methods, structures, features and effects thereof are described in detail below in conjunction with the accompanying drawings and preferred embodiments. In the following description, different "one embodiment" or "another embodiment" does not necessarily refer to the same embodiment. In addition, specific features, structures or characteristics in one or more embodiments may be combined in any suitable form.

[0091] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

[0092] The present invention provides a knowledge graph-driven sales data multidimensional analysis and visualization method and system thereof, aiming to solve the deficiencies of traditional sales data analysis methods in processing complex relationships, dynamic changes, and long-tail data. The present invention achieves a comprehensive, in-depth, and intuitive analysis of sales data by combining knowledge graphs, deep learning, and high-dimensional data visualization technology.

[0093] The specific implementation of the present invention will be described in detail below in conjunction with a sales data analysis scenario of a large e-commerce platform.

[0094] Please refer to Figure 1 The knowledge graph-driven sales data multidimensional analysis and visualization method of the present invention comprises the following steps:

[0095] 1. Knowledge graph construction step. In this step, the system constructs a sales domain knowledge graph containing entities such as products, customers, and transactions. Taking the e-commerce platform as an example, this step includes the following sub-steps:

[0096] 1.1 Data collection and organization: Collect raw data from multiple data sources such as the platform's transaction system, product management system, user management system, etc. For example, collect all transaction records, product information, user information, etc. for the past three years.

[0097] 1.2 Ontology design: Design the ontology structure in the field of e-commerce. The main entity categories include: Product, Customer, Transaction, Category, Brand, etc. The relationship types include: purchase, belong_to, produce, etc.

[0098] 1.3 Entity Recognition and Relation Extraction: Use named entity recognition (NER) technology to identify entities from texts such as product descriptions and user reviews. For example, from the sentence "iPhone12Pro Max" is Apple's flagship phone launched in 2020, identify "iPhone12Pro Max" (Product) and "Apple" (Brand). Then use remote supervision methods to extract the relationship between entities, such as "iPhone12Pro Max" produces "Apple".

[0099] 1.4 Knowledge Fusion: Use entity alignment technology to resolve the ambiguity of entity reference in different data sources. For example, apple, "Apple", Apple Inc., etc. are referred to as the same entity.

[0100] 1.5 Knowledge storage: Use Neo4j graph database to store the constructed knowledge graph, and design an optimized index structure to improve query efficiency.

[0101] 2. Data preprocessing step: Based on the constructed knowledge graph, the original sales data is preprocessed to obtain preprocessed sales data. Specifically, it includes:

[0102] 2.1 Data cleaning: Processing missing values ​​and outliers. Multiple Imputation is used to fill in missing commodity prices. The specific formula is as follows:

[0103] Y imp =Y obs +∈,∈~N(0,σ 2 )

[0104] Among them, Y imp is the value after filling, Y obs is the average price of similar products observed, and ∈ is a random error that follows a normal distribution.

[0105] The Local Outlier Factor (LOF) algorithm is used to identify abnormal transactions. The LOF score calculation formula is:

[0106]

[0107] Among them, N k (p) is the set of k nearest neighbors of point p, lrd k is the local reachable density. Usually points with LOF scores greater than 2 are considered outliers.

[0108] 2.2 Data standardization: Z-score standardization of numerical features (such as product price and sales volume):

[0109]

[0110] Among them, x is the original value, μ is the mean, and σ is the standard deviation.

[0111] One-hot encoding is performed on categorical features (such as product category and brand).

[0112] 2.3 Feature engineering: extract time features, such as the month, season, holiday, etc. of sales. Generate interactive features, such as the product of customer purchase frequency and product unit price.

[0113] 2.4 Data bucketing: Use the equal-frequency bucketing method to discretize continuous variables (such as customer age and product price) to enhance the robustness of the model.

[0114] 2.5 Data balance,For imbalanced category data (such as high-value customers and low-value customers), the SMOTE (Synthetic Minority Over-sampling Technique) algorithm is used for oversampling. The core idea of ​​the SMOTE algorithm is to interpolate between minority class samples to generate new synthetic samples.

[0115] 3. Time-series adaptive multi-view collaborative graph embedding step: Based on the knowledge graph and pre-processed sales data, time-series adaptive multi-view collaborative graph embedding is implemented to obtain graph embedding features. This step includes the following sub-steps:

[0116] 3.1 Temporal attention mechanism, for each entity (such as goods, users) on the e-commerce platform, consider its characteristics that change over time. For example, the sales volume and ratings of a certain mobile phone at different time points. The calculation of the temporal attention mechanism is as follows:

[0117]

[0118] in, represents the embedding representation of node v (such as a certain mobile phone) at time t, N(v) is the set of neighbor nodes of v (such as users who purchased the mobile phone, similar mobile phones, etc.), σ is the LeakyReLU activation function, and its negative slope is usually set to 0.2; represents the embedding representation of node v at time t-1; represents the attention weight of node v to neighbor node u at time t; W h ,W a is a learnable weight matrix; a is a learnable vector parameter; || represents a vector concatenation operation.

[0119] 3.2 Multi-view fusion module, considering different attribute views of the product, such as basic information view, sales performance view, user evaluation view, etc. The calculation of multi-view fusion is as follows:

[0120]

[0121] in, represents the embedding of node v in the i-th view, M is the number of views (such as 3), MLP is a three-layer perceptron, and the activation function is ELU (Exponential Linear Unit); M represents the number of views, β i represents the weight of the i-th view.

[0122] 3.3 Knowledge Distillation Enhanced Explanation Module, we use a rule-based decision tree as a teacher model and transfer its knowledge to our deep learning model. The knowledge distillation loss is calculated as follows:

[0123]

[0124] in, and are the outputs of the teacher model (decision tree) and the student model (deep learning model), τ is the temperature parameter (usually set to 2), and KL is the KL divergence.

[0125] 3.4 Meta-learning anti-long-tail module, in order to deal with long-tail products or new products, a meta-learning strategy is adopted. In each mini-batch, the data is divided into a support set (popular products) and a query set (long-tail products). The model parameter update process is as follows:

[0126]

[0127] Among them, α and β are the learning rates of the inner loop and the outer loop respectively (usually α=0.01, β=0.001), L S and L Q is the loss function for the support set (high-frequency entities) and the query set (low-frequency entities). ′represents the model parameters after the inner loop update, f θ represents the model function with parameter θ, represents the gradient with respect to θ.

[0128] 4. Feature fusion step: fuse the graph embedding features with the traditional sales data features to obtain fused features. This step uses the dynamic cross-modal adaptive feature fusion (DCAFF) algorithm and includes the following sub-steps:

[0129] 4.1 Dynamic feature importance evaluation, for each product, calculate the importance weights of its graph embedding features and traditional features:

[0130]

[0131] Among them, φ G and φ T is the dynamic weight generation function, H G is the graph embedding feature matrix, X T is the feature matrix of traditional sales data; sigmoid represents the sigmoid activation function;

[0132] 4.2 Adaptive feature transformation, nonlinear transformation of graph embedding features and traditional features:

[0133] H G ′=LeakyReLU(W1|H G +b1)

[0134] X T ′=LeakyReLU(W2·X T +b2)

[0135] The slope of LeakyReLU is usually set to 0.2. ′ G represents the transformed graph embedding features, X ′ T represents the transformed traditional features; W1, W2 represent the learnable weight matrices; b1, b2 represent the learnable bias vectors.

[0136] 4.3 Maximizing cross-modal mutual information, introducing the mutual information maximization term to promote the collaborative representation of the two features:

[0137] I(H G ′,X T ′)=E[f(H G ′,X T ′)]-log(E[exp(f(H G ′,X T ″))])

[0138] Among them, f is the discriminator network, X ′ T ′ Yes X ′ T A random permutation of ; E represents expectation;

[0139] 4.4 Dynamic Fusion: For each product i, feature fusion is performed based on dynamic weights:

[0140]

[0141] 4.5 Add residual connection and perform layer normalization:

[0142] F = LayerNorm(F + [H G ∥X T ])

[0143] LayerNorm represents layer normalization operation; ∥ represents vector concatenation operation;

[0144] Optimization goal:

[0145] L=L task +λ1·L MI -λ2·I(H G ′,X T ′)

[0146] Among them, L task is the loss function of the downstream task, L MI is the mutual information maximization loss, λ1 and λ2 are balance coefficients, usually set λ1 = 0.1, λ2 = 0.01.

[0147] 5. Multi-dimensional analysis step: Perform multi-dimensional analysis based on fusion features to obtain analysis results. Specifically include:

[0148] 5.1 Sales trend analysis, using the Prophet time series forecasting model, taking into account holiday effects and long-term trends:

[0149] y(t)=g(t)+s(t)+h(t)+∈ t

[0150] Among them, g(t) is the trend term, s(t) is the season term, h(t) is the holiday term, ∈ t is the error term

[0151] 5.2 Customer clustering: We use the improved DBSCAN algorithm to cluster customers and introduce the Mahalanobis distance to measure similarity:

[0152]

[0153] Among them, d Mis the Mahalanobis distance, and σ is the kernel width parameter (usually set to the median of the nearest neighbor distance in the data set).

[0154] 5.3 Product association analysis, using the improved Apriori algorithm and introducing the time decay factor:

[0155]

[0156] Where δ is the time decay factor (usually set between 0.9 and 0.99), T is the current time, and c t (X) is the number of occurrences of item set X at time t, and N is the total number of transactions.

[0157] 5.4 Customer Lifetime Value Prediction, using RNN model based on attention mechanism:

[0158]

[0159] Among them, α t is the attention weight, x t is the customer feature vector at time t.

[0160] 5.5 Abnormal transaction detection, using the Isolation Forest algorithm. Abnormal score calculation:

[0161]

[0162] Among them, E(h(x)) is the average path length of sample x, and c(n) is the average path length of the tree.

[0163] 6. Visualization presentation step: Visualize the analysis results and generate visual data. Specifically include:

[0164] 6.1 Force-directed graph layout, using the improved ForceAtlas2 algorithm, introduces node importance weights:

[0165]

[0166] Where deg(u) is the degree of node u, d(u,v) is the distance between nodes u and v, and k r is the repulsion coefficient (usually set between 0.1 and 1).

[0167] 6.2t-SNE projection, using the Barnes-Hutt-SNE algorithm to improve the processing efficiency of large-scale data:

[0168]

[0169] Among them, x i and x jis a data point in high-dimensional space, σ i is the bandwidth parameter of the Gaussian kernel.

[0170] 6.3 Multi-level detail display, realizing hierarchical clustering and interactive detail level switching, using the improved BIRCH algorithm.

[0171] 6.4 Dynamic interaction and filtering, realize high-performance interactive visualization based on WebGL, and use quadtree spatial index to optimize the rendering and interaction of large-scale data.

[0172] 6.5 Adaptive color scheme uses genetic algorithm to optimize color harmony:

[0173] F=w1·C+w2·H+w3·B

[0174] Among them, C is contrast, H is harmony, B is balance, and w is i are weights (usually w1=0.4, w2=0.3, w3=0.3). The population size of the genetic algorithm is usually set to 100 and the number of iterations is 50.

[0175] 7. Interactive exploration step: Based on the visualized data, an interactive exploration interface is provided to achieve in-depth analysis and insight into the data. Specifically, it includes:

[0176] 7.1 Intelligent query recommendation, based on user historical behavior and current context, using the seq2seq model to generate query suggestions:

[0177] q t =Decoder(Encoder(h1,...,h t-1 ),c t )

[0178] Among them, h t Encode the user's historical behavior, c t is the current context vector. Both the encoder and decoder use LSTM networks, and the hidden layer size is usually set to 256.

[0179] 7.2 Real-time drilling and summarization, realize OLAP multidimensional data analysis based on GPU acceleration, use improved CUBE algorithm to support incremental update and approximate calculation. For large-scale data, a stratified sampling strategy is adopted, and the sampling rate is usually set to 1% to 10%, which is dynamically adjusted according to the data scale.

[0180] 7.3 Personalized view recommendation uses a multi-armed bandit algorithm to dynamically balance exploration and utilization. The reward update formula is as follows:

[0181] r i (t+1)=(1-α)·r i (t)+α·Ri (t)

[0182] Among them, r i (t) is the estimated reward of arm i at time t, R i (t) is the actual reward, α is the learning rate (usually set to 0.1). The exploration strategy uses the ∈-greedy method, and the ∈ value is initially set to 0.1 and decays over time.

[0183] 7.4 Collaborative analysis and annotation, implement a real-time collaborative editing and annotation system based on WebSocket. Use the improved Operational Transformation algorithm to handle concurrent editing conflicts and ensure consistency when multiple users edit simultaneously.

[0184] 7.5 Knowledge graph assisted interpretation, integrating knowledge graph reasoning engine, provides contextual interpretation of analysis results. Use knowledge graph completion algorithm based on transfer learning to dynamically expand the knowledge base. Knowledge graph completion uses TransE model, and its scoring function is:

[0185] f r (h,t)=∥h+rt∥

[0186] Where h, r, and t are the embedding vectors of the head entity, relation, and tail entity, respectively. The embedding dimension is usually set between 100 and 200.

[0187] Please refer to Figure 2 The knowledge graph-driven sales data multidimensional analysis and visualization system of the present invention includes the following modules:

[0188] Knowledge graph construction module 1, which is responsible for building a sales domain knowledge graph containing entities such as products, customers, and transactions. It uses the Neo4j graph database to store the knowledge graph and provides a RESTful API interface for other modules to call.

[0189] Data preprocessing module 2: This module preprocesses the original sales data based on the constructed knowledge graph. It uses Python's pandas library for data processing and scikit-learn library for feature engineering and standardization.

[0190] Graph embedding module 3, which implements the time-series adaptive multi-view collaborative graph embedding algorithm. It is implemented based on the PyTorch deep learning framework and uses GPU to accelerate computing. The model training uses the Adam optimizer, with the learning rate set to 0.001 and the batch size set to 128.

[0191] Feature fusion module 4: This module implements the dynamic cross-modal adaptive feature fusion algorithm. It is also implemented based on PyTorch and uses TensorBoard for model training visualization and tuning.

[0192] Multidimensional analysis module 5: This module performs multidimensional analysis based on fusion features. Sales trend analysis uses Facebook's Prophet library, customer segmentation and anomaly detection use the scikit-learn library, and product association analysis uses the mlxtend library.

[0193] Visualization presentation module 6: This module visualizes the analysis results, uses D3.js library to implement force-directed graph layout, uses three.js library to implement 3D data visualization, and uses ECharts library to implement regular statistical charts.

[0194] Interactive Exploration Module 7: This module provides an interactive exploration interface. The front-end is developed using the React framework, and the back-end uses Node.js and Express.js to build a RESTful API. The real-time collaboration function is implemented using the Socket.io library.

[0195] The system architecture adopts microservice design, uses Docker containerization deployment, and Kubernetes for container orchestration and management to ensure the scalability and high availability of the system.

[0196] The knowledge graph-driven sales data multidimensional analysis and visualization method and system proposed in the present invention solves the shortcomings of traditional sales data analysis methods in dealing with complex relationships, dynamic changes, and long-tail data by innovatively combining knowledge graphs, deep learning, and high-dimensional data visualization technology. The various modules of the system work closely together to form a complete data analysis closed loop. From data preprocessing, feature learning to multidimensional analysis and visualization, each step has been carefully designed and optimized. The actual application results show that the system can significantly improve the depth, accuracy, and efficiency of sales data analysis, providing strong support for corporate decision-making.

[0197] In order to verify the superiority of the present invention, we conducted a series of comparative experiments. The following is a detailed comparison of Example 1 of the present invention and three comparative examples:

[0198] Example 1: The knowledge graph-driven sales data multidimensional analysis and visualization method of the present invention

[0199] Comparative Example 1: Traditional sales data analysis method (without using knowledge graph and deep learning technology)

[0200] Comparative Example 2: Sales Data Analysis Method Based on Deep Learning (Without Using Knowledge Graph)

[0201] Comparative Example 3: Sales Data Analysis Method Based on Knowledge Graph (Without Using Deep Learning Technology)

[0202] Test indicators and testing methods:

[0203] 1. Sales forecast accuracy: evaluated using mean absolute percentage error (MAPE), in %.

[0204] Detection method: Make sales forecasts for the next 30 days on the test set and calculate MAPE.

[0205] 2. Customer segmentation quality: Evaluated using the Silhouette Coefficient, with a value range of [-1,1].

[0206] Detection method: Group customers and calculate silhouette coefficient.

[0207] 3. Long-tail product analysis accuracy: The F1 score is used for evaluation, with a value range of [0,1].

[0208] Detection method: Forecast sales of long-tail products (sales share less than 1%) and calculate the F1 score.

[0209] 4. Association rule mining effect: Use lift to evaluate, the larger the value, the better.

[0210] Detection method: mine the association rules between products and calculate the average lift.

[0211] 5. Visual response time: Use the average response time evaluation in seconds.

[0212] Detection method: Test 1,000 random data visualization requests and calculate the average response time.

[0213] 6. System scalability: This is evaluated using the number of queries per second (QPS), measured in times per second.

[0214] Detection method: gradually increase the number of concurrent users in the load test environment and record the maximum QPS when the system is running stably. The test results are shown in the following table:

[0215] Test indicators Example 1 Comparative Example 1 Comparative Example 2 Comparative Example 3 Sales Forecast Accuracy (MAPE) 8.2％ 15.7％ 10.5％ 12.3％ Customer segmentation quality 0.68 0.42 0.56 0.51 Long-tail product analysis accuracy 0.73 0.45 0.61 0.58 Association rule mining effect 3.2 1.8 2.5 2.7 Visualization response time (seconds) 0.8 2.3 1.5 1.7 System Scalability (QPS) 12000 3000 8000 6000

[0216] According to the test results in the above table, it can be seen that Example 1 is the best example. The above results are analyzed and explained as follows:

[0217] 1. Sales forecast accuracy: The MAPE of the present invention is 8.2%, which is significantly better than other methods. This shows that the temporal adaptive multi-view collaborative graph embedding algorithm of the present invention can better capture the dynamic characteristics and complex relationships of sales data, thereby improving the forecast accuracy.

[0218] 2. Customer clustering quality: The silhouette coefficient of the present invention is 0.68, which is significantly higher than that of other methods. This shows that the present invention can more accurately characterize customer characteristics by integrating knowledge graph and deep learning technology, thereby achieving more accurate customer clustering.

[0219] 3. Long-tail product analysis accuracy: The F1 score of the present invention is 0.73, which is significantly ahead of other methods. This verifies the effectiveness of the meta-learning anti-long-tail module in the present invention, which can better handle the analysis of long-tail products.

[0220] 4. Association rule mining effect: The average improvement of the present invention is 3.2, which is significantly higher than other methods. This shows that the present invention can discover more valuable commodity association rules by combining knowledge graph and deep learning technology, providing better support for cross-selling and personalized recommendations.

[0221] 5. Visualization response time: The average response time of the present invention is 0.8 seconds, which is much faster than other methods. This is due to the improved ForceAtlas2 algorithm and Barnes-Hutt-SNE algorithm adopted by the present invention, as well as the high-performance rendering technology based on WebGL.

[0222] 6. System scalability: The present invention can stably support 12,000 QPS, which is significantly ahead of other methods. This verifies the effectiveness of the microservice architecture and containerized deployment strategy of the present invention, which can better cope with large-scale data analysis scenarios.

[0223] The present invention is significantly superior to the comparative method in all test indicators, which fully proves the innovation and superiority of the present invention:

[0224] 1. Combination of knowledge graph and deep learning: By integrating these two technologies, the present invention can better handle complex entity relationships and dynamic changes, thereby achieving better results in tasks such as sales forecasting, customer segmentation, and association rule mining.

[0225] 2. Time-Series Adaptive Multi-View Collaborative Graph Embedding Algorithm: This innovative algorithm enables the present invention to better capture the time series characteristics and multi-dimensional attributes of sales data, thereby performing well in tasks such as sales forecasting and long-tail product analysis.

[0226] 3. Dynamic cross-modal adaptive feature fusion algorithm: This algorithm effectively integrates heterogeneous data sources, enabling the present invention to achieve better results in tasks such as customer segmentation and association rule mining.

[0227] 4. Improved visualization technology: The advantage of the present invention in visualization response time reflects its innovation in large-scale data visualization and can provide users with a smoother interactive experience.

[0228] 5. Highly scalable system architecture: The microservice design and containerized deployment strategy of the present invention enable it to better cope with large-scale data analysis scenarios and excel in system scalability.

[0229] These test results not only verify the effectiveness of each innovation of the present invention, but also illustrate the superiority of the present invention as a whole solution in practical applications. It can provide enterprises with more accurate, comprehensive and efficient sales data analysis services, thereby supporting better business decisions.

[0230] It should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A knowledge graph-driven sales data multidimensional analysis and visualization method, characterized in that: The following steps are involved: Build a sales domain knowledge graph containing product, customer, and transaction entities; Based on the knowledge graph, preprocessing the original sales data to obtain preprocessed sales data; According to the knowledge graph and the preprocessed sales data, time-series adaptive multi-view collaborative graph embedding is implemented to obtain graph embedding features; Fusion the graph embedding features with traditional sales data features to obtain fusion features; Perform multi-dimensional analysis based on the fusion features to obtain analysis results; Visually presenting the analysis results to generate visual data; Based on the visualized data, an interactive exploration interface is provided to achieve in-depth analysis and insight into the data.

2. The method according to claim 1, characterized in that The temporal adaptive multi-view collaborative graph embedding comprises the following steps: Execute the temporal attention mechanism to obtain the embedding representation of node v at time t in, represents the embedding representation of node v at time t-1; N(v) represents the set of neighbors of node v; represents the attention weight of node v to neighbor node u at time t; W h ,W a is a learnable weight matrix; a is a learnable vector parameter; σ is the LeakyReLU activation function; || represents a vector concatenation operation; Perform multi-view fusion to obtain the fused node representation h v : in, represents the embedding representation of node v in the i-th view; M represents the number of views, β i represents the weight of the i-th view; MLP represents a multi-layer perceptron; Perform knowledge distillation and calculate the knowledge distillation loss L KD : in, and are the outputs of the teacher model and the student model respectively, τ is the temperature parameter, KL is the KL divergence, and σ represents the softmax function; Execute the meta-learning anti-long-tail module and update the model parameters θ: Among them, θ ′ represents the model parameters after the inner loop is updated, α and β are the learning rates of the inner and outer loops respectively, L S and L Q is the loss function of the support set and query set, f θ represents the model function with parameter θ, represents the gradient with respect to θ.

3. The method according to claim 1, characterized in that: The feature fusion adopts a dynamic cross-modal adaptive feature fusion algorithm, which includes the following steps: Calculate the importance weights of graph embedding features and traditional features: Among them, φ G and φ T is the dynamic weight generation function, H G is the graph embedding feature matrix, X T is the feature matrix of traditional sales data, sigmoid represents the sigmoid activation function; Perform nonlinear transformation on graph embedding features and traditional features: H G ′=LeakyReLU(W1·H G +b1) X T ′=LeakyReLU(W2·X T +b2) Among them, H ′ G represents the transformed graph embedding features, X ′ T represents the transformed traditional features; W1, W2 represent the learnable weight matrices; b1, b2 represent the learnable bias vectors; Introduce the mutual information maximization term: I(H G ′,X T ′)=E[f(H G ′,X T ′)]-log(E[exp(f(H G ′,X T ″))]) Among them, f is the discriminator network, X ′ T ′ Yes X ′ T A random permutation of ; E represents expectation; Feature fusion based on dynamic weights: Add residual connections and perform layer normalization: F=LayerNorm(F+[H G ∥X T ]) LayerNorm represents layer normalization operation; ∥ represents vector concatenation operation; Optimization goal: L=L task +λ1|L MI -λ2|I(H G ′,X T ′) Among them, L task is the loss function of the downstream task, L MI is the mutual information maximization loss, and λ1 and λ2 are balancing coefficients.

4. The method according to claim 1, characterized in that: The multi-dimensional analysis includes sales trend analysis, customer segmentation, product association analysis, customer life cycle value prediction and abnormal transaction detection.

5. The method according to claim 4, characterized in that The sales trend analysis uses the Prophet time series forecasting model: y(t)=g(t)+s(t)+h(t)+∈ t Among them, g(t) is the trend term, s(t) is the season term, h(t) is the holiday term, ∈ t is the error term.

6. The method according to claim 4, characterized in that The customer clustering adopts the improved DBSCAN algorithm and introduces the Mahalanobis distance to measure similarity: Among them, ρ i represents the local density of sample point i, x i ,x j represents the sample points in the data set, d M (x i ,x j ) represents the sample point x i and x j The Mahalanobis distance between them; σ represents the kernel width parameter, which is used to control the range of density estimation; exp represents the exponential function.

7. The method according to claim 4, characterized in that The product association analysis adopts the improved Apriori algorithm and introduces the time decay factor: Where supp(X) represents the support of item set X, X represents the item set to be analyzed; t represents the time when the transaction occurs; T represents the current time; δ represents the time decay factor, which ranges from (0,1); c t (X) represents the number of occurrences of item set X at time t; N represents the total number of transactions.

8. The method according to claim 4, characterized in that The customer lifetime value prediction adopts the RNN model based on the attention mechanism: Among them, CLTV represents the predicted customer lifetime value; T represents the predicted time span, α t represents the attention weight x at time step t t represents the customer feature vector at time step t; RNN represents the recurrent neural network function.

9. The method according to claim 1, characterized in that: The visualization uses the improved ForceAtlas2 algorithm for force-directed graph layout: Among them, F r (u,v) represents the repulsive force between nodes u and v; k r represents the repulsion coefficient, deg(u) and deg(v) represent the degrees of nodes u and v respectively; d(u,v) represents the distance between nodes u and v; And use the Barnes-Hutt-SNE algorithm for high-dimensional data projection: Among them, p j|i represents the conditional probability of point i choosing point j as a neighbor in high-dimensional space; x i ,x j ,x k represents a data point in high-dimensional space, σ i represents the local density parameter of point i; ∥|∥ represents the Euclidean distance; exp represents the exponential function.

10. The knowledge graph-driven sales data multidimensional analysis and visualization system is characterized by: include: The knowledge graph building module is used to build a sales domain knowledge graph containing entities such as products, customers, and transactions; A data preprocessing module, used to preprocess the original sales data based on the knowledge graph to obtain preprocessed sales data; A graph embedding module, used to implement time-series adaptive multi-view collaborative graph embedding according to the knowledge graph and the preprocessed sales data, and obtain graph embedding features; A feature fusion module, used for fusing the graph embedding feature with the traditional sales data feature to obtain a fusion feature; A multi-dimensional analysis module, used to perform multi-dimensional analysis based on the fusion features to obtain analysis results; A visualization module, used to visualize the analysis results and generate visualization data; The interactive exploration module is used to provide an interactive exploration interface based on the visualized data to achieve in-depth analysis and insight into the data.

Images