Sales dynamic prediction method, system and device, and storage medium
By building a knowledge graph and graph attention network combined with a sequence prediction model, the problem of dynamic sales prediction with multi-dimensional features in the private market is solved, and accurate sales trend prediction and market monitoring are achieved.
Patent Information
- Application Number
- CN202510801408.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-16
- Publication Date
- 2025-10-17
AI Technical Summary
Existing technologies lack dynamic sales forecasting methods based on multidimensional features in private markets and are unable to accurately capture long-term trends. Traditional methods rely on static data sets, resulting in large errors and ignoring the correlation between complex features.
By collecting the original features of the sales website, building a knowledge graph, combining the graph attention network with the sequence prediction model, dynamic interactive learning is used to predict sales, and the number of product feedback is used as simulated sales to optimize feature extraction and prediction accuracy.
It enables accurate sales trend forecasting in scenarios where sales figures cannot be directly obtained, enhances the ability to monitor and manage market transactions, and provides structured data support.
Smart Images
Figure CN120807024A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of sales supervision, and particularly relates to a sales dynamic prediction method, system, device and storage medium. BACKGROUND
[0002] Market sales prediction refers to a process of scientifically predicting product sales in a future period by analyzing historical data, market trends and external environment and other factors, and especially for non-public markets or anonymous markets, can assist relevant departments in grasping the healthy transaction of the market.
[0003] In the related art, the traditional method is often directly based on the historical sequence of sales to directly predict, but the publicly disclosed sales data in the market is not accurate enough, and is full of a large number of high-risk transactions not subject to supervision. It is difficult to supervise the non-public market and the sales data cannot be directly obtained. The above situations need other parameters to replace or simulate sales. The existing prediction based on simulated sales all uses single variable analysis, which may cause large errors. In fact, the multi-dimensional features such as suppliers and commodity browsing volume in market transactions can all affect the sales results, and the existing technical means ignore the correlation between these complex features. The current technology using machine learning attempts to consider multi-dimensional feature prediction analysis, but mostly relies on static data sets and lacks dynamic time features, and cannot accurately capture the long-term trends embodied in the sales data.
[0004] In summary, the existing technology lacks a scheme that uses the number of commodity feedback as simulated sales for non-public market prediction, forms a long-term updated knowledge graph based on multi-dimensional data, analyzes complex relationships by means of graph neural network, and outputs long-term prediction sequences by sequence prediction model. SUMMARY
[0005] The present application aims to provide a sales dynamic prediction method, system, device and storage medium, which solves the above problems in the prior art.
[0006] According to a first aspect of the embodiment of the present application, a sales dynamic prediction method is provided, comprising:
[0007] Collecting original features in a sales website, and performing data preprocessing on the original features, wherein the original features include commodity features with the number of commodity feedback as simulated sales;
[0008] Dividing the preprocessed data into sequence groups in time sequence, and constructing a knowledge graph under each time step in the sequence group based on the preprocessed data;
[0009] Inputting the knowledge graph into a graph attention network of the corresponding sequence group, and obtaining a graph embedding representation through dynamic interaction learning of the graph attention network;
[0010] The graph embedding representation and the simulation sales are input into a sequence prediction model corresponding to the sequence group, a sales prediction result of a corresponding prediction period is output through the sequence prediction model, and a final prediction sequence is obtained by splicing sales prediction results of all sequence groups in time sequence.
[0011] According to a second aspect of the embodiment of the present application, a sales dynamic prediction system is provided, comprising:
[0012] The preprocessing module is configured to collect original features in the sales website and perform data preprocessing on the original features, wherein the original features include product features with the number of product feedbacks as simulation sales;
[0013] The knowledge graph construction module is configured to divide the preprocessed data into sequence groups in time sequence, and construct a knowledge graph at each time step within the sequence groups based on the preprocessed data;
[0014] The graph embedding representation module is configured to input the knowledge graph into a graph attention network corresponding to the sequence group, and obtain a graph embedding representation through dynamic interaction learning of the graph attention network;
[0015] The sequence prediction module is configured to input the graph embedding representation and the simulation sales into a sequence prediction model corresponding to the sequence group, output a sales prediction result of a corresponding prediction period through the sequence prediction model, and obtain a final prediction sequence by splicing sales prediction results of all sequence groups in time sequence.
[0016] According to a third aspect of the embodiment of the present application, an electronic device is provided, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program is executed by the processor to implement the steps of the sales dynamic prediction method provided in the first aspect of the present application.
[0017] According to a fourth aspect of the embodiment of the present application, a computer readable storage medium is provided, which stores an information transmission implementation program, wherein the program is executed by a processor to implement the steps of the sales dynamic prediction method provided in the first aspect of the present application.
[0018] The technical scheme provided by the embodiment of the present application has the following beneficial effects: multi-dimensional data relationships are constructed, and the number of product feedbacks is used as simulation sales to match scenarios where sales cannot be directly obtained; the graph attention network is used to dynamically allocate weights between nodes, the feature extraction capability of the knowledge graph is optimized to obtain embedding representation; and further combined with the sequence prediction model, the future sales trend is accurately predicted to provide data support for market monitoring and management.
[0019] It should be understood that the above general description and the following detailed description are only exemplary and explanatory, and cannot limit the present disclosure. BRIEF DESCRIPTION OF DRAWINGS
[0020] In order to more clearly illustrate one or more embodiments of this specification or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments recorded in this specification. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.
[0021] Figure 1 is a flow chart of a sales dynamics forecasting method according to an embodiment of the present invention;
[0022] Figure 2 is a schematic diagram of a knowledge graph according to an embodiment of the present invention;
[0023] Figure 3 is a schematic diagram of the overall implementation framework of an embodiment of the present invention;
[0024] Figure 4 is a schematic diagram of the execution process of an embodiment of the present invention;
[0025] Figure 5 is a schematic diagram of a sales dynamics forecasting system according to an embodiment of the present invention;
[0026] Figure 6 is a schematic diagram of an electronic device according to an embodiment of the present invention. DETAILED DESCRIPTION
[0027] In order to enable those skilled in the art to better understand the technical solutions in one or more embodiments of this specification, the technical solutions in one or more embodiments of this specification will be clearly and completely described below in conjunction with the drawings in one or more embodiments of this specification. Obviously, the described embodiments are only part of the embodiments of this specification, not all of the embodiments. Based on one or more embodiments of this specification, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of this document.
[0028] Method Example
[0029] According to an embodiment of the present invention, a sales dynamics forecasting method is provided. Figure 1 This is a flow chart of a sales dynamics forecasting method according to an embodiment of the present invention. The method is used in scenarios where it is not convenient to directly obtain historical sales, such as non-public markets, anonymous markets or other similar scenarios. Figure 1 As shown, the sales dynamics prediction method according to an embodiment of the present invention specifically includes:
[0030] In step S110, original features are collected in the sales website, and data preprocessing is performed on the original features, wherein the original features include a product feature that takes the number of product feedbacks as the simulated sales, and specifically include:
[0031] The original features are extracted by using an automatic crawler technology to obtain website fields and by using page structuring to extract key fields:
[0032] The browsing volume, price, type, and simulated sales of a product are extracted as product features; the comment content and comment publishing time of a user are extracted as user features; and the product quantity and reputation level of a supplier are extracted as supplier features.
[0033] Preferably, the crawler is automatically adjusted based on market dynamics, for example, if a comment field or supplier exposure field is added to the market, the system will automatically update the configuration of the crawler, and the system will analyze the coverage rate of existing data. If the data of some categories is insufficient, for example, the number of comments is too small, the system will preferentially crawl these features.
[0034] The collected original features are stored in a structured manner to provide data support for subsequent processes.
[0035] The original features are sequentially subjected to data cleaning, missing value filling, and standardization processing to obtain standardized data. In the data cleaning stage, noise and invalid values in the data are removed to ensure the accuracy of the data. The missing value filling uses the mean value method for numerical data, interpolation for time series data, and null filling for other types. Standardization processing makes the data meet the input requirements of the model, avoiding the influence of different numerical values between features on the model effect.
[0036] Based on the comment content in the standardized data, a sentiment score is analyzed using natural language processing technology, the sentiment score is added to the user features, and preprocessed data is obtained.
[0037] The use of natural language processing technology to analyze sentiment scores specifically includes:
[0038] The VADER natural language processing tool is used to obtain sentiment types and corresponding emotion intensity distribution weights, and the sentiment dictionary weights of VADER are dynamically adjusted using context information, wherein the sentiment types include positive, negative, and neutral. The sentiment score is calculated using a weighted summation method, that is, the sentiment score of each comment is multiplied by the weight to generate a weighted sentiment score of the product.
[0039] The features included in the preprocessed data are shown in Table 1:
[0040] Table 1. Multidimensional Feature Content Table
[0041]
[0042] As shown in Table 1, the final feature names contained in each type of feature are shown, in the case of not being convenient to directly obtain sales, the review is often used to represent the real transaction, but in the actual situation, the number of reviews cannot completely equal the number of transactions, especially in the case of false reviews or part of the buyers do not comment, it is more reasonable and accurate to use the feedback number instead of sales.
[0043] In step S120, the preprocessed data is divided into each sequence group in time sequence, and a knowledge graph under each time step in the sequence group is constructed based on the preprocessed data, which specifically includes:
[0044] In the embodiment of the application, time average is used to divide into multiple sequence groups, and the complex long sequence problem is decomposed into controllable short sequence problem, and the strategy of divide and conquer is used to improve the prediction accuracy, flexibility and interpretability, and different time steps are divided under the same sequence group.
[0045] Based on the preprocessed data, entity extraction and relation extraction are performed, and a triple is constructed:
[0046] The commodity, user and supplier are taken as entity nodes, wherein the internal feature vector of the entity node is the corresponding type feature in the preprocessed data, that is, the defined commodity node is E P :{browsing volume, price, simulated sales, commodity type}, the user node is F u :{review content, review publishing time, sentiment score}, and the supplier node is E v :{product quantity, credit rating};
[0047] In the embodiment of the application, each commodity corresponds to a commodity node, each supplier corresponds to a supplier node, and each user comment corresponds to a user node;
[0048] The publishing relationship R ps between the commodity and the entity is extracted, which represents that a commodity is published by a supplier, and the comment relationship between the commodity and the user is taken as the relationship edge R pu , which represents that a user comments on a commodity; the relationship extraction result is converted into a triple (h, r, h) in the knowledge graph, wherein h represents the head entity, r represents the relationship, and t represents the tail entity, and the meaning of the triple (h, r, t) → (commodity A, published by supplier, supplier B) is that commodity A is published by supplier B.
[0049] Based on the triple, a knowledge graph is constructed, Figure 2 is a schematic diagram of the knowledge graph of the embodiment of the application, as shown in Figure 2 , the general structure form of the knowledge graph is shown; the input features of the subsequent graph attention network are provided, and the structured support is provided for sales prediction.
[0050] In step S130, the knowledge graph is input into the graph attention network of the corresponding sequence group, and a graph embedding representation is obtained through dynamic interaction learning of the graph attention network, specifically including:
[0051] In the embodiment of the present application, the number of neighbor nodes is completely determined by the actual transaction relationship of the market, for example, the neighbor nodes of a popular commodity may include multiple suppliers, a large number of user reviews and associated products, while the number of neighbor nodes of a cold commodity is relatively small.
[0052] Data belonging to the same sequence group share a graph attention network;
[0053] The graph attention network dynamically assigns different weights to the relationship between each node and its neighbor nodes through a self-attention mechanism, and the key is to dynamically adjust the information between nodes using the attention mechanism, and the transmission weight gives different importance to different neighbor nodes, so that the model can selectively aggregate the information of neighbor nodes according to the characteristics of the nodes and their relationships, thereby better understanding the interaction relationship;
[0054] The specific implementation is:
[0055] The adjacency matrix A corresponding to the knowledge graph and the node feature matrix X are jointly input into the GAT graph attention network, and the weights between entity nodes are dynamically adjusted through a self-attention mechanism. The attention coefficient is calculated through the self-attention mechanism of formula 1:
[0056]
[0057] Wherein, alpha ij represents the attention coefficient, Leaky ReLU represents the activation function, a represents the learned attention weight vector, W represents the learnable weight matrix, h i and h j represent the feature vectors of node i and node j respectively, N(i) represents the neighbor set of node i, and k represents the index distinction. Unlike existing multi-modal information fusion technology, formula 1 concatenates the feature vectors of node i and neighbor node j, then inputs linear transformation, and then uses Leaky ReLU function to calculate the attention coefficient, and finally normalizes the attention coefficient to a probability distribution through exponential function and summation operation. In this way, the neighbor node that has the greatest impact on the predicted target sales can be focused on in real time, so as to improve the market sensitivity and complex relationship modeling capability of the model;
[0058] The feature vector of the current entity node is updated by using formula 2 to aggregate the information of the neighbor nodes by weighting:
[0059]
[0060] Wherein, denotes the feature of the j-th node in the l-th layer, a ij denotes the attention coefficient between the i-th node and the j-th neighbor node, W (l) denotes the learnable weight matrix of the l-th layer, and denotes the activation function.
[0061] The graph embedding representation of each time step in the sequence group is obtained by multi-layer adjustment aggregation. Since the graph attention network is multi-layer, and each layer further aggregates the information of neighbor nodes, the embedding features of each node gradually fuse the global information of the entire graph. The final graph embedding representation is shown in formula 3:
[0062]
[0063] wherein the embedding vector contains the interaction information and features between the node and its neighbor nodes, and can represent the complex interaction influence of the node, and H (L) denotes the embedding matrix of all nodes, which can be used as the input of the subsequent LSTM model, so as to help the model capture the complex relationship in the time series data.
[0064] In step S140, the graph embedding representation and the simulated sales input corresponding sequence group are input into the sequence prediction model, and the sales prediction result of the corresponding prediction period is output through the sequence prediction model. The sales prediction results of all sequence groups are spliced in time sequence to obtain the final prediction sequence, which specifically includes:
[0065] Research shows that the sales of a commodity are not only affected by the price, the number of comments, the sentiment tendency and other features, but also closely related to the change trend in the historical dimension. Especially for the case where the direct and real sales data cannot be obtained, the comment data has obvious time sequence fluctuation, and the traditional regression model is difficult to capture such multi-dimensional correlation. Therefore, the sequence prediction model is used to obtain the result in the embodiment of the application.
[0066] The specific process is as follows:
[0067] The graph embedding representation g(A, X t ) and the simulated sales corresponding to the time step are spliced into a fusion vector Z t The fusion vectors in the sequence group are combined into a feature sequence in time sequence, that is, the graph embedding representation t1+ simulated sales t1 is the first fusion vector in the feature sequence, the graph embedding representation t2+ simulated sales t2 is the second fusion vector in the feature sequence, and so on.
[0068] The characteristic sequence is input into the LSTM sequence prediction model of the corresponding sequence group, the evolution process of the characteristic sequence in the time dimension is simulated through the gating mechanism, the historical information is selectively forgotten and remembered through the gating mechanism, the hidden state at the corresponding time step is calculated to obtain the sales prediction result, and the core calculation process is shown in formulas 4 to 8:
[0069] f t f ·[h t-1 ,Z t ]+b f ) Formula 4
[0070] i t i ·[h t-1 ,Z t ]+b i ) Formula 5
[0071] C t t ·C t-1 +i t ×tanh(W C ·[h t-1 ,Z t ]+b C ) Formula 6
[0072] o t o ·[h t-1 ,Z t ]+b o ) Formula 7
[0073] h t t ×tanh(C t ) Formula 8
[0074] Wherein, C t represents the memory state of the current time step, which integrates the information of the new historical sales and structural features, f t represents the forgetting gate, which determines which information to discard from the cell state C t-1 at the previous moment, i t represents the input gate, which determines which new information to store, o t represents the output gate, which determines which information to output, W f , W i , W C , W o represent the corresponding weight matrix, b f , b i , b C , b o denotes the corresponding bias term, and finally h t denotes the hidden state of the current time step, which is used for further prediction of the sales trend;
[0075] At each time step, the LSTM unit can remember the structural embedding and sales behavior pattern at the past time point, so as to calculate a new hidden state and memory state in the current prediction, and the prediction result is optimized by the mean square error function MSELoss of formula 9:
[0076]
[0077] wherein loss(x i ,y i ) denotes the mean square error loss, x i denotes the predicted value, and y i denotes the true value.
[0078] If the same time span is predicted for different sequence groups, each sequence group can generate a time prediction value, and the final prediction sequence is obtained by splicing the sales prediction results of all sequence groups in time sequence.
[0079] The above technical solutions of the embodiments of the present application will be illustrated by combining the following drawings.
[0080] Figure 3 is a schematic diagram of the overall implementation framework of the embodiments of the present application, as shown in Figure 3 , which shows the overall implementation framework, each column corresponds to a sequence group, and the GAT model and the LSTM model corresponding to the sequence group, and each sequence group is divided into multiple time steps, and the prediction result h t is the final prediction sequence obtained by splicing.
[0081] Figure 4 is a schematic diagram of the execution process of the embodiments of the present application, as shown in Figure 4 , which shows the complete process including data acquisition, preprocessing, knowledge graph construction, feature learning based on graph attention network GAT, and time series modeling and prediction.
[0082] In summary, in view of the existing problems, the sales dynamic prediction method of the present application constructs a multi-dimensional data relationship and uses the number of product feedback as the simulated sales to match the scenario where the sales cannot be directly obtained; constructs a knowledge graph to structurally depict the product sales distribution and enhances the understanding of the implicit patterns between data; uses the graph attention network GAT to dynamically allocate the weight between nodes, optimizes the feature extraction capability of the knowledge graph to obtain the embedding representation; further combines the sequence prediction model LSTM to accurately predict the future sales trend, and uses the graph embedding representation and the simulated sales corresponding to the time step as the fusion vector to fully consider the influence of other features on the sales, and provides data support for market monitoring and management.
[0083] System embodiment
[0084] According to the embodiment of the present application, a sales dynamic prediction system is provided, Figure 5 is a schematic diagram of the sales dynamic prediction system of the embodiment of the present application, as Figure 5 shown, the sales dynamic prediction system according to the embodiment of the present application specifically comprises:
[0085] The preprocessing module 50 is used for collecting original features in the sales website and performing data preprocessing on the original features, wherein the original features include product features using the number of product feedback as the simulated sales, and is specifically used for:
[0086] The original features are extracted by using the crawler technology to obtain the website field:
[0087] The browsing volume, price, category and simulated sales of the product are extracted as product features; the comment content and comment publishing time of the user are extracted as user features; the product quantity and reputation level of the supplier are extracted as supplier features.
[0088] The original features are sequentially subjected to data cleaning, missing value filling and standardization processing to obtain standardized data;
[0089] Based on the comment content in the standardized data, the sentiment score is analyzed by using the natural language processing technology, the sentiment score is added to the user features, and the preprocessed data is obtained.
[0090] The sentiment score is analyzed by using the natural language processing technology, which specifically includes:
[0091] The sentiment type and corresponding emotion intensity distribution weight are obtained by using the VADER natural language processing tool, wherein the sentiment type includes positive, negative and neutral, and the sentiment score is calculated by using the weighted summation method.
[0092] The knowledge graph construction module 52 is used for dividing the preprocessed data into each sequence group according to the time sequence, constructing the knowledge graph under each time step in the sequence group based on the preprocessed data, and specifically used for:
[0093] Based on the pre-processed data, entity extraction and relation extraction are performed, and triplets are constructed:
[0094] The commodity, the user and the supplier are taken as entity nodes, wherein the internal feature vector of the entity node is the corresponding type feature in the pre-processed data; the publishing relationship between the commodity and the entity, and the comment relationship between the commodity and the user are taken as relationship edges;
[0095] Based on the triplets, a knowledge graph is constructed.
[0096] The graph embedding representation module 54 is configured to input the knowledge graph into a graph attention network of a corresponding sequence group, and obtain a graph embedding representation through dynamic interaction learning of the graph attention network, and is specifically configured to:
[0097] The adjacency matrix and the node feature matrix corresponding to the knowledge graph are jointly input into a GAT graph attention network, the weights between the entity nodes are dynamically adjusted through a self-attention mechanism, and the feature vector of the current entity node is updated by aggregating the information of the neighbor nodes through weighting;
[0098] The graph embedding representation at each time step in the sequence group is obtained through multi-layer adjustment and aggregation.
[0099] The sequence prediction module 56 is configured to input the graph embedding representation and the simulated sales into a sequence prediction model of a corresponding sequence group, and output a sales prediction result for a corresponding prediction period through the sequence prediction model, so as to splice the sales prediction results of all sequence groups in time sequence to obtain a final prediction sequence, and is specifically configured to:
[0100] The graph embedding representation and the simulated sales at the corresponding time step are spliced into a fusion vector, and each fusion vector in the sequence group is combined into a feature sequence in time sequence;
[0101] The feature sequence is input into an LSTM sequence prediction model of a corresponding sequence group, the evolution process of the feature sequence in the time dimension is simulated through a gating mechanism, and a sales prediction result at a corresponding time step is obtained by calculating a hidden state.
[0102] In summary, in view of the existing problems, the sales dynamic prediction system of the present application constructs a multi-dimensional data relationship and uses the number of commodity feedbacks as simulated sales to match scenarios where sales cannot be directly obtained; constructs a knowledge graph to describe the commodity sales distribution in a structured manner, enhances the understanding of the implicit patterns between data; uses a graph attention network GAT to dynamically allocate the weights between nodes, optimizes the feature extraction capability of the knowledge graph to obtain an embedding representation; further combines a sequence prediction model LSTM to accurately predict future sales trends, and takes the graph embedding representation and the simulated sales at the corresponding time step as a fusion vector to fully consider the influence of other features on sales, and provides data support for market monitoring and management.
[0103] Electronic device embodiment
[0104] Figure 6 is a schematic diagram of an electronic device according to an embodiment of the present application. The electronic device 600 can include at least one processor 610 and a memory 620. The processor 610 can execute instructions stored in the memory 620. The processor 610 is communicatively connected to the memory 620 through a data bus. In addition to the memory 620, the processor 610 can be communicatively connected to an input device 630, an output device 640, and a communication device 650 through the data bus.
[0105] The processor 610 can be any conventional processor, such as commercially available CPUs. The processor can also include a Graphic Process Unit (GPU), a Field Programmable Gate Array (FPGA), a System on Chip (SOC), an Application Specific Integrated Circuit (ASIC), or a combination thereof.
[0106] The memory 620 can be implemented by any type of volatile or nonvolatile memory device or a combination thereof, such as a Static Random Access Memory (SRAM), an Electrically Erasable Programmable Read-Only Memory (EEPROM), an Erasable Programmable Read-Only Memory (EPROM), a Programmable Read-Only Memory (PROM), a Read-Only Memory (ROM), a magnetic storage, a flash memory, a magnetic disk, or an optical disk.
[0107] In the embodiments of the present disclosure, the memory 620 stores executable instructions. The processor 610 can read the executable instructions from the memory 620 and execute the instructions to implement all or part of the steps of the sales dynamic prediction method according to any of the above example embodiments.
[0108] Computer readable storage medium embodiment
[0109] In addition to the above method and system, the example embodiments of the present disclosure can also be a computer program product or a computer readable storage medium storing the computer program product, the computer program product including computer program instructions executable by a processor to implement all or part of the steps described in the sales dynamic prediction method according to any of the above example embodiments.
[0110] The computer program product can be written in any combination of one or more programming languages, including an object oriented programming language such as Java, C++ or the like, and conventional procedural programming languages, such as the "C" programming language or similar programming languages, and scripting languages (e.g., Python). The program code can execute entirely on the user's computing device, partly on the user's device, as a stand-alone software package, partly on the user's device and partly on a remote computing device or entirely on the remote computing device or server.
[0111] The computer readable storage medium can be a combination of one or more types of computer readable storage media. The computer readable storage medium can be a computer readable signal medium or a computer readable storage medium. The computer readable storage medium can include, for example, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus or device, or any suitable combination of the above. More specific examples of the computer readable storage medium include static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk, or any suitable combination of the above.
[0112] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.
Claims
1. A sales dynamics forecasting method, characterized in that: include: Collecting original features on the sales website and performing data preprocessing on the original features, wherein the original features include product features using the number of product feedback as simulated sales volume; Dividing the preprocessed data into sequence groups based on time sequence, and constructing a knowledge graph at each time step within the sequence group based on the preprocessed data; Inputting the knowledge graph into the graph attention network of the corresponding sequence group, and obtaining a graph embedding representation through dynamic interactive learning of the graph attention network; The graph is embedded into a sequence prediction model representing the corresponding sequence group and the simulated sales are input. The sales prediction results of the corresponding prediction period are output through the sequence prediction model. The sales prediction results of all sequence groups are spliced in chronological order to obtain a final prediction sequence.
2. The method according to claim 1, characterized in that The collecting of original features on the sales website specifically includes: Use crawler technology to obtain website fields and extract original features: The number of views, price, type and simulated sales of products are extracted as product features; the content of user comments and the time of comment release are extracted as user features; the number of products and reputation level of suppliers are extracted as supplier features.
3. The method according to claim 1, characterized in that The data preprocessing of the original features specifically includes: The original features are sequentially subjected to data cleaning, missing value filling and standardization processing to obtain standardized data; Based on the review content in the standardized data, natural language processing technology is used to analyze the sentiment score, and the sentiment score is added to the user features to obtain preprocessed data.
4. The method according to claim 3, characterized in that The use of natural language processing technology to analyze sentiment scores specifically includes: The VADER natural language processing tool is used to obtain the emotion type and the corresponding emotion intensity distribution weight, wherein the emotion type includes positive, negative and neutral, and the weighted sum method is used to calculate the emotion score.
5. The method according to claim 1, wherein The step of constructing a knowledge graph at each time step within the sequence group based on the preprocessed data specifically includes: Entity extraction and relationship extraction are performed based on the preprocessed data, and triples are constructed: Products, users, and suppliers are used as entity nodes, where the internal feature vectors of the entity nodes are the corresponding type features in the preprocessed data; the publishing relationship between products and entities, and the comment relationship between products and users are extracted as relationship edges; A knowledge graph is constructed based on the triples.
6. The method according to claim 1, characterized in that Inputting the knowledge graph into the graph attention network of the corresponding sequence group, and obtaining the graph embedding representation through dynamic interactive learning of the graph attention network specifically includes: The adjacency matrix and node feature matrix corresponding to the knowledge graph are input into the GAT graph attention network, the weights between entity nodes are dynamically adjusted through the self-attention mechanism, and the feature vector of the current entity node is updated by weighted aggregation of information from neighboring nodes; The graph embedding representation at each time step in the sequence group is obtained through multi-layer adjustment aggregation.
7. The method according to claim 1, characterized in that The step of embedding the graph representation and the simulated sales into a sequence prediction model corresponding to the sequence group, and outputting the sales prediction result corresponding to the prediction period through the sequence prediction model specifically includes: concatenating the graph embedding representation and the simulated sales volume at the corresponding time step into a fusion vector, and combining the fusion vectors in the sequence group into a feature sequence in chronological order; The feature sequence is input into the LSTM sequence prediction model of the corresponding sequence group, the evolution process of the feature sequence in the time dimension is simulated through the gating mechanism, and the hidden state at the corresponding time step is calculated to obtain the sales forecast result.
8. A sales dynamics forecasting system, characterized in that: include: A preprocessing module is used to collect original features from the sales website and perform data preprocessing on the original features, wherein the original features include product features using the number of product feedback as simulated sales volume; A knowledge graph construction module is used to divide the preprocessed data into sequence groups based on time sequence, and to construct a knowledge graph at each time step within the sequence group based on the preprocessed data; A graph embedding representation module is used to input the knowledge graph into the graph attention network of the corresponding sequence group, and obtain the graph embedding representation through dynamic interactive learning of the graph attention network; A sequence prediction module is used to embed the graph representation and the simulated sales input into a sequence prediction model of a corresponding sequence group, output the sales prediction results of the corresponding prediction period through the sequence prediction model, and splice the sales prediction results of all sequence groups in chronological order to obtain a final prediction sequence.
9. An electronic device, characterized in that: include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the computer program is executed by the processor, the steps of the sales dynamics prediction method according to any one of claims 1 to 7 are implemented.
10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores an information transmission implementation program, and when the program is executed by a processor, the steps of the sales dynamics prediction method according to any one of claims 1 to 7 are implemented.