Internet of Things potential customer mining method, device and equipment and storage medium

Abstract

The invention discloses a mining method for potential customers of the Internet of Things. The mining method comprises the following steps: S1, acquiring required data from a data source; s2, specifying a specific execution mode of important parameter operation mined by potential customers; s3, determining a configuration parameter value which enables the model effect to be optimal; s4, carrying out association analysis on all products, and searching an associated product set; s5, mining a sub-model, defining a positive sample and a negative sample, and completing data amplification; s6, extracting and generating data features required by model training; s7, training in combination with the optimal algorithm parameters to obtain a final model; s8, the sub-model processes the enterprise feature data and outputs potential customers of multiple products; and S9, storing configuration and key data, and endowing a unique identifier and a version number to realize version control. According to the method, the association rule between the products can be found by utilizing the characteristic that strong association exists between the products of the internet of things, and the associated product buyer sample is added to realize data amplification, so that the model effect is improved.

Description

Technical Field

[0001] The present invention relates to the field of intelligent data processing technology, and in particular to a method, device, equipment and storage medium capable of efficiently mining potential customers for products in multiple IoT segment industries. Background Art

[0002] The Internet of Things industry is experiencing a booming development. The constantly emerging application needs and directions in society have led to the spawning of multiple sub-industry fields (such as smart home, smart medical care, smart manufacturing, etc.) and promoted the birth of a variety of products (such as smart meters, wearable devices, security monitoring terminals, etc.). On the one hand, there are a large number of products in different sub-industry fields, with different functions and application scenarios. Their users and the needs of these users also have different characteristics. On the other hand, specific products are strongly associated with each other (i.e., the application scenarios, users, etc. are similar; for example, car wifi and driving recorders).

[0003] Combining the characteristics of each product and its target audience and using big data and artificial intelligence technology to build a potential customer mining model suitable for the corresponding product is a key step in carrying out precision marketing activities. However, for companies with a large and growing number of products, it is still difficult to efficiently mine potential customers for products in various industry segments, especially new products.

[0004] At present, the industry mainly uses two independent methods to mine potential customers for IoT products: one is a supervised method represented by classification. This method can combine product and customer characteristics for modeling and provide more accurate and explainable results, but it requires a large amount of labeling of samples. However, some products, especially newly launched products, have a lack of sales data, and the sample size that can be provided is generally insufficient to support the application of these methods; the other is an unsupervised method represented by association rules. This method does not require sample labeling and utilizes the strong correlation between specific IoT products, but cannot make full use of product and customer characteristics, so the effect is limited. In addition, when building a potential customer mining model, the current mainstream technologies only focus on potential customer mining of a single product, which greatly limits the scope and efficiency of mining, and lacks an effect tracking mechanism.

[0005] The above problems often lead to unsatisfactory results in potential customer mining, low efficiency, and difficulty in optimization. These problems are particularly obvious for companies that already have a large number of product types and the number of products is still growing, which may directly lead to the inability to seize market opportunities in a timely and effective manner.

[0006] To this end, the present application specifically proposes a method, device, equipment and storage medium for mining potential customers of the Internet of Things to solve the above technical problems. Summary of the invention

[0007] The main purpose of the present invention is to provide a method, device, equipment and storage medium for mining potential customers of products in segmented industries of the Internet of Things to solve the technical problems raised in the background technology.

[0008] The present invention adopts the following technical solutions to solve the above technical problems:

[0009] A method for mining potential customers of the Internet of Things includes the following specific execution steps:

[0010] S1. Collect various raw data from data sources and organize them into product sales data sets and enterprise information data sets;

[0011] S2. Create a set of configurations to specify how each key operation is performed;

[0012] S3. Determine the configuration parameter values ​​that achieve the best model effect through cross-validation method;

[0013] S4. Conduct association analysis to find other products with strong associations for each product in the data set;

[0014] S5. Mining sub-models for potential customers corresponding to each product, defining positive samples and negative samples, and completing data augmentation;

[0015] S6. Perform feature engineering on the enterprise information dataset to obtain an enterprise feature dataset;

[0016] S7. Combine the optimal algorithm parameters obtained by cross-validation and train to obtain the final model;

[0017] S8. Input the enterprise feature data set into each sub-model respectively, and summarize the output to obtain a list of potential customers;

[0018] S9. Save the configuration and key data, and assign unique identifiers and version numbers to each of them to implement version control.

[0019] Preferably, a device for mining potential customers of the Internet of Things is used to execute the above-mentioned method for mining potential customers of the Internet of Things, and the following unit modules are set corresponding to steps S1 to S9: a data acquisition unit, used to collect required data from a data source; a parameter configuration unit, used to specify important parameters for potential customer mining; a parameter optimization unit, used to determine the optimal values ​​of the specified parameters; an association analysis unit, used to conduct association analysis on all products and find a set of associated products; a sample annotation unit, used to annotate customer samples of various products and complete data augmentation; a feature engineering unit, used to extract and generate data features required for model training; a model training unit, used to train potential customer mining models corresponding to multiple products; a result output unit, used to output the prediction results of the potential customers and mining models of multiple products; a version control unit, used to save and identify configuration parameters, input data and output results, as well as to track effects and perform version rollback operations.

[0020] Preferably, the specific operation steps of searching for the associated product set in the association analysis unit include:

[0021] operate1. Set the set of all IoT products on sale to I = {i 1 ,i 2 ,...,i M}, set all transaction sets to T = {t 1 ,t 2 ,...,t N}, then each transaction t i The products in are all subsets of I;

[0022] operate2. For two different groups of products in I If customers who buy X also tend to buy Y, then the set consisting of them is called a frequent item set f, and the two will form an association rule X→Y. If X∪Y contains a total of k products, then f can be called a frequent k-item set. In this rule, X and Y are called the antecedent and consequent of the rule respectively.

[0023] operate3. Use support and confidence to measure the quality of frequent itemsets and association rules respectively. If σ represents the counting function involving transactions of a specific product, the support s and confidence c can be expressed as and Where N is the number of transactions.

[0024] Preferably, the support of the frequent itemsets in the operate2 step needs to reach the threshold s specified by the configuration. min , so the steps of obtaining and mining the frequent itemsets specifically include:

[0025] a1. First, let k = 1, then generate the set of all frequent k-item sets

[0026] a2. When Repeat the following process: Let k = k + 1, generate all candidate item sets Traverse each transaction t in the transaction set T and generate C k The set of all candidate item sets belonging to t in Traverse the candidate set C t For each item set c in, count item set c by σ(c)=σ(c)+1, and generate the set of all frequent k-item sets at this time

[0027]

[0028] a3. Summarize all frequent item sets except frequent 1-item sets

[0029] Preferably, the confidence of the association rule needs to reach a threshold c specified by the configuration. min Therefore, the steps of obtaining and mining the association rules specifically include:

[0030] b1. Traverse F 总 For each frequent k-itemset f in k ;

[0031] b2. Let m = 1, generate the 1-term consequent H of all rules at this time 1 ={i|i∈f k}, if k>m+1, repeat the following process: generate all candidate rule consequent sets Traverse the rule consequence set H m+1 Each consequent h in m+1 , calculate the confidence c = σ(f k ) / σ(f k -h m+1 ), if c ≥ c min , then the output rule (f k -h m+1 )→h m+1 , otherwise from H m+1 Delete h m+1 , finally set m=m+1;

[0032] b3. Summarize all association rules to R 总 .

[0033] Preferably, the specific operation steps of labeling customer samples of various products in the sample labeling unit include:

[0034] L1. The item set consisting of the specified product is recorded as X. In this case, the item set contains only one product, that is, |X|=1;

[0035] L2. All customers with X purchase records in the product sales dataset are classified as positive samples;

[0036] L3. Classify other customers in the product sales dataset as negative samples;

[0037] L4. Since each IoT product has its own corresponding positive samples and negative samples, the two types of samples can be merged to form a customer label dataset for the corresponding product, thereby completing the labeling process of the customer samples.

[0038] Preferably, if the number of positive samples in the customer sample is lower than the sample size threshold specified in the configuration, the following operations are performed before the L3 step:

[0039] 1) Traverse all association rules R 总 , filter out the association rule R whose rule antecedent (or consequent) is equal to X 关联X = {Y|X→Y∨Y→X}, and R 关联X All Y in are regarded as the associated product set (|Y|≥1);

[0040] 2) Centralize the product sales data and store it in R 关联X Customers with any purchase record of Y in are classified as positive samples to achieve data augmentation.

[0041] Preferably, the parameter optimization unit adopts a cross-validation method to determine the optimal values ​​of all range parameters for which unique values ​​have not yet been determined as the configuration parameter values ​​for achieving the optimal model effect, and adopts a cross-validation method to evaluate and select parameter combinations for the model trained by the model training unit.

[0042] On the other hand, the present invention also discloses a computer device, including a memory storage and a processor, wherein the memory storage is used to provide high-bandwidth, low-latency, non-persistent storage functions, the memory storage stores a computer program, and the processor implements the functions of each unit of the above-mentioned device by reading and executing the program in the memory storage.

[0043] On the other hand, the present invention further discloses a computer-readable storage medium which persistently stores programs and data. When the computer program is read and executed by a processor, the processor implements the functions of each unit of the above-mentioned device.

[0044] It can be seen from the above technical solution that the present invention provides a method, device, equipment and storage medium for mining potential customers of the Internet of Things. Compared with the prior art, the present invention has the following advantages:

[0045] 1. The present invention can propose a data amplification method for unsupervised association analysis by setting an association analysis unit and a sample annotation unit, and combine it with a supervised classification algorithm to complete data amplification before training the potential customer mining model, so that the sample size can be increased, thereby effectively solving the problem of insufficient training data and low effect of the potential customer mining model.

[0046] 2. The present invention, by setting up an association analysis unit and a sample labeling unit, can utilize the fact that there are often strong associations between IoT products, and proposes to use an association analysis method to find association rules between products. For products with a small number of buyers, samples of buyers of related products are added to achieve data amplification, so that the sample size meets the requirements of the supervised learning algorithm, thereby improving the effect of the model.

[0047] 3. Based on model establishment and training, the present invention can integrate association analysis into the cross-validation process in the association analysis unit by combining an extended cross-validation method, so that the effect of the data augmentation method based on association rules can be evaluated and optimized, which ultimately helps to improve the model effect.

[0048] 4. The method of the present invention can be used to construct a potential customer mining model that covers multiple industry segment products at the same time, improve the coverage and efficiency of potential customer mining, and simultaneously output a list of potential customers for each product, with a wider range of applicability.

[0049] It should be understood that the contents described in this section are not intended to identify the key or important features of the embodiments of the present invention, nor are they intended to limit the scope of the present invention. Other features of the present invention will become easily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] The drawings constituting a part of the present application are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0051] Figure 1 It is a schematic diagram of the overall process of the method of the present invention;

[0052] Figure 2 It is a schematic diagram of a unit frame of the device of the present invention;

[0053] Figure 3 It is a schematic diagram of the data set processing flow of the present invention;

[0054] Figure 4 It is a schematic diagram of parameter optimization of the present invention;

[0055] Figure 5 It is a plane example diagram for explaining the effect of data amplification on sample size in the present invention;

[0056] Figure 6 This is a plane example diagram illustrating the impact of data augmentation on the model effect of the present invention;

[0057] Figure 7 It is a schematic diagram of the computing device and storage medium of the present invention. DETAILED DESCRIPTION

[0058] 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 part of the embodiments of the present invention, rather than all of the embodiments. In the absence of conflict, the embodiments in this application and the features in the embodiments can be combined with each other. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0059] In the embodiment, see Figures 1 to 7 .

[0060] like Figure 1 The method for mining potential customers of the Internet of Things proposed in the embodiment of the present invention is used to solve the problems of insufficient sample quantity, unsatisfactory model effect, single and inefficient mining scope, lack of effect tracking mechanism, etc. by combining the characteristics of many types of products in the Internet of Things industry and strong correlation between specific types of products. The method specifically includes the following execution steps:

[0061] S1. Collect various raw data from data sources and organize them into product sales data sets and enterprise information data sets;

[0062] S2. Create a set of configurations to specify how each key operation is performed;

[0063] S3. Determine the configuration parameter values ​​that achieve the best model effect through cross-validation method;

[0064] S4. Conduct association analysis to find other products with strong associations for each product in the data set;

[0065] S5. Mining sub-models for potential customers corresponding to each product, defining positive samples and negative samples, and completing data augmentation;

[0066] At this time, in order to solve the problem of insufficient training samples for certain products (especially newly launched products) and the inconvenience of using machine learning algorithms for modeling, data augmentation can be completed before training the potential customer mining model, so that the sample size can be increased. In addition, in view of the problem that association analysis belongs to unsupervised learning and the cross-validation optimization rules cannot be directly applied, the cross-validation range can be improved to cover the association analysis and data augmentation processes, and the direct evaluation of the effects of the association rule algorithm parameters can be achieved. The effect of the mining model can be further improved based on the data augmentation of the association analysis.

[0067] S6. Perform feature engineering on the enterprise information dataset to obtain an enterprise feature dataset;

[0068] S7. Combine the optimal algorithm parameters obtained by cross-validation and train to obtain the final model;

[0069] S8. Input the enterprise feature data set into each sub-model respectively, and summarize the output to obtain a list of potential customers;

[0070] S9. Save the configuration and key data, and assign unique identifiers and version numbers to achieve version control;

[0071] At this time, in order to address the problem of lack of effect tracking and result tracing in general potential customer mining methods (especially potential customer mining involving multiple products), a version control mechanism can be introduced to achieve the marking, preservation and rollback of historical version results.

[0072] On the other hand, based on the above method, the present invention further proposes a mining device for potential customers of the Internet of Things, which is used to execute the above mining method for potential customers of the Internet of Things, and can carry out construction processes for a variety of products, complete parameter selection and effect verification, thereby improving the efficiency of modeling. Figure 2 As shown, the device specifically sets the following unit modules in sequence corresponding to steps S1 to S9:

[0073] (1) a data acquisition unit, used to collect required data from a data source;

[0074] This unit is used to collect various raw data from data sources and organize them into product sales data sets and enterprise information data sets. The product sales data sets come from the product dealer management database, which includes the sales volume of various products and purchasing customers, etc. The enterprise information data sets come from the organizational industrial and commercial database, which contains the industrial and commercial information data of enterprises across the country. The relevant data sets and data flows here are as follows: Figure 3 As shown;

[0075] In addition, it should be pointed out that this technical solution limits the scope of potential customers of IoT products to enterprises. In addition, the enterprise information data set also includes enterprises that have purchased any products and those that have not purchased any products. The data of the former will be used for model training later, while the data of the latter will be used for model prediction and output of potential customers.

[0076] (2) a parameter configuration unit, used to specify important parameters for potential customer mining;

[0077] This unit is used to create a set of configurations to specify the specific execution methods of each key operation in this technical solution, where the configuration includes multiple configuration items and corresponding parameter values ​​or parameter value ranges, which are used to specify the implementation conditions of specific operations in subsequent steps (such as association analysis, feature engineering, model training, etc.);

[0078] At the same time, the value of the parameter in the configuration content can be uniquely determined or can be an optional value range;

[0079] For configuration parameters that specify a unique value, subsequent processes will be executed directly according to the parameter value;

[0080] For configuration parameters with a specified value range, the cross-validation method is used in the subsequent process to select the best unique value from the range;

[0081] (3) a parameter optimization unit, used to determine the optimal value of a specified parameter;

[0082] This unit part uses the cross-validation method to determine the configuration parameter values ​​that make the model effect optimal. It is mainly used to determine the optimal values ​​of all range parameters that have not yet determined unique values;

[0083] Specifically, for the potential customer mining model corresponding to each product, the cross-validation method can be used to evaluate and optimize the parameter combination;

[0084] The cross-validation method specifically includes the steps of generating candidate parameter combinations, dividing the data set, training the test model, calculating the evaluation index, and selecting the optimal parameter combination according to the index. Figure 4 As shown;

[0085] (4) a correlation analysis unit, used to perform correlation analysis on all products and find a set of related products;

[0086] It should be noted that in some scenarios of the prior art, customers often need to use multiple different IoT products at the same time to achieve a common demand, or combine different products to make the application more perfect; for these products, their potential customers are the same or similar, for example, multiple vehicle-mounted devices such as driving recorders, smart rearview mirrors, and vehicle locators can be installed in the same vehicle, and customers are likely to purchase them at the same time; therefore, to some extent, the potential customers of one product (such as a driving recorder) are also the potential customers of another product (such as a smart rearview mirror);

[0087] Therefore, in order to find similar product relationships, this unit is used to conduct association analysis on the product sales dataset to find other products with strong associations for each product, and the results will be used to guide the division of positive and negative samples in the model training samples in the subsequent steps;

[0088] Specifically, the specific operation steps of searching for a related product set in the related analysis unit include:

[0089] operate1. Set the set of all IoT products on sale to I = {i 1 ,i 2 ,...,i M}, set all transaction sets to T = {t 1 ,t 2 ,...,t N}, then each transaction t i The products in are all subsets of I;

[0090] operate2. For two different groups of products in I If customers who buy X also tend to buy Y, then the set consisting of them is called a frequent item set f, and the two will form an association rule X→Y. If X∪Y contains a total of k products, then f can be called a frequent k-item set. In this rule, X and Y are called the antecedent and consequent of the rule respectively.

[0091] Since the support of frequent itemsets needs to reach the threshold s specified by the configuration min , so the steps to obtain and mine frequent itemsets specifically include:

[0092] a1. First, let k = 1, then generate the set of all frequent k-item sets

[0093] a2. When Repeat the following process: Let k = k + 1, generate all candidate item sets Traverse each transaction t in the transaction set T and generate C k The set of all candidate item sets belonging to t in Traverse the candidate set C t For each item set c in, count item set c by σ(c)=σ(c)+1, and generate the set of all frequent k-item sets at this time

[0094]

[0095] a3. Summarize all frequent item sets except frequent 1-item sets

[0096] In addition, since the confidence of the association rule needs to reach the configured threshold c min , so the steps of obtaining and mining association rules specifically include:

[0097] b1. Traverse F 总 For each frequent k-itemset f in k ;

[0098] b2. Let m = 1, generate the 1-term consequent H of all rules at this time 1 ={i|i∈f k}, if k>m+1, repeat the following process: generate all candidate rule consequent sets Traverse the rule consequence set H m+1 Each consequent h in m+1 , calculate the confidence c = σ(f k ) / σ(f k -h m+1 ), if c ≥ c min , then the output rule (f k -h m+1 )→h m+1 , otherwise from H m+1 Delete h m+1 , finally set m=m+1;

[0099] b3. Summarize all association rules to R 总

[0100] operate3. Use support and confidence to measure the quality of frequent itemsets and association rules respectively. If σ represents the counting function involving transactions of a specific product, the support s and confidence c can be expressed as and Where N is the number of transactions;

[0101] (5) Sample labeling unit, used to label customer samples of various products and complete data augmentation;

[0102] This unit is used to define positive and negative samples for the potential customer mining sub-model corresponding to each product and complete data augmentation;

[0103] For example, the positive samples of some products (e.g., smart rearview mirrors) include not only customers who have purchased the product, but also customers who have purchased related products (e.g., driving recorders). Since customers of related products are added to the samples, this method can increase training samples and data amplification for products with fewer existing customers, so that the model can achieve better results.

[0104] Specifically, taking some products as an example, the specific steps for marking customer samples of various products in this unit include:

[0105] L1. The item set consisting of the specified product is recorded as X. In this case, the item set contains only one product, that is, |X|=1;

[0106] L2. All customers with X purchase records in the product sales dataset are classified as positive samples;

[0107] At this time, if the number of positive samples in the customer sample is lower than the sample size threshold specified in the configuration, the following operations are performed:

[0108] 1) Traverse all association rules R 总 , filter out the association rule R whose antecedent (or consequent) is equal to X 关联X = {Y|X→Y∨Y→X}, and R 关联X All Y in are regarded as the associated product set (|Y|≥1);

[0109] 2) Centralize the product sales data and store it in R 关联X Customers with any purchase record of Y in are classified as positive samples to achieve data augmentation

[0110] L3. Classify other customers in the product sales dataset as negative samples;

[0111] L4. Since each IoT product has its own corresponding positive samples and negative samples, the two types of samples can be merged to form a customer label dataset for the corresponding product, thereby completing the labeling process of the customer samples.

[0112] At this point, the effect can be further explained:

[0113] Since supervised machine learning algorithms require a large number of labeled samples, but for some products (especially newly launched products), labeled sample data is difficult to obtain, which makes it difficult for supervised machine learning to play a good role in mining potential customers for these products; on the other hand, unsupervised association rule algorithms cannot fully utilize product and customer characteristics, and cannot achieve good results when used alone;

[0114] Therefore, the present invention can propose a data augmentation method for unsupervised association analysis by setting an association analysis unit and a sample annotation unit, and combine it with a supervised classification algorithm to complete data augmentation before training the potential customer mining model, so that the sample size can be increased, thereby effectively solving the problem of insufficient training data and low effect of the potential customer mining model. In addition, the present invention can also use the characteristics of strong associations between IoT products by setting an association analysis unit and a sample annotation unit, and propose to use an association analysis method to find association rules between products. For products with a small number of buyers, samples of buyers of related products are added to achieve data augmentation, so that the sample size meets the requirements of the supervised learning algorithm, thereby improving the effect of the model;

[0115] (6) Feature engineering unit, used to extract and generate data features required for model training;

[0116] This unit part is used to perform feature engineering on the enterprise information dataset to obtain the enterprise feature dataset, which is the same as the enterprise information dataset. The enterprise feature dataset contains both enterprises that have purchased any product and enterprises that have not purchased any product.

[0117] Specifically, for the numerical fields (such as registered capital) and categorical fields (such as enterprise type) in the enterprise information dataset, this unit part converts them into floating-point and integer variables to form features;

[0118] Specifically, for text fields (including company name and business scope), this unit will construct features through natural language processing technology;

[0119] It should be noted that all the above features need to be standardized or normalized to remove the dimension and keep the order of magnitude similar;

[0120] (7) a model training unit, used to train potential customer mining models corresponding to multiple products;

[0121] This unit is used to combine the optimal algorithm parameters obtained by cross-validation to train the final model. At this time, since the potential customer model corresponding to each product is equivalent to a sub-model, it is possible to mine potential customers for a product separately.

[0122] It can be further explained that before training the model, it is necessary to first associate the enterprise feature dataset and the customer label dataset through the unified social credit code, and keep the record of successful association to obtain the training dataset, and then train according to the specified algorithm and hyperparameters to obtain the model, in which the training dataset contains both enterprise features and sample type labels;

[0123] In addition, the optimal configuration parameter values ​​determined by the cross-validation method in the above parameter optimization unit can also be used to evaluate and select the parameter combination of the model trained by the model training unit by the cross-validation method;

[0124] In the prior art, since conventional cross-validation methods cannot be directly applied to unsupervised methods such as association analysis, it is difficult to achieve effect evaluation and parameter optimization. Therefore, the present invention proposes an extended cross-validation method, which enables the data augmentation method based on association analysis proposed by the present invention to effectively improve the effect of the model. By combining the extended cross-validation method on the basis of model establishment and training, the association analysis can be integrated into the cross-validation process in the association analysis unit, so that the effect of the data augmentation method based on association rules can be evaluated and optimized, which ultimately helps to improve the model effect.

[0125] (8) a result output unit, used to output the potential customers of multiple products and the prediction results of the mining model;

[0126] This unit is used to input the enterprise feature data set into each sub-model and summarize the output to obtain a potential customer list. The summarized potential customer list here includes potential customers of all products.

[0127] It can also be added that in this unit part, each product sub-model will output a value between 0 and 1 for each enterprise, which represents the probability of the enterprise being a potential customer of the product. Therefore, according to the probability threshold parameter specified in the configuration, the enterprises with probabilities exceeding the threshold are screened out and regarded as potential customers of the corresponding products. Then the results of all products are summarized to obtain the final result.

[0128] In addition, since the prior art can only focus on the construction of a single industry product model when building a model, the output results only cover the potential customers of a single product and have a single application surface; if multiple products need to be covered, the construction process must be repeated for each product. Therefore, the method of the present invention can improve the coverage and efficiency of potential customer mining by constructing a potential customer mining model that covers multiple subdivided industry products at the same time, and simultaneously outputs a list of potential customers for each product, which has a wider application surface;

[0129] (9) Version control unit, used to save and identify configuration parameters, input data, and output results, as well as to track effects and perform version rollback operations;

[0130] This unit is used to save configuration and key data, and assign unique identifiers and version numbers to achieve version control. The materials saved here will be used for version control to achieve version rollback and other functions;

[0131] In addition, after obtaining a new version of the model, if the new model is found to be less effective than the historical version by comparing the model evaluation indicators, you can specify the historical version number to use the historical version data, parameters and programs, re-run the process and output the results;

[0132] In summary, since version control can be used for the process construction of the entire model, it helps to track the model effect and the necessary result backtracking. Therefore, the present invention introduces version control into the potential customer mining model construction process, and can save and identify the historical versions of model input, parameter configuration, and model output, so as to realize the functions of tracing and rolling back the historical versions.

[0133] This function plays an important role in urgently fixing result errors, summarizing effects, and providing improvement directions. It is especially important for application scenarios involving multiple potential customer models of different products.

[0134] Based on the above-mentioned method and device for mining potential customers of the Internet of Things, a specific embodiment can be proposed to further illustrate the technical solution of the present invention.

[0135] Through the data collection unit of the device proposed by the present invention, this embodiment collects raw data from the data source and integrates and processes it, and the obtained product sales data set and enterprise information data set are shown in the following tables respectively:

[0136] Product sales data set table (partial example after sensitive information is obfuscated):

[0137]

[0138]

[0139] Enterprise information data set table:

[0140]

[0141] Through the parameter configuration unit of the device proposed by the present invention, the configuration and parameters created in this embodiment are shown in the following table:

[0142] Configuration items and parameter tables:

[0143]

[0144]

[0145] For several range parameters in the configuration, this embodiment uses completely random sampling within the specified value range to generate 30 parameter combinations for the modeling process corresponding to each product for cross-validation and determines the optimal parameter combination. Taking a product numbered 1 as an example, the combination of non-unique parameters involved and the cross-validation results are shown in the following table (unique parameters are not listed in the table) (in this embodiment, the cross-validation fold is 5, the evaluation index is the area under the ROC curve, and the final evaluation index is the average of the index obtained by 5 verifications):

[0146] Cross validation result table:

[0147]

[0148]

[0149] In the above results, select the row with the largest area under the ROC curve, and the corresponding parameter combination is the optimal combination; when training the final model corresponding to this product, the above parameters will be used to complete the entire process. At the same time, for other products, the above parameter optimization process needs to be completed separately to obtain their respective optimal parameter combinations.

[0150] For the product sales data set, the method described in the technical solution of this application is used to mine product combinations that are associated with the sales objects. Continuing with a product numbered 1 as an example, according to the method described in this technical solution, and the optimal support threshold and confidence threshold parameters used in this embodiment, the associated product rules finally mined are shown in the following table:

[0151] Product sales data set association rule table:

[0152] rule Support Confidence {Product 2} -> {Product 1} 0.221 0.523 {Product 3} -> {Product 1} 0.167 0.435 {Product 5, Product 6} -> {Product 1} 0.061 0.362

[0153] The above results show that customers who purchase product 2 or product 3 alone, as well as customers who purchase product 5 and product 6 at the same time, are very likely to purchase product 1. Since the number of customers who have purchased product 1 is small and does not reach the positive sample size threshold specified in the configuration, according to the sample annotation and data augmentation method proposed in this technical solution, customers of related products can be regarded as potential customers of product 1 and included in the positive samples.

[0154] Specifically, the positive sample labeling rules of the potential customer mining model of product 1 are:

[0155] Customers that exist in the product sales dataset and meet any of the following conditions:

[0156] (1) Purchased product 1;

[0157] (2) Purchased product 2;

[0158] (3) Purchased product 3;

[0159] (4) purchased both product 5 and product 6;

[0160] The negative sample labeling rule for the potential customer mining model of product 1 is: customers who exist in the product sales dataset and are not labeled as positive samples.

[0161] For other products except product 1, the above-mentioned association analysis and sample labeling process are required to obtain corresponding results. In this embodiment, the obtained customer label data set is shown in the following table:

[0162] Customer label dataset table:

[0163] Unified social credit code Product Code Sample Type 91652200***L 1 1 91652200***L 2 0 …… …… ……

[0164] In this embodiment, a total of 151 products need to build a potential customer mining model. Before using the association analysis and sample annotation method proposed in this technical solution, the number of positive samples of 37 products did not reach the sample size threshold specified in the configuration. After using the method, related products were found for 29 of the products and data was amplified, which increased the mean number of positive samples from 263 to 598 (an increase of 127%), and finally increased the overall positive sample size compliance rate from 75.9% to 94.7% ( Figure 5 shown).

[0165] For the enterprise information data set, feature engineering is performed according to the method described in the technical solution of this application, and the enterprise feature data set shown in the following table is obtained:

[0166] Enterprise characteristic data set table (partial example):

[0167]

[0168] After obtaining the enterprise feature data set, combined with the sample annotation results of each product, the model is trained according to the model training parameters (configuration items and parameter tables) specified in the configuration. It should be noted here that both the parameter optimization stage and the final model training stage involve model training operations. The difference is that in the parameter optimization stage, multiple parameter combinations will be used to train multiple models for each product on the cross-validation data set, and the optimal parameter combination will be determined by comparing evaluation indicators; while in the final model training stage, the determined optimal parameter combination will be used to train the model on the full data set to obtain the final model for each product.

[0169] For the 37 products in this embodiment whose positive sample quantity does not meet the standard, after the association analysis and data amplification operation introduced in this technical solution, not only the number of positive samples is increased ( Figure 5As shown in Figure 2, the model effect was also improved. According to the cross-validation results, the mean area under the ROC curve corresponding to the 37 products increased from 0.574 to 0.693, an increase of 20.7% ( Figure 6 shown).

[0170] In the model prediction stage, the enterprise feature data set is input into each product model to obtain the corresponding output; according to the binary classification threshold specified in the configuration, the records with probability exceeding the threshold are screened out, regarded as potential customers and summarized into the final result table:

[0171] Prediction results table:

[0172] Unified social credit code Product Code Prediction results 91360802***J 1 0.865 91330109***C 2 0.535 91360481***C 3 0.664 91330621***D 4 0.699 91330110***T 5 0.571 …… …… ……

[0173] Save configuration and key data, and assign unique identifiers and version numbers to them to implement version control. The saved materials will be used for version control to implement functions such as version rollback.

[0174] In this embodiment, various objects saved in a certain version are as follows:

[0175] Historical version table:

[0176] Object Type Object Identification Version Number Configuration 224e2ccda861c2514faa683b3683c361 2024041200 Product Sales Dataset 894b23e4fb786a236f32f561312445f4 2024041200 Enterprise Information Dataset 502a556680e694834bb1200e235df9bc 2024041200 Enterprise feature dataset 357890fc997cb40244b709dd9cd6a07e 2024041200 Sample annotation dataset 1b01d20e024f75d63aecb1907a253b53 2024041200 Cross validation results 9b92f7b578be38f778d11ec854eff6ee 2024041200 Prediction results dc21b725561c66bd16bc8ebb2db6113c 2024041200

[0177] After obtaining the new version of the model, if you find that the effect of the new model is lower than that of the historical version of the model by comparing the model evaluation indicators, you can specify the historical version number to use the historical version data, parameters and programs, re-run the process and output the results.

[0178] On the other hand, reference Figure 7 The present invention also discloses a computer device, including a memory storage and a processor, wherein the memory storage is used to provide high-bandwidth, low-latency, non-persistent storage functions, the memory storage stores computer programs, and the processor realizes the functions of each unit of the above-mentioned device by reading and executing the programs in the memory storage.

[0179] On the other hand, refer to Figure 7 The present invention also discloses a computer-readable storage medium that persistently stores programs and data. When the computer program is read and executed by a processor, the processor implements the functions of each unit of the above-mentioned device.

[0180] In another embodiment provided by the present application, a computer program product including instructions is also provided, which, when executed on a computer, enables the computer to execute any of the methods for mining potential customers of the Internet of Things in the above-mentioned embodiments.

[0181] It is understandable that the system provided by the embodiment of the present invention corresponds to the method provided by the embodiment of the present invention, and the explanation, examples and beneficial effects of the relevant contents can refer to the corresponding parts in the above method.

[0182] The embodiment of the present application also provides an electronic computing device, including a processor, a storage medium, a communication network interface, a memory storage and a communication bus, wherein the processor, the storage medium, the communication network interface and the memory storage complete mutual communication interaction through the communication bus:

[0183] The storage medium stores the potential customer mining model calculation program proposed by the invention, as well as various signal data required by the program.

[0184] Memory storage, used to store computer programs;

[0185] Processor, which can read and write computing programs and signal data in storage media, and transfer programs and data to memory storage to achieve high-performance access;

[0186] In addition, the computing device also comes with a network interface for interacting with an external system platform to realize the collection of signal data.

[0187] The communication bus mentioned in the above electronic computing device can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. The communication bus can be divided into an address bus, a data bus, a control bus, etc.

[0188] The communication interface is used for communication between the above electronic device and other devices.

[0189] The memory storage may include a random access memory (RAM) or a non-volatile memory (NVM), such as at least one disk storage. Optionally, the memory may also be at least one storage device located away from the aforementioned processor.

[0190] The above-mentioned processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components.

[0191] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware or any combination thereof. When implemented using software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the process or function described in the embodiment of the present application is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium, for example, the computer instructions can be transmitted from a website site, computer, server or data center by wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) mode to another website site, computer, server or data center. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server, data center, etc. that contains one or more available media integrated. The available medium can be a magnetic medium, (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium (e.g., a solid-state drive So lid State Dis sk (SSD)), etc.

[0192] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

[0193] In addition, if there are descriptions involving "first", "second", etc. in the embodiments of the present invention, the descriptions of "first", "second", etc. are only used for descriptive purposes and cannot be understood as indicating or suggesting their relative importance or implicitly indicating the number of technical features indicated. In addition, the meaning of "and / or" appearing in the full text includes three parallel schemes. Taking "A and / or B" as an example, it includes scheme A, or scheme B, or a scheme in which A and B are satisfied at the same time. In addition, in the embodiments of the present invention, "multiple" refers to more than two. In addition, the technical solutions between the various embodiments can be combined with each other, but it must be based on the ability of ordinary technicians in the field to implement. When the combination of technical solutions is contradictory or cannot be implemented, it should be deemed that such combination of technical solutions does not exist and is not within the scope of protection required by the present invention.

Claims

1. A method for mining potential customers of the Internet of Things, characterized in that: The specific implementation steps include the following: S1. Collect various raw data from data sources and organize them into product sales data sets and enterprise information data sets; S2. Create a set of configurations to specify how each key operation is performed; S3. Determine the configuration parameter values ​​that achieve the best model effect through cross-validation method; S4. Conduct association analysis to find other products with strong associations for each product in the data set; S5. Mining sub-models for potential customers corresponding to each product, defining positive samples and negative samples, and completing data augmentation; S6. Perform feature engineering on the enterprise information dataset to obtain an enterprise feature dataset; S7. Combine the optimal algorithm parameters obtained by cross-validation and train to obtain the final model; S8. Input the enterprise feature data set into each sub-model respectively, and summarize the output to obtain a list of potential customers; S9. Save the configuration and key data, and assign unique identifiers and version numbers to each of them to implement version control.

2. A device for mining potential customers of the Internet of Things, characterized in that: Used to execute the method for mining potential customers of the Internet of Things as described in claim 1 above, and the following unit modules are set corresponding to steps S1 to S9: A data collection unit, used for collecting required data from a data source; Parameter configuration unit, used to specify important parameters for potential customer mining; A parameter optimization unit, used to determine the optimal value of a specified parameter; The association analysis unit is used to conduct association analysis on all products and find related product sets; The sample annotation unit is used to annotate customer samples of various products and complete data augmentation; Feature engineering unit, used to extract and generate data features required for model training; Model training unit, used to train potential customer mining models corresponding to multiple products; A result output unit, used to output the potential customers of multiple products and the prediction results of the mining model; The version control unit is used to save and identify configuration parameters, input data, and output results, as well as track effects and perform version rollback operations.

3. The device for mining potential customers of the Internet of Things as claimed in claim 2, characterized in that: The specific operation steps of searching for the associated product set in the association analysis unit include: operate1. Set the set of all IoT products on sale to I = {i1,i2,...,i M }, set the set of all transactions to T = {t1, t2, ..., t N }, then each transaction t i The products in are all subsets of I; operation2. For two different groups of products X in I, If customers who buy X also tend to buy Y, then the set consisting of them is called a frequent item set f, and the two will form an association rule X→Y. If X∪Y contains a total of k products, then f can be called a frequent k-item set. In this rule, X and Y are called the antecedent and consequent of the rule respectively. operate3. Use support and confidence to measure the quality of frequent itemsets and association rules respectively. If σ represents the counting function involving transactions of a specific product, the support s and confidence c can be expressed as and Where N is the number of transactions.

4. The device for mining potential customers of the Internet of Things as claimed in claim 3, characterized in that: The support of the frequent itemsets in the operate2 step needs to reach the threshold s specified in the configuration. min , so the steps of obtaining and mining the frequent itemsets specifically include: a1. First, let k = 1, then generate the set of all frequent k-item sets a2. When Repeat the following process: Let k = k + 1, generate all candidate item sets Traverse each transaction t in the transaction set T and generate C k The set of all candidate item sets belonging to t in Traverse the candidate set C t For each item set c in, count item set c by σ(c)=σ(c)+1, and generate the set of all frequent k-item sets at this time a3. Summarize all frequent item sets except frequent 1-item sets 5. The device for mining potential customers of the Internet of Things as claimed in claim 4, characterized in that: The confidence of the association rule needs to reach the threshold c specified by the configuration min Therefore, the steps of obtaining and mining the association rules specifically include: b1. Traverse F 总 For each frequent k-itemset f in k ; b2. Let m = 1, generate the 1-term consequent H1 = {i|i∈f k }, if k>m+1, repeat the following process: generate all candidate rule consequent sets Traverse the rule consequence set H m+1 Each consequent h in m+1 , calculate the confidence c = σ(f k ) / σ(f k -h m+1 ), if c ≥ c min , then the output rule (f k -h m+1 )→h m+1 , otherwise from H m+1 Delete h m+1 , finally set m=m+1; b3. Summarize all association rules to R 总 .

6. The device for mining potential customers of the Internet of Things as claimed in claim 3, characterized in that: The specific operation steps of labeling customer samples of various products in the sample labeling unit include: L1. The item set consisting of the specified product is recorded as X. In this case, the item set contains only one product, that is, |X|=1; L2. All customers with X purchase records in the product sales dataset are classified as positive samples; L3. Classify other customers in the product sales dataset as negative samples; L4. Since each IoT product has its own corresponding positive samples and negative samples, the two types of samples can be merged to form a customer label dataset for the corresponding product, thereby completing the labeling process of the customer samples.

7. The device for mining potential customers of the Internet of Things as claimed in claim 6, characterized in that: If the number of positive samples in the customer sample is lower than the sample size threshold specified in the configuration, the following operations are performed before the L3 step: 1) Traverse all association rules R 总 , filter out the association rule R whose antecedent (or consequent) is equal to X 关联X = {Y|X→Y∨Y→X}, and R 关联X All Y in are regarded as the associated product set (|Y|≥1); 2) Centralize the product sales data and store it in R 关联X Customers with any purchase record of Y in are classified as positive samples to achieve data augmentation.

8. The device for mining potential customers of the Internet of Things as claimed in claim 2, characterized in that: The parameter optimization unit adopts a cross-validation method to determine the optimal values ​​of all range parameters that have not yet determined unique values ​​as the configuration parameter values ​​for achieving the optimal model effect, and adopts a cross-validation method to evaluate and select parameter combinations for the model trained by the model training unit.

9. A computer device, characterized in that: It includes a memory storage and a processor, wherein the memory storage is used to provide high-bandwidth, low-latency, non-persistent storage functions, the memory storage stores a computer program, and the processor implements the functions of each unit of the device as described in any one of claims 2 to 8 by reading and executing the program in the memory storage.

10. A computer-readable storage medium, characterized in that: Programs and data are persistently stored, and when the computer program is read and executed by a processor, the processor implements the functions of each unit of the device as claimed in any one of claims 2 to 8.

Images