Product recommendation method and device, storage medium and processor

Abstract

The invention provides a product recommendation method and device, a storage medium and a processor, and the method comprises the steps: obtaining a plurality of dialogue data and a plurality of click behavior data of a user, the dialogue data being a dialogue, obtained before the current moment, of the user in application software, and the click behavior data being a click behavior, obtained at the current moment, of the user in the application software; the dialogue data and the click behavior data are input into a bidirectional encoder characterization method model so as to at least extract the semantic relation between the dialogue data and the click behavior data, and the dialogue sequence is used for characterizing semantic information corresponding to the dialogue data; inputting the dialogue sequence into a first converter model, and inputting the behavior sequence into a second converter model; performing weighted splicing processing on the first vector and the second vector to obtain a third vector; and generating prompt information according to the third vector, and recommending to the user according to the prompt information, thereby solving the problem of inaccurate product recommendation to the user in the prior art.

Description

Technical Field

[0001] This application relates to information mining technology, and more specifically, to a product recommendation method, a product recommendation device, a computer-readable storage medium, and a processor. Background Technology

[0002] Existing technologies utilize interactive information and individual user behavior to recommend products to users.

[0003] However, if a user purchases a product after interacting with the system, or performs a subsequent action that contradicts their initial intention after a single action, the system cannot recognize this, thus reducing the accuracy of product recommendations and increasing errors.

[0004] The information disclosed above in the background section is only intended to enhance the understanding of the background art of the art described herein. Therefore, the background art may contain certain information that does not constitute prior art known to those skilled in the art in this country. Summary of the Invention

[0005] The main objective of this application is to provide a product recommendation method, a product recommendation device, a computer-readable storage medium, and a processor to at least solve the problem of insufficient accuracy in product recommendations in the prior art.

[0006] To achieve the above objectives, according to one aspect of this application, a product recommendation method is provided, comprising: acquiring multiple dialogue data and multiple click behavior data of a user, wherein the dialogue data is dialogues of the user in an application software acquired before the current time, and the click behavior data is click behavior of the user in the application software acquired at the current time; inputting the dialogue data and the click behavior data into a bidirectional encoder representation model to extract at least the semantic relationship between the dialogue data and the click behavior data, obtaining a dialogue sequence and a behavior sequence, wherein the dialogue sequence is used to represent the semantic information corresponding to the dialogue data, and the behavior sequence is used to represent the semantic information of the click behavior data; inputting the dialogue sequence into a first converter model to extract the relationship between the data in the dialogue sequence, obtaining a first vector, and inputting the behavior sequence into a second converter model to extract the relationship between the data in the behavior sequence, obtaining a second vector; performing weighted concatenation processing on the first vector and the second vector to obtain a third vector; generating prompt information based on the third vector, and making recommendations to the user based on the prompt information.

[0007] Optionally, acquiring multiple dialogue data of a user includes: acquiring multiple dialogue texts of the user, each dialogue text including at least one Chinese character to be mapped; acquiring a Chinese character sequence number mapping relationship, and determining a target sequence number based on the Chinese character sequence number mapping relationship and the Chinese character to be mapped, wherein the Chinese character sequence number mapping relationship is a pre-set mapping relationship between Chinese characters and sequences, and the target sequence number is the sequence number corresponding to the Chinese character that is the same as the Chinese character to be mapped in the Chinese character sequence number mapping relationship; acquiring a sequence number vector mapping relationship, and determining a target vector based on the target sequence number and the sequence number vector mapping relationship, wherein multiple target vectors constitute the dialogue data, wherein the sequence number vector mapping relationship is a pre-set mapping relationship between sequences and vectors, and the target vector is the vector corresponding to the sequence number that is the same as the target sequence number in the sequence number vector mapping relationship.

[0008] Optionally, the dialogue data and the click behavior data are input into a bidirectional encoder representation model to extract at least the semantic relationship between the dialogue data and the click behavior data, obtaining a dialogue sequence and a behavior sequence. This includes: inputting multiple dialogue data sets into a first model to perform intent recognition on the dialogue data, obtaining an intent sequence, wherein the intent sequence is used to represent the intent of the dialogue data; and inputting the intent sequence and the click behavior data into a second model to extract at least the semantic relationship within the intent sequence and between the click behavior data, obtaining the dialogue sequence and the behavior sequence. The bidirectional encoder representation model includes the first model and the second model.

[0009] Optionally, multiple dialogue data are input into a first bidirectional encoder representation model to perform intent recognition on the dialogue data and obtain an intent sequence, including: an input step, in which one dialogue data is sequentially input into multiple encoders to classify the dialogue data and obtain at least one intent label, wherein the intent label is used to represent the intent of the dialogue data, and when there are multiple intent labels, the order of the intent labels is arranged from smallest to largest according to the predicted probability; and a repetition step, in which the input step is repeated multiple times until multiple dialogue data are input and multiple intent labels are obtained to form the intent sequence.

[0010] Optionally, the first converter model includes a first multi-head self-attention layer, a first feedforward network layer, and a first residual connection layer connected in sequence. Inputting the dialogue sequence into the first converter model to extract the relationship between data in the dialogue sequence and obtain a first vector includes: inputting the dialogue sequence sequentially into the first multi-head self-attention layer to obtain first output data, wherein the first multi-head self-attention layer is used to calculate the relationship between data in the dialogue sequence; inputting the first output data into the first feedforward network layer to obtain second output data, wherein the first feedforward network layer is used to perform nonlinear transformation and feature extraction on the first output data; and inputting the second output data into the first residual connection layer to obtain the first vector. The second converter model includes a second multi-head self-attention layer, a second feedforward network layer, and a second residual connection layer connected in sequence. The behavior sequence is input into the second converter model to extract the relationship between data in the behavior sequence and obtain a second vector. This includes: inputting the behavior sequence into the second multi-head self-attention layer to obtain third output data, wherein the second multi-head self-attention layer is used to calculate the relationship between data in the behavior sequence; inputting the third output data into the second feedforward network layer to obtain fourth output data, wherein the second feedforward network layer is used to perform nonlinear transformation and feature extraction on the third output data; and inputting the fourth output data into the second residual connection layer to obtain the second vector.

[0011] Optionally, a weighted concatenation process is performed on the first vector and the second vector to obtain a third vector, including: obtaining a first hyperparameter and a second hyperparameter, wherein the first hyperparameter is the weight of the first vector, the second hyperparameter is the weight of the second vector, and the sum of the first hyperparameter and the second hyperparameter is 1;

[0012] Calculate the product of the first hyperparameter and the first vector to obtain the first vector to be concatenated, and calculate the product of the second hyperparameter and the second vector to obtain the second vector to be concatenated; concatenate the first vector to be concatenated and the second vector to be concatenated in a predetermined dimension to obtain the third vector, wherein the predetermined dimension is the last dimension of the first vector and the second vector.

[0013] Optionally, generating prompt information based on the third vector includes: inputting the third vector into a fully connected layer to transform the vector space of the third vector, obtaining a fourth vector and its predicted probability; comparing the predicted probability of the fourth vector with a predetermined probability, and determining the fourth vector as the target intent if the predicted probability of the fourth vector is greater than or equal to the predetermined probability; obtaining an intent tree, and determining the prompt information based on the target intent and the intent tree, wherein the intent tree is used to characterize the mapping relationship between the user's intent and clue content, and the prompt information is the clue content corresponding to the intent in the intent tree that is the same as the target intent.

[0014] According to another aspect of this application, a product recommendation device is provided, comprising: an acquisition unit, configured to acquire multiple dialogue data and multiple click behavior data of a user, wherein the dialogue data are dialogues of the user in an application software acquired before the current time, and the click behavior data are click behaviors of the user in the application software acquired at the current time; a first input unit, configured to input the dialogue data and the click behavior data into a bidirectional encoder representation model to at least extract the semantic relationship between the dialogue data and the click behavior data, obtaining a dialogue sequence and a behavior sequence, wherein the dialogue sequence is used to represent the semantic information corresponding to the dialogue data, and the behavior sequence is used to represent the semantic information of the click behavior data; a second input unit, configured to input the dialogue sequence into a first converter model to extract the relationship between the data in the dialogue sequence to obtain a first vector, and input the behavior sequence into a second converter model to extract the relationship between the data in the behavior sequence to obtain a second vector; a processing unit, configured to perform weighted concatenation processing on the first vector and the second vector to obtain a third vector; and a recommendation unit, configured to generate prompt information based on the third vector, and make recommendations to the user based on the prompt information.

[0015] According to another aspect of this application, a computer-readable storage medium is provided, the computer-readable storage medium including a stored program, wherein, when the program is executed, it controls the device on which the computer-readable storage medium is located to perform any of the methods described.

[0016] According to another aspect of this application, a processor is provided for running a program, wherein the program, when running, performs any of the methods described.

[0017] The technical solution of this application first acquires multiple dialogue data and multiple click behavior data of the user. The dialogue data consists of user conversations within the application software acquired up to the current moment, and the click behavior data consists of user click behaviors within the application software acquired at the current moment. Then, the dialogue data and click behavior data are input into a bidirectional encoder representation model to extract at least the semantic relationship between the dialogue data and the click behavior data. The dialogue sequence is used to represent the semantic information corresponding to the dialogue data. Next, the dialogue sequence is input into a first converter model, and the behavior sequence is input into a second converter model. Finally, the first and second vectors are weighted and concatenated to obtain a third vector. Based on the third vector, a prompt message is generated, and recommendations are made to the user based on the prompt message. By combining the dialogue data and click behavior data, and then performing intent recognition to obtain the dialogue sequence and behavior sequence, sequence modeling is then performed to obtain vector representations. These vectors are then weighted and concatenated to obtain a combined vector, i.e., the third vector. Finally, the customer's intent is predicted based on the combined vector, thereby generating prompt messages, which helps to achieve accurate recommendations and solves the problem of inaccurate product recommendations to users in existing technologies. Attached Figure Description

[0018] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0019] Figure 1 A hardware structure block diagram of a mobile terminal performing a product recommendation method according to an embodiment of this application is shown;

[0020] Figure 2 A schematic flowchart of a product recommendation method provided according to an embodiment of this application is shown;

[0021] Figure 3 A detailed flowchart of a product recommendation method according to an embodiment of this application is shown;

[0022] Figure 4 A flowchart illustrating a training process for obtaining an intent sequence is shown according to an embodiment of this application;

[0023] Figure 5 A detailed flowchart of a product recommendation method according to an embodiment of this application is shown;

[0024] Figure 6 A schematic flowchart of a product recommendation device provided according to an embodiment of this application is shown.

[0025] The above figures include the following reference numerals:

[0026] 102. Processor; 104. Memory; 106. Transmission device; 108. Input / output device. Detailed Implementation

[0027] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0028] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0029] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate for the embodiments of this application described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0030] As described in the background section, product recommendations in the prior art are not accurate enough. To solve the above problems, embodiments of this application provide a product recommendation method, a product recommendation apparatus, a computer-readable storage medium, and a processor.

[0031] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0032] The methods and embodiments provided in this application can be executed on a mobile terminal, computer terminal, or similar computing device. Taking running on a mobile terminal as an example, Figure 1 This is a hardware structure block diagram of a mobile terminal for a product recommendation method according to an embodiment of the present invention. Figure 1 As shown, a mobile terminal may include one or more ( Figure 1Only one is shown in the diagram. A processor 102 (which may include, but is not limited to, a microprocessor MCU or a programmable logic device FPGA, etc.) and a memory 104 for storing data are also shown. The mobile terminal may further include a transmission device 106 for communication functions and an input / output device 108. Those skilled in the art will understand that... Figure 1 The structure shown is for illustrative purposes only and does not limit the structure of the mobile terminal described above. For example, the mobile terminal may also include components that are more... Figure 1 The more or fewer components shown, or having the same Figure 1 The different configurations shown.

[0033] The memory 104 can be used to store computer programs, such as application software programs and modules, like the computer program corresponding to the product recommendation method in this embodiment of the invention. The processor 102 executes various functional applications and data processing by running the computer program stored in the memory 104, thereby implementing the above-described method. The memory 104 may include high-speed random access memory and non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory 104 may further include memory remotely located relative to the processor 102, and these remote memories can be connected to the mobile terminal via a network. Examples of the aforementioned networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof. The transmission device 106 is used to receive or send data via a network. Specific examples of the aforementioned networks may include wireless networks provided by the mobile terminal's communication provider. In one example, the transmission device 106 includes a network interface controller (NIC), which can be connected to other network devices via a base station to communicate with the Internet. In one example, the transmission device 106 may be a radio frequency (RF) module, which is used to communicate with the Internet wirelessly.

[0034] This embodiment provides a product recommendation method that runs on a mobile terminal, computer terminal, or similar computing device. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Also, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0035] Figure 2 This is a flowchart of a product recommendation method according to an embodiment of this application. For example... Figure 2 As shown, the method includes the following steps:

[0036] Step S201: Obtain multiple dialogue data and multiple click behavior data of the user. The dialogue data is the dialogue of the user in the application software obtained before the current time, and the click behavior data is the click behavior of the user in the application software obtained at the current time.

[0037] Specifically, dialogue data refers to the conversation information between the user and bank consultants, while click behavior is the collection of user actions when using the client system. "Before the current moment" typically refers to the cutoff time for data acquisition, i.e., the latest time of data collection during clue analysis. Dialogue content may include text and voice, but if voice data is included, it needs to be converted to text format for subsequent text analysis. Click behavior data refers to any user actions performed within the application software, such as clicking on product details, purchasing products, or checking balances. Similar to dialogue data, click behavior data should also be data prior to the "current moment" to reflect the user's latest operating habits and actual needs. In this invention, considering the dynamic and real-time nature of behavioral data, we focus on the user's click behavior within the application software acquired at the "current moment." This helps to promptly capture changes in user behavior, especially those behaviors that may immediately generate push notification opportunities, such as product purchases.

[0038] Step S202: Input the dialogue data and the click behavior data into the bidirectional encoder representation model to extract at least the semantic relationship between the dialogue data and the click behavior data, and obtain a dialogue sequence and a behavior sequence. The dialogue sequence is used to represent the semantic information corresponding to the dialogue data, and the behavior sequence is used to represent the semantic information of the click behavior data.

[0039] Specifically, all conversations between the user and the account manager within a specific time window are used as input. These conversations are converted into text format to ensure they can be processed by a natural language processing model like BERT. The BERT model analyzes the semantics of each sentence, understanding the intent or emotion within, thus providing rich semantic information for subsequent sequence modeling. Every click the user makes in the banking application, such as browsing product details, checking account balances, or making transfers, is also converted into a text description. The conversation data processed by the BERT model is then transformed into a series of vectors, each vector corresponding to the semantic information of one conversation. This series of vectors constitutes the conversation sequence, reflecting the continuity and variability of the user-account manager interaction.

[0040] Step S203: Input the above dialogue sequence into the first converter model to extract the relationship between the data in the dialogue sequence and obtain a first vector; and input the above behavior sequence into the second converter model to extract the relationship between the data in the behavior sequence and obtain a second vector.

[0041] Specifically, first, we feed the dialogue sequence obtained from the BERT model as input to the first transformer model. This sequence consists of a series of vectors, each representing the semantic information of a dialogue. The Transformer model is particularly well-suited for processing this type of sequential data because it effectively captures the dependencies between different elements in the sequence. When processing the dialogue sequence, the Transformer uses a self-attention mechanism to evaluate the strength of the association between each vector in the sequence and other vectors, thereby understanding the dynamic evolution and contextual dependencies of the dialogue. The behavior sequence is then fed into the second transformer model. Similar to the dialogue sequence, the behavior sequence consists of a series of vectors, each corresponding to the semantic information of a click behavior. The second transformer model also utilizes a self-attention mechanism to understand the internal relationships within the behavior sequence. It analyzes how the user's click behaviors are related at different time points. Ultimately, the second transformer model generates another vector, which we call the second vector. This vector summarizes the overall characteristics of the user's click behavior, and through the model's sequence modeling capabilities, it also reflects the changing trends of user behavior over time and implicit preferences.

[0042] Step S204: Perform a weighted concatenation process on the first vector and the second vector to obtain the third vector.

[0043] Specifically, the first vector represents the semantic information of the user's dialogue sequence, while the second vector reflects the semantic information of the user's click behavior. Both vectors are numerical representations of the user behavior features extracted by the model. The weighted first and second vectors are then concatenated to form a longer vector, the third vector. The concatenation operation is performed on the last dimension of the vectors, meaning that the two vectors are connected along this dimension to form a new vector that incorporates comprehensive information about dialogue intent and click behavior. The third vector is a vector representation that integrates user dialogue intent and behavioral patterns, providing a more comprehensive understanding of user behavior than a single intent vector or behavior vector.

[0044] Step S205: Generate prompt information based on the third vector mentioned above, and make recommendations to the user based on the prompt information mentioned above.

[0045] Specifically, the third vector at this point contains information about the user's expressed intent in the dialogue and their click behavior on the application. After weighted concatenation, it more accurately reflects the user's true needs and preferences. By inputting the third vector into the subsequent classification model, we can predict the user's most likely current needs and interests. Based on the output of the classification model, we can filter out the highest probability prompts, i.e., which products or services the user is most likely to be interested in.

[0046] In practical applications, such as Figure 3 As shown, the above steps can be implemented as follows: The second-stage training process is divided into an input layer, an intermediate layer, and an output layer. The input layer takes in the dialogue text intent sequence and click behavior sequence. The intermediate layer receives the input from the input layer, then processes it through the BERT2 model, Transformer1 model, and Transformer2 model to generate intent sequence vectors and click behavior sequence vectors. These vectors are then weighted and concatenated, and finally connected to a fully connected layer to transform the vector dimensions. The output layer outputs the intent prediction.

[0047] This embodiment first acquires multiple dialogue data and multiple click behavior data of the user. The dialogue data consists of user conversations within the application software acquired up to the current moment, while the click behavior data consists of user click behaviors within the application software acquired at the current moment. Then, the dialogue data and click behavior data are input into a bidirectional encoder representation model to extract at least the semantic relationship between them. The dialogue sequence is used to represent the semantic information corresponding to the dialogue data. Next, the dialogue sequence is input into a first converter model, and the behavior sequence is input into a second converter model. Finally, the first and second vectors are weighted and concatenated to obtain a third vector. Based on the third vector, a prompt message is generated, and recommendations are made to the user based on the prompt message. By combining the dialogue data and click behavior data, and then performing intent recognition to obtain the dialogue sequence and behavior sequence, sequence modeling is then performed to obtain vector representations. These vectors are then weighted and concatenated to obtain a combined vector, i.e., the third vector. Finally, the customer's intent is predicted based on the combined vector, thereby generating a prompt message, which helps in accurate recommendations and solves the problem of inaccurate product recommendations to users in existing technologies.

[0048] In the specific implementation process, the above step S201 can be implemented through the following steps: Step S2011, obtain multiple dialogue texts of the above user, each of the above dialogue texts includes at least one Chinese character to be mapped; Step S2012, obtain the Chinese character sequence number mapping relationship, and determine the target sequence number according to the above Chinese character sequence number mapping relationship and the above Chinese character to be mapped, wherein the above Chinese character sequence number mapping relationship is a pre-set mapping relationship between Chinese characters and sequence numbers, and the above target sequence number is the above sequence number corresponding to the above Chinese character that is the same as the above Chinese character to be mapped in the above Chinese character sequence number mapping relationship; Step S2013, obtain the sequence number vector mapping relationship, and determine the target vector according to the above target sequence number and the above sequence number vector mapping relationship, multiple above target vectors constitute the above dialogue data, wherein the above sequence number vector mapping relationship is a pre-set mapping relationship between sequence numbers and vectors, and the above target vector is the above vector corresponding to the above sequence number that is the same as the above target sequence number in the above sequence number vector mapping relationship. This method, through Chinese character sequence number mapping and sequence number vector mapping, transforms and preserves the semantic information in the original dialogue text, thus avoiding information loss during data preprocessing and ensuring the accuracy and effectiveness of model training. The sequence number vector mapping relationship provides a standard input format for deep learning models, enabling them to process dialogue data from any user without requiring special adjustments for different text formats or language characteristics.

[0049] Specifically, the system first collects multiple dialogue texts generated by users during their interactions with the banking application. These texts contain various intentions and needs expressed by users, serving as a key source of information for generating prompts. To convert the text data into numerical vectors, the system uses a pre-set mapping relationship between Chinese character numbers. After obtaining the number of each Chinese character in the dialogue text, the system converts each character number into a fixed-dimensional vector according to the pre-set number vector mapping relationship. Multiple target vectors are then integrated into a vector representation of the dialogue data, i.e., the dialogue sequence. This process ensures that the semantic information of the entire dialogue is fully encoded, providing a data foundation for subsequent deep learning model analysis.

[0050] In specific implementation, step S202 can be achieved through the following steps: Step S2021, inputting multiple dialogue data into a first model to perform intent recognition on the dialogue data and obtain an intent sequence, wherein the intent sequence is used to represent the intent of the dialogue data; Step S2022, inputting the intent sequence and the click behavior data into a second model to extract at least the semantic relationship between the intent sequence and the click behavior data, obtaining the dialogue sequence and the behavior sequence. The bidirectional encoder representation model includes the first model and the second model. By inputting user intent and click behavior data together into the second model for analysis, this invention can achieve a deep integration and understanding of user needs, which is beneficial for generating prompts that better match the user's actual needs. Since the second model considers not only the user's dialogue intent but also the user's actual behavior, the generated prompts can more accurately reflect the user's real needs and current purchasing tendencies, thereby improving the targeting and timeliness of the push strategy. This invention can generate personalized prompts based on user dialogue and click behavior, which helps recommend the most relevant products or services to users, thereby improving user experience and satisfaction. It also benefits banks by improving customer conversion and retention rates.

[0051] The process begins by inputting collected dialogue data into a first model. This dialogue data refers to the conversation records between the user and the account manager within the application software, containing multiple intentions and needs expressed by the user. The primary task of the first model is to identify the intent within this dialogue data, that is, to identify and understand the specific purpose or need expressed by the user in the dialogue. These intents may include information queries, product inquiries, service requests, etc. The first model is typically a deep learning model, such as the bidirectional encoder representation model, which can process natural language text. Through pre-training and fine-tuning, it can accurately identify the intents conveyed by the user in the dialogue and output them in a structured form, namely, an intent sequence. The intent sequence contains each user intent identified from the dialogue data. After obtaining the intent sequence, the next step is to input the intent sequence and the user's click behavior data together into a second model for comprehensive analysis. The second model is also built based on the bidirectional encoder representation model, but its design and training are aimed at understanding the semantic relationship between the intent sequence and click behavior, and how to integrate this information to generate a more comprehensive representation of user behavior. The purpose of the second model is to capture the correlation between user intent and actual user behavior, which helps to more accurately predict user behavior and needs, thereby generating high-quality prompts. Figure 4 As shown, the first model mentioned above includes an input layer, an intermediate layer, and an output layer. The input layer is used to receive dialogue text messages, the intermediate layer is used to process the received messages using the BERT1 model, and the output layer is used to output intent labels.

[0052] In specific implementation, step S2021 can be achieved through the following steps: Step S20211, input step: input the dialogue data one by one into multiple encoders to classify the dialogue data and obtain at least one intent label. The intent label represents the intent of the dialogue data. If there are multiple intent labels, their order is arranged from smallest to largest predicted probability. Step S20212, repeat step: repeat the above input step multiple times until multiple dialogue data are input, resulting in an intent sequence composed of multiple intent labels. By inputting each dialogue data piece into multiple encoders for processing, the system can understand user intent from different dimensions and depths. The construction of the intent sequence allows the system to track the coherence and evolution of user intent over time. Once the intent sequence is constructed, the system can generate prompts based on the information within the sequence. These prompts can reveal the products or services that users are most likely to be interested in, providing personalized recommendations to customers and improving the bank's push efficiency and customer satisfaction.

[0053] Specifically, the focus is on deep analysis of single dialogue data, aiming to identify and extract the user's intentions or needs expressed in the dialogue—that is, intent recognition. Multiple encoders within a deep learning framework are used to process the dialogue data. These encoders can extract features of the message from different angles or levels, such as grammatical structure, emotional tone, and contextual dependencies, ensuring a comprehensive understanding of the text's meaning. Each encoder may focus on different aspects of the dialogue data, working together to generate a more comprehensive vector representation. Once the dialogue data is encoded and converted into vector form, it is fed into a classification component for processing. The classification component can predict the intent labels contained in the dialogue data based on the knowledge learned by the model. When multiple intent labels are predicted, they are arranged in ascending order of predicted probability, meaning that the later labels in the dialogue data represent the user's intent that the model deems more likely to be expressed.

[0054] In specific implementation, the first converter model includes a first multi-head self-attention layer, a first feedforward network layer, and a first residual connection layer connected in sequence, and the second converter model includes a second multi-head self-attention layer, a second feedforward network layer, and a second residual connection layer connected in sequence. Step S203 can be implemented through the following steps: Step S2031, input the dialogue sequence into the first converter model to extract the relationship between the data in the dialogue sequence and obtain a first vector; Step S2032, input the dialogue sequence into the first multi-head self-attention layer to obtain first output data, wherein the first multi-head self-attention layer is used to calculate the relationship between the data in the dialogue sequence; Step S2033, input the first output data into the first feedforward network layer to obtain... The process proceeds as follows: First output data is processed by a first feedforward network layer, which performs nonlinear transformation and feature extraction. Second output data is then fed into a first residual connection layer to obtain a first vector. Third output data is generated by sequentially feeding the behavior sequence into a second multi-head self-attention layer, which calculates the relationships between data points in the behavior sequence. Fourth output data is generated by feeding the third output data into a second feedforward network layer, which performs nonlinear transformation and feature extraction. Fifth output data is generated by feeding the fourth output data into a second residual connection layer to obtain a second vector. This multi-head self-attention mechanism allows the model to understand the relationships between data points in the sequence more precisely, extending beyond adjacent data points to include long-range dependencies. This is crucial for revealing deeper user intentions and behavioral patterns. By combining the analysis results of intent sequences and behavior sequences, prompts that are closer to users' actual needs can be generated. This helps banks or financial institutions optimize their push strategies, improve the efficiency of push notifications by account managers, and increase customer satisfaction.

[0055] Specifically, when performing in-depth analysis of the identified intent sequence, it is first passed as input to the first multi-head self-attention layer in the first converter model. The multi-head self-attention mechanism allows the model to focus on different parts of the sequence data simultaneously, which is equivalent to understanding the sequence from multiple perspectives, improving the model's ability to analyze complex sequences. The first output data is then fed into the first feedforward network layer. This layer is responsible for performing non-linear transformations and feature extraction on the data, i.e., processing the vector through activation functions and fully connected layers to capture higher-order features. Finally, the second output data is processed by the first residual connection layer to obtain the first vector. The introduction of residual connections is to alleviate the gradient vanishing problem during neural network training, ensuring that the model can learn long-term dependencies. The processing flow of the second converter model is similar to that of the first converter model, but it targets user behavior sequence data. Behavior sequences may include a series of user actions on the application software, such as browsing, clicking, and purchasing. The behavior sequence data enters the second multi-head self-attention layer, which also focuses on calculating the relationships between different operations in the sequence and the impact of the operation time order on user behavior, ultimately obtaining the third output data, which reflects the relationships between the data in the behavior sequence. Similar to dialogue sequence processing, the third output data, after undergoing a nonlinear transformation in the second feedforward network layer, yields the fourth output data. Here, the nonlinear transformation further extracts features from the behavior sequence. The fourth output data is then processed through the second residual connection layer to output the second vector.

[0056] In specific implementation, step S204 can be achieved through the following steps: Step S2041, obtain the first hyperparameter and the second hyperparameter, wherein the first hyperparameter is the weight of the first vector, the second hyperparameter is the weight of the second vector, and the sum of the first hyperparameter and the second hyperparameter is 1; Step S2042, calculate the product of the first hyperparameter and the first vector to obtain the first vector to be concatenated, and calculate the product of the second hyperparameter and the second vector to obtain the second vector to be concatenated; Step S2043, concatenate the first vector to be concatenated and the second vector to be concatenated along a predetermined dimension to obtain the third vector, wherein the predetermined dimension is the last dimension of the first vector and the second vector. This balances the influence of different data sources. By adjusting the vector weights using the first and second hyperparameters, this step ensures fair treatment of dialogue sequences and behavior sequences in comprehensive analysis. This is crucial for generating prompts that consider both the user's expressed intent and actual behavioral patterns. Feature fusion and concatenation effectively combine feature representations from different data sources into a single vector, providing a unified data input format for subsequent multi-task learning or deep learning. This fusion not only increases the vector's dimensionality but, more importantly, carries more comprehensive user information, improving the accuracy and richness of prompts. The flexibility of the weighted concatenation strategy lies in its ability to adapt to different user analysis needs by adjusting hyperparameters α and β.

[0057] Specifically, hyperparameters play a role in adjusting the relative importance of the first and second vectors in the final composite representation. The hyperparameters are set based on the experience of domain experts or through experimental optimization, and their sum is fixed at 1. This ensures that there is no imbalance in weight ratios during the concatenation process. The first hyperparameter α represents the proportion of the dialogue sequence vector in the final representation, while the second hyperparameter β indicates the weight of the behavior sequence vector. By adjusting α and β, the priority of one data source relative to the other can be flexibly emphasized. The first vector to be concatenated is concatenated with the second vector along a predetermined dimension to obtain the third vector. The choice of the predetermined dimension ensures that the concatenation operation does not destroy the original structure of the vectors, but rather smoothly integrates them.

[0058] In specific implementation, step S205 can be achieved through the following steps: Step S2051, input the third vector into the fully connected layer to transform the vector space of the third vector, obtaining the fourth vector and its predicted probability; Step S2052, compare the predicted probability of the fourth vector with the predetermined probability, and determine the fourth vector as the target intent if the predicted probability of the fourth vector is greater than or equal to the predetermined probability; Step S2053, obtain the intent tree, and determine the prompt information based on the target intent and the intent tree. The intent tree is used to represent the mapping relationship between the user's intent and the clue content, and the prompt information is the clue content corresponding to the intent in the intent tree that is the same as the target intent. By generating prompt information based on the third vector, banks can quickly locate users with specific needs, thereby improving the targeting and efficiency of push notifications. Compared to traditional prompt information generation methods that rely on manual analysis, this method greatly reduces time and manpower costs. Personalized recommendation strategies can improve user experience because they ensure that the recommended content highly matches the user's interests and needs. Users who feel that the bank is paying attention to their needs may increase their trust and loyalty.

[0059] Specifically, generating prompts using third vectors involves transforming vector representations into specific user needs or behavioral predictions. This typically requires designing or invoking an additional model layer, such as a fully connected layer or a custom neural network component, that maps the third vector to a range of possible prompt categories. Numerous prompts may be generated, covering potential user needs and interests. The system filters out less likely prompts based on certain thresholds and prioritizes the remaining prompts according to their predictions, ensuring that the prompts most relevant to the user's current situation are adopted by the account manager first. The prompts are ultimately transformed into actionable information that directly guides the account manager in taking action.

[0060] To enable those skilled in the art to better understand the technical solution of this application, the implementation process of the product recommendation method of this application will be described in detail below with reference to specific embodiments.

[0061] This embodiment relates to a specific product recommendation method, such as... Figure 5 As shown, it includes the following steps:

[0062] Step S1: Call the API to obtain the user's conversations in the past three days and call the API to obtain the user's click behavior tag sequence on the bank application software in the past three days;

[0063] Step S2: Call BERT model 1 to identify the intent and obtain the intent label sequence. Use BERT model 2 to obtain the initial vector representation of the label and the initial vector representation of each click behavior label.

[0064] Step S3: Use Transformer1 to model the sequence to obtain a vector representation of the intent sequence, and use Transformer2 to model the sequence to obtain a vector representation of the click behavior sequence;

[0065] Step S4: Weighted concatenation of vectors;

[0066] Step S5: Predict the intent label;

[0067] Step S6: Map intent tags to marketing leads.

[0068] This application also provides a product recommendation device. It should be noted that the product recommendation device of this application can be used to execute the product recommendation method provided in this application. This device is used to implement the above embodiments and preferred embodiments; details already described will not be repeated. As used below, the term "module" can refer to a combination of software and / or hardware that implements a predetermined function. Although the device described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0069] The following describes the product recommendation device provided in the embodiments of this application.

[0070] Figure 6 This is a schematic diagram of a product recommendation device according to an embodiment of this application. Figure 6 As shown, the device includes:

[0071] The acquisition unit 10 is used to acquire multiple dialogue data and multiple click behavior data of the user. The dialogue data is the dialogue of the user in the application software acquired before the current time, and the click behavior data is the click behavior of the user in the application software acquired at the current time.

[0072] Specifically, dialogue data refers to the conversation information between the user and bank consultants, while click behavior is the collection of user actions when using the client system. "Before the current moment" typically refers to the cutoff time for data acquisition, i.e., the latest time of data collection during clue analysis. Dialogue content may include text and voice, but if voice data is included, it needs to be converted to text format for subsequent text analysis. Click behavior data refers to any user actions performed within the application software, such as clicking on product details, purchasing products, or checking balances. Similar to dialogue data, click behavior data should also be data prior to the "current moment" to reflect the user's latest operating habits and actual needs. In this invention, considering the dynamic and real-time nature of behavioral data, we focus on the user's click behavior within the application software acquired at the "current moment." This helps to promptly capture changes in user behavior, especially those behaviors that may immediately generate push notification opportunities, such as product purchases.

[0073] The first input unit 20 is used to input the dialogue data and the click behavior data into the bidirectional encoder representation model to at least extract the semantic relationship between the dialogue data and the click behavior data to obtain a dialogue sequence and a behavior sequence. The dialogue sequence is used to represent the semantic information corresponding to the dialogue data, and the behavior sequence is used to represent the semantic information of the click behavior data.

[0074] Specifically, all conversations between the user and the account manager within a specific time window are used as input. These conversations are converted into text format to ensure they can be processed by a natural language processing model like BERT. The BERT model analyzes the semantics of each sentence, understanding the intent or emotion within, thus providing rich semantic information for subsequent sequence modeling. Every click the user makes in the banking application, such as browsing product details, checking account balances, or making transfers, is also converted into a text description. The conversation data processed by the BERT model is then transformed into a series of vectors, each vector corresponding to the semantic information of one conversation. This series of vectors constitutes the conversation sequence, reflecting the continuity and variability of the user-account manager interaction.

[0075] The second input unit 30 is used to input the dialogue sequence into the first converter model to extract the relationship between the data in the dialogue sequence and obtain a first vector, and to input the behavior sequence into the second converter model to extract the relationship between the data in the behavior sequence and obtain a second vector.

[0076] Specifically, first, we feed the dialogue sequence obtained from the BERT model as input to the first transformer model. This sequence consists of a series of vectors, each representing the semantic information of a dialogue. The Transformer model is particularly well-suited for processing this type of sequential data because it effectively captures the dependencies between different elements in the sequence. When processing the dialogue sequence, the Transformer uses a self-attention mechanism to evaluate the strength of the association between each vector in the sequence and other vectors, thereby understanding the dynamic evolution and contextual dependencies of the dialogue. The behavior sequence is then fed into the second transformer model. Similar to the dialogue sequence, the behavior sequence consists of a series of vectors, each corresponding to the semantic information of a click behavior. The second transformer model also utilizes a self-attention mechanism to understand the internal relationships within the behavior sequence. It analyzes how the user's click behaviors are related at different time points. Ultimately, the second transformer model generates another vector, which we call the second vector. This vector summarizes the overall characteristics of the user's click behavior, and through the model's sequence modeling capabilities, it also reflects the changing trends of user behavior over time and implicit preferences.

[0077] The processing unit 40 is used to perform weighted concatenation processing on the first vector and the second vector to obtain the third vector.

[0078] Specifically, the first vector represents the semantic information of the user's dialogue sequence, while the second vector reflects the semantic information of the user's click behavior. Both vectors are numerical representations of the user behavior features extracted by the model. The weighted first and second vectors are then concatenated to form a longer vector, the third vector. The concatenation operation is performed on the last dimension of the vectors, meaning that the two vectors are connected along this dimension to form a new vector that incorporates comprehensive information about dialogue intent and click behavior. The third vector is a vector representation that integrates user dialogue intent and behavioral patterns, providing a more comprehensive understanding of user behavior than a single intent vector or behavior vector.

[0079] Recommendation unit 50 is used to generate prompt information based on the third vector mentioned above, and to make recommendations to the user based on the prompt information mentioned above.

[0080] Specifically, the third vector at this point contains information about the user's expressed intent in the dialogue and their click behavior on the application. After weighted concatenation, it more accurately reflects the user's true needs and preferences. By inputting the third vector into the subsequent classification model, we can predict the user's most likely current needs and interests. Based on the output of the classification model, we can filter out the highest probability prompts, i.e., which products or services the user is most likely to be interested in.

[0081] In practical applications, such as Figure 3As shown, the above steps can be implemented as follows: The second-stage training process is divided into an input layer, an intermediate layer, and an output layer. The input layer takes in the dialogue text intent sequence and click behavior sequence. The intermediate layer receives the input from the input layer, then processes it through the BERT2 model, Transformer1 model, and Transformer2 model to generate intent sequence vectors and click behavior sequence vectors. These vectors are then weighted and concatenated, and finally connected to a fully connected layer to transform the vector dimensions. The output layer outputs the intent prediction.

[0082] In this embodiment, the acquisition unit acquires multiple dialogue data and multiple click behavior data of the user. The dialogue data consists of dialogues between the user and the user within the application software acquired up to the current moment, while the click behavior data consists of click behaviors acquired at the current moment. The first input unit inputs the dialogue data and click behavior data into a bidirectional encoder representation model to extract at least the semantic relationship between the dialogue data and click behavior data. The dialogue sequence is used to represent the semantic information corresponding to the dialogue data. The second input unit inputs the dialogue sequence into a first converter model and the behavior sequence into a second converter model. The processing unit performs weighted concatenation processing on the first and second vectors to obtain a third vector. The recommendation unit generates prompt information based on the third vector and makes recommendations to the user based on the prompt information. By combining the dialogue data and click behavior data, and then performing intent recognition to obtain the dialogue sequence and behavior sequence, sequence modeling is then performed to obtain vector representations. The vectors are then weighted and concatenated to obtain a combined vector, i.e., the third vector. Finally, the customer's intent is predicted based on the combined vector, thereby generating prompt information, which helps to make accurate recommendations and solves the problem of inaccurate product recommendations to users in the prior art.

[0083] In the specific implementation process, the above-mentioned acquisition unit includes a first acquisition module, a second acquisition module, and a third acquisition module. The first acquisition module is used to acquire multiple dialogue texts of the user, each of which includes at least one Chinese character to be mapped. The second acquisition module is used to acquire the Chinese character sequence number mapping relationship and determine the target sequence number based on the Chinese character sequence number mapping relationship and the Chinese character to be mapped. The Chinese character sequence number mapping relationship is a pre-set mapping relationship between Chinese characters and sequences, and the target sequence number is the sequence number corresponding to the Chinese character that is the same as the Chinese character to be mapped in the Chinese character sequence number mapping relationship. The third acquisition module is used to acquire the sequence number vector mapping relationship and determine the target vector based on the target sequence number and the sequence number vector mapping relationship. Multiple target vectors constitute the dialogue data. The sequence number vector mapping relationship is a pre-set mapping relationship between sequences and vectors, and the target vector is the vector corresponding to the sequence number that is the same as the target sequence number in the sequence number vector mapping relationship. This method, through Chinese character sequence number mapping and sequence number vector mapping, transforms and preserves the semantic information in the original dialogue text, thus avoiding information loss during data preprocessing and ensuring the accuracy and effectiveness of model training. The sequence number vector mapping relationship provides a standard input format for deep learning models, enabling them to process dialogue data from any user without requiring special adjustments for different text formats or language characteristics.

[0084] Specifically, the system first collects multiple dialogue texts generated by users during their interactions with the banking application. These texts contain various intentions and needs expressed by users, serving as a key source of information for generating prompts. To convert the text data into numerical vectors, the system uses a pre-set mapping relationship between Chinese character numbers. After obtaining the number of each Chinese character in the dialogue text, the system converts each character number into a fixed-dimensional vector according to the pre-set number vector mapping relationship. Multiple target vectors are then integrated into a vector representation of the dialogue data, i.e., the dialogue sequence. This process ensures that the semantic information of the entire dialogue is fully encoded, providing a data foundation for subsequent deep learning model analysis.

[0085] In specific implementation, the first input unit includes a first input module and a second input module. The first input module is used to input multiple dialogue data into a first model to perform intent recognition on the dialogue data and obtain an intent sequence, wherein the intent sequence represents the intent of the dialogue data. The second input module is used to input the intent sequence and the click behavior data into a second model to extract at least the semantic relationship between the intent sequence and the click behavior data, obtaining the dialogue sequence and the behavior sequence. The bidirectional encoder representation model includes the first model and the second model. By inputting user intent and click behavior data together into the second model for analysis, this invention can achieve a deep integration and understanding of user needs, which is beneficial for generating prompts that better match the user's actual needs. Since the second model considers not only the user's dialogue intent but also the user's actual behavior, the generated prompts can more accurately reflect the user's real needs and current purchasing tendencies, thereby improving the targeting and timeliness of the push strategy. This invention can generate personalized prompts based on user dialogue and click behavior, which helps to recommend the most relevant products or services to users, thereby improving user experience and satisfaction, and also benefits banks in improving customer conversion and retention rates.

[0086] The process begins by inputting collected dialogue data into a first model. This dialogue data refers to the conversation records between the user and the account manager within the application software, containing multiple intentions and needs expressed by the user. The primary task of the first model is to identify the intent within this dialogue data, that is, to identify and understand the specific purpose or need expressed by the user in the dialogue. These intents may include information queries, product inquiries, service requests, etc. The first model is typically a deep learning model, such as the bidirectional encoder representation model, which can process natural language text. Through pre-training and fine-tuning, it can accurately identify the intents conveyed by the user in the dialogue and output them in a structured form, namely, an intent sequence. The intent sequence contains each user intent identified from the dialogue data. After obtaining the intent sequence, the next step is to input the intent sequence and the user's click behavior data together into a second model for comprehensive analysis. The second model is also built based on the bidirectional encoder representation model, but its design and training are aimed at understanding the semantic relationship between the intent sequence and click behavior, and how to integrate this information to generate a more comprehensive representation of user behavior. The purpose of the second model is to capture the correlation between user intent and actual user behavior, which helps to more accurately predict user behavior and needs, thereby generating high-quality prompts. Figure 4As shown, the first model mentioned above includes an input layer, an intermediate layer, and an output layer. The input layer is used to receive dialogue text messages, the intermediate layer is used to process the received messages using the BERT1 model, and the output layer is used to output intent labels.

[0087] In the specific implementation, the first input module includes an input submodule and a repetition submodule. The input submodule is used for the input step, sequentially inputting the dialogue data into multiple encoders to classify the dialogue data and obtain at least one intent label. This intent label represents the intent of the dialogue data, and when there are multiple intent labels, their order is arranged from smallest to largest predicted probability. The repetition submodule is used for the repetition step, repeating the input step multiple times until multiple dialogue data are input, resulting in an intent sequence composed of multiple intent labels. By inputting each dialogue data piece into multiple encoders for processing, the system can understand user intent from different dimensions and depths. The construction of the intent sequence allows the system to track the coherence and evolution of user intent over time. Once the intent sequence is constructed, the system can generate prompts based on the information within the sequence. These prompts can reveal the products or services the user is most likely to be interested in, providing personalized recommendations and improving the bank's push efficiency and customer satisfaction.

[0088] Specifically, the focus is on deep analysis of single dialogue data, aiming to identify and extract the user's intentions or needs expressed in the dialogue—that is, intent recognition. Multiple encoders within a deep learning framework are used to process the dialogue data. These encoders can extract features of the message from different angles or levels, such as grammatical structure, emotional tone, and contextual dependencies, ensuring a comprehensive understanding of the text's meaning. Each encoder may focus on different aspects of the dialogue data, working together to generate a more comprehensive vector representation. Once the dialogue data is encoded and converted into vector form, it is fed into a classification component for processing. The classification component can predict the intent labels contained in the dialogue data based on the knowledge learned by the model. When multiple intent labels are predicted, they are arranged in ascending order of predicted probability, meaning that the later labels in the dialogue data represent the user's intent that the model deems more likely to be expressed.

[0089] In specific implementation, the first converter model includes a first multi-head self-attention layer, a first feedforward network layer, and a first residual connection layer connected in sequence; the second converter model includes a second multi-head self-attention layer, a second feedforward network layer, and a second residual connection layer connected in sequence; the second input unit includes a third input module, a fourth input module, a fifth input module, a sixth input module, a seventh input module, an eighth input module, and a ninth input module. The third input module is used to input the dialogue sequence into the first converter model to extract the relationship between the data in the dialogue sequence and obtain a first vector; the fourth input module is used to input the dialogue sequence into the first multi-head self-attention layer to obtain first output data, wherein the first multi-head self-attention layer is used to calculate the relationship between the data in the dialogue sequence; the fifth input module is used to... The first output data is input to the first feedforward network layer to obtain the second output data, wherein the first feedforward network layer is used to perform nonlinear transformation and feature extraction on the first output data; the sixth input module is used to input the second output data to the first residual connection layer to obtain the first vector; the seventh input module is used to input the behavior sequence sequentially to the second multi-head self-attention layer to obtain the third output data, wherein the second multi-head self-attention layer is used to calculate the relationship between the data in the behavior sequence; the eighth input module is used to input the third output data to the second feedforward network layer to obtain the fourth output data, wherein the second feedforward network layer is used to perform nonlinear transformation and feature extraction on the third output data; the ninth input module is used to input the fourth output data to the second residual connection layer to obtain the second vector. In this way, by using the multi-head self-attention mechanism, the model can understand the relationship between the data in the sequence in greater detail, not only limited to adjacent data points, but also including long-distance dependencies, which is crucial for revealing the user's deep-seated intentions and behavioral patterns. By combining the analysis results of intent sequences and behavior sequences, prompts that are closer to users' actual needs can be generated. This helps banks or financial institutions optimize their push strategies, improve the efficiency of push notifications by account managers, and increase customer satisfaction.

[0090] Specifically, when performing in-depth analysis of the identified intent sequence, it is first passed as input to the first multi-head self-attention layer in the first converter model. The multi-head self-attention mechanism allows the model to focus on different parts of the sequence data simultaneously, which is equivalent to understanding the sequence from multiple perspectives, improving the model's ability to analyze complex sequences. The first output data is then fed into the first feedforward network layer. This layer is responsible for performing non-linear transformations and feature extraction on the data, i.e., processing the vector through activation functions and fully connected layers to capture higher-order features. Finally, the second output data is processed by the first residual connection layer to obtain the first vector. The introduction of residual connections is to alleviate the gradient vanishing problem during neural network training, ensuring that the model can learn long-term dependencies. The processing flow of the second converter model is similar to that of the first converter model, but it targets user behavior sequence data. Behavior sequences may include a series of user actions on the application software, such as browsing, clicking, and purchasing. The behavior sequence data enters the second multi-head self-attention layer, which also focuses on calculating the relationships between different operations in the sequence and the impact of the operation time order on user behavior, ultimately obtaining the third output data, which reflects the relationships between the data in the behavior sequence. Similar to dialogue sequence processing, the third output data, after undergoing a nonlinear transformation in the second feedforward network layer, yields the fourth output data. Here, the nonlinear transformation further extracts features from the behavior sequence. The fourth output data is then processed through the second residual connection layer to output the second vector.

[0091] In the specific implementation process, the above processing unit includes a fourth acquisition module, a calculation module, and a splicing module. The fourth acquisition module acquires a first hyperparameter and a second hyperparameter, where the first hyperparameter is the weight of the first vector, the second hyperparameter is the weight of the second vector, and the sum of the first and second hyperparameters is 1. The calculation module calculates the product of the first hyperparameter and the first vector to obtain a first vector to be spliced, and calculates the product of the second hyperparameter and the second vector to obtain a second vector to be spliced. The splicing module splices the first and second vectors to be spliced ​​along a predetermined dimension to obtain a third vector, where the predetermined dimension is the last dimension of the first and second vectors. This balances the influence of different data sources. By adjusting the vector weights using the first and second hyperparameters, this step ensures fair treatment of dialogue sequences and behavior sequences in the comprehensive analysis. This is crucial for generating prompts that consider both the user's expressed intent and actual behavioral patterns. Feature fusion and concatenation effectively combine feature representations from different data sources into a single vector, providing a unified data input format for subsequent multi-task learning or deep learning. This fusion not only increases the vector's dimensionality but, more importantly, carries more comprehensive user information, improving the accuracy and richness of prompts. The flexibility of the weighted concatenation strategy lies in its ability to adapt to different user analysis needs by adjusting hyperparameters α and β.

[0092] Specifically, hyperparameters play a role in adjusting the relative importance of the first and second vectors in the final composite representation. The hyperparameters are set based on the experience of domain experts or through experimental optimization, and their sum is fixed at 1. This ensures that there is no imbalance in weight ratios during the concatenation process. The first hyperparameter α represents the proportion of the dialogue sequence vector in the final representation, while the second hyperparameter β indicates the weight of the behavior sequence vector. By adjusting α and β, the priority of one data source relative to the other can be flexibly emphasized. The first vector to be concatenated is concatenated with the second vector along a predetermined dimension to obtain the third vector. The choice of the predetermined dimension ensures that the concatenation operation does not destroy the original structure of the vectors, but rather smoothly integrates them.

[0093] In its specific implementation, the recommendation unit includes a tenth input module, a comparison module, and a fifth module. The tenth input module inputs the third vector into a fully connected layer to transform its vector space, obtaining a fourth vector and its predicted probability. The comparison module compares the predicted probability of the fourth vector with a predetermined probability; if the predicted probability is greater than or equal to the predetermined probability, the fourth vector is identified as the target intent. The fifth module obtains an intent tree and, based on the target intent and the intent tree, determines the prompt information. The intent tree represents the mapping relationship between the user's intent and the clue content, and the prompt information is the clue content corresponding to the intent in the intent tree that is identical to the target intent. By generating prompt information based on the third vector, banks can quickly locate users with specific needs, thereby improving the targeting and efficiency of push notifications. Compared to traditional prompt generation methods that rely on manual analysis, this method significantly reduces time and manpower costs. Personalized recommendation strategies enhance user experience because they ensure that recommended content closely matches the user's interests and needs. Users who perceive the bank's attention to their needs may increase their trust and loyalty.

[0094] Specifically, generating prompts using third vectors involves transforming vector representations into specific user needs or behavioral predictions. This typically requires designing or invoking an additional model layer, such as a fully connected layer or a custom neural network component, that maps the third vector to a range of possible prompt categories. Numerous prompts may be generated, covering potential user needs and interests. The system filters out less likely prompts based on certain thresholds and prioritizes the remaining prompts according to their predictions, ensuring that the prompts most relevant to the user's current situation are adopted by the account manager first. The prompts are ultimately transformed into actionable information that directly guides the account manager in taking action.

[0095] The aforementioned product recommendation device includes a processor and a memory. The acquisition unit, first input unit, second input unit, processing unit, and recommendation unit are all stored as program units in the memory. The processor executes these program units stored in the memory to achieve the corresponding functions. All of the above modules are located in the same processor; alternatively, the modules may be located in different processors in any combination.

[0096] The processor contains a kernel, which retrieves the corresponding program units from memory. One or more kernels can be configured, and product recommendations can be made by adjusting kernel parameters.

[0097] The memory may include non-permanent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM, and the memory includes at least one memory chip.

[0098] This invention provides a computer-readable storage medium including a stored program, wherein, when the program is executed, it controls the device containing the computer-readable storage medium to perform the product recommendation method.

[0099] This invention provides a processor for running a program, wherein the program executes the product recommendation method.

[0100] It is obvious to those skilled in the art that the modules or steps of the present invention described above can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. They can be implemented using computer-executable program code, and thus can be stored in a storage device for execution by a computing device. In some cases, the steps shown or described can be performed in a different order than those described herein, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, the present invention is not limited to any particular combination of hardware and software.

[0101] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0102] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0103] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0104] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0105] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0106] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0107] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0108] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0109] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0110] As can be seen from the above description, the embodiments of this application achieve the following technical effects:

[0111] 1) The product recommendation method of this application first acquires multiple dialogue data and multiple click behavior data of the user. The dialogue data consists of dialogues between the user and the user within the application software acquired before the current moment, and the click behavior data consists of click behaviors of the user within the application software acquired at the current moment. Then, the dialogue data and click behavior data are input into a bidirectional encoder representation model to extract at least the semantic relationship between the dialogue data and the click behavior data. The dialogue sequence is used to represent the semantic information corresponding to the dialogue data. Next, the dialogue sequence is input into a first converter model, and the behavior sequence is input into a second converter model. Finally, the first and second vectors are weighted and concatenated to obtain a third vector. Based on the third vector, a prompt message is generated, and recommendations are made to the user based on the prompt message. By combining the dialogue data and click behavior data, and then performing intent recognition to obtain the dialogue sequence and behavior sequence, sequence modeling is then performed to obtain vector representations. These vectors are then weighted and concatenated to obtain a combined vector, i.e., the third vector. Finally, the customer's intent is predicted based on the combined vector, thereby generating prompt messages, which helps to achieve accurate recommendations and solves the problem of inaccurate product recommendations to users in existing technologies.

[0112] 2) The product recommendation device of this application includes an acquisition unit that acquires multiple dialogue data and multiple click behavior data of the user. The dialogue data consists of dialogues between the user and the user in the application software acquired before the current moment, and the click behavior data consists of click behaviors of the user in the application software acquired at the current moment. A first input unit inputs the dialogue data and click behavior data into a bidirectional encoder representation model to extract at least the semantic relationship between the dialogue data and the click behavior data. The dialogue sequence is used to represent the semantic information corresponding to the dialogue data. A second input unit inputs the dialogue sequence into a first converter model and the behavior sequence into a second converter model. A processing unit performs weighted concatenation processing on the first vector and the second vector to obtain a third vector. A recommendation unit generates prompt information based on the third vector and makes recommendations to the user based on the prompt information. By combining the dialogue data and click behavior data, and then performing intent recognition to obtain the dialogue sequence and behavior sequence, sequence modeling is then performed to obtain vector representations using the sequence model. The vectors are then weighted and concatenated to obtain a combined vector, i.e., the third vector. Finally, the customer's intent is predicted based on the combined vector, and prompt information is generated to facilitate accurate recommendations, thereby solving the problem of inaccurate product recommendations to users in the prior art.

[0113] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A product recommendation method characterized by, The method comprises the following steps: obtaining a plurality of dialogue data and a plurality of click behavior data of a user, the dialogue data being dialogue of the user in an application software obtained before a current time, and the click behavior data being click behavior of the user in the application software obtained at the current time; inputting the dialogue data and the click behavior data into a bidirectional encoder representation method model to at least extract semantic relationships between the dialogue data and the click behavior data, obtaining a dialogue sequence and a behavior sequence, the dialogue sequence being used to represent semantic information corresponding to the dialogue data, and the behavior sequence being used to represent semantic information of the click behavior data; inputting the dialogue sequence into a first converter model to extract relationships between data in the dialogue sequence, obtaining a first vector, and inputting the behavior sequence into a second converter model to extract relationships between data in the behavior sequence, obtaining a second vector; performing weighted splicing processing on the first vector and the second vector to obtain a third vector; generating prompt information according to the third vector, and recommending to the user according to the prompt information.

2. The method of claim 1, wherein, The method comprises the following steps: obtaining a plurality of dialogue data of a user, comprising: obtaining a plurality of dialogue texts of the user, each of the dialogue texts comprising at least one Chinese character to be mapped; obtaining a Chinese character sequence number mapping relationship, and determining a target sequence number according to the Chinese character sequence number mapping relationship and the Chinese character to be mapped, wherein the Chinese character sequence number mapping relationship is a pre-set mapping relationship between Chinese characters and sequence numbers, and the target sequence number is a sequence number corresponding to the Chinese character to be mapped in the Chinese character sequence number mapping relationship; 3. The method of claim 1, wherein, obtaining a sequence number vector mapping relationship, and determining a target vector according to the target sequence number and the sequence number vector mapping relationship, a plurality of target vectors constituting the dialogue data, wherein the sequence number vector mapping relationship is a pre-set mapping relationship between sequence numbers and vectors, and the target vector is a vector corresponding to the target sequence number in the sequence number vector mapping relationship. Inputting the dialogue data and the click behavior data into a bidirectional encoder representation method model to at least extract semantic relationships between the dialogue data and the click behavior data, obtaining a dialogue sequence and a behavior sequence, comprising: inputting a plurality of dialogue data into a first model to perform intent recognition on the dialogue data, obtaining an intent sequence, wherein the intent sequence is used to represent the intent of the dialogue data; 4. The method of claim 3, wherein, inputting the intent sequence and the click behavior data into a second model to at least extract semantic relationships within the intent sequence and between the click behavior data, obtaining the dialogue sequence and the behavior sequence, the bidirectional encoder representation method model comprising the first model and the second model. Inputting a plurality of dialogue data into a first model to perform intent recognition on the dialogue data, obtaining an intent sequence, comprising: Inputting a dialogue data into an encoder to classify the dialogue data and obtain an intent label, wherein the intent label represents an intent of the dialogue data, and when there are multiple intent labels, the intent labels are arranged in descending order of prediction probability. Repeating the inputting step until all dialogue data are inputted to obtain an intent sequence composed of multiple intent labels.

5. The method of claim 1, wherein The first transformer model comprises a first multi-head self-attention layer, a first feedforward network layer and a first residual connection layer connected in sequence, and the dialogue sequence is inputted into the first transformer model to extract the relationship between data in the dialogue sequence and obtain a first vector, comprising: The dialogue sequence is inputted into the first multi-head self-attention layer in sequence to obtain first output data, wherein the first multi-head self-attention layer is used to calculate the relationship between data in the dialogue sequence; The first output data is inputted into the first feedforward network layer to obtain second output data, wherein the first feedforward network layer is used to perform nonlinear transformation and feature extraction on the first output data; The second output data is inputted into the first residual connection layer to obtain the first vector; The second transformer model comprises a second multi-head self-attention layer, a second feedforward network layer and a second residual connection layer connected in sequence, and the action sequence is inputted into the second transformer model to extract the relationship between data in the action sequence and obtain a second vector, comprising: The action sequence is inputted into the second multi-head self-attention layer in sequence to obtain third output data, wherein the second multi-head self-attention layer is used to calculate the relationship between data in the action sequence; The third output data is inputted into the second feedforward network layer to obtain fourth output data, wherein the second feedforward network layer is used to perform nonlinear transformation and feature extraction on the third output data; The fourth output data is inputted into the second residual connection layer to obtain the second vector.

6. The method of claim 1, wherein, The first vector and the second vector are weighted and spliced to obtain a third vector, comprising: A first hyperparameter and a second hyperparameter are obtained, wherein the first hyperparameter is a weight of the first vector, the second hyperparameter is a weight of the second vector, and the sum of the first hyperparameter and the second hyperparameter is 1; A product of the first hyperparameter and the first vector is calculated to obtain a first to-be-spliced vector, and a product of the second hyperparameter and the second vector is calculated to obtain a second to-be-spliced vector; The first to-be-spliced vector and the second to-be-spliced vector are spliced in a predetermined dimension to obtain the third vector, wherein the predetermined dimension is the last dimension in the first vector and the second vector.

7. The method of claim 1, wherein, According to the third vector, a prompt information is generated, comprising: inputting the third vector into a full connection layer to convert a vector space of the third vector to obtain a fourth vector and a predicted probability of the fourth vector; comparing the predicted probability of the fourth vector with a predetermined probability, and determining that the fourth vector is a target intent in a case that the predicted probability of the fourth vector is greater than or equal to the predetermined probability; obtaining an intent tree, and determining the prompt information according to the target intent and the intent tree, the intent tree being used to represent a mapping relationship between the intent of the user and clue content, and the prompt information being the clue content corresponding to the intent identical to the target intent in the intent tree.

8. A product recommendation device characterized by comprising: comprise: an obtaining unit, configured to obtain a plurality of dialogue data and a plurality of click behavior data of a user, the dialogue data being dialogue of the user in an application software obtained before a current time, and the click behavior data being click behavior of the user in the application software obtained at the current time; a first input unit, configured to input the dialogue data and the click behavior data into a bidirectional encoder representation method model to at least extract semantic relationships between the dialogue data and the click behavior data, to obtain a dialogue sequence and a behavior sequence, the dialogue sequence being used to represent semantic information corresponding to the dialogue data, and the behavior sequence being used to represent semantic information of the click behavior data; a second input unit, configured to input the dialogue sequence into a first converter model to extract relationships between data in the dialogue sequence to obtain a first vector, and input the behavior sequence into a second converter model to extract relationships between data in the behavior sequence to obtain a second vector; a processing unit, configured to perform weighted splicing processing on the first vector and the second vector to obtain a third vector; a recommendation unit, configured to generate prompt information according to the third vector, and recommend to the user according to the prompt information.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium comprises a stored program, wherein the program controls a device in which the computer-readable storage medium is located to execute the method in any one of claims 1 to 7 when the program is running.

10. A processor, comprising: The processor is configured to run a program, wherein the program executes the method in any one of claims 1 to 7 when the program is running.

Images