Short message recommendation method and device, equipment and storage medium

Abstract

The invention discloses a short message recommendation method and device, equipment and a storage medium, and relates to the technical field of big data, and the method comprises the steps: carrying out the feature extraction of multi-source behavior data of a user, and obtaining a document preference dimension; generating candidate short message copywriting matched with the big language model and the cue word template through the big language model and the cue word template; and performing multi-target multi-classification prediction on the user, and performing association processing on the candidate short message copywriting through a double-flow recommendation model according to the obtained customer group segmentation result and the multi-source behavior data to obtain a short message copywriting recommendation result. According to the method, the candidate short message copywriting matched with the copywriting preference dimension of the user is automatically generated through the large language model, and then the multi-target multi-classification prediction is performed on the user to obtain the customer group segmentation result with different emphasis on multiple targets, so that the customer group segmentation result with different emphasis on multiple targets can be obtained even in various user data. And short message copywriting recommendation results matched with user behaviors can be accurately generated, the diversity of customer groups can be accurately considered, and the marketing conversion rate of short messages is effectively improved.

Description

Technical Field

[0001] This application relates to the field of big data technology, and in particular to a method, apparatus, device and storage medium for SMS recommendation. Background Technology

[0002] The current method of creating and sending marketing SMS messages typically involves operations staff manually crafting messages based on simple information such as product selling points and current events. After internal review, the operations staff configures the SMS template. Once the template is configured, the message is sent to selected users based on simple customer tags (such as age). The process is repeated for each new user group.

[0003] However, in the above process, the creation and recommendation of marketing SMS copy basically rely on the experience of the operations staff. Due to the large amount of user data, it is difficult to accurately adapt SMS recommendations to the diversity of the customer base. Summary of the Invention

[0004] The main purpose of this application is to provide a method, apparatus, device, and storage medium for recommending SMS messages, which aims to solve the technical problem that traditional marketing SMS copy recommendation methods rely mainly on the experience of operators to make selections, and are difficult to accurately account for the diversity of customer groups due to the large amount of user data.

[0005] To achieve the above objectives, this application proposes a text message recommendation method, the method comprising: Feature extraction is performed on the user's multi-source behavioral data to obtain the user's copywriting preference dimension; By using a large language model and preset prompt word templates, candidate SMS texts that match the aforementioned text preference dimensions are generated. Perform multi-objective multi-classification prediction on the users to obtain customer segmentation results; Based on the customer segmentation results and the multi-source behavioral data, the candidate SMS messages are correlated using a preset dual-stream recommendation model to obtain SMS message recommendation results.

[0006] In one embodiment, the step of associating the candidate SMS messages with the customer segmentation results and the multi-source behavioral data using a preset dual-stream recommendation model to obtain SMS message recommendation results includes: Based on the multi-source behavioral data, the candidate SMS messages corresponding to each customer group in the customer segmentation results are associated to obtain a serialized recommendation list; The single-layer correlation and deep multi-dimensional correlation between customer behavior and SMS copy are determined by a pre-set dual-stream recommendation model to obtain a dual-stream recommendation list. The serialized recommendation list and the dual-stream recommendation list are weighted and fused to obtain the SMS text recommendation result.

[0007] In one embodiment, the step of associating candidate SMS messages for each customer group in the customer segmentation results based on the multi-source behavioral data to obtain a serialized recommendation list includes: Determine the customer group categories in the customer segmentation results; Based on the multi-source behavioral data, the candidate SMS messages corresponding to each customer group category are vectorized to obtain SMS template vectors; Determine the correlation matrix between pairwise SMS template vectors; Based on the attention score of the relevance matrix, a serialized recommendation list of candidate SMS messages is determined.

[0008] In one embodiment, the step of extracting features from the user's multi-source behavioral data to obtain the user's copywriting preference dimension includes: The user's multi-source behavioral data is preprocessed to obtain standardized structured data. The preprocessing includes missing value imputation and outlier replacement. The standardized structured data is converted into a long text feature sequence; The user attribute prediction of the long text feature sequence is performed by a preset dual-weight factor extraction network to determine user activity and funding needs. The user activity level and the funding requirement level are used as dimensions of the user's copywriting preference.

[0009] In one embodiment, the dual-weight factor extraction network is constructed from a deep network model. The dual-weight factor extraction network includes a feature extraction layer, a hidden layer, and a softmax output layer. The hidden layer is connected to the feature extraction layer and the softmax output layer, respectively. The dual-weight factor extraction network is used to predict user activity and funding demand based on the long text feature sequence.

[0010] In one embodiment, the step of generating candidate SMS text messages that match the text preference dimension using a large language model and preset prompt word templates includes: The long text feature sequence is filled into a preset prompt word template, which includes at least one of the following: background introduction section, core word limitation section, style prompt section, customer characteristics section, and positive and negative case section. The filled prompt word template is input into the large language model to obtain multiple candidate SMS text messages that match the text preference dimension.

[0011] In one embodiment, the step of performing multi-objective multi-classification prediction on the user to obtain customer segmentation results includes: The multi-source behavioral data is input into a preset multi-objective multi-classification model for classification prediction to obtain customer segmentation results. The multi-objective multi-classification model is composed of the dual-weight factor extraction network and the machine learning model. The multi-objective multi-classification model is used to identify scene words in the multi-source behavioral data to obtain classification results.

[0012] Furthermore, to achieve the above objectives, this application also proposes a text message recommendation device, the device comprising: The feature extraction module is used to extract features from the user's multi-source behavioral data to obtain the user's copywriting preference dimension; The SMS generation module is used to generate candidate SMS texts that match the text preference dimension using a large language model and preset prompt word templates. The customer segmentation module is used to perform multi-objective multi-classification prediction on the users to obtain customer segmentation results. The SMS recommendation module is used to perform correlation processing on the candidate SMS texts based on the customer segmentation results and the multi-source behavioral data, and obtain SMS text recommendation results through a preset dual-stream recommendation model.

[0013] In addition, to achieve the above objectives, this application also proposes a text message recommendation device, the device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the text message recommendation method as described above.

[0014] In addition, to achieve the above objectives, this application also proposes a storage medium, which is a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the steps of the SMS recommendation method described above.

[0015] One or more technical solutions proposed in this application have at least the following technical effects: The SMS recommendation method of this application includes: extracting features from the user's multi-source behavioral data to obtain the user's text message preference dimension; generating candidate SMS messages that match the text message preference dimension through a large language model and a preset prompt word template; performing multi-objective multi-classification prediction on the user to obtain customer segmentation results; and performing association processing on the candidate SMS messages based on the customer segmentation results and the multi-source behavioral data through a preset dual-stream recommendation model to obtain SMS message recommendation results.

[0016] Because this application targets users' copywriting preferences, it can automatically generate matching candidate SMS copywriting through a large language model. Then, by performing multi-objective, multi-classification predictions on users, it can obtain customer segmentation results with different emphases on multiple objectives. Thus, even with vast amounts of user data, a dual-stream recommendation model can accurately generate SMS copywriting recommendations that match user behavior, accurately accommodating customer diversity and effectively improving SMS marketing conversion rates. Attached Figure Description

[0017] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating the SMS recommendation method in an embodiment of this application; Figure 2 This is a specific implementation diagram of SMS text generation in the embodiments of this application; Figure 3 This is a specific implementation diagram of customer group segmentation in the embodiments of this application; Figure 4 The following are specific implementation diagrams recommended in the embodiments of this application; Figure 5 This is a schematic diagram of the module structure of the SMS recommendation device according to an embodiment of this application; Figure 6 This is a schematic diagram of the device structure of the hardware operating environment involved in the SMS recommendation method in this application embodiment.

[0020] The realization of the purpose, functional features and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0021] To make the above-mentioned objectives, features, and advantages of this application more apparent and understandable, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0022] It should be noted that the executing entity of this application embodiment can be a computing service device with features extraction, SMS copy generation, customer segmentation, and SMS recommendation functions, such as a personal computer, a server, etc., or an electronic device capable of implementing the above functions, an SMS recommendation device executing the SMS recommendation method of this application, etc. This embodiment does not limit this. The following uses an SMS recommendation device as an example to describe this embodiment and the following embodiments.

[0023] This application provides a method for recommending SMS messages, as described in the embodiments below. Figure 1 , Figure 1 This is a flowchart illustrating the SMS recommendation method in an embodiment of this application. In this embodiment, the SMS recommendation method may include the following steps: Step S10: Extract features from the user's multi-source behavioral data to obtain the user's copywriting preference dimension.

[0024] Among them, multi-source behavioral data is user behavior information collected from multiple data sources, including basic customer behavioral characteristics (such as age, province, marital status, education, etc.), People's Bank of China behavioral data (credit loan information, credit card usage, job change information, housing provident fund payment information, etc.), Sensors Data behavioral data (customer login to the APP, browsing pages, registration for activities, etc.), marketing outreach data (historical SMS marketing records, telephone marketing records, targeted coupon marketing records, etc.), historical response records, and other data sources (product usage records, external third-party credit reports, etc.). This embodiment does not impose any restrictions on this.

[0025] It should be noted that the user-related data involved in this application (such as the aforementioned basic customer behavior data, pedestrian behavior data, and Sensors data) were all obtained with the user's permission or consent; that is to say, when this application is used in specific products or technologies, user permission is required to obtain and process the relevant data, and the processing of the relevant data must comply with the relevant laws, regulations, and regulatory standards of the relevant countries and regions.

[0026] As for the copywriting preference dimension, it refers to the user's preference characteristics in SMS copywriting content identified by the SMS recommendation device based on the user's multi-source behavioral data, such as sensitivity to interest rates, preference for holiday greetings, different activity levels, and financial needs.

[0027] Specifically, after obtaining multi-source user behavior data, SMS recommendation devices can judge and extract some user behavioral attributes based on this data. For example, activity level and financial needs are two major concerns for operations. People with different activity levels and financial needs have different preferences for copywriting. Therefore, a preliminary judgment can be made on user activity level and user financial needs to determine the user's copywriting preference dimensions.

[0028] Step S20: Generate candidate SMS texts that match the text preference dimension using a large language model and preset prompt word templates.

[0029] Among them, the large language model is a pre-trained language model (such as Qwen3) with powerful natural language understanding and generation capabilities. It can generate text content that is context-appropriate and stylistically diverse based on the input prompts. The prompt template guides the large language model to generate SMS copy that meets business needs. The pre-designed structured input template can include background information, style prompts, customer characteristics, etc., to ensure that the generated content matches user preferences and business scenarios.

[0030] In this embodiment, after determining the above-mentioned copywriting preference dimensions, the large language model can automatically generate a set of personalized SMS copywriting candidates that conform to the user's copywriting preference dimensions based on the preset prompt word templates, for subsequent recommendation model screening and sorting.

[0031] Step S30: Perform multi-objective multi-classification prediction on the user to obtain customer segmentation results.

[0032] It should be noted that the customer segmentation results further divide the original coarse-grained user group into multiple smaller groups with similar behaviors and attributes (such as highly active low-demand customers, low-active high-debt customers, etc.), providing a foundation for precision marketing.

[0033] Specifically, after obtaining the aforementioned multi-source behavioral data, the SMS recommendation device can simultaneously classify and predict users across multiple dimensions (such as activity level, funding needs, willingness to participate, debt level, etc.), and output multiple tags (such as high activity, low demand, married, undergraduate, etc.) to obtain customer segmentation results, thereby achieving a refined user profile.

[0034] Step S40: Based on the customer segmentation results and the multi-source behavioral data, the candidate SMS messages are correlated using a preset dual-stream recommendation model to obtain SMS message recommendation results.

[0035] The SMS message recommendation results are output by the dual-stream recommendation model, representing the most likely SMS messages to elicit a response from the user. In this process, the dual-stream recommendation model matches candidate SMS messages with the user's behavioral characteristics, customer tags, and historical response records, evaluating the suitability of each message for the user, and thus selecting the optimal recommendation.

[0036] Furthermore, in the dual-stream recommendation model proposed in this embodiment, its output can integrate two recommendation strategies. For example, the model can predict the user's preference order for different text messages based on the user's historical SMS response behavior to obtain a sequential recommendation; or it can perform deep cross-calculation based on user behavior variables (such as login frequency, withdrawal records, etc.) to recommend the most matching text message to obtain a behavioral feature recommendation. Either recommendation strategy can be selected, or a combination of both can be determined to obtain the final SMS text message recommendation result.

[0037] This application provides an SMS recommendation method. The method includes extracting features from multi-source behavioral data of users to obtain the user's copywriting preference dimension; generating candidate SMS copywriting that matches the copywriting preference dimension using a large language model and preset prompt word templates; performing multi-objective multi-classification prediction on the user to obtain customer segmentation results; and performing association processing on the candidate SMS copywriting using a preset dual-stream recommendation model based on the customer segmentation results and the multi-source behavioral data to obtain SMS copywriting recommendation results. Since this embodiment targets the user's copywriting preference dimension, the large language model can automatically generate multiple candidate SMS copywriting sets that match the user's copywriting preference dimension for subsequent recommendation model filtering and sorting. Then, by performing multi-objective multi-classification prediction on the user, customer segmentation results with different emphases on multiple objectives can be obtained, achieving a refined user profile. Thus, even with a large amount of user data, the dual-stream recommendation model can accurately generate SMS copywriting recommendation results that match user behavior, accurately accommodating the diversity of customer groups, thereby effectively improving the marketing conversion rate of SMS.

[0038] In one feasible embodiment, the step of extracting features from the user's multi-source behavioral data to obtain the user's text preference dimension may include: Step S11: Preprocess the user's multi-source behavioral data to obtain standardized structured data. The preprocessing includes missing value imputation and outlier replacement.

[0039] Step S12: Convert the standardized structured data into a long text feature sequence.

[0040] It should be noted that the contents of this application embodiment that are the same as or similar to the above embodiments can be referred to the above description, and will not be repeated hereafter.

[0041] Standardized structured data involves filling missing data in multi-source behavioral data with multiple imputations, replacing outliers with quantile limiting or nearest neighbor methods, to ensure that every record entering the model is complete, resulting in data that can be directly used for subsequent processing.

[0042] For example, refer to Figure 2 , Figure 2This diagram illustrates a specific implementation of SMS message generation in this application. During the preparation phase of various data sources (such as basic attribute data preparation, People's Bank of China data preparation, Sensors Data preparation, marketing outreach data preparation, and other types of data preparation), the data structures are complex due to the different sources, including both structured and unstructured data. Furthermore, the data quality varies significantly, with some data being severely incomplete or containing special characters. Therefore, data processing of multi-source behavioral data is necessary. This can be divided into two stages for data processing.

[0043] The first stage is basic processing, including handling missing values ​​(such as using multiple imputation and nearest neighbor imputation to fill missing values, making the filling more realistic), special value replacement, handling excessively long characters (such as truncating excessively long text), and handling time variables (such as converting dates to days). After completing the basic processing, it can be determined whether the data is anomaly-free; if there is anomaly, this stage is repeated; if there is no anomaly, the subsequent judgment of activity level and funding needs is carried out.

[0044] Step S13: Use a preset dual-weight factor extraction network to predict user attributes of the long text feature sequence to determine user activity and funding needs.

[0045] Step S14: Use the user activity level and the funding requirement as the user's copywriting preference dimensions.

[0046] It should be noted that the dual-weighted factor extraction network is a model for judging and refining user behavioral attributes, which can be used to predict and judge user activity and user funding needs. User activity measures the frequency of a user's recent interactions with the platform, while funding needs measure the user's current willingness to borrow or withdraw money.

[0047] In the second stage of data processing, a deep model based on a dual-weight factor extraction network is used to judge and refine user behavior attributes to determine user activity and funding needs. The model results are then stored in a data warehouse for subsequent models to convert into a long text, which is then used by a large language model for recognition and analysis.

[0048] Furthermore, the dual-weight factor extraction network is constructed from a deep network model. The dual-weight factor extraction network includes a feature extraction layer, a hidden layer, and a softmax output layer. The hidden layer is connected to the feature extraction layer and the softmax output layer, respectively. The dual-weight factor extraction network is used to predict user activity and funding demand based on the long text feature sequence.

[0049] Specifically, FM (Factorization Machine) models can be improved to construct a deep network model based on dual-weight factor extraction. This model extracts factor features by combining fixed and variable weights, and then uses improved hidden layer parameter design to determine user activity and funding needs. The specific formula is as follows: (1) Factor feature extraction layer with fixed weights and variable weights: ; in, Indicates the original Multiplication of variable matrix Linear transformations of matrices It is a parameter matrix that can be trained; and The weights of fixed variables. and This represents a cross-multiplication of two variables. Here, a fixed weight is assigned to certain specific variables in the cross-multiplication, with different weights for different variables. This allows the model to focus more on these variables in high-dimensional cross-multiplication. and The weights here are variable weights, meaning they can be trained. This involves assigning weights to other non-specific variables using a training method. Finally, the sum of these parts yields the final result. .

[0050] (2) Improve the deep network with hidden layers: ; in, It is the first A hidden layer, It is the first There are several hidden layers; this is where the deep neural network is trained, and the number of layers is... Yes, it can be set. W is the trained parameter, and b is the trained constant. In order to prevent the constant from being set to 0 or too small, through practice, adding an e for adjustment can achieve better results.

[0051] (3) The final output is changed to a softmax output layer: ; in, It is a probability mapping process that adds the results of the previous two steps to obtain the final result.

[0052] User activity and funding demand are trained separately. Taking user activity as an example: First, the Y label, indicating user activity, is obtained from historical data. Then, multiple label data for the user are calculated, namely X1, X2, ..., Xi. n variables are assigned fixed weights with a default weight; the other m variables are assigned variable weights, which are trained. After training, a prediction result for factor feature extraction combining fixed and variable weights is obtained. Simultaneously, a deep network with improved hidden layers is trained, yielding parameters W and prediction results. Finally, the two results are summed and passed through a softmax output layer to obtain the final activity prediction result. The prediction process for funding demand is similar and will not be elaborated further.

[0053] This embodiment provides a method for understanding copywriting preferences. For copywriting preferences, user activity and financial needs are two key considerations. People with different levels of activity and financial needs will have different copywriting preferences. Preliminary assessments of user performance in these two dimensions can reduce the data dimensionality, allowing for more focused thinking within larger models, and also help operations personnel better understand the process of generating personalized copy.

[0054] In one feasible embodiment, considering that manual writing during the SMS creation stage relies heavily on the experience of operations personnel, resulting in inconsistent copywriting quality and a long production cycle; the step in this embodiment of generating candidate SMS copywriting that matches the copywriting preference dimension through a large language model and preset prompt word templates may include: Step S21: Fill the long text feature sequence into a preset prompt word template, wherein the prompt word template includes at least one of the following: background introduction section, core word limitation section, style prompt section, customer characteristics section, and positive and negative case section.

[0055] Step S22: Input the filled prompt word template into the large language model to obtain multiple candidate SMS texts that match the text preference dimension.

[0056] After predicting user activity and funding needs, the next step is to generate candidate SMS messages. For example... Figure 2 As shown, the process of generating candidate SMS text in this application embodiment is divided into two stages. The first stage is the best prompt word template experiment, and the second stage is the fine-tuning training of the large language model by combining excellent text.

[0057] In Phase 1, the prompts were continuously modified and the business team tagged the results. After multiple rounds of experiments, the optimal prompt template writing formula was obtained: Background introduction paragraph (i.e., use 1-2 sentences to explain the SMS sending scenario and set the context for the large model), core keyword restriction paragraph (i.e., specify the keywords that must appear or are prohibited to ensure that the generated copy meets the selling point requirements), style prompt paragraph (i.e., specify the tone and word count so that the large model can output the corresponding SMS tone), customer characteristics paragraph (i.e., insert the activity level + funding needs + scenario tags in natural language here, for example, high activity, low funding needs, sensitivity to holiday greetings, so that the large model knows who to speak to), positive case paragraph and negative case paragraph (provide positive excellent copy and negative poor copy examples to guide the large model to move towards a high-scoring style and avoid low-scoring expressions).

[0058] In the second stage, for example, a basic large language model, Qwen3, can be used. The parameters of the intermediate layers of the model are frozen, and only the output layer is fine-tuned. After training in this way, multiple candidate SMS messages with different styles that meet the requirements can be output according to the prompt word template.

[0059] With its five-segment prompt templates and lightweight, fine-tuned large model, the system can automatically generate multiple high-quality SMS messages that match user preferences, significantly shortening the SMS message production cycle.

[0060] In one feasible embodiment, the step of performing multi-objective multi-classification prediction on the user to obtain customer segmentation results as described in this example may include: Step S31: Input the multi-source behavioral data into a preset multi-objective multi-classification model for classification prediction to obtain customer segmentation results. The multi-objective multi-classification model is composed of the dual-weight factor extraction network and the machine learning model. The multi-objective multi-classification model is used to identify scene words in the multi-source behavioral data to obtain classification results.

[0061] In the traditional customer selection phase, operators typically segment customers based on simple user tags, resulting in a coarse granularity. It is difficult for operators to manually identify a large number of tag features and to segment customers into more detailed categories while taking into account the diversity of customer groups, thus failing to achieve multi-customer segmentation under multiple objectives.

[0062] To solve the above problems, such as Figure 3 As shown, Figure 3This is a specific implementation diagram of customer group segmentation in this application embodiment. This application embodiment proposes a multi-objective, multi-classification model (i.e., a tree partitioning model based on probability value dual-model fusion). This model can receive the aforementioned multi-source behavioral data, and then output multiple attribute judgments (such as high activity, moderate demand, moderate debt, married with family) at once based on various objectives (e.g., cash withdrawal, activity level, application submission) and store the model results. This allows for rapid and comprehensive user profiling. It can also cluster customers based on attribute similarity, thereby automatically dividing customers into multiple categories. The overall structure of this model consists of two stages: In the first stage, the aforementioned dual-weight factor extraction network is creatively fused with classic machine learning models (such as Wide & Deep). Instead of using the traditional voting method to select the classification result, the weights are fused based on probability values ​​before tree model partitioning. The multi-objective multi-classification model is as follows: ; in, It is the number of categories. These are weighting coefficients; different models assign different weights. This represents a dual-weighted factor extraction network. This represents a machine learning model.

[0063] The second stage of data processing involves training and outputting a multi-objective, multi-classification model. This process is based on scene discrimination and long-sequence network partitioning models, and the specific steps are as follows: First, the results of the single-objective model are processed. After the multi-source behavioral data is converted into continuous long text sentences, scene labeling is performed. Each sentence can be considered to consist of m scenes: ; Here, Scene refers to scene words, meaning that each continuous long text sentence can be summarized and described by multiple scene words, and different segments within a continuous long text sentence can be described by different scene words. For each continuous long text sentence, m scene words are labeled, with the number of scene words varying for each continuous long text sentence.

[0064] (2) Then, the original long text data or other supplementary data are processed. The continuous long text sentences X1 are cut into blocks, the scene words are added into the blocks, and then the attention mechanism is combined to perform linear transformation to obtain the vector matrix of the scene, such as Q1 matrix (i.e. query matrix), K1 matrix (i.e. key matrix), and V1 matrix (i.e. value matrix).

[0065] (3) Based on the above vector matrix, the single-head attention mechanism and the multi-head attention mechanism in multiple scenarios can be calculated, and the model automatically learns the relationship between each sentence block and each scene.

[0066] (4) Train a deep neural network model to obtain multi-objective multi-classification results.

[0067] As mentioned earlier, scene-related words have been integrated into the original continuous long text sentences to form new continuous long text sentences. During training, a method similar to a cloze test is used, randomly removing scene-related words to allow the deep neural network to learn and recognize them. The weights are gradually adjusted to ensure the model can accurately identify these words, which represent the multi-objective, multi-classification results (e.g., highly educated, moderately active, moderate financial needs, 45 years old, etc.). Given another original continuous long text sentence, this allows the model to predict its multi-objective, multi-classification results, achieving a one-time output of multi-label customer segmentation results.

[0068] In a feasible embodiment, the step of associating the candidate SMS messages with the customer segmentation results and the multi-source behavioral data using a preset dual-stream recommendation model to obtain SMS message recommendation results may include: Step S41: Based on the multi-source behavioral data, perform association processing on the candidate SMS messages corresponding to each customer group in the customer segmentation results to obtain a serialized recommendation list.

[0069] The serialized recommendation list is a list of recommended templates that takes into account the order in which users clicked on SMS messages in the past.

[0070] Furthermore, in a feasible embodiment, the step of associating candidate SMS messages corresponding to each customer group in the customer segmentation results based on the multi-source behavioral data to obtain a serialized recommendation list includes: determining the customer group categories in the customer segmentation results; vectorizing the candidate SMS messages corresponding to each customer group category based on the multi-source behavioral data to obtain SMS template vectors; determining the correlation matrix between each pair of SMS template vectors; and determining the serialized recommendation list of the candidate SMS messages based on the attention score of the correlation matrix.

[0071] Considering that traditional methods can only send one SMS template to one customer group, and cannot adapt different SMS templates based on differences in user behavior attributes, this embodiment provides a precise SMS copy recommendation method. Here, combining customer lifecycle and operational habits, users are divided into 7 sub-customer groups, and a model is trained for each group separately. The candidate SMS copy and multi-source behavioral data are used as variables and input into the model for training. An adaptive model for integrating and adjusting dual-stream recommendation is proposed here, the specific process of which is as follows: (1) First, perform vector transformation on the candidate SMS texts of users’ historical browsing and conversion in the multi-source behavioral data (such as using the T5-Sentence large model for transformation).

[0072] (2) Then calculate the correlation between each pair of SMS template vectors to form a correlation matrix: ; in, SMS templates and SMS templates The correlation coefficient (which can be calculated using cosine similarity).

[0073] (3) Creatively using the correlation matrix as the initial WQ matrix to calculate the attention score: ; ; Where X is the matrix of consecutive long text sentences X1; This is the relevance matrix mentioned earlier, used as the initial value for the Q1 query matrix in the attention mechanism. K is the key matrix in the attention mechanism, V is the value matrix in the attention mechanism, and W... K Q and Wv are the initial weights of matrices K and V, respectively. Then, the attention value is calculated based on Q, K, and V.

[0074] (4) Nest an output layer to obtain the serialized recommendation result: ; ; in, Indicates the current number The final hidden state of the layer, obtained after passing through the activation function σ, contains both the attention context and deep nonlinear features, which are used for subsequent scoring. c represents the output layer bias term, with a dimension equal to the number of candidate templates, ensuring that the Softmax score distribution has degrees of freedom for translation and avoiding all scores crowding around 0.

[0075] Here, a deep neural network is designed. Attention is the attention result of step (3), α is the hidden layer parameter, and b is a constant. Similarly, to prevent the constant from being 0, the optimal parameter - √e - was creatively added through practice. After training, this network can recommend a serialized recommendation list of SMS text messages, such as list=[emb4,emb1…,embn]. That is, the model believes that the Emb4 template should be used for recommendation next time, the Emb1 template should be used for recommendation the next time, and so on. This is done in order to fully learn the features of the historical sequence and recommend a serialized recommendation list. .

[0076] Step S42: Determine the single-layer correlation and deep multi-dimensional correlation between customer behavior and SMS copy through the preset dual-stream recommendation model, and obtain the dual-stream recommendation list.

[0077] The dual-stream recommendation model employs a parallel structure. The upper branch is a single-layer correlation model, which calculates the cosine of each pair of user behavior vectors and then performs a weighted sum to obtain a fast linear score. The lower branch is a deep multi-dimensional correlation model, which feeds the behavior vectors into a multi-layer MLP for higher-order cross-multiplication to obtain a deep score. Finally, a softmax function is used to obtain the output of the dual-stream recommendation model, forming the dual-stream recommendation list. .

[0078] Step S43: Weighted fusion of the serialized recommendation list and the dual-stream recommendation list to obtain the SMS text recommendation result.

[0079] The two results are weighted together to form the final adaptive SMS copy recommendation result: ; in, For serializing the recommendation list, The recommendation list is a dual-stream recommendation list, where g is an adaptive parameter that can be trained and has a value between 0 and 1. By adjusting g, the results of the two previous recommendations can be effectively combined to obtain a final SMS message recommendation. This deep sequential recommendation integrates and adjusts the adaptive model of dual-stream recommendations, enabling recommendations based on both customer behavioral characteristics and historical behavior, and combining the advantages of both methods to achieve accurate SMS message recommendations.

[0080] Example, for reference Figure 3 , Figure 4 This is a specific implementation diagram of the text recommendation in this application. Multi-source behavioral data and candidate SMS texts are input into the aforementioned deep sequential recommendation integrated adjustment dual-stream recommendation adaptive model. Seven customer groups are modeled according to business operation rules (e.g., no application - pure text text customer group, no application - coupon customer group, credit granted but not withdrawn - pure text text customer group, credit granted but not withdrawn - coupon customer group, loan in progress - pure text text customer group, loan in progress - coupon customer group, settled but not re-loan - coupon customer group). The model can then output the SMS text recommendation results.

[0081] This application proposes a personalized SMS generation and precise copywriting recommendation scheme based on a large-scale model integrating multiple adaptive algorithms. For SMS copywriting creation, multi-source behavioral data is submitted to a large language model for analysis and processing, automatically generating candidate SMS copywriting that matches copywriting preference dimensions, thus achieving the ability to automatically create different SMS copywriting based on different users. For user segmentation, an algorithm is proposed that can simultaneously form multiple categories based on multiple goals, thereby further dividing a large customer group into multiple smaller customer groups, each with different focuses on multiple goals. For copywriting recommendation, an integrated recommendation algorithm combining classic recommendation techniques and large-scale model recommendation techniques is adopted. This algorithm can recommend different SMS models based on different customer groups, and can also recommend SMS templates based on customers' historical SMS response patterns. Overall, this scheme can effectively improve the conversion rate of SMS marketing.

[0082] It should be noted that the above examples are only for understanding this application and do not constitute a limitation on the SMS recommendation method of this application. Any simple modifications based on this technical concept are within the protection scope of this application.

[0083] This application also provides a text message recommendation device, please refer to... Figure 5 , Figure 5 This is a schematic diagram of the module structure of the SMS recommendation device according to an embodiment of this application; the SMS recommendation device includes: The feature extraction module 501 is used to extract features from the user's multi-source behavioral data to obtain the user's copywriting preference dimension. The SMS generation module 502 is used to generate candidate SMS text that matches the text preference dimension by using a large language model and preset prompt word templates. The customer segmentation module 503 is used to perform multi-objective multi-classification prediction on the user to obtain the customer segmentation result. The SMS recommendation module 504 is used to perform correlation processing on the candidate SMS texts based on the customer segmentation results and the multi-source behavioral data through a preset dual-stream recommendation model to obtain SMS text recommendation results.

[0084] In one implementation, the feature extraction module 501 is further configured to preprocess the user's multi-source behavioral data to obtain standardized structured data, wherein the preprocessing includes missing value imputation and outlier replacement; convert the standardized structured data into a long text feature sequence; perform user attribute prediction on the long text feature sequence through a preset dual-weight factor extraction network to determine user activity and funding needs; and use the user activity and funding needs as the user's copywriting preference dimensions.

[0085] In one implementation, the feature extraction module 501 uses a dual-weight factor extraction network constructed from a deep network model. The dual-weight factor extraction network includes a feature extraction layer, a hidden layer, and a softmax output layer. The hidden layer is connected to the feature extraction layer and the softmax output layer, respectively. The dual-weight factor extraction network is used to predict user activity and funding demand based on the long text feature sequence.

[0086] As one implementation, the SMS generation module 502 is further configured to fill the long text feature sequence into a preset prompt word template, wherein the prompt word template includes at least one of the following: background introduction section, core word limitation section, style prompt section, customer characteristic section, and positive and negative case section; and input the filled prompt word template into a large language model to obtain multiple candidate SMS texts that match the copywriting preference dimension.

[0087] In one implementation, the customer segmentation module 503 is further used to input the multi-source behavioral data into a preset multi-objective multi-classification model for classification prediction to obtain customer segmentation results. The multi-objective multi-classification model is composed of the dual-weight factor extraction network and the machine learning model. The multi-objective multi-classification model is used to identify scene words in the multi-source behavioral data to obtain classification results.

[0088] As one implementation method, the SMS recommendation module 504 is further configured to perform association processing on the candidate SMS texts corresponding to each customer group in the customer segmentation results based on the multi-source behavioral data to obtain a serialized recommendation list; determine the single-layer correlation and deep multi-dimensional correlation between customer behavior and SMS texts through a preset dual-stream recommendation model to obtain a dual-stream recommendation list; and weight and fuse the serialized recommendation list and the dual-stream recommendation list to obtain the SMS text recommendation result.

[0089] In one implementation, the SMS recommendation module 504 is further configured to determine the customer group category in the customer group segmentation result; vectorize the candidate SMS text corresponding to each customer group category based on the multi-source behavioral data to obtain SMS template vectors; determine the correlation matrix between each pair of SMS template vectors; and determine the serialized recommendation list of the candidate SMS text based on the attention score of the correlation matrix.

[0090] Other embodiments or specific implementations of the SMS recommendation device of this application can be found in the above-described method embodiments, and will not be repeated here.

[0091] The SMS recommendation device provided in this application, employing the SMS recommendation method described in the above embodiments, can solve the technical problem that traditional marketing SMS copy recommendation methods rely heavily on the experience of operators for selection, and due to the large amount of user data, it is difficult to accurately consider the diversity of the customer base. Compared with the prior art, the beneficial effects of the SMS recommendation device provided in this application are the same as those of the SMS recommendation method provided in the above embodiments, and other technical features in the SMS recommendation device are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0092] This application provides a text message recommendation device, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, which are executed by the at least one processor to enable the at least one processor to perform the text message recommendation method in the above embodiments.

[0093] The following is for reference. Figure 6 , Figure 6 This is a schematic diagram of the hardware operating environment involved in the SMS recommendation method in the embodiments of this application, showing a structural diagram of a device suitable for implementing the SMS recommendation method in the embodiments of this application. The SMS recommendation device in the embodiments of this application may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Descriptions), PMPs (Portable Media Players), in-vehicle terminals (such as in-vehicle navigation terminals), and fixed terminals such as digital TVs and desktop computers. Figure 6 The SMS recommendation device shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0094] like Figure 6As shown, the SMS recommendation device may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in a read-only memory 1002 or a program loaded from a storage device 1003 into a random access memory 1004. The random access memory 1004 also stores various programs and data required for the operation of the SMS recommendation device. The processing unit 1001, the read-only memory 1002, and the random access memory 1004 are interconnected via a bus 1005. An input / output interface 1006 is also connected to the bus. Typically, the following systems can be connected to the input / output interface 1006: input devices 1007 including, for example, a touchscreen, touchpad, keyboard, mouse, image sensor, microphone, accelerometer, gyroscope, etc.; output devices 1008 including, for example, a liquid crystal display (LCD), speaker, vibrator, etc.; storage devices 1003 including, for example, magnetic tape, hard disk, etc.; and communication devices 1009. Communication device 1009 allows the SMS recommendation device to communicate wirelessly or wiredly with other devices to exchange data. Although SMS recommendation devices with various systems are shown in the figures, it should be understood that implementation or possession of all the systems shown is not required. More or fewer systems may be implemented alternatively.

[0095] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from read-only memory 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0096] The SMS recommendation device provided in this application, employing the SMS recommendation method described in the above embodiments, solves the technical problem that traditional marketing SMS copy recommendation methods rely heavily on the experience of operators, making it difficult to accurately consider the diversity of customer groups due to the large amount of user data. Compared with the prior art, the beneficial effects of the SMS recommendation device provided in this application are the same as those of the SMS recommendation method provided in the above embodiments, and other technical features of this SMS recommendation device are the same as those disclosed in the previous embodiment method, and will not be repeated here.

[0097] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0098] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0099] This application provides a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, the computer-readable program instructions being used to execute the SMS recommendation method in the above embodiments.

[0100] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems or devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0101] The aforementioned computer-readable storage medium may be included in the SMS recommendation device; or it may exist independently and not assembled into the SMS recommendation device.

[0102] The aforementioned computer-readable storage medium carries one or more programs. When these programs are executed by the SMS recommendation device, the SMS recommendation device performs the following actions: extracts features from the user's multi-source behavioral data to obtain the user's text message preference dimension; generates candidate SMS messages that match the text message preference dimension using a large language model and preset prompt word templates; performs multi-objective multi-classification prediction on the user to obtain customer segmentation results; and performs association processing on the candidate SMS messages using a preset dual-stream recommendation model based on the customer segmentation results and the multi-source behavioral data to obtain SMS message recommendation results.

[0103] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a Local Area Network (LAN) or a Wide Area Network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0104] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0105] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0106] The readable storage medium provided in this application is a computer-readable storage medium that stores computer-readable program instructions (i.e., a computer program) for executing the above-described SMS recommendation method. This addresses the technical problem that traditional marketing SMS copy recommendation methods rely heavily on the experience of operators, making it difficult to accurately consider the diversity of customer groups due to the large amount of user data. Compared to existing technologies, the beneficial effects of the computer-readable storage medium provided in this application are the same as those of the SMS recommendation method provided in the above embodiments, and will not be elaborated upon here.

[0107] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the SMS recommendation method described above.

[0108] The computer program product provided in this application solves the technical problem that traditional marketing SMS copy recommendation methods rely heavily on the experience of operators, making it difficult to accurately consider the diversity of customer groups due to the large amount of user data. Compared with the prior art, the beneficial effects of the computer program product provided in this application are the same as those of the SMS recommendation method provided in the above embodiments, and will not be repeated here.

[0109] The above description is only a part of the embodiments of this application and does not limit the scope of protection of this application. All equivalent structural transformations made under the technical concept of this application and using the content of this application specification and drawings, or direct / indirect applications in other related technical fields, are included in the scope of protection of this application.

Claims

1. A method for recommending SMS messages, characterized in that, The method includes: Feature extraction is performed on the user's multi-source behavioral data to obtain the user's copywriting preference dimension; By using a large language model and preset prompt word templates, candidate SMS messages that match the aforementioned copywriting preference dimension are generated. Perform multi-objective multi-classification prediction on the users to obtain customer segmentation results; Based on the customer segmentation results and the multi-source behavioral data, the candidate SMS messages are correlated using a preset dual-stream recommendation model to obtain SMS message recommendation results.

2. The method as described in claim 1, characterized in that, The step of associating candidate SMS messages with the customer segmentation results and the multi-source behavioral data using a preset dual-stream recommendation model to obtain SMS message recommendation results includes: Based on the multi-source behavioral data, the candidate SMS messages corresponding to each customer group in the customer segmentation results are associated to obtain a serialized recommendation list; The single-layer correlation and deep multi-dimensional correlation between customer behavior and SMS copy are determined by a pre-set dual-stream recommendation model to obtain a dual-stream recommendation list. The serialized recommendation list and the dual-stream recommendation list are weighted and fused to obtain the SMS text recommendation result.

3. The method as described in claim 2, characterized in that, The step of associating candidate SMS messages for each customer group in the customer segmentation results based on the multi-source behavioral data to obtain a serialized recommendation list includes: Determine the customer group categories in the customer segmentation results; Based on the multi-source behavioral data, the candidate SMS messages corresponding to each customer group category are vectorized to obtain SMS template vectors; Determine the correlation matrix between pairwise SMS template vectors; Based on the attention score of the relevance matrix, a serialized recommendation list of candidate SMS messages is determined.

4. The method as described in claim 1, characterized in that, The step of extracting features from the user's multi-source behavioral data to obtain the user's text preference dimension includes: The user's multi-source behavioral data is preprocessed to obtain standardized structured data. The preprocessing includes missing value imputation and outlier replacement. The standardized structured data is converted into a long text feature sequence; The user attribute prediction of the long text feature sequence is performed by a preset dual-weight factor extraction network to determine user activity and funding needs. The user activity level and the funding requirement level are used as dimensions of the user's copywriting preference.

5. The method as described in claim 4, characterized in that, The dual-weight factor extraction network is constructed from a deep network model. The dual-weight factor extraction network includes a feature extraction layer, a hidden layer, and a softmax output layer. The hidden layer is connected to the feature extraction layer and the softmax output layer, respectively. The dual-weight factor extraction network is used to predict user activity and funding demand based on the long text feature sequence.

6. The method as described in claim 4, characterized in that, The step of generating candidate SMS text messages that match the text preference dimension using a large language model and preset prompt word templates includes: The long text feature sequence is filled into a preset prompt word template, which includes at least one of the following: background introduction section, core word limitation section, style prompt section, customer characteristics section, and positive and negative case section. The filled prompt word template is input into the large language model to obtain multiple candidate SMS text messages that match the text preference dimension.

7. The method as described in claim 4, characterized in that, The step of performing multi-objective multi-class prediction on the user to obtain customer segmentation results includes: The multi-source behavioral data is input into a preset multi-objective multi-classification model for classification prediction to obtain customer segmentation results. The multi-objective multi-classification model is composed of the dual-weight factor extraction network and the machine learning model. The multi-objective multi-classification model is used to identify scene words in the multi-source behavioral data to obtain classification results.

8. A text message recommendation device, characterized in that, The device includes: The feature extraction module is used to extract features from the user's multi-source behavioral data to obtain the user's copywriting preference dimension; The SMS generation module is used to generate candidate SMS texts that match the text preference dimension using a large language model and preset prompt word templates. The customer segmentation module is used to perform multi-objective multi-classification prediction on the users to obtain customer segmentation results. The SMS recommendation module is used to perform correlation processing on the candidate SMS texts based on the customer segmentation results and the multi-source behavioral data, and obtain SMS text recommendation results through a preset dual-stream recommendation model.

9. A text message recommendation device, characterized in that, The device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the SMS recommendation method as described in any one of claims 1 to 7.

10. A storage medium, characterized in that, The storage medium is a computer-readable storage medium, and a computer program is stored on the storage medium. When the computer program is executed by a processor, it implements the steps of the SMS recommendation method as described in any one of claims 1 to 7.

Images