Recommendation method for stealing gas users based on deep learning

Abstract

The present invention relates to the field of gas data processing technology and discloses a method for recommending users of gas theft based on deep learning. The method comprises the following steps: constructing a preliminary screening model for gas theft to screen out users who are initially identified as abnormal; the preliminary screening model includes a continuous zero gas usage indicator, conventional statistical indicators, a linear fit goodness-of-fit indicator, and an outlier indicator; constructing a dataset based on the initially identified abnormal users for training a deep learning model; inputting the user name and the user's historical gas usage data into the trained deep learning model, and outputting recommended users who are identified as gas theft users. The present invention aims to improve the accuracy of recommending users who are identified as gas theft users based on historical gas usage data.

Description

Technical Field

[0001] The present invention relates to the technical field of gas data processing, and in particular to a method for recommending gas theft users based on deep learning. Background Art

[0002] Recommending information about gas theft based on a user's historical gas usage data is a mathematically multifaceted problem. On the one hand, gas theft is bound to cause certain data phenomena; on the other hand, similar data phenomena cannot be definitively attributed to gas theft. This recommendation system is similar to today's audience-focused software. Recommended content is likely to be of interest to the user, but there's also a chance it won't. However, as the service develops and the model undergoes several rounds of iterative learning, its accuracy improves. Therefore, after iterative learning based on historical gas usage data, the challenge is to improve the accuracy of recommendations about gas theft. Summary of the Invention

[0003] The purpose of the present invention is to improve the accuracy of recommending users who steal gas based on historical gas usage data, and to provide a method for recommending users who steal gas based on deep learning.

[0004] In order to achieve the above-mentioned object of the invention, the embodiment of the present invention provides the following technical solutions:

[0005] The method for recommending users of stolen gas based on deep learning includes the following steps:

[0006] Step 1: construct a gas theft screening model to screen out abnormal users; the gas theft screening model includes a continuous zero gas usage index, conventional statistical index, linear fit goodness of fit index, and outlier index;

[0007] Step 2: Build a data set based on the initial screening of abnormal users for training the deep learning model;

[0008] Step 3: Input the user name and the user's historical gas usage data into the trained deep learning model and output the recommended gas theft users.

[0009] Compared with the prior art, the present invention has the following beneficial effects:

[0010] The gas theft screening model constructed by the present invention is used for the initial launch of the recommendation system. Generally speaking, at this stage, there is still a lack of labeled data, and the machine learning model cannot be directly established. Manual logical indicators are mainly used to make recommendations. The gas theft screening model defaults to continuous zero gas usage indicators, conventional statistical indicators, linear fit goodness of fit indicators, and outlier indicators. These indicator algorithms are combined to form a gas theft screening model, which can screen out abnormal users. The model accuracy at this stage is not high, but it covers all gas theft behaviors as much as possible. At this stage, the gas company organizes a comprehensive manual investigation based on the abnormal users in the initial screening to verify the real gas theft behavior and normal gas usage behavior.

[0011] The deep learning model is trained to recommend gas theft behaviors based on manual verification results. Currently, the model's data dimensions include historical gas usage and user names. The user names are introduced to enable the model to learn these characteristics. This information serves as a filter to exclude users without a motive for gas theft (such as government users). It also serves as an attention mechanism, helping the model focus on gas usage during relevant months (such as non-holiday usage for school users). BRIEF DESCRIPTION OF THE DRAWINGS

[0012] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments. It should be understood that the following drawings only illustrate certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other relevant drawings can be obtained based on these drawings without paying any creative work.

[0013] Figure 1 Flow chart of the method of the present invention;

[0014] Figure 2 This is a schematic diagram of the deep learning network model of the present invention;

[0015] Figure 3 This is a schematic diagram of the comparative training of deep learning models in the present invention. DETAILED DESCRIPTION

[0016] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, not all of the embodiments. The components of the embodiments of the present invention generally described and shown in the drawings herein can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the claimed invention, but merely represents selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making creative work are within the scope of protection of the present invention.

[0017] It should be noted that similar numbers and letters represent similar items in the following figures. Therefore, once an item is defined in one figure, it does not need to be further defined or explained in subsequent figures. At the same time, in the description of the present invention, the terms "first", "second", etc. are used only to distinguish the description and are not to be understood as indicating or implying relative importance, or implying any actual relationship or order between these entities or operations. In addition, the terms "connected" and "connected" can refer to direct connection between elements or indirect connection via other elements.

[0018] Example 1:

[0019] The present invention is achieved through the following technical solutions: Figure 1 As shown, a method for recommending users of stolen gas based on deep learning includes the following steps:

[0020] Step 1: construct a preliminary screening model for stolen gas to screen out abnormal users; the preliminary screening model for stolen gas includes a continuous petty gas usage index, conventional statistical indexes, a linear fit goodness of fit index, and an outlier index.

[0021] The gas theft screening model consists of four indicators: continuous petty gas usage, conventional statistical indicators, linear fit indicators, and outlier indicators. The model sorts and filters users based on logical indicators. Each indicator extracts up to cnt outliers, which are used to generate data labels for training the deep learning model.

[0022] (1) Continuous zero gas consumption index.

[0023] The continuous zero gas usage index is ranked based on the number of consecutive zero gas usage by the user and the maximum gas usage in a single day.

[0024] In this embodiment, a time period of one year (365 days) is given, and the calculation unit is 1 day. If the gas consumption for a whole day is 0, it is recorded as a continuous zero gas consumption. First, calculate the number of consecutive zero gas consumption times for each user, z iand the average number of consecutive zero gas usage for all users :

[0025]

[0026] Among them, z i represents the number of consecutive zero gas usage by the i-th user, i=1,2,...,n; n represents the total number of users managed by a gas company; Indicates the average number of consecutive zero gas usage for all users.

[0027] Secondly, assuming that n users are all users of a gas company, according to Calculate the Poisson distribution of the gas company's petty gas consumption segments and give a probability value P zero , calculate the lower limit of the number of consecutive zero gas consumption T zero :

[0028]

[0029] Among them, j is an index variable, which has no actual physical meaning and its value range is [0,100].

[0030] Then, the number of times the user uses zero gas is determined by each user: i Is it greater than the lower limit T zero , the number of consecutive zero gas consumption z i Greater than T zero The user is regarded as an abnormal user:

[0031]

[0032] Among them, user zero Represents a collection of abnormal users.

[0033] Finally, according to the principle of "grasping the big and letting go of the small", zero The median of the maximum daily gas consumption of users in zero ) is the threshold, and users with large gas consumption are further screened out among the users to be checked. After sorting, at most cnt abnormal users are recommended:

[0034]

[0035] Among them, u i Represents the collection user zero The maximum daily gas consumption of the i-th user in the past year; Abnormal zero Indicates that the maximum daily gas consumption in a year is greater than the median (user zero ) user collection. For example, the collection user zeroThere are 10 users in total, each of whom has 365 daily gas usages. From these, we can select the maximum daily gas usage of the user, and then we can get a total of 10 maximum daily gas usages. After sorting by size, we select the median as the threshold median(user zero ); Then set the user zero The maximum daily gas consumption per user u i and median(user zero ) for comparison, if it is greater than median(user zero ), then the user is included in the set Abnormal zero It should be noted that if the Abnormal zero The number of people in the game will exceed cnt, then u i After sorting from largest to smallest, select the corresponding first cnt users and add them to the set Abnormal zero .

[0036] (2) Conventional statistical indicators.

[0037] The conventional statistical indicators include a skewness index and a user's daily gas consumption index.

[0038] First, calculate the skewness index s for each user in one year i :

[0039]

[0040] Among them, x i,t represents the daily gas consumption of the i-th user on the t-th day in the past year; Represents the average daily gas consumption of the i-th user in the past year.

[0041] Then, the user's skewness index s is determined user by user i Is it less than 0? If it is less than 0, it means that the small flow data in the daily gas consumption of the user exceeds the number of standard normal distribution, which may be caused by gas theft. i Users with a value less than 0 are considered abnormal users:

[0042]

[0043] Among them, user stas Represents a collection of abnormal users.

[0044] Finally, according to the principle of "grasping the big and letting go of the small", we can collect user stas The median of the maximum daily gas consumption of users in stas) is the threshold, and users with large gas consumption are further screened out among the users to be checked. After sorting, at most cnt abnormal users are recommended:

[0045]

[0046] Among them, u i Represents the collection user stas The maximum daily gas consumption of the i-th user in the past year; Abnormal stas Indicates that the maximum daily gas consumption in a year is greater than the median (user stas ) user collection. For example, the collection user stas There are 11 users in total, each of whom has 365 daily gas usages. From these, we can select the maximum daily gas usage of the user, so a total of 11 maximum daily gas usages can be obtained. After sorting by size, we select the median as the threshold median(user stas ); Then set the user stas The maximum daily gas consumption per user u i and median(user stas ) for comparison, if it is greater than median(user stas ), then the user is included in the set Abnormal stas It should be noted that if the Abnormal stas The number of people in the game will exceed cnt, then u i After sorting from largest to smallest, select the corresponding first cnt users and add them to the set Abnormal stas .

[0047] (3) Linear fit goodness of fit index.

[0048] Generally speaking, if you draw a gas consumption curve with days as the horizontal axis and cumulative gas consumption (days) as the vertical axis, the curve is likely to be a straight line. Therefore, we can use the first reporting period as the zero point, fit the date and cumulative gas consumption, and use the goodness of fit r i As an evaluation index. Goodness of fit r i Between 0 and 1, the goodness of fit r i The larger the value, the higher the linearity.

[0049] First, calculate the goodness of fit r for each user i :

[0050]

[0051]

[0052]

[0053]

[0054]

[0055] Among them, y i,t represents the actual daily gas consumption of the i-th user on the t-th day; T is the day number, in this embodiment, T=365; f i,t represents the daily gas consumption of the i-th user on the t-th day predicted by the fitting model; , SS reg , SS tot , SS res All are intermediate parameters.

[0056] Secondly, exclude the goodness of fit r i For users with a goodness of fit greater than 0.95, the goodness of fit r i Users with a score less than or equal to 0.95 are considered abnormal users:

[0057]

[0058] Among them, user fit Represents a collection of abnormal users.

[0059] Finally, according to the principle of "grasping the big and letting go of the small", fit The median of the maximum daily gas consumption of users in fit ) is the threshold, and users with large gas consumption are further screened out among the users to be checked. After sorting, at most cnt abnormal users are recommended:

[0060]

[0061] Among them, u i Represents the collection user fit The maximum daily gas consumption of the i-th user in the past year; Abnormal fit Indicates that the maximum daily gas consumption in a year is greater than the median (user fit ) user collection. For example, the collection user fit There are 12 users in total, each of whom has 365 daily gas usages. From these, we can select the maximum daily gas usage of the user, and then we can get 12 maximum daily gas usages in total. After sorting by size, we select the median as the threshold median(user fit ); Then set the user fit The maximum daily gas consumption per user u i and median(userzfit ) for comparison, if it is greater than median(user fit ), then the user is included in the set Abnormal fit It should be noted that if the Abnormal fit The number of people in the game will exceed cnt, then u i After sorting from largest to smallest, select the corresponding first cnt users and add them to the set Abnormal fit .

[0062] (4) Outlier indicators.

[0063] Some macro factors (such as holidays and weather) can affect a group of users simultaneously. These users share similar gas usage characteristics, forming a characteristic group. Users with relatively isolated gas usage characteristics can be included in the abnormal user list.

[0064] First, the cosine similarity of each user's gas usage with another user is calculated based on the daily gas usage data:

[0065]

[0066] Among them, Similar a,b Represents the cosine similarity between user a and user b; a t represents the daily gas consumption of user a on day t; b t Represents the daily gas consumption of user b on day t.

[0067] Secondly, for a certain user i, the minimum value of its cosine similarity with all remaining users is used as the outlier indicator sim of the user i :

[0068]

[0069] Among them, sim i Represents the outlier indicator of the i-th user; {Similar i,n-1} represents the cosine similarity set between the i-th user and all remaining users; n represents the total number of users.

[0070] Then, users are simulated by outlier indicators i Sort from largest to smallest and extract the first 2×cnt users as abnormal users:

[0071]

[0072] Among them, user sim Represents the outlier indicator sim for all users i After sorting, the set of the first 2×cnt users; top(simi ,2×cnt) represents the sim i The first 2×cnt users sorted from largest to smallest.

[0073] Finally, according to the principle of "grasping the big and letting go of the small", sim The median of the maximum daily gas consumption of users in sim ) is the threshold, and users with large gas consumption are further screened out among the users to be checked. After sorting, at most cnt abnormal users are recommended:

[0074]

[0075] Among them, u i Represents the collection user sim The maximum daily gas consumption of the i-th user in the past year; Abnormal sim Indicates that the maximum daily gas consumption in a year is greater than the median (user sim ) user collection. For example, the collection user sim There are 13 users in total, each of which has 365 daily gas usages. From these, we can select the maximum daily gas usage of the user, so a total of 13 maximum daily gas usages can be obtained. After sorting by size, we select the median as the threshold median(user sim ); Then set the user sim The maximum daily gas consumption per user u i and median(user sim ) for comparison, if it is greater than median(user sim ), then the user is included in the set Abnormal sim It should be noted that if the Abnormal sim The number of people in the game will exceed cnt, then u i After sorting from largest to smallest, select the corresponding first cnt users and add them to the set Abnormal sim .

[0076] (5) Initial screening model for gas theft.

[0077] The users selected by the above four indicators are combined and counted as the initial screening abnormal users A:

[0078]

[0079] Among them, A represents the initial screening of abnormal users.

[0080] Step 2: Build a data set based on the initial screening of abnormal users for training the deep learning model.

[0081] Suppose there are B users (B = n), of whom A are initially identified as abnormal users. Manual verification of these A users reveals that a1 of them are indeed gas theft users, while the remaining a2 are normal gas users. The gas usage data and labels of these B users are used as the dataset for training the deep learning model, with a1 being positive samples labeled "gas theft users" and a2 + (BA) being negative samples labeled "normal gas users." However, due to the model's initial screening and manual verification, there may be some gas theft users among the a2 + (BA) normal gas users. Therefore, deep learning model classification in step 2 is required to improve classification accuracy.

[0082] See Figure 2 The deep learning model includes a backbone network and a classification network. In the backbone network, the user name is input into the BERT language model, and a 768-dimensional feature vector is extracted. This feature vector is then passed through a fully connected layer to generate a 32-dimensional user information feature vector. Furthermore, the user's 28-day daily gas usage for a month is arranged as a column into a 12-column, 28-dimensional user gas usage table. Each month in the user gas usage table is processed through the embedding layer to form a 7-dimensional feature vector. This 7-dimensional feature vector is then concatenated with the user gas usage table to form a 12-column, 35-dimensional feature vector, with each column containing one token. These 12 tokens are passed through a 6-layer transformer layer and then a fully connected layer to extract a 32-dimensional user gas usage feature vector. The 32-dimensional user information feature vector and the 32-dimensional user gas usage feature vector are then concatenated to form a 64-dimensional feature vector. This 64-dimensional feature vector is then passed through a fully connected layer to output the 32-dimensional feature vector for the user.

[0083] In the classification network, the 32-dimensional feature vector output by the backbone network is passed through the fully connected layer to be reduced to an 8-dimensional feature vector, and then through the fully connected layer to be reduced to a 2-dimensional feature vector. Finally, it is passed through the softmax layer to output the user gas usage classification result, which is "1" for gas theft users or "0" for normal gas users.

[0084] When training the deep learning model, due to the early stage of business development, there are usually fewer training samples, so the entire training process is divided into two stages. The first stage is comparative training, during which the classification network is disconnected, such as Figure 3 As shown, the 32-dimensional feature vectors of each two users output by the backbone network are compared to make users with the same label as similar as possible and users with different labels dissimilar. The loss function used for comparison is cosine similarity f similary :

[0085]

[0086] Where a and b represent the 32-dimensional feature vectors of user a and user b, respectively, output by the backbone network. User a and user b are any two of B users, so theoretically the model will learn B×(B-1) cosine similarities.

[0087] Furthermore, samples with the same label in the dataset used to train the deep learning model are defined as similar samples, while samples with different labels are defined as dissimilar samples. If, among B users, a1 are labeled as gas theft users and a2 are labeled as normal gas users or users who do not require verification, without considering sample balance, to enhance the sample, a1×a1+a2×a2 positive samples and a1×(B-a1) negative samples can be generated. Because gas theft users are relatively rare, samples randomly selected from the total sample can be considered samples of normal gas users.

[0088] The second stage is classification training. During classification training, the backbone network and the classification network are opened at the same time to perform user classification training. The loss function used is the cross entropy loss function f classify :

[0089]

[0090] Where N represents the number of samples; y i represents the true label of the i-th sample (y i =0 indicates normal gas users, y i =1 indicates a user who steals gas); represents the predicted probability.

[0091] Step 3: Input the user name and the user's historical gas usage data into the trained deep learning model and output the recommended gas theft users.

[0092] When applying, the user's name and the user's gas consumption data in the past year are input into the trained deep learning model. When , the model recommends the user as a gas thief, otherwise it recommends the user as a normal gas user.

[0093] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

Claims

1. A method for recommending users of stolen gas based on deep learning, characterized in that: The following steps are involved: Step 1: construct a gas theft screening model to screen out abnormal users; the gas theft screening model includes a continuous zero gas usage index, conventional statistical index, linear fit goodness of fit index, and outlier index; In step 1, the steps of screening out abnormal users through the continuous zero gas usage index are as follows: Given a historical period of T days, if the gas consumption on a day is zero, it is recorded as a continuous zero gas consumption; calculate the number of consecutive zero gas consumption times z for each user i and the average number of consecutive zero gas usage for all users : Among them, z i is the number of consecutive zero gas usage of the i-th user, i=1,2,...,n; n is the total number of users; The average number of consecutive zero gas usage for all users; according to Calculate the Poisson distribution of the number of gas consumption segments and set the probability value P zero , calculate the lower limit of the number of consecutive zero gas consumption T zero : Among them, j is an index variable with a value range of [0,100]; The number of consecutive zero gas consumption is z i Greater than T zero The user is regarded as an abnormal user: Among them, user zero Represents a collection of abnormal users; To collect user zero The median of the maximum daily gas consumption of users in zero ) is the threshold, and after sorting, at most cnt abnormal users are recommended: Among them, u i Represents the collection user zero The maximum daily gas consumption of the i-th user in the historical period; Abnormal zero Indicates that the maximum daily gas consumption in the historical period is greater than the median (user zero )’s user collection; In step 1, the steps of screening out abnormal users through conventional statistical indicators are as follows: Calculate the skewness index s for each user in the historical time period i : Among them, x i,t represents the daily gas consumption of the i-th user on the t-th day in the past year, t=1,2,...,T; represents the average daily gas consumption of the i-th user in the past year; The skewness index s i Users with a value less than 0 are considered abnormal users: Among them, user stas Represents a collection of abnormal users; To collect user stas The median of the maximum daily gas consumption of users in stas ) is the threshold, and after sorting, at most cnt abnormal users are recommended: Among them, u i Represents the collection user stas The maximum daily gas consumption of the i-th user in the historical period; Abnormal stas Indicates that the maximum daily gas consumption in the historical period is greater than the median (user stas )’s user collection; In step 1, the steps of screening out abnormal users by using the linear fit goodness index are as follows: Calculate the goodness of fit r for each user i : Among them, y i,t represents the actual daily gas consumption of the i-th user on the t-th day; f i,t represents the daily gas consumption of the i-th user on the t-th day predicted by the fitting model; , SS reg , SS tot , SS res All are intermediate parameters; The goodness of fit r i Users with a score less than or equal to 0.95 are considered abnormal users: Among them, user fit Represents a collection of abnormal users; To collect user fit The median of the maximum daily gas consumption of users in fit ) is the threshold, and after sorting, at most cnt abnormal users are recommended: Among them, u i Represents the collection user fit The maximum daily gas consumption of the i-th user in the historical period; Abnormal fit Indicates that the maximum daily gas consumption in the historical period is greater than the median (user fit )’s user collection; In step 1, the steps of filtering out abnormal users by outlier indicators are as follows: Calculate the cosine similarity of each user's gas usage with another user: Among them, Similar a,b Represents the cosine similarity between user a and user b; a t represents the daily gas consumption of user a on day t; b t represents the daily gas consumption of user b on day t; For user i, the minimum value of its cosine similarity to all remaining users is used as the user's outlier indicator sim i : Among them, sim i Represents the outlier indicator of the i-th user; {Similar i,n-1 } represents the cosine similarity set between the i-th user and all remaining users; Simulate the user's outlier index i Sort from largest to smallest and extract the first 2×cnt users as abnormal users: Among them, user sim Represents the outlier indicator sim for all users i After sorting, the set of the first 2×cnt users; top(sim i ,2×cnt) represents the sim i The first 2×cnt users sorted from largest to smallest; To collect user sim The median of the maximum daily gas consumption of users in sim ) is the threshold, and after sorting, at most cnt abnormal users are recommended: Among them, u i Represents the collection user sim The maximum daily gas consumption of the i-th user in the historical period; Abnormal sim Indicates that the maximum daily gas consumption in the historical period is greater than the median (user sim )’s user collection; Step 2: Build a data set based on the initial screening of abnormal users for training the deep learning model; In step 2, the step of constructing a data set based on the initial screening of abnormal users includes: Assume there are B users in total, B = n, and A users are initially screened as abnormal users. Then, through manual verification of these A users, it is found that a1 of them are gas theft users and the remaining a2 are normal gas users. The gas usage data and labels of these n users are used as the dataset for training the deep learning model, with a1 being positive samples and a2 + (BA) being negative samples. Step 3: Input the user name and the user's historical gas usage data into the trained deep learning model and output the recommended gas theft users.

2. The method for recommending users of stolen gas based on deep learning according to claim 1 is characterized in that: In step 1, the abnormal users screened out are: Among them, A represents the initial screening of abnormal users.

3. The method for recommending users of stolen gas based on deep learning according to claim 1, characterized in that: The deep learning model includes a backbone network and a classification network. The operation process of the backbone network is as follows: The user name is input into the BERT language model of the backbone network to extract a 768-dimensional feature vector. This vector is then passed through a fully connected layer to generate a 32-dimensional user information feature vector. The user's 28-day daily gas usage for a month is taken as a column and arranged into 12 columns of 28-dimensional user gas usage tables. The month of each column in the user gas usage table is passed through the embedding layer to form a 7-dimensional feature vector. This 7-dimensional feature vector is concatenated with the user gas usage table to form 12 columns of 35-dimensional feature vectors, with each column containing one token. These 12 tokens are passed through a 6-layer transformer layer and then a fully connected layer to extract a 32-dimensional user gas usage feature vector. The 32-dimensional user information feature vector and the 32-dimensional user gas usage feature vector are concatenated to form a 64-dimensional feature vector. This 64-dimensional feature vector is then passed through a fully connected layer to output the 32-dimensional feature vector for the user. The operation process of the classification network is: The 32-dimensional feature vector output by the backbone network is passed through a fully connected layer to be reduced to an 8-dimensional feature vector, and then through a fully connected layer to be reduced to a 2-dimensional feature vector. Finally, it is passed through a softmax layer to output the user gas usage classification result, which is classified as a gas theft user or a normal gas user.

Images