A risk-aware stream identification method and system for power grid marketing event risks
Through technologies such as Dirichrey process, Fisher coding and CLIP model, the shortcomings of existing technologies in cross-domain migration, concept drift and complex data flow processing are solved, and efficient and reliable grid marketing event risk identification is achieved, improving the risk of improving grid marketing event The risk detection capabilities and platform security of the power industry.
Patent Information
- Application Number
- CN202411039222.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-31
- Publication Date
- 2024-11-29
- Estimated Expiration
- 2044-07-31
AI Technical Summary
The existing marketing event risk identification methods have shortcomings in cross-domain migration, coping with concept drift, relying on artificial feature engineering, and processing complex data flows, resulting in inefficiency and limited scalability.
Dirichlet process and Fisher encoding are used for data clustering and feature encoding, combined with the CLIP model to extract text and image features, and multimodal risk identification model and model pooling mechanism are used to realize adaptive enhanced domain and batch abnormal detection. Continuously update and adapt the model to deal with data flow changes.
The model's cross-domain identification ability of grid marketing event risks has been improved, effectively responding to concept drift, improving the efficiency and reliability of marketing event risk identification in the power industry, and enhancing the platform's security and user experience.
Smart Images

Figure CN118569908B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of marketing event risk identification in the electric power industry, and specifically relates to a risk-aware power grid marketing event risk stream identification method and system. Background Art
[0002] With the digital transformation of the power industry, various marketing events are developing rapidly on the platform, which also brings potential risks. These risks may include fraud, data leakage, malicious attacks, etc., which pose a threat to the security and user experience of the platform. Traditional marketing event risk identification methods often face challenges of diversity and variability.
[0003] There are many problems with the anomaly detection algorithms of existing marketing event risk identification methods: (1) Limited effectiveness of cross-domain risk identification model migration: Existing algorithms mainly focus on the risk identification prediction performance of specific fields, and perform poorly in risk identification of power grid marketing events. There are limitations in the effect of cross-domain risk identification model migration. (2) Difficulty in coping with concept drift: The continuous development of data streams leads to concept drift, that is, the attributes of the target domain change arbitrarily. Existing algorithms usually adopt the method of incrementally updating the model, but this method only adapts to the latest data points and may not be effective for arbitrary concept drift in the data stream. (3) Reliance on artificial feature engineering: such as dimensionality reduction, random subsampling, and linear feature transformation to process complex data. This dependence leads to suboptimal results and limited scalability. (4) Difficulty in processing complex data streams: Traditional methods use pre-trained fixed models or repeatedly create new models, which cannot effectively process complex and evolving data streams. This either leads to inefficiency or the model cannot cope with changes in the data stream.
[0004] These problems limit the effectiveness of existing anomaly detection algorithms in identifying risks of marketing events with ever-changing data streams and diversity. Summary of the invention
[0005] In view of the shortcomings of existing anomaly detection algorithms, such as poor cross-domain effectiveness, difficulty in dealing with concept drift, reliance on artificial feature engineering, and difficulty in processing complex data streams, the present invention provides a risk-aware power grid marketing event risk streaming identification method and system, which can effectively combat the concept drift generated in the process of data stream development, so as to meet the challenges of marketing event risk identification in the power industry.
[0006] To achieve the above object, the present invention adopts the following technical solutions.
[0007] In a first aspect, the present invention provides a risk-aware marketing event risk stream identification method, which comprises:
[0008] 1) Scalable marketing event coding: After preprocessing the power grid marketing event data, the Dirichlet process is used for clustering, and then the eigenvector is calculated through Fisher coding to adapt to the real distribution of the data, improve the generalization ability, and achieve efficient coding;
[0009] 2) Adaptive enhancement domain: The CLIP model is used to extract text and image features, and after weighted fusion, principal component analysis is used for dimensionality reduction and feature selection to improve the CLIP model's ability to identify risks of power grid marketing events;
[0010] 3) Initialize the model pool: Use the source domain and part of the target domain data, pre-train the language representation model and use the hierarchical attention mechanism to extract features, and then concatenate and generate vectors E , train the risk classifier and freeze the representation vector, then train the domain classifier and build a model pool with a shared architecture;
[0011] 4) Batch anomaly detection: preprocessing of marketing event features (handling missing values, standardization and coding), risk prediction using multimodal risk identification models, setting risk score thresholds based on historical data, and detecting and marking anomalies that exceed the risk score thresholds;
[0012] 5) Domain resampling: Randomly select samples that are important for target domain prediction from the source domain, and randomly sample some samples γ from the target domain enhanced dataset to form a new target domain enhanced dataset;
[0013] 6) Model pool adaptation and update: Monitor the reliability of the model pool. If there is a high-reliability model, update its parameters. Otherwise, create a new model and merge it. Continue to perform anomaly detection and adaptive updates through concept-driven reasoning.
[0014] Furthermore, the Dirichlet process specifically includes:
[0015] The Dirichlet process is a nonparametric Bayesian method used to define infinite-dimensional discrete probability distributions; in practical applications, it is used to cluster groups of unknown number, i.e., clustering of the Dirichlet process.
[0016] Furthermore, the Fisher coding specifically includes:
[0017] Fisher coding is a feature representation technology. The basic idea is to map a set of features to points in a high-dimensional space and encode them by considering the gradients of these feature points with respect to the probability distribution parameters. Based on the Dirichlet process, Fisher coding can adapt to the actual distribution of the data, and effectively deal with the overfitting problem by automatically adjusting the number and features of clusters, making marketing event coding more stable on new or unknown data sets, thereby improving the scalability of marketing event coding and achieving fast and efficient encoding in streaming data.
[0018] Furthermore, the adaptive enhancement domain specifically includes:
[0019] 1) Feature extraction
[0020] For text Perform text feature extraction:
[0021]
[0022] in, is the text encoder of the CLIP model;
[0023] For images Perform image feature extraction:
[0024]
[0025] in, It is the image encoder of the CLIP model;
[0026] 2) Feature Fusion
[0027] Weighted fusion of text features and image features:
[0028]
[0029] in, and is the adjusted weight;
[0030] 3) Feature cutting and reorganization
[0031] Use principal component analysis to reduce feature dimensionality:
[0032]
[0033] And select based on feature importance, using feature importance ranking of random forest:
[0034]
[0035] in, criteria is the information gain or other statistical measure of the feature.
[0036] Furthermore, the initialization model pool specifically includes:
[0037] 1) Representation of marketing content: Use a pre-trained language representation model and fine-tune the marketing content of the adaptive enhancement domain Encode and generate representation vectors of multimodal marketing content;
[0038] 2) Representation of user comments on marketing events: Use a hierarchical attention mechanism + a feedforward neural network for pre-training. After training, delete the feedforward neural network and only use the hierarchical attention mechanism to encode user comments on marketing events.
[0039] 3) Representation of user-marketing interactions: using a feed-forward neural network that takes as input a binary vector of user-marketing interactions and returns a representation that contains important information about these interactions;
[0040] 4) After constructing the representation network and pre-training the language representation model and the hierarchical attention mechanism, the outputs of these three components are concatenated into a vector , and generate the representation vector of the entire marketing event through a feedforward neural network .
[0041] Furthermore, the multimodal risk identification model is specifically described as follows:
[0042] The multimodal risk identification model uses the D-Cox-Time algorithm as the basic component and performs a deep learning-based survival analysis model to identify risks in power grid marketing events. The D-Cox-Time algorithm introduces a deep learning method based on the traditional Cox proportional hazard model, performs nonlinear mapping of the original data through a neural network, and then applies the loss function of the Cox proportional hazard model.
[0043] Furthermore, the batch anomaly detection is specifically as follows:
[0044] 1) Event feature preprocessing: perform feature preprocessing on each event in the batch data set and encode the classification features to eliminate noise and inconsistency in the data and improve the accuracy of prediction;
[0045] 2) Risk prediction: Use a multimodal risk identification model to assess the risk of each event in the batch dataset and calculate a risk score that represents the probability of a specific risk;
[0046] 3) Set anomaly thresholds, set risk score thresholds based on historical data or business needs to identify a small number of high-risk events, and adjust the thresholds according to business needs to balance the false positive and false negative rates;
[0047] 4) Abnormal event detection: traverse the batch data set, detect and mark abnormal events that exceed the risk score threshold according to the set risk score threshold, help identify and intervene in potential high-risk events, and reduce risks.
[0048] Furthermore, the model pool is updated, and the specific contents are as follows:
[0049] Monitor the reliability of the model pool and use the significance level Trigger the update of the pool; when a model pool has at least one highly reliable model, select the most reliable model and update the parameters of the model using the current batch data; but when the models in the pool have only neutral reliability values, create a new model using the current batch and merge the new model into the model pool.
[0050] Furthermore, the model pool is adapted, and the specific contents are as follows:
[0051] 1) Estimate the overall reliability of the model pool, that is, the probability that at least one model in the model pool is reliable;
[0052] 2) If the reliability of the model pool exceeds a preset threshold, the most reliable model is selected and the parameters of the model are updated using the current batch data;
[0053] 3) If the reliability of the model pool is lower than the preset threshold, create a new model and merge the new model into the model pool.
[0054] In a second aspect, the present invention provides a risk-aware marketing event risk stream identification system, which includes:
[0055] Coding unit: After preprocessing the power grid marketing event data, the Dirichlet process is used for clustering, and then the eigenvector is calculated through Fisher coding to adapt to the real distribution of the data;
[0056] Adaptive Enhanced Domain Unit: Use the CLIP model to extract text and image features, perform weighted fusion, and perform principal component analysis dimensionality reduction and feature selection to improve the CLIP model's ability to identify risks of power grid marketing events;
[0057] Model pool initialization unit: Use source domain and part of target domain data to pre-train language representation model and extract features using hierarchical attention mechanism, and then concatenate and generate vectors E , train the risk classifier and freeze the representation vector, then train the domain classifier and build a model pool with a shared architecture;
[0058] Batch anomaly detection unit: preprocesses marketing event features, uses a multimodal risk identification model to predict risks, sets risk thresholds based on historical data, and detects and marks anomalies that exceed the risk threshold;
[0059] Domain resampling unit: randomly select samples that are important for target domain prediction from the source domain, and randomly sample some samples γ from the target domain enhanced dataset to form a new target domain enhanced dataset;
[0060] Model pool adaptation and update unit: monitors the reliability of the model pool, updates the parameters of high-reliability models if there are any, otherwise creates new models and merges them; continuously performs anomaly detection and adaptive updates through concept-driven reasoning.
[0061] The beneficial effects of the present invention are as follows: the present invention effectively enhances the model's risk detection capability for power grid marketing events, improves cross-domain effectiveness, adaptively processes ever-changing data streams, and effectively combats concept drift generated during the development of data streams, and can effectively respond to the ever-changing marketing risk challenges in the power industry. It provides the power industry with an efficient and reliable marketing event risk identification method, enhancing the platform's security and user experience. BRIEF DESCRIPTION OF THE DRAWINGS
[0062] Figure 1 A flow chart of a risk-aware marketing event risk stream identification method of the present invention;
[0063] Figure 2 The schematic diagram of the expandable marketing event coding principle of the present invention;
[0064] Figure 3 This is a diagram of the adaptive enhancement domain coding architecture of the present invention;
[0065] Figure 4 This is the ADEC-MRD model architecture diagram of the present invention;
[0066] Figure 5 Schematic diagram of the process of the D-Cox-Time algorithm of the present invention;
[0067] Figure 6 The overall process diagram of the model pool adaptation and update of the present invention;
[0068] Figure 7 This is a structural diagram of a risk-aware marketing event risk streaming identification system of the present invention. DETAILED DESCRIPTION
[0069] The present invention is further described in detail below with reference to the accompanying drawings and specific implementation methods.
[0070] Marketing events refer to specific activities or situations related to marketing activities, while marketing event data streams refer to the process of continuous flow and recording of data related to marketing activities. Marketing event risk streaming identification refers to the real-time discovery and prevention of risks related to marketing events in marketing event data streams. The data involved in marketing event risk streaming identification includes real-time data generated by various user behaviors, transaction behaviors or system events in marketing activities. These data include information such as user clicks, purchases, and feedback. These event data are captured and serialized into the standard format of JSON for easy transmission and processing.
[0071] Risk perception of power grid marketing events refers to the identification, assessment and response to potential risks throughout the life cycle of power grid marketing events. Specifically, risk perception includes the following aspects: Identification of potential risks: At different stages of power grid marketing activities, various risks may be faced, such as market competition, policy changes, technical risks, etc. Risk factor analysis: Identify the main risk factors that affect power grid marketing events, including external environment (such as policies and regulations, changes in market demand) and internal factors (such as technical capabilities, operational management). Risk probability assessment: Assess the likelihood of occurrence of each identified risk, usually using quantitative or qualitative methods. Risk prioritization: Prioritize risks based on their probability and impact in order to focus on managing high-priority risks.
[0072] Concept drift refers to the impact of changes in market conditions, consumer demand, competitive landscape and other factors, which require adjustments or updates to the strategies for identifying marketing risk events that were previously effective. For example, changes in regulations and policies (changing the price mechanism for purchasing electricity, etc.), the emergence of new technologies (smart grids, the maturity of renewable energy technologies, etc.) and changes in consumer behavior. In terms of form, if the input data point x The joint probability of its label is t Time changes, then t Concept drift occurs in time, i.e. .
[0073] Once concept drift occurs, the current risk detection model becomes outdated and should be updated to learn new concepts. Since drift can occur in various forms, such as “sudden”, “gradual”, “incremental”, and “recurring”, it is important to have a general mechanism to handle all these forms equally.
[0074] Example 1
[0075] This embodiment provides a risk-aware marketing event risk stream identification method, such as Figure 1 As shown, the specific steps are as follows:
[0076] 1) Extensible marketing event coding
[0077] The scalable marketing event encoding step performs Fisher encoding on the power grid marketing data based on the Dirichlet process, such as Figure 2 As shown. The Dirichlet process is a nonparametric Bayesian method that can be used to define infinite-dimensional discrete probability distributions. In practical applications, it is often used to cluster unknown numbers of groups, namely the Dirichlet process clustering (DPC) model. Fisher coding is a feature representation technique. The basic idea is to map a set of features to points in a high-dimensional space and encode them by considering the gradient of these feature points with respect to the probability distribution parameters.
[0078] First, the power grid marketing event data is preprocessed to extract key features. These features include user behavior data, transaction time, transaction amount, etc. Then, the Dirichlet process modeling is used to cluster the data. This step is shown in formula (1-1). By selecting a suitable concentration parameter and basic distribution, the data is distributed in different clusters.
[0079] Assume that clusters, each class The center of Definition, and each data point The cluster labels Follows a multinomial distribution:
[0080] (1-1)
[0081] in, are the mixture weights generated by the Dirichlet process.
[0082] Next, for each data point, the corresponding Fisher vector is calculated. This vector is a function of the probability gradient of the data point at each cluster center, and the calculation formula is shown in formulas (1-2) and (1-3):
[0083] set up It is The characteristic function corresponding to the cluster center is is the Fisher fraction, calculated as:
[0084] (1-2)
[0085] in is a given parameter hour The probability density function of .
[0086] The entire Fisher vector can be expressed as:
[0087] (1-3)
[0088] With the help of the Dirichlet process, Fisher coding can adapt to the real distribution of data, which helps the generalization ability of marketing event coding on unseen data. By automatically adjusting the number and characteristics of clusters, the overfitting problem can be effectively handled, making the marketing event coding more stable on new or unknown data sets, thereby improving the scalability of marketing event coding and achieving fast and efficient coding in streaming data.
[0089] 2) Adaptive Enhancement Domain
[0090] In order to improve the stability of marketing event encoding on new data sets, that is, to improve the model's ability to identify new types of power grid marketing event risks, the adaptive enhancement domain uses the CLIP (Contrastive Language–Image Pre-training) model to crop and reorganize marketing event features, mainly using the ability of CLIP to understand and process cross-modal content (such as text and images). In marketing data analysis, this can be used to analyze and reorganize visual and text data related to products, advertisements, or user feedback, and reorganize to generate marketing event features. Figure 3 As shown, the specific implementation of the adaptive enhancement domain includes the following steps:
[0091] Step 1: Feature Extraction
[0092] For text Perform text feature extraction:
[0093] (1-4)
[0094] in, is the text encoder of the CLIP model;
[0095] For images Perform image feature extraction:
[0096] (1-5)
[0097] in, It is the image encoder of the CLIP model;
[0098] 2) Feature Fusion
[0099] Weighted fusion of text features and image features:
[0100] (1-6)
[0101] in, and is the adjusted weight;
[0102] 3) Feature cutting and reorganization
[0103] Use principal component analysis (PCA) or other dimensionality reduction techniques to reduce feature dimensions:
[0104] (1-7)
[0105] And select based on feature importance, using feature importance ranking of random forest:
[0106] (1-8)
[0107] in, criteria is the information gain or other statistical measure of the feature.
[0108] The following is the pseudo code for adaptive enhancement domain:
[0109]
[0110] 3) Initialize the model pool
[0111] The input of the ADEC-MRD (Adaptive Domain Enhancement Coding-marketing risk model) model for marketing event risk detection is the marketing event dataset from the source domain and enhanced marketing event datasets for target domains A part of (γ) is used to build an enhanced domain adaptive learning marketing event risk identification model (ADEC-MRD) from the first batch of data streams and initialize the model pool. The ADEC-MRD model mainly consists of two parts: (1) Adaptive enhanced domain encoding module (2) Multimodal risk identification model, such as Figure 3 and Figure 4 shown.
[0112] Representation of marketing content: Use the pre-trained language representation model BERT and fine-tune the marketing content of the adaptive enhancement domain Encode and generate representation vectors of multimodal marketing content.
[0113] Representation of user comments on marketing events: Use the hierarchical attention mechanism (HAN) + feedforward neural network for pre-training. After training, delete the feedforward neural network and only use HAN to encode user comments on marketing events.
[0114] Representation of user-marketing interactions: Use a feed-forward neural network that takes as input a binary vector of user-marketing interactions and returns a representation that contains important information about these interactions.
[0115] After constructing the representation network and pre-training BERT and HAN, the outputs of these three components are concatenated into a vector , and generate the representation vector of the entire marketing event through the feedforward network Once the representation network is stacked with a feed-forward neural network classifier, the source domain dataset can be used and target domain dataset Part of γ trains the marketing event risk classifier. After training the marketing event risk classifier After that, freeze the representation vector and use the representation vector of the marketing event Training domain classifier .
[0116] in and Represents a set of marketing events that are binary operations with source and target domain labels. Each marketing event (such as a wrongly sent text message) contains a content, which is a length of sequence (recipient account does not match), a group of user reviews (recipient feedback) and user-marketing interactions , where user-marketing interaction Is a binary vector representing the number of posts, reposts, or likes about marketing events The user of the tweet (e.g. a text message sent by mistake). Represented as source domain flag; Represented as the target domain flag; express Is it a risk event label? Represents the total number of users who posted, retweeted, or liked tweets about all marketing events in the provided data (including the aggregation of various interaction data).
[0117] Definition 1. Model pool: The model pool is a collection of ADEC-MRD models , they share the same architecture but have different parameters.
[0118] Definition 2. MRD model: The MRD model uses the D-Cox-Time algorithm as a basic component to perform a deep learning-based survival analysis model to identify risks in power grid marketing events.
[0119] The D-Cox-Time algorithm process is as follows Figure 5 As shown in the figure, the D-Cox-Time algorithm introduces a deep learning method based on the traditional Cox proportional hazard model, maps the original data nonlinearly through a neural network, and then applies the loss function of the Cox proportional hazard model.
[0120] The Cox proportional hazards model is a semiparametric model whose partial likelihood function is used to estimate the regression coefficients. For a given training data set :in It is The survival time or observed time of an individual, It is The survival status of each individual (1 indicates the event occurred, 0 indicates right censoring), It is A vector of covariates for each individual.
[0121] The Cox proportional hazards model assumes a hazard function It can be expressed as:
[0122] (1-9)
[0123] in is the benchmark risk function, is the regression coefficient.
[0124] The partial likelihood function is defined as:
[0125] (1-10)
[0126] The corresponding negative log partial likelihood function (i.e. loss function) is:
[0127] (1-11)
[0128] in, Indicated in T i Time remains a collection of individuals at risk.
[0129] The pseudo code of the D-Cox-Time algorithm is shown in Algorithm 2:
[0130]
[0131] 4) Batch Anomaly Detection
[0132] 4.1 Event feature preprocessing
[0133] Performing feature preprocessing on each event in a batch dataset is a critical step. First, you need to ensure that the feature format of the batch dataset is aligned with the training data. This step usually includes handling missing values and performing standardization or normalization to ensure that the data is on the same scale. In addition, any non-numerical features (such as categorical features) need to be encoded (such as one-hot encoding). These preprocessing steps help eliminate noise and inconsistencies in the dataset, thereby improving the accuracy of model predictions.
[0134] 4.2 Risk Prediction
[0135] Use the trained MRD model to predict the risk of each event in the batch dataset. In this step, the MRD model will evaluate the features of each event and calculate a risk score. This score represents the probability of a specific risk occurring in that event. Because the D-Cox-Time algorithm is used, the model is able to capture the complex relationship between features and survival time, thereby providing accurate risk predictions.
[0136] 4.3 Setting anomaly thresholds
[0137] Setting a risk score threshold based on historical data or business needs is a key step in anomaly detection. This threshold is usually determined by analyzing the risk distribution in historical data. Choosing a high percentile (95%) as the threshold means that only a few high-risk events will be marked as anomalies. This threshold needs to be adjusted according to actual business needs to balance the false positive and false negative rates.
[0138] 4.4 Abnormal Event Detection
[0139] In this step, the batch data set is traversed to detect and mark abnormal events according to the set risk score threshold. For each event, if its risk score exceeds the set threshold, it is marked as abnormal. This step can help identify those events with potential high risks, so as to facilitate subsequent prevention and intervention measures, thereby reducing the risks in power grid marketing events.
[0140] 5) Domain Resampling
[0141] According to the characteristics of the source domain and the target domain, A certain number of samples are randomly selected from the dataset, and those samples with higher importance for the prediction of the target domain are selected. The specific contribution can be determined by the sample contribution after model pre-training. Select a part of samples γ from All available samples in are randomly sampled, and some samples γ are combined to form a new target domain enhanced dataset .
[0142] 6) Model pool adaptation and update
[0143] Given an infinite sequence of marketing events, i.e., data points Arrives at a data stream, a data stream with parameters Anomaly detection model (i.e., ADEC-MRD model) calculates the anomaly score for a single data point, , continuously updating parameters in an unsupervised manner , and report data points with scores exceeding the threshold as anomalies.
[0144] Reliability of the model: The concepts reflected in a given model may be different from those of the current batch, especially when concept drift occurs. The reliability of the model is estimated by comparing the concepts learned by the model with those of the current batch.
[0145] Will Represented as the anomaly score set of the current batch, is the last batch of data used to update the model. Given a model , and The statistical significance of the difference between indicates how reliable the model is for the current batch.
[0146] Theorem 1. Hoeffding's mean difference based on inequality restrictions: Given a range of values in Two independent random variables between and The sample is and The probability of the mean difference between is constrained by:
[0147] (1-12)
[0148] in, is the difference threshold, , is the number of samples.
[0149] Applying Theorem 1, and The statistical significance of the differences is given in Corollary 1.
[0150] Corollary 1. Conceptual gap: Assume and For Model The independent anomaly scores returned are and Respectively from and The sampled anomaly score set. At this time,
[0151] (1-13)
[0152] in is the batch size.
[0153] From Theorem 1, we can see that the reliability of a single model as shown in Formula (7) can be derived using the probability difference bound in Corollary 1. Therefore, the sample mean difference of the anomaly score is greater than or equal to The probability is at most .
[0154] Concept-driven anomaly scoring: Given a set of models In the model pool, there is a corresponding reliability , current batch variable A data point in Concept-driven anomaly scoring for:
[0155] (1-14)
[0156] Model pool reliability: Given a model pool ,reliability for yes .
[0157] Drift risk-aware model pool update: monitor the reliability of the model pool and update the model pool at a significant level Trigger pool update. When a model pool has at least one highly reliable model (i.e. ), the most reliable model is selected and the parameters of the model are updated using the current batch data. However, when the models in the pool have only neutral reliability values (i.e. ), first create a new model using the current batch and merge the new model into the model pool.
[0158] The overall process of model pool adaptation and updating is as follows: Figure 6 As shown, an overview is given in Algorithm 3. Once the model pool is initialized using the models created for the first batch, concept-driven reasoning is repeated for anomaly detection on each batch, and concept drift-aware updates are used for model pool adaptation. The anomaly detection step calculates the anomaly score of the data point.
[0159]
[0160] Example 2
[0161] This embodiment provides a risk-aware marketing event risk stream identification system. Figure 7 As shown in the figure, it consists of an encoding unit, an adaptive enhancement domain unit, a model pool initialization unit, a batch anomaly detection unit, a domain resampling unit and a model pool adaptation and update unit.
[0162] The encoding unit: after pre-processing the power grid marketing event data, clustering is performed using the Dirichlet process, and then the eigenvector is calculated through Fisher coding to adapt to the real distribution of the data;
[0163] The Dirichlet process is a nonparametric Bayesian method for defining an infinite-dimensional discrete probability distribution; in practical applications, it is used for clustering groups of unknown number, i.e., clustering of the Dirichlet process.
[0164] The Fisher coding specifically includes:
[0165] Fisher coding is a feature representation technology. The basic idea is to map a set of features to points in a high-dimensional space and encode them by considering the gradients of these feature points with respect to the probability distribution parameters. Based on the Dirichlet process, Fisher coding can adapt to the actual distribution of the data, and effectively deal with the overfitting problem by automatically adjusting the number and features of clusters, making marketing event coding more stable on new or unknown data sets, thereby improving the scalability of marketing event coding and achieving fast and efficient encoding in streaming data.
[0166] The adaptive enhancement domain unit: uses the CLIP model to extract text and image features, performs dimensionality reduction and feature selection through principal component analysis after weighted fusion, and improves the recognition ability of the CLIP model on the risks of new power grid marketing events;
[0167] The adaptive enhancement domain specifically includes:
[0168] 1) Feature extraction
[0169] For text t Perform text feature extraction:
[0170]
[0171] in, f text is the text encoder of the CLIP model;
[0172] For images I Perform image feature extraction:
[0173]
[0174] in, f image It is the image encoder of the CLIP model;
[0175] 2) Feature Fusion
[0176] Weighted fusion of text features and image features:
[0177]
[0178] in, and is the adjusted weight;
[0179] 3) Feature cutting and reorganization
[0180] Use principal component analysis to reduce feature dimensionality:
[0181]
[0182] And select based on feature importance, using feature importance ranking of random forest:
[0183]
[0184] in, criteria is the information gain or other statistical measure of the feature.
[0185] The model pool initialization unit described above uses source domain and part of target domain data to pre-train the language representation model and extract features using a hierarchical attention mechanism, and then concatenates the generated vectors. E , train the risk classifier and freeze the representation vector, then train the domain classifier and build a model pool with a shared architecture.
[0186] The initialization model pool specifically includes:
[0187] 1) Representation of marketing content: Use a pre-trained language representation model and fine-tune the marketing content of the adaptive enhancement domain Encode and generate representation vectors of multimodal marketing content;
[0188] 2) Representation of user comments on marketing events: Use a hierarchical attention mechanism + a feedforward neural network for pre-training. After training, delete the feedforward neural network and only use the hierarchical attention mechanism to encode user comments on marketing events.
[0189] 3) Representation of user-marketing interactions: using a feed-forward neural network that takes as input a binary vector of user-marketing interactions and returns a representation that contains important information about these interactions;
[0190] 4) After constructing the representation network and pre-training the language representation model and the hierarchical attention mechanism, the outputs of these three components are concatenated into a vector , and generate the representation vector of the entire marketing event through a feedforward neural network .
[0191] The batch anomaly detection unit: preprocesses marketing event features, uses a multimodal risk identification model to predict risks, sets risk thresholds based on historical data, and detects and marks abnormal events that exceed the risk threshold.
[0192] The batch anomaly detection is specifically as follows:
[0193] 1) Event feature preprocessing: perform feature preprocessing on each event in the batch dataset and encode the classification features;
[0194] 2) Risk prediction: Use a multimodal risk identification model to assess the risk of each event in the batch dataset and calculate a risk score that represents the probability of a specific risk;
[0195] 3) Set anomaly thresholds, set risk score thresholds based on historical data or business needs to identify a small number of high-risk events, and adjust the thresholds according to business needs to balance the false positive and false negative rates;
[0196] 4) Abnormal event detection: traverse the batch data set, detect and mark abnormal events that exceed the risk score threshold according to the set risk score threshold, help identify and intervene in potential high-risk events, and reduce risks.
[0197] The multimodal risk identification model uses the D-Cox-Time algorithm as a basic component and performs a deep learning-based survival analysis model to identify risks in power grid marketing events; the D-Cox-Time algorithm introduces a deep learning method based on the traditional Cox proportional risk model, performs nonlinear mapping of the original data through a neural network, and then applies the loss function of the Cox proportional risk model.
[0198] The domain resampling unit randomly selects samples important for target domain prediction from the source domain, and randomly samples part of the samples γ from the target domain enhanced dataset to form a new target domain enhanced dataset.
[0199] The domain resampling unit specifically comprises: according to the characteristics of the source domain and the target domain, A certain number of samples are randomly selected from the dataset, and those samples with higher importance for the prediction of the target domain are selected. The specific contribution can be determined by the sample contribution after model pre-training. Select a part of samples γ from All available samples in are randomly sampled, and some samples γ are combined to form a new target domain enhanced dataset .
[0200] The model pool adaptation and update unit: monitors the reliability of the model pool, updates the parameters of a high-reliability model if there is one, otherwise creates a new model and merges it; continuously performs anomaly detection and adaptive updates through concept-driven reasoning.
[0201] The model pool is updated, and the specific contents are as follows:
[0202] Monitor the reliability of the model pool and use the significance level Trigger the update of the pool; when a model pool has at least one highly reliable model, select the most reliable model and update the parameters of the model using the current batch data; but when the models in the pool have only neutral reliability values, first create a new model using the current batch and merge the new model into the model pool.
[0203] The model pool adaptation is specifically as follows:
[0204] 1) Estimate the overall reliability of the model pool, that is, the probability that at least one model in the model pool is reliable;
[0205] 2) If the reliability of the model pool exceeds a preset threshold, the most reliable model is selected and the parameters of the model are updated using the current batch data;
[0206] 3) If the reliability of the model pool is lower than the preset threshold, create a new model and merge the new model into the model pool.
[0207] For the parts not described in detail in this embodiment, please refer to Example 1.
[0208] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Ordinary technicians in the relevant field can still modify or make equivalent substitutions to the specific implementation methods of the present invention with reference to the above embodiments. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention are within the scope of protection of the claims of the present invention to be approved.
Claims
1. A risk-aware stream identification method for power grid marketing events, It is characterized in that include: 1) Scalable marketing event coding: After preprocessing the power grid marketing event data, the Dirichlet process is used for clustering, and then the eigenvector is calculated through Fisher coding to adapt to the actual distribution of the data; 2) Adaptive enhancement domain: The CLIP model is used to extract text and image features, and after weighted fusion, principal component analysis is used for dimensionality reduction and feature selection to improve the CLIP model's ability to identify risks of power grid marketing events; 3) Initialize the model pool: Use the source domain and part of the target domain data, pre-train the language representation model and use the hierarchical attention mechanism to extract features, and then concatenate and generate vectors E , train the risk classifier and freeze the representation vector, then train the domain classifier and build a model pool with a shared architecture; 4) Batch anomaly detection: preprocessing marketing event features, using multimodal risk identification models for risk prediction, setting risk score thresholds based on historical data, and detecting and marking anomalies that exceed the risk score thresholds; 5) Domain resampling: Randomly select samples that are important for target domain prediction from the source domain, and randomly sample some samples γ from the target domain enhanced dataset to form a new target domain enhanced dataset; 6) Model pool adaptation and update: monitor the reliability of the model pool, update the parameters of the high reliability model if there is one, otherwise create a new model and merge it; Continuously perform anomaly detection and adaptive updates through concept-driven reasoning; The input of the marketing event risk identification model with enhanced domain adaptive learning for marketing event risk detection is the marketing event dataset from the source domain. and enhanced marketing event datasets for target domains As part of this, we build an enhanced domain adaptive learning marketing event risk identification model from the first batch of data streams and initialize the model pool; The enhanced domain adaptive learning marketing event risk identification model consists of two parts: adaptive enhanced domain encoding module and multimodal risk identification model; The multimodal risk identification model is specifically described as follows: the multimodal risk identification model uses the D-Cox-Time algorithm as a basic component to perform a deep learning-based survival analysis model to identify risks in power grid marketing events; the D-Cox-Time algorithm introduces a deep learning method based on the traditional Cox proportional risk model, performs nonlinear mapping of the original data through a neural network, and then applies the loss function of the Cox proportional risk model.
2. According to the risk-aware power grid marketing event risk stream identification method described in claim 1, It is characterized in that The adaptive enhancement domain specifically includes: 1) Feature extraction For text Perform text feature extraction: , in, is the text encoder of the CLIP model; For images Perform image feature extraction: , in, It is the image encoder of the CLIP model; 2) Feature Fusion Weighted fusion of text features and image features: , in, and is the adjusted weight; 3) Feature cutting and reorganization Use principal component analysis to reduce feature dimensionality: , And select based on feature importance, using feature importance ranking of random forest: , in, criteria is the information gain or other statistical measure of the feature.
3. According to the risk-aware stream identification method for power grid marketing events according to claim 1, It is characterized in that The initialization model pool specifically includes: 1) Representation of marketing content: Use a pre-trained language representation model and fine-tune the marketing content of the adaptive enhancement domain Encode and generate representation vectors of multimodal marketing content; 2) Representation of user comments on marketing events: Use a hierarchical attention mechanism + a feedforward neural network for pre-training. After training, delete the feedforward neural network and only use the hierarchical attention mechanism to encode user comments on marketing events. 3) Representation of user-marketing interactions: using a feed-forward neural network that takes as input a binary vector of user-marketing interactions and returns a representation that contains important information about these interactions; 4) After constructing the representation network and pre-training the language representation model and the hierarchical attention mechanism, the outputs of these three components are concatenated into a vector , and generate the representation vector of the entire marketing event through a feedforward neural network .
4. According to the risk-aware power grid marketing event risk stream identification method of claim 1, It is characterized in that The batch anomaly detection is specifically as follows: 1) Event feature preprocessing: perform feature preprocessing on each event in the batch dataset and encode the classification features; 2) Risk prediction: Use a multimodal risk identification model to assess the risk of each event in the batch dataset and calculate a risk score that represents the probability of a specific risk; 3) Set anomaly thresholds, set risk score thresholds based on historical data or business needs to identify a small number of high-risk events, and adjust the thresholds according to business needs to balance the false positive and false negative rates; 4) Abnormal event detection: traverse the batch data set, detect and mark abnormal events that exceed the risk score threshold according to the set risk score threshold, help identify and intervene in potential high-risk events, and reduce risks.
5. According to the risk-aware power grid marketing event risk stream identification method of claim 1, It is characterized in that The model pool is updated, and the specific contents are as follows: Monitor the reliability of the model pool and use the significance level Trigger the update of the pool; when a model pool has at least one highly reliable model, select the most reliable model and update the parameters of the model using the current batch data; but when the models in the pool have only neutral reliability values, create a new model using the current batch and merge the new model into the model pool.
6. According to the risk-aware power grid marketing event risk stream identification method of claim 1, It is characterized in that The model pool adaptation is specifically as follows: 1) Estimate the overall reliability of the model pool, that is, the probability that at least one model in the model pool is reliable; 2) If the reliability of the model pool exceeds a preset threshold, the most reliable model is selected and the parameters of the model are updated using the current batch data; 3) If the reliability of the model pool is lower than the preset threshold, create a new model and merge the new model into the model pool.
7. A risk-aware stream identification system for power grid marketing events, It is characterized in that include: Coding unit: After preprocessing the power grid marketing event data, the Dirichlet process is used for clustering, and then the eigenvector is calculated through Fisher coding to adapt to the real distribution of the data; Adaptive Enhanced Domain Unit: Use the CLIP model to extract text and image features, perform weighted fusion, and perform principal component analysis dimensionality reduction and feature selection to improve the CLIP model's ability to identify risks of power grid marketing events; Model pool initialization unit: Use source domain and part of target domain data to pre-train language representation model and extract features using hierarchical attention mechanism, and then concatenate and generate vectors E , train the risk classifier and freeze the representation vector, then train the domain classifier and build a model pool with a shared architecture; Batch anomaly detection unit: preprocesses marketing event features, uses a multimodal risk identification model to predict risks, sets risk thresholds based on historical data, and detects and marks anomalies that exceed the risk threshold; Domain resampling unit: randomly select samples that are important for target domain prediction from the source domain, and randomly sample some samples γ from the target domain enhanced dataset to form a new target domain enhanced dataset; Model pool adaptation and update unit: monitors the reliability of the model pool, updates the parameters of high-reliability models if any, otherwise creates a new model and merges it; Continuously perform anomaly detection and adaptive updates through concept-driven reasoning; The input of the marketing event risk identification model with enhanced domain adaptive learning for marketing event risk detection is the marketing event dataset from the source domain. and enhanced marketing event datasets for target domains As part of this, we build an enhanced domain adaptive learning marketing event risk identification model from the first batch of data streams and initialize the model pool; The enhanced domain adaptive learning marketing event risk identification model consists of two parts: adaptive enhanced domain encoding module and multimodal risk identification model; The multimodal risk identification model uses the D-Cox-Time algorithm as a basic component and performs a deep learning-based survival analysis model to identify risks in power grid marketing events; the D-Cox-Time algorithm introduces a deep learning method based on the traditional Cox proportional risk model, performs nonlinear mapping of the original data through a neural network, and then applies the loss function of the Cox proportional risk model.
8. The power grid marketing event risk stream identification system according to claim 7, It is characterized in that The model pool initialization unit is used to: 1) Representation of marketing content: Use a pre-trained language representation model and fine-tune the marketing content of the adaptive enhancement domain Encode and generate representation vectors of multimodal marketing content; 2) Representation of user comments on marketing events: Use a hierarchical attention mechanism + a feedforward neural network for pre-training. After training, delete the feedforward neural network and only use the hierarchical attention mechanism to encode user comments on marketing events. 3) Representation of user-marketing interactions: using a feed-forward neural network that takes as input a binary vector of user-marketing interactions and returns a representation that contains important information about these interactions; 4) After constructing the representation network and pre-training the language representation model and the hierarchical attention mechanism, the outputs of these three components are concatenated into a vector , and generate the representation vector of the entire marketing event through a feedforward neural network .
Citation Information
Patent Citations
Visual-word-bag-mode-based shoe print image feature extraction and retrieval method
CN108932518A
Self-positioning method and system during movement of depth sensor
CN111739096A
