Package recommendation method, device, equipment and storage medium

Abstract

The application discloses a package recommendation method and device, equipment and storage medium, relate to package recommendation technical field, method includes: obtaining the related historical data set of target user, related historical data set includes multiple user related data;The related historical data set is input into the feature extraction module of the package recommendation model, and long sequence feature coding is generated;The feature extraction module includes a multi-granularity scanning module of a deep forest neural network;The long sequence feature coding is input into the package matching module of the package recommendation model, and a package matching result of a to-be-recommended package and the target user is obtained;The package matching module includes a recurrent neural network.The application makes the recurrent neural network output the package matching result of the to-be-recommended package and the target customer more accurately according to the long sequence feature coding, solves the technical problem that the existing package recommendation model has low accuracy, and improves the accuracy of package recommendation.

Description

Technical Field

[0001] This invention relates to the field of package recommendation technology, and in particular to a package recommendation method, apparatus, device, and storage medium. Background Technology

[0002] In related technologies, telecommunications operators can build package recommendation models based on user feature profile data and historical sample label data, using appropriate feature engineering and algorithm models to achieve high-quality marketing of various package products. However, the data features obtained from existing feature engineering are noisy, resulting in low accuracy of package recommendation models. Summary of the Invention

[0003] The main objective of this invention is to provide a method, apparatus, device, and storage medium for recommending service packages, aiming to solve the technical problem of low accuracy in existing service package recommendation models.

[0004] To achieve the above objectives, the present invention adopts the following technical solution:

[0005] In a first aspect, the present invention provides a method for recommending meal packages, the method comprising:

[0006] Obtain a set of relevant historical data for the target user, which includes various types of user-related data.

[0007] The relevant historical data set is input into the feature extraction module of the package recommendation model to generate long sequence feature codes; the feature extraction module includes a multi-granularity scanning module of a deep forest neural network.

[0008] The long sequence feature encoding is input into the package matching module of the package recommendation model to obtain the package matching result between the package to be recommended and the target user; the package matching module includes a recurrent neural network.

[0009] Optionally, before inputting the relevant historical data set into the feature extraction module of the package recommendation model to generate long sequence feature codes, the method further includes:

[0010] Obtain training and test sample sets, construct the first loss function for the deep forest neural network, and the second loss function for the recurrent neural network; the training sample set includes multiple training-related data sets and multiple sample label information corresponding to the multiple training-related data sets, and the test sample set includes multiple test-related data sets;

[0011] The training sample set is used to train the package recommendation model, and the network parameters of the current deep forest neural network in the package recommendation model are updated using the first loss function to obtain the model after the deep forest parameters are updated.

[0012] The model with updated deep forest parameters is trained using the training sample set, and the network parameters of the current recurrent neural network in the model with updated deep forest parameters are updated using the second loss function to obtain the model with updated package matching parameters.

[0013] Using multiple test samples from the test sample set, the model after updating the package matching parameters is tested to obtain the training effect evaluation value;

[0014] If the training effect evaluation value meets the round-robin training conditions, the model with updated package matching parameters will be used as the recommended package model to be trained.

[0015] Return to the previous step and train the package recommendation model using the training sample set, update the network parameters of the current deep forest neural network in the package recommendation model, obtain the model after the deep forest parameters are updated, and continue until the training effect evaluation value no longer meets the training conditions, and obtain the package recommendation model.

[0016] Optionally, the steps of constructing the first loss function for the deep forest neural network and the second loss function for the recurrent neural network include:

[0017] Construct a second loss function based on the sample label information and the second prediction output of the recurrent neural network;

[0018] Based on the sample label information, the first prediction output of the deep forest neural network, and the second loss function, construct the first loss function.

[0019] Optionally, the step of training the package recommendation model using the training sample set includes:

[0020] Based on the first preset number of iterations, the training sample set is used to train the package recommendation model;

[0021] The steps for training the model after updating the parameters of the deep forest using the training sample set include:

[0022] Based on the second preset number of iterations, the model with updated deep forest parameters is trained using the training sample set.

[0023] Optionally, the training performance evaluation value includes the AUC value;

[0024] If the training performance evaluation value meets the round-robin training conditions, the model with updated parameters of the package matching module will be used as the package recommendation model to be trained, including:

[0025] ACU volatility is obtained based on the current AUC value and historical AUC values;

[0026] If the ACU volatility is greater than the preset volatility threshold, the model with updated package matching module parameters will be used as the package recommendation model to be trained.

[0027] Optionally, before the steps of obtaining the training sample set and the test sample set, the method further includes:

[0028] Preprocessing is performed on the initial training sample set and the initial test sample set to obtain the training sample set and the test sample set; the preprocessing includes at least one of the following: data format conversion, null value handling, outlier handling, and high single value handling.

[0029] Optionally, the types of user-related data are determined based on the type of the package to be recommended. User-related data includes call data, SMS data, data usage data, consumption data, and / or package usage data.

[0030] Secondly, the present invention also provides a package recommendation device, the device comprising:

[0031] The data acquisition module is used to acquire relevant historical data sets of the target user, which include various types of user-related data.

[0032] The feature encoding module is used to input relevant historical data sets into the feature extraction module of the package recommendation model to generate long sequence feature codes; the feature extraction module includes a multi-granularity scanning module of a deep forest neural network.

[0033] The recommendation module is used to input long sequence feature encoding into the package matching module of the package recommendation model to obtain the package matching result between the package to be recommended and the target user; the package matching module includes a recurrent neural network.

[0034] Thirdly, the present invention also provides a package recommendation device, the device comprising: a memory, a processor, and a package recommendation program stored in the memory and executable on the processor, the package recommendation program being configured to implement the steps of any of the package recommendation methods described above.

[0035] Fourthly, the present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the package recommendation method as described above.

[0036] This invention provides a method, apparatus, device, and storage medium for recommending service packages. The method involves acquiring a set of relevant historical data for a target user, including various types of user-related data; inputting the relevant historical data set into a feature extraction module of a service package recommendation model to generate a long-sequence feature code; the feature extraction module includes a multi-granularity scanning module of a deep forest neural network; inputting the long-sequence feature code into a service package matching module of the service package recommendation model to obtain the matching result between the recommended service package and the target user's service package; the service package matching module includes a recurrent neural network.

[0037] Therefore, this invention encodes various relevant historical data sets of the target user using a multi-granularity scanning module of a deep forest neural network to obtain long-sequence feature codes. A recurrent neural network then uses these long-sequence feature codes to obtain a matching result between the proposed service plan and the target user's service plan, and recommends the proposed service plan to the target user. The multi-granularity scanning module of the deep forest neural network automatically converts the noisy and diverse historical data of the target user into continuous and uniformly distributed long-sequence feature codes, making it more suitable for the input of the recurrent neural network. By integrating the multi-granularity scanning module of the deep forest neural network and the recurrent neural network for service plan recommendations, the matching result between the proposed service plan and the target customer's service plan output by the recurrent neural network based on the long-sequence feature codes becomes more accurate, thus improving the accuracy of service plan recommendations. Attached Figure Description

[0038] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0039] Figure 1 This is a schematic diagram of the structure of the recommended equipment in the package of this invention;

[0040] Figure 2 This is a flowchart illustrating the first embodiment of the package recommendation method of the present invention;

[0041] Figure 3 This is a flowchart illustrating the second embodiment of the package recommendation method of the present invention;

[0042] Figure 4 This is a technical architecture diagram of the fusion training of deep forest encoding and recurrent neural networks in this invention;

[0043] Figure 5 for Figure 3 Detailed flowchart of step S100;

[0044] Figure 6 The performance metrics for using the DF neural network alone for package recommendations;

[0045] Figure 7 The performance metrics for using RNNs alone to recommend service packages;

[0046] Figure 8 The performance metrics for recommending packages using the package recommendation model of this invention;

[0047] Figure 9 This is a schematic diagram of the modules of the first embodiment of the package recommendation device of the present invention.

[0048] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0049] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0050] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0051] In this invention, the terms "comprising," "including," or any other variations thereof are intended to cover a non-exclusive inclusion, such that an apparatus or system comprising a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such an apparatus or system. Without further limitation, an element defined by the phrase "comprising..." does not exclude the presence of other identical elements in the apparatus or system that includes that element.

[0052] Telecommunications operators possess powerful data collection, storage, and processing capabilities. Furthermore, their data is timely, diverse, accurate, and massive, enabling them to effectively construct user profiles from multiple perspectives, including basic identity information, call and SMS data, internet browsing behavior data, consumption behavior data, and location data.

[0053] Building marketing models for service packages based on data from telecommunications operators is currently an important business for these operators. Under the guidance of relevant laws and regulations, telecommunications operators can build marketing models for service packages based on their own user profile data and historical sample tag data, utilizing appropriate feature engineering and algorithmic model tools to achieve high-quality marketing of products such as data packages.

[0054] In terms of feature engineering, this mainly includes two aspects: business scenario data feature engineering and mathematically derived feature engineering. Business scenario feature engineering requires guidance from industry experts to redefine and derive from basic data. For example, based on users' periodic receipt of housing provident fund SMS messages, a stable employment label can be defined for users; based on users' group network information, business classification can be performed to define user industry levels, etc.

[0055] Mathematical feature engineering involves data derivation through analysis of historical data such as mean, variance, trend, and year-on-year / month-on-month comparisons. For example, it can obtain information on the stability of a user's app usage by analyzing data from the past three months.

[0056] The data modification, aggregation, and derivation involved in the two feature engineering processes mentioned above all require human intervention. Data processing is carried out based on the subjective experience of modelers or industry experts. Feature engineering without subjective participation often fails to achieve optimal feature processing.

[0057] Regarding algorithm models: For package marketing models, the mainstream algorithm models used in the industry include logistic regression, scorecards, traditional trees, gradient boosting trees, neural networks, etc. In practice, user feature profile data is correlated with historical sample label data to obtain a dataset for model training. After model training and parameter tuning are completed, predictions are made on the entire user base to realize the formulation of package marketing plans.

[0058] Recently, due to the powerful sequential feature processing capabilities of Recurrent Neural Networks (RNNs), an increasing number of operators are choosing RNNs as one of the preferred solutions for building their marketing packages. For example, Financial A uses RNNs to assess users' borrowing capacity, thereby determining marketing targets and loan amounts. They utilize customer transaction data from partner banks to create a time-series dataset for model building, achieving good overall results. Operator B has also repeatedly used RNN models in various marketing operations, refining user profile data to daily or weekly granularity, constructing time-series feature datasets for model training, and assisting in product marketing efforts, demonstrating positive results.

[0059] However, existing RNN models still have some shortcomings or technical problems in scenarios where data from telecommunications operators is used to build marketing packages, as follows:

[0060] 1. Regarding the input data of RNN models, there are two main problems: First, the feature metrics within a single cycle of the input samples are extremely noisy. Specifically, telecommunications operator data exhibits great diversity, encompassing many dimensions such as calls, SMS messages, behavior, consumption, internet access, and location. These metrics differ significantly in distribution characteristics, interpretability, data scale, and quantification range. Directly combining them as the model input features for a single cycle is inappropriate because, in its original design, RNNs aim for the content within each cycle of the serialized input data to be coherent, consistently distributed, and interpretable as a unified whole—a characteristic clearly demonstrated in natural language processing.

[0061] On the other hand, the available metrics from telecommunications operators are limited. As a deep neural network model, RNN requires a sufficient number of feature metrics to achieve better results. However, after basic data processing such as removing outliers, high null values, and high single values, the usable metrics retained from telecommunications operator data are usually less than one hundred. This number of metrics may be sufficient for traditional machine learning models, but it is barely enough for deep learning models.

[0062] 2. Regarding Feature Engineering: Current feature engineering methods involve data modification, aggregation, and derivation, all of which require human intervention. Data processing relies on the subjective experience of modelers or industry experts, which presents several problems. First, subjective feature engineering often fails to optimize feature processing. Second, the number of metrics derived by modelers or industry experts is often limited, hindering high scalability. Third, new metrics require further filtering for validity and label relevance, a complex and time-consuming process. Finally, existing feature engineering solutions and RNN models may not be compatible; if the metric noise issue mentioned in point 1 persists, a better feature engineering solution should be sought.

[0063] In summary, there are still many problems in the construction of existing package marketing models in the current telecommunications operation scenarios, resulting in low accuracy of the package marketing models.

[0064] In view of the technical problem that existing package recommendation methods have low accuracy, the present invention provides a package recommendation method, the overall idea of ​​which is as follows:

[0065] The method includes: acquiring a set of relevant historical data for the target user, which includes various types of user-related data; inputting the relevant historical data set into the feature extraction module of the package recommendation model to generate long-sequence feature codes; the feature extraction module includes a multi-granularity scanning module of a deep forest neural network; inputting the long-sequence feature codes into the package matching module of the package recommendation model to obtain the package matching results between the package to be recommended and the target user; the package matching module includes a recurrent neural network.

[0066] This invention provides a package recommendation method. It encodes multiple relevant historical data sets of a target user using a multi-granularity scanning module of a deep forest neural network to obtain long-sequence feature codes. A recurrent neural network then uses these long-sequence feature codes to obtain a matching result between the proposed package and the target user's package, and recommends the proposed package to the target user. The multi-granularity scanning module of the deep forest neural network automatically converts the noisy and diverse historical data of the target user into continuous and uniformly distributed long-sequence feature codes, making it more suitable for the input of the recurrent neural network. By integrating the multi-granularity scanning module of the deep forest neural network and the recurrent neural network for package recommendation, the matching result between the proposed package and the target customer's package is more accurate, thus improving the accuracy of package recommendations.

[0067] The following provides a detailed description of the package recommendation method, apparatus, equipment, and storage medium used in the technical implementation of this invention:

[0068] Reference Figure 1 , Figure 1 This is a schematic diagram of the recommended hardware operating environment package involved in the embodiments of this application.

[0069] like Figure 1 As shown, the device may include: a processor 1001, such as a CPU; a user interface 1003; a memory 1005; and a communication bus 1002. The communication bus 1002 is used to enable communication between these components. The user interface 1003 may include user devices such as smartphones, tablets, and PDAs (Personal Digital Assistants). Optionally, the user interface 1003 may also be a display screen or an input unit such as a keyboard. The memory 1005 may be high-speed RAM or non-volatile memory, such as disk storage. Optionally, the memory 1005 may also be a storage device independent of the aforementioned processor 1001.

[0070] It is understood that the device may also include a network interface 1004, which may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface). Optionally, the device may also include RF (Radio Frequency) circuitry, sensors, audio circuitry, a Wi-Fi module, etc.

[0071] Those skilled in the art will understand that Figure 1 The device structure shown does not constitute a limitation on the device and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0072] The method, apparatus, device, and storage medium for recommending packages according to the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0073] Based on, but not limited to, the above hardware structure, refer to Figure 2 , Figure 2 This is a flowchart illustrating the first embodiment of the package recommendation method of the present invention.

[0074] This embodiment provides a method for recommending service packages, which may include:

[0075] Step S700: Obtain the relevant historical data set of the target user, which includes various types of user-related data.

[0076] In this embodiment, the executing entity can be, for example, Figure 1 The recommended equipment shown in the package can be a physical server consisting of a standalone host, or a virtual server hosted in a host cluster.

[0077] The target user can be any user of a telecommunications operator. The relevant historical data set can include various user-related data within a preset number of time periods (T). The time periods can include the current time period and the historical time periods preceding the current time period (T-1). The user-related data can be various telecommunications operation data. The preset number T can be set according to actual needs, and the time periods can be one month, one week, or one day, or selected based on actual requirements. For example, if the preset number is 10 and the time period is one month, the relevant historical data set can include the target user's telecommunications operation data for the current month and the telecommunications operation data for the nine months preceding the current month.

[0078] Specifically, as one implementation method, the types of user-related data are determined according to the type of the package to be recommended. User-related data may include call data, SMS data, data usage data, consumption data and / or package usage data.

[0079] In this embodiment, the recommended packages may include data packages, voice packages, and basic consumption packages. The data types can be set according to different package products. For example, for data packages, a large amount of data related to data usage, package details, and consumption can be used. Specifically, call data can be the duration of outgoing calls within a single time period; SMS data can be the number of SMS messages within a single time period; data usage data can be the data usage and number of times the app is used within a single time period; consumption data can include call charges, data usage, and out-of-package consumption within a single time period; and package usage data can include the package usage within a single time period.

[0080] Step S800: Input the relevant historical data set into the feature extraction module of the package recommendation model to generate long sequence feature codes; the feature extraction module includes a multi-granularity scanning module of a deep forest neural network.

[0081] In this embodiment, each layer of the Deep Forest (DF) neural network uses multiple simple tree models as evaluators to automatically perform tasks such as grouping, cross-referencing, aggregation, and higher-order derivation of communication operation data, thereby mining broader and deeper information from the communication operation data. Furthermore, the output encoding of each layer is strongly correlated with the training samples of the DF neural network. Thus, the DF neural network can transform the representation of communication operation data, converting the original noisy communication operation data into continuous, uniformly distributed long codes, mapping the feature information of scattered and independent communication operation data into high-dimensional, long-sequence codes, and simultaneously expanding the data volume of communication operation data. It can be understood that the long-sequence feature encoding is obtained by sequentially inputting user-related data from relevant historical datasets into the multi-granularity scanning module of the Deep Forest neural network and then arranging them. The encoding length of the long-sequence feature encoding is preset according to actual usage requirements. Continuing with the example of a preset quantity of 10 and a time period of one month, if the number of data in a single time period of the relevant historical data set is 100, then the tensor of the relevant historical data set is 10×100. If the encoding length is 500, then after encoding by the multi-granularity scanning module, the number of data in a single time period can be expanded from 100 to 500, which greatly enriches the input data of the RNN.

[0082] Step S900: Input the long sequence feature encoding into the package matching module of the package recommendation model to obtain the package matching result between the package to be recommended and the target user; the package matching module includes a recurrent neural network.

[0083] In this embodiment, the Recurrent Neural Network (RNN) can output a matching score between the recommended service package and the target user based on long sequence feature encoding. A higher matching score indicates a higher value in recommending the service package to the target user, allowing for the recommendation of a more expensive service package to the target user. The matching score can be in the range [0, 1]. For example, a matching score of 0.88 for the target user can be considered high, indicating that the target user is a potential customer for the recommended service package. Alternatively, the recommendation of the service package can be determined based on whether the matching score exceeds a preset recommendation score threshold.

[0084] This embodiment provides a package recommendation method. Through a multi-granularity scanning module of a deep forest neural network, it automatically converts the noisy and diverse historical data of the target user into continuous and consistently distributed long-sequence feature codes, making it more suitable for input to recurrent neural networks. By integrating the multi-granularity scanning module of the deep forest neural network and the recurrent neural network for package recommendation, the recurrent neural network, based on the long-sequence feature codes, outputs a more accurate matching result between the recommended package and the target customer's package, thus improving the accuracy of package recommendations. Furthermore, it eliminates the need for human intervention in feature engineering, avoiding the subjective influence of human intervention on the optimization of feature processing and improving the efficiency of feature engineering.

[0085] Furthermore, such as Figures 3 to 8 As shown, Figure 3 This is a flowchart illustrating the second embodiment of the package recommendation method of the present invention. Figure 4 This is a technical architecture diagram of the fusion training of deep forest encoding and recurrent neural networks in this invention. Figure 5 for Figure 3 Detailed flowchart of step S100, Figure 6 The metrics for recommending meal packages using the DF neural network alone. Figure 7 This refers to the performance metrics for using RNNs alone to recommend meal plans. Figure 8 The metrics used to recommend packages using the package recommendation model of this invention are as follows.

[0086] like Figure 3 As shown, based on the above embodiment one, before step S100, the method may further include:

[0087] Step S100: Obtain the training sample set and the test sample set, construct the first loss function of the deep forest neural network, and the second loss function of the recurrent neural network.

[0088] The training sample set includes multiple training-related data sets and multiple sample label information corresponding to the multiple training-related data sets, while the test sample set includes multiple test-related data sets.

[0089] In this embodiment, multiple training-related datasets and multiple test-related datasets can both include multiple users' relevant historical datasets. The sample label information is determined based on the results of multiple users purchasing or upgrading the recommended package. The sample label information can be set as positive samples for users who purchased or upgraded the recommended package and negative samples for users who did not purchase or upgrade the recommended package.

[0090] like Figure 4 As shown, a training sample set of size D can be represented as For the i-th training sample, y iFor the sample label information corresponding to the i-th training sample, a single training sample It can be represented as:

[0091]

[0092] Where T represents the number of time periods, i.e., the preset quantity. This refers to user-related data within the current time period. This refers to user-related data within the (T-1)th historical time period prior to the current time period. User-related data within a single time period can be represented as:

[0093]

[0094] For telecommunications operation data, This can be visualized as shown in the table below:

[0095]

[0096] Specifically, such as Figure 5 As shown, in one specific implementation, step S100 may include:

[0097] Step S110: Construct a second loss function based on the sample label information and the second prediction output of the recurrent neural network.

[0098] In this embodiment, as Figure 4 As shown, during the RNN training phase, the long sequence encoding feature sequence of the samples output by the mGS (Multi-Grained Scanning) module of the DF neural network is input. i This allows us to obtain the second predicted output of the RNN. The cross-entropy function can be used to determine the sample label information y. i Second prediction output Obtain the second loss function of the recurrent neural network.

[0099]

[0100] Step S120: Construct the first loss function based on the sample label information, the first prediction output of the deep forest neural network, and the second loss function.

[0101] In this embodiment, during the DF neural network training phase, the mGS (Multi-Grained Scanning) module outputs a long sequence encoding of the feature sequence. i It is to use a single training sample User-related data within T time periods The sample long sequence encoding feature sequence is obtained by sequentially inputting the mGS module of the DF neural network and then arranging it. i It can be represented as:

[0102]

[0103] In the Cascade Forest Structure (CF) module of a DF neural network, only the communication operation data within a single time period is typically considered to be the most representative. Furthermore, since the communication operation data from the most recent time period most closely reflects user behavior, the mGS module output is selected during the DF neural network training phase. As input to the CF module, the first predicted output of the DF neural network is obtained. Using the mean squared error function, the loss function of the DF neural network itself is obtained as follows:

[0104]

[0105] Furthermore, in order to enable the long sequence feature encoding generated by the mGS module to better serve the deep feature processing of the RNN, a second loss function of the RNN can be introduced during the training stage of the DF neural network to construct the first loss function of the DF neural network. This will create a correlation between the training process of the mGS module and the RNN in the high-dimensional and complex feature space, making the feature encoding process of the mGS module more targeted.

[0106] Therefore, based on the sample label information y i The first prediction output of the DF neural network Second loss function The first loss function for constructing a DF neural network for:

[0107]

[0108] Here, α and β are both loss weights, which can be set according to actual needs. Preferably, both α and β are 0.5.

[0109] Step S200: Train the package recommendation model to be trained using the training sample set, update the network parameters of the current deep forest neural network in the package recommendation model to be trained using the first loss function, and obtain the model after the deep forest parameters are updated.

[0110] Step S300: Train the deep forest model after parameter update using the training sample set, and update the network parameters of the current recurrent neural network in the deep forest model after parameter update using the second loss function to obtain the model after parameter update of package matching.

[0111] In this embodiment, the deep forest neural network and recurrent neural network in the package recommendation model are trained using a multi-round training method. Loss functions need to be constructed separately for the deep forest neural network and the recurrent neural network to update their model parameters. In a single training round, the RNN is first fixed, and the deep forest neural network (DF) is trained using the training sample set. The network parameters of the mGS and CF modules of the DF neural network are updated using the first loss function. Then, the DF neural network is fixed, and the RNN is trained using the training sample set. The network parameters of the RNN are updated using the second loss function.

[0112] Specifically, as another specific implementation, step S200 may include: training the training sample set for the package recommendation model according to the first preset number of iterations.

[0113] Step S300 may include: training the model with updated deep forest parameters using the training sample set according to the second preset number of iterations.

[0114] In this embodiment, during a single training iteration, the number of iterations for the DF neural network can be preset to a first preset number of iterations, and the number of iterations for the RNN can be preset to a second preset number of iterations. The first and second preset number of iterations can be preset according to specific usage conditions.

[0115] Step S400: Using multiple test samples from the test sample set, test the model after updating the package matching parameters to obtain the training effect evaluation value.

[0116] Step S500: If the training effect evaluation value meets the round-robin training conditions, then the model with updated package matching parameters is used as the recommended model for the package to be trained.

[0117] Step S600: Return to step S200 until the training effect evaluation value no longer meets the round-robin training conditions, and obtain the package recommendation model.

[0118] In this embodiment, the sample size is D. test The test sample set can be represented as Let j be the j-th test sample. The types and number of time periods of user-related data in the test sample are the same as those in the training sample. After each round of training, the model is updated with the updated package matching parameters and used for test prediction. This yields multiple test results for multiple test samples, resulting in a training performance evaluation value to assess the training results of the model after the package matching parameters are updated. The round-robin conditions can be set according to the type of training performance evaluation value.

[0119] Specifically, as another specific implementation method, the training effect evaluation value may include the AUC value.

[0120] Step S600 may include: obtaining ACU volatility based on the current AUC value and historical AUC values; if the ACU volatility is greater than a preset volatility threshold, then using the updated model of the package matching module as the package recommendation model to be trained.

[0121] In this embodiment, the AUC value is the area under the ROC curve (Receiver Operating Characteristic Curve), which can be used to compare the performance of the model after different training rounds and parameter updates for the package matching module. The current AUC value is the AUC value obtained by the model after the current training round and parameter updates for the package matching module when predicting multiple test samples in the test sample set; the historical AUC value is the AUC value obtained by the model after the previous training round and parameter updates for the package matching module when predicting multiple test samples in the test sample set. ACU volatility is the rate of change of the AUC value. The preset volatility threshold is set according to actual needs; for example, the preset volatility threshold can be 0.1%.

[0122] like Figures 6 to 8 As shown, the advantages and disadvantages of various package recommendation models are compared using three indicators: ks value, auc value, and PR curve. Figure 6 The KS value was 0.3008, and the AUC value was 0.7027. Figure 7 The KS value was 0.3343, and the AUC value was 0.7202. Figure 8 The KS value is 0.3933, and the AUC value is 0.7635. It can be seen that the package recommendation model in this embodiment integrates the mGS module of the DF neural network and the RNN for package recommendation. Compared with using the DF neural network or the RNN alone, the package recommendation model in this embodiment improves the KS value by 0.059, the AUC value by 0.0433, and the PR curve has a larger coverage area. Therefore, the package recommendation model in this embodiment performs better than using the DF neural network or the RNN alone for package recommendation.

[0123] Furthermore, as another implementation, before step S100, the method may further include: preprocessing the initial training sample set and the initial test sample set to obtain the training sample set and the test sample set; the preprocessing includes at least one of data format conversion processing, null value processing, outlier processing and high single value processing.

[0124] In this embodiment, the training and test sample sets need to be preprocessed before model training. The training sample set undergoes format standardization and data cleaning, as does the test sample set. The specific processing methods for data format conversion, null value handling, outlier handling, and handling of high single-value errors can be selected according to actual usage requirements.

[0125] This embodiment provides a package recommendation method. Before recommending a package to the target user, a round-robin training scheme is used to train the feature extraction module and the package matching module of the package recommendation model respectively. When training the feature extraction module, the loss function of the package matching module is considered, so that the feature extraction module and the package matching module are correlated in the high-dimensional and complex feature space. The long sequence feature code generated by the feature extraction module is more consistent with the deep feature processing of the package matching module, thereby improving the accuracy and recommendation efficiency of the package recommendation model.

[0126] Based on the same inventive concept, referring to Figure 9 , Figure 9 This is a schematic diagram of a module of a first embodiment of the package recommendation device of the present invention. This embodiment provides a package recommendation device, which may include:

[0127] Data acquisition module 10 is used to acquire relevant historical data sets of the target user, including various types of user-related data;

[0128] Feature encoding module 20 is used to input relevant historical data sets into the feature extraction module of the package recommendation model to generate long sequence feature codes; the feature extraction module includes a multi-granularity scanning module of a deep forest neural network;

[0129] The recommendation module 30 is used to input long sequence feature encoding into the package matching module of the package recommendation model to obtain the package matching result between the package to be recommended and the target user; the package matching module includes a recurrent neural network.

[0130] Furthermore, as one embodiment, the device may further include:

[0131] The construction module is used to obtain the training sample set and the test sample set, construct the first loss function of the deep forest neural network, and the second loss function of the recurrent neural network; wherein, the training sample set includes multiple training-related data sets and multiple sample label information corresponding to the multiple training-related data sets, and the test sample set includes multiple test-related data sets;

[0132] The deep forest training module is used to train the package recommendation model under training using the training sample set, and to update the network parameters of the current deep forest neural network in the package recommendation model under training using the first loss function, so as to obtain the model after the deep forest parameters are updated.

[0133] The recurrent network training module is used to train the model after updating the parameters of the deep forest using the training sample set, and to update the network parameters of the current recurrent neural network in the model after updating the parameters of the deep forest using the second loss function, so as to obtain the model after updating the parameters of the package matching.

[0134] The model testing module is used to test the model after updating the package matching parameters using multiple test samples in the test sample set, and obtain the training effect evaluation value.

[0135] The model evaluation module is used to use the model with updated package matching parameters as the package recommendation model to be trained if the training effect evaluation value meets the round-robin training conditions. It then returns to execute the steps of training the package recommendation model to be trained using the training sample set, updating the network parameters of the current deep forest neural network in the package recommendation model to be trained, and obtaining the model with updated deep forest parameters, until the training effect evaluation value no longer meets the round-robin training conditions, and finally obtaining the package recommendation model.

[0136] Specifically, the building blocks can include:

[0137] The first construction unit is used to construct the second loss function based on the sample label information and the second prediction output of the recurrent neural network;

[0138] The second construction unit is used to construct the first loss function based on the sample label information, the first prediction output of the deep forest neural network, and the second loss function.

[0139] Specifically, the deep forest training module is also used to train the training sample set of the package recommendation model to be trained according to the first preset number of iterations.

[0140] The recurrent network training module is also used to train the model with updated deep forest parameters using the training sample set according to the second preset number of iterations.

[0141] Specifically, the evaluation values ​​for training effectiveness include the AUC value;

[0142] The model evaluation module is also used to obtain ACU volatility based on the current AUC value and historical AUC values; if the ACU volatility is greater than the preset volatility threshold, the model with updated parameters from the package matching module will be used as the recommended package model to be trained.

[0143] Furthermore, as another embodiment, the device may further include:

[0144] The data processing module is used to preprocess the initial training sample set and the initial test sample set to obtain the training sample set and the test sample set. The preprocessing includes at least one of the following: data format conversion processing, null value processing, outlier value processing, and high single value processing.

[0145] For more details on the specific implementation of the above-mentioned package recommendation device, please refer to the description of the specific implementation of the package recommendation method in Embodiment 1 or 2 above. For the sake of brevity, these details will not be repeated here.

[0146] Furthermore, embodiments of the present invention also propose a computer storage medium storing a computer program, which, when executed by a processor, implements the steps of the package recommendation method described above. Therefore, further details will not be repeated here. Additionally, the beneficial effects of using the same method will also not be repeated. For technical details not disclosed in the embodiments of the computer-readable storage medium involved in this application, please refer to the description of the method embodiments of this application. As an example, program instructions can be deployed to execute on a single computing device, or on multiple computing devices located in one location, or on multiple computing devices distributed across multiple locations and interconnected via a communication network.

[0147] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A method for recommending meal packages, characterized in that, The method includes: Obtain a set of relevant historical data for the target user, wherein the set of relevant historical data includes various types of user-related data; The relevant historical data set is input into the feature extraction module of the package recommendation model to generate long sequence feature codes; the feature extraction module includes a multi-granularity scanning module of a deep forest neural network. The long sequence feature encoding is input into the package matching module of the package recommendation model to obtain the package matching result between the package to be recommended and the target user's package; the package matching module includes a recurrent neural network. Before inputting the relevant historical data set into the feature extraction module of the package recommendation model to generate long sequence feature codes, the method further includes: Obtain a training sample set and a test sample set, construct a first loss function for the deep forest neural network, and a second loss function for the recurrent neural network; the training sample set includes multiple training-related data sets and multiple sample label information corresponding to the multiple training-related data sets, and the test sample set includes multiple test-related data sets; The training sample set is used to train the package recommendation model to be trained, and the network parameters of the current deep forest neural network in the package recommendation model to be trained are updated using the first loss function to obtain the model after the deep forest parameters are updated. The deep forest parameter-updated model is trained using the training sample set, and the network parameters of the current recurrent neural network in the deep forest parameter-updated model are updated using the second loss function to obtain the package matching parameter-updated model. Using multiple test samples from the test sample set, the model after updating the package matching parameters is tested to obtain a training effect evaluation value; If the training effect evaluation value meets the round-robin training condition, then the model with the updated package matching parameters is used as the recommended model for the package to be trained. Return to the previous step of training the package recommendation model using the training sample set, updating the network parameters of the current deep forest neural network in the package recommendation model, and obtaining the model after the deep forest parameters are updated, until the training effect evaluation value no longer meets the round-robin training condition, and obtain the package recommendation model.

2. The method as described in claim 1, characterized in that, The steps of constructing the first loss function of the deep forest neural network and the second loss function of the recurrent neural network include: The second loss function is constructed based on the sample label information and the second prediction output of the recurrent neural network. The first loss function is constructed based on the sample label information, the first prediction output of the deep forest neural network, and the second loss function.

3. The method as described in claim 1, characterized in that, The step of training the package recommendation model using the training sample set includes: The training sample set is used to train the package recommendation model according to the first preset number of iterations. The step of training the model after updating the deep forest parameters using the training sample set includes: The deep forest model with updated parameters is trained using the training sample set according to the second preset number of iterations.

4. The method as described in claim 1, characterized in that, The training performance evaluation value includes the AUC value; If the training effect evaluation value meets the round-robin training condition, then the model with updated parameters of the package matching module is used as the package recommendation model to be trained, including: ACU volatility is obtained based on the current AUC value and historical AUC values; If the ACU volatility is greater than a preset volatility threshold, the model with updated parameters of the package matching module will be used as the recommended model for the package to be trained.

5. The method as described in claim 1, characterized in that, Before the step of obtaining the training sample set and the test sample set, the method further includes: The initial training sample set and the initial test sample set are preprocessed to obtain the training sample set and the test sample set; the preprocessing includes at least one of data format conversion processing, null value processing, outlier value processing and high single value processing.

6. The method as described in claim 1, characterized in that, The types of user-related data are determined based on the type of the package to be recommended. The user-related data includes call data, SMS data, data usage data, consumption data, and / or package usage data.

7. A package recommendation device, characterized in that, The device includes: The data acquisition module is used to acquire a set of relevant historical data of the target user, wherein the set of relevant historical data includes various types of user-related data; The feature encoding module is used to input the relevant historical data set into the feature extraction module of the package recommendation model to generate long sequence feature codes; the feature extraction module includes a multi-granularity scanning module of a deep forest neural network. The recommendation module is used to input the long sequence feature encoding into the package matching module of the package recommendation model to obtain the package matching result between the package to be recommended and the target user; the package matching module includes a recurrent neural network; The device further includes: The construction module is used to obtain the training sample set and the test sample set, construct the first loss function of the deep forest neural network, and the second loss function of the recurrent neural network; wherein, the training sample set includes multiple training-related data sets and multiple sample label information corresponding to the multiple training-related data sets, and the test sample set includes multiple test-related data sets; The deep forest training module is used to train the package recommendation model under training using the training sample set, and to update the network parameters of the current deep forest neural network in the package recommendation model under training using the first loss function, so as to obtain the model after the deep forest parameters are updated. The recurrent network training module is used to train the model after updating the parameters of the deep forest using the training sample set, and to update the network parameters of the current recurrent neural network in the model after updating the parameters of the deep forest using the second loss function, so as to obtain the model after updating the parameters of the package matching. The model testing module is used to test the model after updating the package matching parameters using multiple test samples in the test sample set, and obtain the training effect evaluation value. The model evaluation module is used to use the model with updated package matching parameters as the package recommendation model to be trained if the training effect evaluation value meets the round-robin training conditions. It then returns to execute the steps of training the package recommendation model to be trained using the training sample set, updating the network parameters of the current deep forest neural network in the package recommendation model to be trained, and obtaining the model with updated deep forest parameters, until the training effect evaluation value no longer meets the round-robin training conditions, and finally obtaining the package recommendation model.

8. A package recommendation device, characterized in that, The device includes: a memory, a processor, and a package recommendation program stored in the memory and executable on the processor, configured to implement the steps of the package recommendation method as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the steps of the package recommendation method as described in any one of claims 1 to 6.

Images