Customer behavior trajectory prediction method, device, equipment, medium and program product

Abstract

Embodiments of the present application provide a customer behavior trajectory prediction method, device, equipment, medium and program product, relating to the field of financial technology. The method comprises: obtaining structured interaction data, unstructured audio and video data and third-party customer data of a plurality of preset channels; performing feature extraction on the unstructured audio and video data to obtain customer intention features; associating and matching the customer intention features with corresponding structured interaction data to obtain cross-channel behavior association data; performing hierarchical processing on the cross-channel behavior association data to generate real-time wide tables and merged wide tables; integrating the third-party customer data, the real-time wide tables and the merged wide tables in chronological order to generate customer behavior trajectory data; and inputting the customer behavior trajectory data into a preset trajectory prediction model to obtain a customer behavior prediction trajectory. The method of the present application increases the fusion processing capability of multi-channel and multi-modal data, and improves the accuracy and continuity of customer behavior trajectory prediction.

Description

Technical Field

[0001] This application relates to the field of financial technology, and in particular to a method, apparatus, device, medium, and program product for predicting customer behavior patterns. Background Technology

[0002] With the development of fintech, the ways banks interact with customers have expanded from single branches to multiple channels such as mobile banking, online banking, self-service terminals, and customer service hotlines. The behavioral data generated by customers on different channels includes both structured operation records and unstructured audio and video data, exhibiting characteristics of diverse channels, heterogeneous formats, and dispersed time sequences.

[0003] In existing technologies, data from various banking channels is typically managed by independent systems, resulting in fragmented processing chains for structured operational data and unstructured audio / video data. For structured data, systems perform only localized analysis within each channel; for audio / video data, post-event retrospective analysis relies heavily on manual sampling or simple keyword matching. The lack of a unified time benchmark and correlation mechanism between different channels makes cross-channel alignment difficult based on customer identity and time sequence, leading to isolated customer behavior records across different channels that cannot be pieced together into a continuous behavioral sequence.

[0004] Existing data processing architectures are unable to integrate multi-source, heterogeneous data from across channels, resulting in fragmented customer behavior patterns and a lack of a unified data foundation. Consequently, predictive models built upon this foundation suffer from incomplete data, making it difficult to guarantee the continuity and accuracy of their predictions. Summary of the Invention

[0005] This application provides a customer behavior trajectory prediction method, apparatus, equipment, medium, and program product to improve the continuity and accuracy of customer behavior prediction.

[0006] Firstly, this application provides a method for predicting customer behavior patterns, including:

[0007] Acquire structured interaction data, unstructured audio and video data, and third-party customer data from multiple preset channels;

[0008] Feature extraction is performed on the unstructured audio and video data to obtain customer intent features;

[0009] The customer intent features are associated and matched with the corresponding structured interaction data to obtain cross-channel behavior association data;

[0010] The cross-channel behavior-related data is processed in layers to generate real-time wide tables and merged wide tables;

[0011] The third-party customer data, the real-time wide table, and the merged wide table are integrated in chronological order to generate customer behavior trajectory data;

[0012] The customer behavior trajectory data is input into a preset trajectory prediction model to obtain the customer behavior prediction trajectory.

[0013] Secondly, this application provides a customer behavior trajectory prediction device, comprising:

[0014] The acquisition module is used to acquire structured interactive data, unstructured audio and video data, and third-party customer data from multiple preset channels;

[0015] The feature extraction module is used to extract features from the unstructured audio and video data to obtain customer intent features;

[0016] The association module is used to associate and match the customer intent features with the corresponding structured interaction data to obtain cross-channel behavior association data;

[0017] The processing module is used to perform hierarchical processing on the cross-channel behavior-related data to generate real-time wide tables and merged wide tables;

[0018] The processing module is also used to integrate the third-party customer data, the real-time wide table, and the merged wide table in chronological order to generate customer behavior trajectory data;

[0019] The prediction module is used to input the customer behavior trajectory data into a preset trajectory prediction model to obtain the customer behavior prediction trajectory.

[0020] Thirdly, this application provides an electronic device, including: a processor, and a memory communicatively connected to the processor;

[0021] The memory stores computer-executed instructions;

[0022] The processor executes computer execution instructions stored in the memory, causing the processor to perform the first aspect and / or various possible implementations of the first aspect as described above.

[0023] Fourthly, this application provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the first aspect and / or various possible embodiments of the first aspect.

[0024] Fifthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the first aspect and / or various possible implementations of the first aspect.

[0025] The customer behavior trajectory prediction method, device, equipment, medium, and program products provided in this application acquire structured interaction data, unstructured audio and video data, and third-party customer data from multiple preset channels. They extract customer intent features from the unstructured audio and video data, correlate and match these features with corresponding structured interaction data to obtain cross-channel behavior association data, perform hierarchical processing on the cross-channel behavior association data to generate real-time wide tables and merged wide tables, integrate the third-party customer data, real-time wide tables, and merged wide tables in chronological order to generate customer behavior trajectory data, and input this data into a preset trajectory prediction model to obtain the customer behavior prediction trajectory. This approach achieves the fusion of fragmented multimodal data across channels into a continuous and complete customer behavior trajectory, and then uses this trajectory for prediction, thereby improving the continuity and accuracy of customer behavior prediction. Attached Figure Description

[0026] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0027] Figure 1 A schematic diagram illustrating a scenario for the customer behavior trajectory prediction method provided in this application;

[0028] Figure 2 A flowchart illustrating the customer behavior trajectory prediction method provided in this application;

[0029] Figure 3 A schematic diagram of the customer behavior trajectory prediction device provided in this application;

[0030] Figure 4 A schematic diagram of the structure of the electronic device provided in this application.

[0031] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0032] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0033] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, storage, use, processing, transmission, provision, disclosure, and application of the relevant data all comply with the relevant laws, regulations, and standards of the relevant countries and regions, have taken necessary confidentiality measures, do not violate public order and good morals, and provide corresponding operation access points for users to choose to authorize or refuse.

[0034] Furthermore, the technical solution involved in this application, which involves big data analysis of user information (including but not limited to personal biometrics, identity data, consumption data, asset data, electronic terminal operation data, etc.) and the use of artificial intelligence technology for automated decision-making, and makes decisions that have a significant impact on personal rights based on the results of automated decision-making, provides users with corresponding operation entry points for users to choose to agree to or reject the results of automated decision-making; if the user chooses to reject, the process will proceed to the expert decision-making process.

[0035] It should be noted that the customer behavior trajectory prediction method, device, equipment, medium and program products provided in this application can be used in the field of financial technology, or in any field other than financial technology. The application field of the customer behavior trajectory prediction method, device, equipment, medium and program products in this application is not limited.

[0036] In existing technologies, predictive models are built by collecting structured interaction data from different channels and then simply splicing them together. However, this approach has limitations in terms of the ability to efficiently integrate and align heterogeneous data from multiple sources across channels in a hierarchical manner, making it difficult to generate continuous and complete customer behavior trajectories. Consequently, the accuracy and real-time performance of customer behavior predictions are insufficient.

[0037] The specific application scenario of this application is multi-channel customer service and operation management in financial institutions such as banks, covering various online and offline customer interaction links. For example, it is a scenario for integrating and processing structured interaction data and unstructured audio and video data generated when customers interact with the bank through multiple preset channels such as mobile banking, online banking, offline branches, and customer service hotlines. This is used to support customer demand prediction, service process optimization, and business resource adjustment.

[0038] The customer behavior trajectory prediction method provided in this application solves the technical problem that cross-channel multi-source heterogeneous data cannot be efficiently integrated in a layered manner, making it difficult to generate continuous and complete customer behavior trajectories, thus leading to insufficient prediction accuracy and real-time performance. This method simultaneously acquires multi-channel structured interaction data, unstructured audio and video data, and third-party customer data, extracts customer intent features from unstructured audio and video data and associates and matches them with corresponding structured data, performs hierarchical processing on the associated data to generate real-time wide tables and merged wide tables, integrates multi-source data in chronological order to generate customer behavior trajectory data, and finally inputs it into the prediction model to obtain the predicted customer behavior trajectory.

[0039] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will be described below with reference to the accompanying drawings.

[0040] Figure 1 This is a schematic diagram illustrating the application scenario of the customer behavior trajectory prediction method provided in this application, such as... Figure 1 As shown, it includes: terminal 101 and server 102.

[0041] Terminal 101 is used to generate multi-channel structured interactive data and unstructured audio and video data, and transmit the collected interactive content data externally. Server 102 is used to receive the interactive data transmitted by terminal 101 and third-party customer data, and execute the complete processing flow corresponding to customer behavior trajectory prediction to generate customer behavior prediction trajectory.

[0042] Figure 2 This is a flowchart illustrating the customer behavior trajectory prediction method provided in this application. The execution entity in this embodiment can be... Figure 1 The server 102 in the illustrated embodiment can also be other computer-related devices, and this embodiment is not particularly limited.

[0043] like Figure 2 As shown, this customer behavior trajectory prediction method includes the following steps:

[0044] Step S201: Obtain structured interaction data, unstructured audio and video data, and third-party customer data from multiple preset channels.

[0045] Specifically, structured interaction data refers to customer operation information recorded in a preset field format, including click behavior, page browsing history, and corresponding timestamps and channel identifiers. Unstructured audio and video data refers to customer interaction records in audio or video format, including multimedia data such as customer service call recordings and video consultation recordings that have not undergone field processing. Third-party customer data refers to supplementary customer information imported from outside the banking system through standardized interfaces, including credit data and social media behavior data.

[0046] Through a unified data access layer, structured interactive data is collected from online service channels and offline business outlets, unstructured audio and video data is collected from interactive scenarios such as customer service calls and video consultations, and standardized third-party customer data is obtained from cooperative institutions that meet the preset requirements. The heterogeneous data scattered across various channels is uniformly incorporated into the same processing link, so that subsequent processing links can be operated on a complete data view, avoiding the interruption of customer behavior analysis due to data loss.

[0047] Step S202: Extract features from unstructured audio and video data to obtain customer intent features.

[0048] Specifically, the unstructured audio and video data is first identified and transformed into text data. Then, initial intent features and customer emotion features are extracted from the text data. Finally, the two types of features are merged to obtain customer intent features that integrate semantic and emotional information. By transforming the customer needs and emotional information implicit in unstructured audio and video data into structured feature expressions, audio and video records that were originally impossible for computers to process directly are transformed into feature data that can be quantified and analyzed, providing semantic dimension input for subsequent association matching.

[0049] Step S203: Match customer intent features with corresponding structured interaction data to obtain cross-channel behavior association data.

[0050] Specifically, timestamps and channel identifiers are extracted from structured interaction data. Based on the timestamps, the corresponding time range of unstructured audio and video data is determined. Combining channel identifiers and customer identifiers, customer intent features are matched and associated with structured interaction data of the same customer within the corresponding time range, forming unified cross-channel behavior association data. This associates the originally isolated structured operation records with unstructured intent features in both customer and time dimensions, enabling the splicing of customer behavior fragments on different channels into a contextually related behavior sequence. This allows the needs expressed by customers in customer service calls to be correlated with their browsing operations on mobile banking, providing a foundation for constructing a complete behavior trajectory.

[0051] Step S204: Perform layered processing on cross-channel behavior-related data to generate real-time wide tables and merged wide tables.

[0052] Specifically, the preset stream processing engine refers to the processing component used to perform real-time calculations on each continuously arriving data stream; the preset batch processing engine refers to the processing component used to perform periodic offline calculations on a dataset accumulated over a certain time range; the real-time wide table is a wide table structure generated by the stream processing engine to perform real-time aggregation of related data; and the merged wide table is a wide table structure generated by the batch processing engine to perform periodic full calculations on historical related data.

[0053] The preset stream computing engine performs real-time cleaning, correlation, and indicator calculation on continuously generated cross-channel behavior-related data to generate a real-time wide table that is updated in real time; the preset batch processing engine performs multi-dimensional aggregation and in-depth calculation on all cross-channel behavior-related data at fixed intervals to generate a merged wide table containing all historical information.

[0054] Through a hierarchical processing approach integrating stream and batch processing, the stream computing engine performs calculations and instant aggregation on each piece of real-time related data to meet the low-latency data timeliness requirements; the batch processing engine performs periodic full calculations on historical related data to meet the high-throughput data integrity requirements; it can both instantly perceive the latest customer behavior and provide comprehensive data support based on complete historical data accumulation, balancing the timeliness and integrity of data processing. Real-time wide tables support instant responses to current customer behavior, while merged wide tables support in-depth analysis of long-term customer behavior patterns, complementing each other to form a complete data view.

[0055] Step S205: Integrate third-party customer data, real-time wide tables, and merged wide tables in chronological order to generate customer behavior trajectory data.

[0056] Specifically, using the customer identifier as the primary key, third-party customer data is integrated with data from the real-time wide table and the merged wide table, arranged in timestamp order to form complete customer behavior trajectory data organized by time series. Customer behavior trajectory data refers to the data set of all behavioral events of a customer interacting with the bank on multiple channels, recorded in chronological order.

[0057] By aligning and stitching multi-source data over time, pre-processed real-time wide tables, merged wide tables, and third-party customer data are uniformly sorted and integrated to construct a continuous behavioral record centered on the customer and based on time. This aggregates the dispersed multi-source data into unified customer behavior trajectory data, providing complete and continuous data input for subsequent trajectory prediction. It achieves time-series integration of multi-source heterogeneous data, enabling customers' interactions across different channels and time points to form a continuous behavioral trajectory, eliminating information gaps caused by different data sources or processing paths.

[0058] Step S206: Input customer behavior trajectory data into the preset trajectory prediction model to obtain the customer behavior prediction trajectory.

[0059] Specifically, a pre-built trajectory prediction model refers to a pre-constructed and trained computational model used to predict future behavioral sequences based on historical behavioral sequences, such as Long Short-Term Memory (LSTM) networks and Temporal Fusion Transformers (TFTs). A customer behavior prediction trajectory refers to the sequence data of possible behavioral events of a customer within a future time period, output by the model. The customer behavior trajectory data is input into the pre-built trajectory prediction model, which then predicts the possible behavioral nodes of the customer within the future time period based on the historical behavioral sequences, and outputs the customer behavior prediction trajectory.

[0060] By modeling the temporal dependencies in a behavioral sequence using a pre-defined trajectory prediction model, the system infers potential subsequent behavioral nodes for customers and outputs a continuous behavioral prediction sequence. This effectively transforms historical behavioral trajectory data into predictions of future customer behavior, enabling banks to anticipate potential behavioral trends before customers actually take action. This allows for proactive deployment of service resources and preparation for service before customers reach key behavioral nodes.

[0061] The customer behavior trajectory prediction method provided in this invention extracts features from unstructured audio and video data and performs cross-channel correlation matching with structured interactive data. After hierarchical processing, it generates real-time wide tables and merged wide tables. Third-party customer data and the two types of wide tables are integrated into customer behavior trajectory data in chronological order and then input into a preset trajectory prediction model. This achieves efficient and unified integration of multi-source heterogeneous data across channels, generates continuous and complete customer behavior trajectories, and improves the accuracy and real-time performance of customer behavior prediction.

[0062] This embodiment provides a detailed explanation of the process for extracting features from unstructured audio and video data to obtain customer intent features as described in the above embodiments. The specific implementation of this process includes the following steps:

[0063] Step a1: Recognize and convert unstructured audio and video data to obtain text data.

[0064] Specifically, a pre-trained speech recognition model processes raw unstructured audio and video data. First, the speech signal is extracted from the audio and video; then, noise reduction, frame segmentation, and feature extraction preprocessing are performed on the speech signal. Finally, the model calculates and outputs the corresponding structured text sequence. The speech recognition model is a deep learning model that converts human speech signals into natural language text. It maps preprocessed speech features into phoneme sequences using an acoustic model, and then converts the phoneme sequences into grammatically correct natural language text using a language model.

[0065] This process converts audio and video signals that cannot be directly processed by algorithms into text data that can be semantically analyzed, completing the initial structuring of unstructured audio and video data, fully preserving language information during customer interactions, and eliminating processing obstacles caused by differences in audio and video data formats from different channels.

[0066] Step a2: Extract intent from text data using a preset intent recognition model to obtain initial intent features.

[0067] Specifically, a pre-trained intent recognition model refers to a deep learning model capable of extracting core user needs information from natural language text. By learning the semantic features of a large amount of labeled text, it establishes a mapping relationship between text content and different intent categories, thereby achieving automatic recognition of unknown text intent, such as the BERT (Bidirectional Encoder Representations from Transformers) model. The transformed text data is input into the pre-trained intent recognition model, which performs word segmentation, word embedding encoding, and contextual semantic encoding on the text, and then outputs the corresponding initial intent feature vector through a classification layer.

[0068] By extracting core customer needs from text data, the system transforms customer expressions in natural language into quantifiable and calculable feature representations, providing core semantic basis for subsequent cross-channel behavior association. This enables automated and standardized extraction of customer intent, allowing for rapid acquisition of core customer needs from large amounts of text data and improving the efficiency of customer information processing.

[0069] Step a3: Process the text data using a preset emotion recognition model to obtain customer emotion characteristics.

[0070] Specifically, an emotion recognition model refers to a deep learning model that can identify a user's emotional state from text content. By learning the emotional vocabulary, sentence structure, and contextual semantic features contained in the text, it establishes a mapping relationship between text content and different emotional states, thereby automatically judging the user's emotional inclination. The converted text data is input into a pre-trained emotion recognition model, which performs semantic analysis and emotion feature extraction, outputting the corresponding customer emotion feature vector through an emotion classification layer.

[0071] By extracting emotional state information during customer interactions and supplementing implicit customer state information beyond intent features, the emotional information in customer calls is made explicit, thereby distinguishing different emotional states when customers express the same intent and providing an emotional reference for differentiated services.

[0072] Step a4: Merge the customer's emotional characteristics with the initial intent characteristics to obtain the customer's intent characteristics.

[0073] Specifically, feature dimension concatenation is an operation that joins multiple vectors with the same number of samples but different feature dimensions end-to-end along the feature dimension to form a higher-dimensional feature vector. This integrates feature information from different sources without losing any original feature information. Using dimension concatenation, the initial intent feature vector and the customer sentiment feature vector are concatenated according to feature dimensions to form a unified customer intent feature vector that integrates semantic and emotional information.

[0074] This approach unifies and integrates feature information scattered across the intent recognition and emotion recognition processing branches, providing a more comprehensive feature input with higher information density for subsequent cross-channel behavior association matching. It achieves the organic fusion of intent semantics and emotional state, enabling customer intent features to more comprehensively depict the customer's true state during the interaction process, thereby improving the accuracy of subsequent association matching and behavior prediction.

[0075] This invention achieves complete extraction and structured expression of customer intent and emotion information from unstructured audio and video data by converting unstructured audio and video data into text data through speech recognition, and then extracting intent features and emotion features from the text data through a preset intent recognition model and a preset emotion recognition model, respectively, and then fusing them. This improves the information completeness and expression accuracy of customer intent features.

[0076] This embodiment details the process of hierarchically processing cross-channel behavioral correlation data to generate real-time wide tables and merge wide tables as described in the above embodiments. The specific implementation of this process includes the following steps:

[0077] Step b1: Real-time processing of cross-channel behavior-related data is performed using a preset stream computing engine to generate a real-time wide table.

[0078] Specifically, a stream computing engine refers to a distributed computing engine capable of real-time computation processing of continuously generated data streams. It employs an event-driven computing model, performing incremental computation on each arriving data piece without waiting for the full dataset to load, and can output the results immediately. A real-time wide table refers to a wide table structure generated by the stream computing engine through the instant aggregation of related data arriving in real time.

[0079] By continuously accessing cross-channel behavioral data streams through a pre-defined stream computing engine, the system performs real-time cleaning and format standardization on the data streams, associates them in real-time according to customer identifiers, calculates the latest behavioral metrics, writes the processed results into a real-time wide table, and updates the table data periodically. This enables real-time processing and updating of cross-channel behavioral data, promptly capturing the latest changes in customer behavior, significantly reducing data processing latency, and providing data support for business scenarios with high real-time requirements.

[0080] Step b2: Use a preset batch processing engine to process cross-channel behavior-related data in batches and generate a merged wide table.

[0081] Specifically, a batch processing engine refers to a distributed computing engine capable of performing batch computations on large-scale static datasets. Its core principle is to divide the entire dataset into multiple computational tasks, distribute them to multiple computing nodes for parallel processing, and finally aggregate the results. A merged wide table refers to a wide table structure generated by the batch processing engine through periodic full computations of historical correlated data, covering the entire historical time span.

[0082] By using a pre-set batch processing engine to read all cross-channel behavioral correlation data at fixed intervals, historical data is subjected to multi-dimensional correlation and aggregation calculations. Behavioral information from different time periods is integrated to generate a merged wide table containing all historical behavioral characteristics. This achieves deep integration and full calculation of historical cross-channel behavioral data, constructing a complete view of customer historical behavior. It ensures the integrity and comprehensiveness of customer behavior data, covering behavioral information throughout the entire customer lifecycle, providing a comprehensive data foundation for long-term behavioral analysis and prediction, and improving the accuracy of subsequent trajectory prediction.

[0083] This invention implements a hierarchical data processing architecture that combines stream processing and batch processing, achieving both real-time performance and completeness of data processing. It uses a preset stream computing engine to process cross-channel behavior-related data in real time to generate a real-time wide table, and a preset batch processing engine to process cross-channel behavior-related data in batches to generate a merged wide table.

[0084] This embodiment details the process described in the above embodiment of associating and matching customer intent features with corresponding structured interaction data to obtain cross-channel behavioral association data. The specific implementation of this process includes the following steps:

[0085] Step c1: Obtain the timestamp and channel identifier from the structured interaction data.

[0086] Specifically, a timestamp is a time stamp that records the moment a customer's action occurs. A channel identifier is a coded marker that distinguishes different customer interaction channels; each channel has a unique identifier. The timestamp and channel identifier for each action record are extracted from the standard fields of the structured interaction data. The timestamp determines the moment the customer's action occurred, and the channel identifier distinguishes the channel type through which the action occurred. The extracted fields undergo format validation and standardization to unify the precision of the timestamp and the encoding rules for the channel identifier. By parsing the fields of the structured interaction data, the time and channel dimension metadata fields are located and extracted, providing benchmark information for subsequent time-series alignment and channel classification.

[0087] Step c2 involves matching customer intent features with structured interaction data of the same customer within the corresponding time range based on timestamps and channel identifiers to obtain cross-channel behavioral association data.

[0088] Specifically, all data is grouped according to customer identifiers, and the associated time window is determined based on the timestamps corresponding to the unstructured audio and video data (e.g., 30 minutes before and after the customer service call). In the structured interaction data of the same customer, behavioral records (e.g., clicks, transactions) whose timestamps fall within the associated time window and whose channel identifiers (e.g., mobile banking, customer service hotline) match are filtered out. The customer intent characteristics are then associated with the filtered behavioral records.

[0089] For example, a time-series alignment algorithm can be used, with customer identity as the primary key and timestamp as the sorting axis. By using a sliding time window, asynchronously generated intent features and structured operation records are aligned and matched on the time series, establishing a correlation mapping between the two types of data. This achieves a preliminary association between unstructured intent features and structured behavioral data, establishes a basic correspondence between intent information and behavioral events, and forms a unified data record containing semantic and behavioral information.

[0090] In some optional implementations, step c2 above further includes:

[0091] Step c21: Construct a customer behavior knowledge graph based on structured interaction data and customer intent features.

[0092] Specifically, a knowledge graph is a data model that uses a graph structure to represent entities and the relationships between them, providing an intuitive view of the network of connections between different entities. By using customers, behavioral events, intent categories, and channels as nodes in the knowledge graph, and the relationships between customers and behavioral events, behavioral events and intents, and behavioral events and channels as edges, a structured graph data representation is constructed.

[0093] By organizing scattered customer behavior and intent information into a graph structure and preserving the relationships between entities, discrete behavior and intent information is transformed into a structured graph representation, realizing the networked organization of customer behavior information and fully preserving the relationship information between different entities, which facilitates in-depth relationship analysis.

[0094] Step c22: Process the customer behavior knowledge graph using a pre-trained graph neural network model to obtain cross-channel implicit association features.

[0095] Specifically, pre-trained graph neural network models are deep learning models pre-trained on graph-structured data, possessing the ability to extract features from node relationships and topological patterns in graph data. Examples include Graph Convolutional Networks (GCNs) and Graph Attention Networks (GATs). Cross-channel implicit association features refer to vectorized representations of potential cross-channel behavioral associations mined from customer behavior knowledge graphs using graph neural network models, which do not directly rely on timestamps and channel identifiers.

[0096] Based on the message passing mechanism of graph neural networks, this method propagates and integrates neighborhood information on the graph structure through multi-layer node aggregation and edge feature updates. It learns high-order association patterns between customer behavior nodes and extracts implicit association features that transcend explicit time and channel rules. By leveraging the graph structure learning capabilities of graph neural networks, it supplements the deep associations that cannot be captured by simply relying on timestamps and channel identifiers for matching. It can discover implicit associations between customer behaviors across different channels, such as semantic associations outside the time window that may exist between a customer's browsing behavior on one channel and their consultation behavior on another channel. This provides additional dimensions of reference information for association matching, improving its coverage and accuracy.

[0097] Step c23 involves matching customer intent features with structured interaction data of the same customer within the corresponding time range, based on timestamps, channel identifiers, and cross-channel implicit association features.

[0098] Specifically, based on the association matching of timestamps and channel identifiers, implicit cross-channel association features are introduced as supplementary matching criteria. These three types of features are used together as judgment conditions for association matching, performing association matching between customer intent features and structured interaction data. By jointly using time-dimensional features, channel-dimensional features, and implicit association features obtained through graph structure learning, the association matching process retains both the interpretability based on explicit rules and the deep association discovery capability based on graph learning. The combined effect of multiple features improves the accuracy and recall of association matching.

[0099] This invention obtains timestamps and channel identifiers from structured interaction data, and then matches customer intent features with structured interaction data of the same customer within the corresponding time range based on the timestamps and channel identifiers. In optional embodiments, customer behavior knowledge graphs and pre-trained graph neural network models are further introduced to mine implicit cross-channel association features as supplementary matching criteria, thereby realizing multi-dimensional and multi-level association matching between customer intent features and structured interaction data, and improving the accuracy and completeness of cross-channel behavior association.

[0100] In some optional implementations, the specific implementation of the above embodiments further includes the following steps:

[0101] Step d1: Extract at least one future behavior node from the customer behavior prediction trajectory.

[0102] Specifically, after the preset trajectory prediction model outputs the customer behavior prediction trajectory, the trajectory is analyzed to identify and extract one or more key future behavior nodes from the future behavior sequence contained in the prediction trajectory. A future behavior node refers to a data unit in the customer behavior prediction trajectory that identifies a specific interactive behavior that a customer may engage in at a specific point in the future, including information such as behavior type, predicted time of occurrence, and predicted channel of occurrence.

[0103] By structurally decomposing the predicted trajectory and extracting behavioral node units with independent business meaning from the sequence according to the preset behavioral node definition rules, each node corresponds to a customer behavior event that can be matched by subsequent strategies. The continuous behavioral prediction trajectory is discretized into several independently processable behavioral nodes, enabling subsequent service strategy matching to make refined recommendations based on specific behavioral events.

[0104] Step d2: Input the future action node into the pre-trained deep reinforcement learning model to obtain the service policy recommendation sequence corresponding to the future action node.

[0105] Specifically, a pre-trained deep reinforcement learning model refers to a deep reinforcement learning model that has been pre-trained through interaction with the environment and possesses the ability to output the optimal action policy based on the input state. A service policy recommendation sequence refers to a series of service actions recommended by the model for future customer behavior nodes, including information such as service action type, execution timing, and execution channel. The future behavior node is taken as the state input and fed into the pre-trained deep reinforcement learning model. The model analyzes the current state based on the learned policy function and outputs the service policy recommendation sequence corresponding to the future behavior node.

[0106] Deep reinforcement learning models analyze the future state of behavioral nodes by learning value and policy functions, and select the service action sequence with the maximum cumulative benefit from the action space as the output. This enables a leap from passively perceiving customer behavior to proactively recommending service strategies, transforming customer behavior prediction results into executable service strategy sequences. This allows bank employees to obtain system-recommended service action suggestions before customers reach key behavioral nodes, improving the initiative and accuracy of services.

[0107] This invention extracts future behavior nodes from customer behavior prediction trajectories and inputs them into a pre-trained deep reinforcement learning model to generate corresponding service strategy recommendation sequences, thereby achieving a technical closed loop from behavior prediction to strategy recommendation and improving the initiative of customer service and the accuracy of strategy recommendation.

[0108] This embodiment provides a detailed description of the training process of the pre-trained deep reinforcement learning model in the above embodiments. The specific implementation of this process includes the following steps:

[0109] Step e1: Obtain sample data of the behavioral trajectories of local bank branch customers.

[0110] Specifically, behavioral trajectory sample data refers to the training sample set containing historical customer behavior sequences and corresponding service result labels used to train deep reinforcement learning models. By extracting historical customer behavior trajectory data and corresponding service effect data from the local data warehouse, the extracted data is cleaned, deduplicated, and standardized in format to filter invalid and abnormal samples, thus constructing a behavioral trajectory sample dataset that meets the requirements for model training, providing a data foundation for the local training of deep reinforcement learning models.

[0111] Step e2: Train the preset initial deep reinforcement learning model based on the behavioral trajectory sample data to obtain the local model parameters.

[0112] Specifically, the preset initial deep reinforcement learning model refers to a deep reinforcement learning model framework with a basic network structure that has not yet been trained. Local model parameters refer to the set of model weight parameters that reflect the local customer behavior characteristics obtained after training on local training data.

[0113] By dividing the behavioral trajectory sample dataset into training and validation sets, and using service performance as the reward function, the initial deep reinforcement learning model is iteratively trained on the training set. The model's performance is evaluated and hyperparameters are adjusted on the validation set. After training, the weight and bias parameters of the local model are saved. The deep reinforcement learning model continuously learns the optimal decision-making strategy through trial and error by interacting with the environment, updating the model parameters with the goal of maximizing cumulative rewards. This trains an initial model adapted to the local business scenario, enabling the model to learn the behavioral characteristics and service preferences of local customers.

[0114] Step e3: Send local model parameters to the preset federated learning coordination node.

[0115] Specifically, the pre-defined federated learning coordination node is the central coordination server in the federated learning architecture responsible for receiving model parameters uploaded by each participant, performing parameter aggregation operations, and returning the aggregation results. The locally trained model parameters are encrypted and sent to the pre-defined federated learning coordination node through a secure communication channel, while simultaneously reporting the number of local training samples.

[0116] Step e4: Receive the aggregated model parameters returned by the preset federated learning coordination node. The aggregated model parameters are obtained by the preset federated learning coordination node by weighting and aggregating the model parameters according to the number of local training samples of multiple participants based on the federated averaging algorithm.

[0117] Specifically, the pre-defined federated learning coordinating node collects encrypted model parameters uploaded by all participants. It then uses a federated averaging algorithm to weight and aggregate the model parameters from all participants, generating global aggregated model parameters. These encrypted global aggregated model parameters are then returned to each participant. The federated averaging algorithm is the most commonly used parameter aggregation method in federated learning. It obtains global model parameters that incorporate the data features of all participants by weighting the model parameters of each participant according to the number of samples.

[0118] By aggregating model parameters from multiple stakeholders, it is possible to integrate the data distribution characteristics of different institutions, compensate for the deficiencies of insufficient data volume and data distribution bias of a single institution, improve the generalization ability and performance of the model, and enable the model to learn more comprehensive customer behavior patterns.

[0119] Step e5: Update the initial deep reinforcement learning model based on the aggregated model parameters to obtain the pre-trained deep reinforcement learning model.

[0120] Specifically, the received aggregated model parameters are decrypted, and the corresponding parameters of the local initial deep reinforcement learning model are replaced with the decrypted global aggregated model parameters to complete the model update, and finally the pre-trained deep reinforcement learning model is obtained.

[0121] The aggregated model parameters integrate training knowledge from multiple participants, enabling the local model to acquire prediction and decision-making capabilities by incorporating knowledge from multiple parties. This effectively improves the performance and generalization ability of the local model, allowing it to perform well in different business scenarios.

[0122] This invention provides an embodiment of the invention that obtains local model parameters by acquiring local behavioral trajectory sample data to train an initial deep reinforcement learning model locally. The local model parameters are then sent to a preset federated learning coordination node and aggregated with the model parameters of at least one other participant to obtain aggregated model parameters. Based on the aggregated model parameters, a pre-trained deep reinforcement learning model is updated. This achieves cross-institutional collaborative modeling while protecting the data privacy of each participant, and improves the model's adaptability to diverse customer behavior patterns.

[0123] Figure 3 A schematic diagram of the customer behavior trajectory prediction device provided in this application. Figure 3 As shown, the customer behavior trajectory prediction device 30 includes:

[0124] The acquisition module 301 is used to acquire structured interactive data, unstructured audio and video data, and third-party customer data from multiple preset channels;

[0125] The feature extraction module 302 is used to extract features from unstructured audio and video data to obtain customer intent features;

[0126] The association module 303 is used to associate and match customer intent features with corresponding structured interaction data to obtain cross-channel behavior association data;

[0127] Processing module 304 is used to perform hierarchical processing on cross-channel behavior-related data to generate real-time wide tables and merged wide tables;

[0128] The processing module 304 is also used to integrate third-party customer data, real-time wide tables and merged wide tables in chronological order to generate customer behavior trajectory data;

[0129] The prediction module 305 is used to input customer behavior trajectory data into a preset trajectory prediction model to obtain the customer behavior prediction trajectory.

[0130] In one possible design, the feature extraction module 302 is specifically used for:

[0131] Unstructured audio and video data is identified and converted to obtain text data; the intent of the text data is extracted using a preset intent recognition model to obtain initial intent features; the text data is processed using a preset emotion recognition model to obtain customer emotion features; the customer emotion features are merged with the initial intent features to obtain customer intent features.

[0132] In one possible design, processing module 304 is specifically used for:

[0133] The preset stream processing engine processes cross-channel behavioral correlation data in real time to generate a real-time wide table; the preset batch processing engine processes cross-channel behavioral correlation data in batches to generate a merged wide table.

[0134] In one possible design, the associated module 303 is specifically used for:

[0135] Obtain the timestamp and channel identifier from the structured interaction data; based on the timestamp and channel identifier, associate and match the customer intent features with the structured interaction data of the same customer within the corresponding time range to obtain cross-channel behavior association data.

[0136] In one possible design, the associated module 303 is specifically used for:

[0137] A customer behavior knowledge graph is constructed based on structured interaction data and customer intent features. The customer behavior knowledge graph is processed by a pre-trained graph neural network model to obtain cross-channel implicit association features. Based on timestamps, channel identifiers, and cross-channel implicit association features, customer intent features are associated and matched with the structured interaction data of the same customer within the corresponding time range.

[0138] In one possible design, a service generation module 306 is also included, specifically for:

[0139] Extract at least one future behavior node from the customer behavior prediction trajectory; input the future behavior node into a pre-trained deep reinforcement learning model to obtain a service strategy recommendation sequence corresponding to the future behavior node.

[0140] In one possible design, the service generation module 306 is also used for:

[0141] Acquire behavioral trajectory sample data of local bank branch customers; train a preset initial deep reinforcement learning model based on the behavioral trajectory sample data to obtain local model parameters; send the local model parameters to a preset federated learning coordination node; receive the aggregated model parameters returned by the preset federated learning coordination node, wherein the aggregated model parameters are obtained by the preset federated learning coordination node by weighting and aggregating the model parameters according to the number of local training samples of multiple participants based on the federated averaging algorithm; update the initial deep reinforcement learning model based on the aggregated model parameters to obtain a pre-trained deep reinforcement learning model.

[0142] The customer behavior trajectory prediction device provided in this embodiment can be used to execute the customer behavior trajectory prediction method described above. Its implementation principle and technical effect are similar, and will not be described again in this embodiment.

[0143] Figure 4 A schematic diagram of the hardware structure of the electronic device provided in this application, such as... Figure 4 As shown, the electronic device 40 includes at least one processor 401 and a memory 402. Optionally, the electronic device 40 also includes a communication component 403. The processor 401, memory 402, and communication component 403 are connected via a bus 404.

[0144] In the specific implementation process, at least one processor 401 executes computer execution instructions stored in memory 402, causing at least one processor 401 to perform the above method.

[0145] The specific implementation process of processor 401 can be found in the above method embodiments, and its implementation principle and technical effect are similar. It will not be repeated here.

[0146] In the above embodiments, it should be understood that the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor.

[0147] The memory may include random access memory (RAM) and may also include non-volatile memory (NVM), such as at least one disk storage device.

[0148] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of illustration, the buses shown in the accompanying drawings are not limited to a single bus or a single type of bus.

[0149] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0150] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the above-described method.

[0151] The aforementioned readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The readable storage medium can be any available medium accessible to a general-purpose or special-purpose computer.

[0152] An exemplary readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can reside in an Application Specific Integrated Circuit (ASIC). Alternatively, the processor and the readable storage medium can exist as discrete components in the device.

[0153] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are all optional embodiments, and the actions and modules involved are not necessarily essential to this application.

[0154] It should be further noted that although the steps in the flowchart are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowchart may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0155] It should be understood that the above-described device embodiments are merely illustrative, and the device of this application can also be implemented in other ways. For example, the division of units / modules in the above embodiments is only a logical functional division, and there may be other division methods in actual implementation. For example, multiple units, modules, or components may be combined, or integrated into another system, or some features may be ignored or not executed.

[0156] Furthermore, unless otherwise specified, the functional units / modules in the various embodiments of this application can be integrated into one unit / module, or each unit / module can exist physically separately, or two or more units / modules can be integrated together. The integrated units / modules described above can be implemented in hardware or as software program modules.

[0157] When integrated units / modules are implemented in hardware, the hardware can be digital circuits, analog circuits, etc. The physical implementation of the hardware structure includes, but is not limited to, transistors, memristors, etc. Unless otherwise specified, the processor can be any suitable hardware processor, such as a CPU, GPU, FPGA, DSP, and ASIC, etc. Unless otherwise specified, the storage unit can be any suitable magnetic or magneto-optical storage medium, such as Resistive Random Access Memory (RRAM), Dynamic Random Access Memory (DRAM), Static Random Access Memory (SRAM), Enhanced Dynamic Random Access Memory (EDRAM), High-Bandwidth Memory (HBM), Hybrid Memory Cube (HMC), etc.

[0158] If the integrated unit / module is implemented as a software program module and sold or used as an independent product, it can be stored in a computer-readable storage device (CMD). Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a memory and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned memory includes various media capable of storing program code, such as a USB flash drive, read-only memory (ROM), random access memory (RAM), portable hard drive, magnetic disk, or optical disk.

[0159] In the above embodiments, the descriptions of each embodiment have their own emphasis. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments. The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification.

[0160] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this application are indicated by the following claims.

[0161] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.

Claims

1. A method for predicting customer behavior trajectories, characterized in that, include: Acquire structured interaction data, unstructured audio and video data, and third-party customer data from multiple preset channels; Feature extraction is performed on the unstructured audio and video data to obtain customer intent features; The customer intent features are associated and matched with the corresponding structured interaction data to obtain cross-channel behavior association data; The cross-channel behavior-related data is processed in layers to generate real-time wide tables and merged wide tables; The third-party customer data, the real-time wide table, and the merged wide table are integrated in chronological order to generate customer behavior trajectory data; The customer behavior trajectory data is input into a preset trajectory prediction model to obtain the customer behavior prediction trajectory.

2. The method according to claim 1, characterized in that, Feature extraction is performed on the unstructured audio and video data to obtain customer intent features, including: The unstructured audio and video data is identified and converted to obtain text data; The text data is subjected to intent extraction using a preset intent recognition model to obtain initial intent features; The text data is processed by a preset emotion recognition model to obtain customer emotion characteristics; The customer emotional characteristics are combined with the initial intent characteristics to obtain the customer intent characteristics.

3. The method according to claim 1, characterized in that, The cross-channel behavioral correlation data is processed in layers to generate real-time wide tables and merged wide tables, including: The cross-channel behavior-related data is processed in real time by a preset stream computing engine to generate the real-time wide table; The cross-channel behavior-related data is processed in batches using a preset batch processing engine to generate the merged wide table.

4. The method according to claim 1, characterized in that, The customer intent features are correlated and matched with the corresponding structured interaction data to obtain cross-channel behavior correlation data, including: Obtain the timestamp and channel identifier from the structured interactive data; Based on the timestamp and the channel identifier, the customer intent features are associated and matched with the structured interaction data of the same customer within the corresponding time range to obtain the cross-channel behavior association data.

5. The method according to claim 4, characterized in that, Associating and matching the customer intent features with the structured interaction data of the same customer within the corresponding time range also includes: A customer behavior knowledge graph is constructed based on the structured interaction data and the customer intent features; The customer behavior knowledge graph is processed by a pre-trained graph neural network model to obtain cross-channel implicit association features; Based on the timestamp, the channel identifier, and the cross-channel implicit association features, the customer intent features are associated and matched with the structured interaction data of the same customer within the corresponding time range.

6. The method according to claim 1, characterized in that, Also includes: Extract at least one future behavior node from the customer behavior prediction trajectory; The future behavior node is input into a pre-trained deep reinforcement learning model to obtain a service strategy recommendation sequence corresponding to the future behavior node.

7. The method according to claim 6, characterized in that, Also includes: Obtain sample data of customer behavior patterns at local bank branches; The preset initial deep reinforcement learning model is trained based on the behavioral trajectory sample data to obtain local model parameters; Send the local model parameters to the preset federated learning coordination node; The system receives aggregated model parameters returned by the preset federated learning coordination node, wherein the aggregated model parameters are obtained by the preset federated learning coordination node by weighting and aggregating the model parameters according to the number of local training samples of multiple participants based on the federated averaging algorithm. The initial deep reinforcement learning model is updated based on the aggregated model parameters to obtain the pre-trained deep reinforcement learning model.

8. A customer behavior trajectory prediction device, characterized in that, include: The acquisition module is used to acquire structured interactive data, unstructured audio and video data, and third-party customer data from multiple preset channels; The feature extraction module is used to extract features from the unstructured audio and video data to obtain customer intent features; The association module is used to associate and match the customer intent features with the corresponding structured interaction data to obtain cross-channel behavior association data; The processing module is used to perform hierarchical processing on the cross-channel behavior-related data to generate real-time wide tables and merged wide tables; The processing module is also used to integrate the third-party customer data, the real-time wide table, and the merged wide table in chronological order to generate customer behavior trajectory data; The prediction module is used to input the customer behavior trajectory data into a preset trajectory prediction model to obtain the customer behavior prediction trajectory.

9. An electronic device, characterized in that, include: Memory, processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory, causing the processor to perform the method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the method as described in any one of claims 1-7.

11. A computer program product, characterized in that, Includes a computer program that, when executed by a processor, implements the method described in any one of claims 1-7.

Images