Communication user off-network early warning method, medium and device based on multi-modal fusion

Abstract

This application provides a method, medium, and device for early warning of user churn based on multimodal fusion. The method includes: acquiring and processing time-series data and text data of user communication to obtain a multi-dimensional time-series matrix and a text token sequence; inputting the multi-dimensional time-series matrix into a time-series coding tower to output a hidden layer sequence vector containing temporal context information; inputting the text token sequence into a semantic coding tower to output a global semantic vector representing the user's complaint intent; performing dimensionality reduction projection on the global semantic vector to align it with the dimension of the hidden layer sequence vector; using the global semantic vector as a query vector and the hidden layer sequence vector as a key vector to obtain attention weights in the time dimension, and using these attention weights to perform a weighted summation of the hidden layer sequence vectors to generate a fused feature vector; inputting the fused feature vector into a fully connected classification layer to output the probability value of user churn. This application can significantly improve the prediction accuracy of user churn early warning.

Description

Technical Field

[0001] This application belongs to the field of artificial intelligence and big data mining technology, and in particular relates to the field of communication user churn early warning processing technology based on multimodal fusion. Background Technology

[0002] In the customer lifecycle management of telecommunications operators, churn warning is a key means to reduce customer churn rate. Operators have accumulated massive amounts of user data, mainly including two categories: one is structured time-series data, such as daily data consumption, call duration, signal strength, etc.; the other is unstructured text data, such as user complaint records from customer service calls, online consultation dialogues, etc.

[0003] Existing technologies have three main technical shortcomings when processing the above data that are difficult to overcome:

[0004] 1) Spatial alignment difficulties caused by feature heterogeneity: Text data is characterized by high dimensionality and sparsity, with its semantic information implicit in the combination of words; while communication behavior records belong to low dimensional and dense time-series data. Existing technologies usually use a simple splicing method to combine the two. This coarse fusion will cause the gradient update direction to be dominated by low-dimensional and dense features during model training, making it difficult for the model to converge and unable to truly utilize the deep semantics in the text.

[0005] 2) The Average Trap Caused by Time-Series Aggregation: Existing solutions for processing time-series data typically use statistical measures such as monthly mean and variance for dimensionality reduction. However, users' thoughts of leaving the network often stem from a very short period of poor experience, such as a few minutes of network outage or lag. Statistical averages smooth out these sudden abnormal peaks, leading to the loss of key risk signals and causing the model to miss detections.

[0006] 3) Black-box models lack interpretability: Traditional machine learning models (such as XGBoost) can only output a probability value of churn, and cannot trace back to which indicator on which day caused the risk to increase. When faced with a high-risk list, operations and maintenance personnel cannot know the specific reasons, and therefore cannot develop targeted recovery strategies, resulting in poor intervention effects after the warning. Summary of the Invention

[0007] This application provides a method, medium, and device for early warning of communication user churn based on multimodal fusion, which can improve the accuracy of early warning of communication user churn.

[0008] In a first aspect, embodiments of this application provide a method for early warning of user churn based on multimodal fusion, comprising: acquiring time-series data and text data of user communication, and processing the time-series data and the text data respectively to obtain a multi-dimensional time-series matrix and a text token sequence; inputting the multi-dimensional time-series matrix into a preset time-series coding tower, and outputting a hidden layer sequence vector containing time context information through the time-series coding tower; inputting the text token sequence into a preset semantic coding tower, and outputting a global semantic vector representing the user's complaint intent through the semantic coding tower; performing dimensionality reduction projection on the global semantic vector based on a fully connected layer, so that the dimension of the global semantic vector is aligned with the dimension of the hidden layer sequence vector; using the dimension-aligned global semantic vector as a query vector, using the hidden layer sequence vector as a key vector, obtaining attention weights in the time dimension based on the query vector and the key vector, and using the attention weights to perform weighted summation on the hidden layer sequence vector to generate a fused feature vector; inputting the fused feature vector into a fully connected classification layer, and outputting a probability value of user churn.

[0009] In one implementation of the first aspect, processing the time-series data to obtain a multi-dimensional time-series matrix includes: extracting daily communication behavior data of the user over the past T days, wherein the communication behavior data includes multiple types of traffic data, voice data, communication quality data, and communication anomaly data; filtering out D key business dimensions from the communication behavior data, wherein the D key business dimensions include multiple types of traffic indicators, voice indicators, quality indicators, and anomaly indicators; performing daily aggregation and normalization processing on the filtered D key business dimension data to obtain D-dimensional business indicator aggregated and normalized data for the user each day; and constructing a multi-dimensional time-series matrix of dimension T×D based on the D-dimensional business indicator aggregated and normalized data.

[0010] In one implementation of the first aspect, processing the text data to obtain a text token sequence includes: acquiring text data of a user complaint; using a word segmenter to segment the text data of the user complaint into character-level segments, breaking down the text data of the user complaint into the smallest character units; inserting a classification tag CLS for carrying global information at the beginning of the smallest character unit; inserting a separator tag SEP for defining sentence boundaries at the end of the effective text of the character unit sequence to obtain an initial token sequence; performing a fixed-length conversion on the initial token sequence based on a preset maximum length threshold to obtain a fixed-length token sequence; mapping the fixed-length token sequence to a discrete integer identifier sequence, and determining the discrete integer identifier sequence as the final text token sequence.

[0011] In one implementation of the first aspect, the time-series coding tower employs an LSTM network model, which includes an input gate, a forget gate, and an output gate. The step of inputting the multidimensional time-series matrix into a preset time-series coding tower and outputting a hidden layer sequence vector containing temporal context information through the time-series coding tower includes: inputting the multidimensional time-series matrix into the LSTM network model, whereby the LSTM network model reads and processes the data of the multidimensional time-series matrix row by row in chronological order; for any given day, the LSTM network model extracts the user communication behavior feature vector for that day and the hidden layer state of the previous day from the data of the multidimensional time-series matrix, and outputs the vector through the input gate, forget gate, and output gate. The collaborative processing of the forget gate and the output gate obtains the hidden layer vector for that day: the forget gate uses the Sigmoid activation function to calculate the information retention ratio and remove invalid historical memories from the core cell state; the input gate, in conjunction with the Tanh activation function, extracts new features with high information entropy from the user communication behavior feature vector for that day and linearly superimposes them onto the core cell state to complete the dynamic update of the memory bank; the output gate performs nonlinear filtering based on the updated core cell state, calculates and outputs the hidden layer vector for that day; captures the hidden layer vectors generated by the LSTM network model for each day, and concatenates the hidden layer vectors of all days to output a hidden layer sequence vector composed of features from multiple days.

[0012] In one implementation of the first aspect, the forget gate uses the Sigmoid activation function to calculate the information retention ratio and remove invalid historical memories from the core cell state; the input gate, in conjunction with the Tanh activation function, extracts new features with high information entropy from the user communication behavior feature vector of the day and linearly superimposes them onto the core cell state to complete the dynamic update of the memory bank; the output gate performs nonlinear filtering based on the updated core cell state to calculate and output the hidden layer vector of the day, including: the forget gate maps the user communication behavior feature vector of the day and the hidden layer state of the previous day to the 0-1 interval through the Sigmoid activation function to obtain the information retention ratio to remove invalid historical memories; the input gate first filters the valid information of the current day from the user communication behavior feature vector of the day through the Sigmoid activation function, and then, in conjunction with the Tanh activation function, maps the valid information to the -1-1 interval to generate a high information entropy feature vector, and linearly superimposes the new feature vector with high information entropy onto the core cell state of the LSTM network model to achieve dynamic update of the memory bank; the output gate performs nonlinear filtering on the updated core cell state and calculates and outputs the hidden layer vector of the day.

[0013] In one implementation of the first aspect, the semantic encoding tower employs a pre-trained BERT model; the step of inputting the text token sequence into the preset semantic encoding tower and outputting a global semantic vector representing the user's complaint intent through the semantic encoding tower includes: inputting the text token sequence into the BERT model, whereby the BERT model maps the text token sequence into a composite embedding vector matrix that integrates the lexical features and absolute position features of each character in the token sequence; feeding the composite embedding vector matrix into a multi-layered Transformer bidirectional encoder of the BERT model, whereby, within each layer of the network structure, a multi-head self-attention mechanism is used to calculate the forward and backward dependency weights between any character nodes within the composite embedding vector matrix in parallel, so that the feature representations of each character are deeply integrated with the global context of the entire complaint text; after processing by the multi-layered Transformer bidirectional encoder, in the sequence output by the top layer of the BERT model, the hidden state vector corresponding to the classification label CLS at the starting position is extracted in a directional manner, and this hidden state vector is used as a global semantic vector representing the user's complaint intent, which is directly output by the semantic encoding tower.

[0014] In one implementation of the first aspect, the formula used to perform dimensionality reduction projection on the global semantic vector based on the fully connected layer, so that the dimension of the global semantic vector is aligned with the dimension of the hidden layer sequence vector, is as follows:

[0015]

[0016] in, This is the global semantic vector after dimensionality reduction. This is the global semantic vector before dimensionality reduction. It is the weight matrix, a two-dimensional real matrix whose dimension is equal to the product of the input semantic vector dimension and the target temporal feature dimension. It is the bias term, which is a one-dimensional parameter vector corresponding to the dimension of the target feature space. ReLU is the modified linear unit activation function.

[0017] In one implementation of the first aspect, obtaining the attention weights in the time dimension based on the query vector and the key vector includes: calculating the dot product correlation between the query vector and the key vector, wherein the calculation method is as follows:

[0018]

[0019] in, This represents the correlation value of the dot product. For query vector, The key vector;

[0020] The attention weights in the time dimension are obtained by normalizing the dot product correlation values ​​using the Softmax function.

[0021] Secondly, embodiments of this application provide a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the communication user churn warning method based on multimodal fusion as described in any one of the first aspects of this application.

[0022] Thirdly, embodiments of this application provide an electronic device, the electronic device including a processor and a memory; the memory stores program instructions; the processor is used to run the program instructions to execute the communication user churn warning method based on multimodal fusion as described in any one of the first aspects of this application.

[0023] The communication user churn early warning method based on multimodal fusion provided in this application has the following beneficial effects:

[0024] 1. This application constructs a dual-encoding tower collaborative processing architecture by integrating user communication time-series data and complaint text data. This effectively overcomes the representational limitations of traditional single-modal early warning methods and significantly improves the prediction accuracy and reliability of user churn early warning. The time-series encoding tower deeply mines the temporal contextual relationships of users' daily communication behavior through an LSTM network, fully preserving the micro-evolutionary characteristics of user behavior fluctuations and accurately capturing instantaneous anomalies masked by the mean. The semantic encoding tower extracts the core intent of user complaint texts based on a BERT pre-trained model, achieving accurate interpretation of user dissatisfaction and business demands. The combination of the two constructs a user churn risk profile from both behavioral data and subjective demands dimensions, fundamentally solving the problem of the mean trap. In particular, the prediction accuracy is significantly improved for sudden churn scenarios caused by sudden network failures and short-term service degradation, providing accurate data support for early intervention in churn risks.

[0025] 2. This application achieves deep fusion of temporal and semantic features through dimensional alignment and cross-attention mechanisms, enabling not only accurate prediction of churn risk but also strong business applicability and interpretability. The fully connected layer aligns the dimensions of semantic vectors with those of the temporal hidden layer, ensuring effective fusion of multimodal features. The attention mechanism generates attention weights for the time dimension, allowing for intuitive identification of key dates and core causes of user churn, automating the root cause analysis without requiring manual review of massive amounts of communication logs and complaint texts. This significantly reduces operational and maintenance costs and improves the response efficiency and targeted handling of churn risk warnings. Simultaneously, the fully connected classification layer outputs quantified churn probability values, providing a scientific basis for business personnel to develop differentiated intervention strategies, effectively reducing user churn rates and improving user retention and service quality for telecommunications companies. Attached Figure Description

[0026] Figure 1 The diagram shown is an overall flowchart of a communication user churn warning method based on multimodal fusion according to an embodiment of this application.

[0027] Figure 2 The flowchart shown is a process for obtaining a multi-dimensional time series matrix in a communication user churn warning method based on multimodal fusion according to an embodiment of this application.

[0028] Figure 3 The flowchart shown is a process for obtaining a text token sequence in a communication user churn warning method based on multimodal fusion according to an embodiment of this application.

[0029] Figure 4 The flowchart shown is a process for outputting hidden layer sequence vectors through a time-series coding tower in a communication user churn warning method based on multimodal fusion according to an embodiment of this application.

[0030] Figure 5 The diagram shown illustrates the data processing principles of the input gate, forget gate, and output gate in the LSTM network model of the communication user churn warning method based on multimodal fusion, according to an embodiment of this application.

[0031] Figure 6 The flowchart shown is a process for outputting a global semantic vector through a semantic coding tower in a communication user churn warning method based on multimodal fusion according to an embodiment of this application.

[0032] Figure 7 The diagram shown illustrates the implementation architecture of a communication user churn warning method based on multimodal fusion, according to an embodiment of this application.

[0033] Figure 8 The diagram shown is a structural schematic of an electronic device according to an embodiment of this application.

[0034] Component designation explanation

[0035] 100 electronic devices 101 memory 102 processor 103 monitor S100~S500 step S111~S114 step S121~S124 step S211~S213 step S221~S223 step Detailed Implementation

[0036] The following specific examples illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. This application can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. It should be noted that, unless otherwise specified, the following embodiments and features in the embodiments can be combined with each other.

[0037] Existing web security IP blacklists have the following problems: from the occurrence of abnormal IP access to the blocking of blacklisted IPs, the update is delayed and cannot take effect in real time; in IP-sharing scenarios, collateral blocking poses a risk of false positives; complex filtering methods affect the processing efficiency of web services and cause performance loss.

[0038] This embodiment provides a communication user churn early warning method, medium, and device based on multimodal fusion. By designing a semantically guided dual-tower spatiotemporal fusion architecture, text semantics are used as dynamic probes. An attention mechanism is used to retrieve and highlight abnormal features in long-term time-series data, thereby achieving high-precision prediction and interpretable attribution of communication user churn.

[0039] The following will refer to the appendices in the embodiments of this application. Figure 1 To be continued Figure 7 This application provides a detailed description of the technical solution of the communication user churn early warning method based on multimodal fusion in the embodiments of this application. This allows those skilled in the art to understand and implement the communication user churn early warning method based on multimodal fusion in this embodiment without inventive effort.

[0040] Figure 1 The flowchart shown is a presentation of a communication user churn warning method based on multimodal fusion, as described in an embodiment of this application. Figure 1 As shown, the communication user churn warning method based on multimodal fusion provided in this application includes the following steps S100 to S500.

[0041] Step S100: Obtain time-series data and text data of user communication, and process the time-series data and the text data respectively to obtain a multi-dimensional time-series matrix and a text token sequence;

[0042] Step S200: Input the multi-dimensional temporal matrix into a preset temporal coding tower, and output a hidden layer sequence vector containing temporal context information through the temporal coding tower; input the text token sequence into a preset semantic coding tower, and output a global semantic vector representing the user's complaint intent through the semantic coding tower.

[0043] Step S300: Perform dimensionality reduction projection on the global semantic vector based on the fully connected layer, so that the dimension of the global semantic vector is aligned with the dimension of the hidden layer sequence vector;

[0044] Step S400: Use the dimension-aligned global semantic vector as the query vector and the hidden layer sequence vector as the key vector. Obtain the attention weights in the time dimension based on the query vector and the key vector, and use the attention weights to perform a weighted summation on the hidden layer sequence vector to generate a fused feature vector.

[0045] Step S500: Input the fused feature vector into the fully connected classification layer and output the probability value of the user leaving the network.

[0046] This embodiment of the communication user churn warning method based on multimodal fusion accurately acquires user communication time-series data and text data, and obtains a multi-dimensional time-series matrix and text token sequence through normalization processing. Relying on the time-series coding tower and semantic coding tower, it deeply extracts the hidden layer sequence vector containing time context association and the global semantic vector representing the core intent of the user complaint, respectively. Then, it accurately achieves the dimensional alignment of the two-modal vectors through a fully connected layer. Using the semantic vector as the query vector and the time-series hidden layer vector as the key vector, it accurately obtains the time dimension attention weight through an attention mechanism and generates a fusion feature vector that deeply integrates two-modal information. Finally, it outputs a accurately quantified user churn probability value through a fully connected classification layer. This method innovatively achieves deep fusion of two-modal features, effectively overcomes the representation limitations of traditional single-modal warning methods, completely solves the warning bias problem caused by the average value trap, and significantly improves the accuracy and reliability of user churn warning.

[0047] The following is a combination of Figure 2 and Figure 3. Figure 7 The steps S100 to S500 of the communication user churn warning method based on multimodal fusion in this embodiment will be described in detail.

[0048] Step S100: Obtain time-series data and text data of user communication, and process the time-series data and text data respectively to obtain a multi-dimensional time-series matrix and a text token sequence.

[0049] This embodiment collects two types of core user communication data: one is user communication time-series data (such as daily data usage, call duration, signal strength, number of network outages, and other behavioral data that change over time); the other is user communication text data (mainly user complaint texts, which carry users' subjective demands and dissatisfaction).

[0050] In this embodiment, a time series matrix is ​​constructed for the user's daily communication behavior data (data traffic, calls, signal strength, etc.). For each target user, their communication behavior logs for the past T days (e.g., T=30) are extracted. D key business dimensions are selected (e.g., D=16), specifically including: traffic-related indicators (such as daily cumulative data traffic, uplink and downlink rates), voice-related indicators (such as call duration, call frequency), quality-related indicators (such as average RSRP signal strength, TCP retransmission rate), and anomaly-related indicators (such as network outages, call drop rate).

[0051] The original data was aggregated daily and normalized using Min-Max to eliminate the influence of different units, ultimately constructing a multidimensional time series matrix. Each row of this matrix represents the user's full-dimensional behavioral state on a given day, serving as the input for the subsequent LSTM time-series coding tower.

[0052] Figure 2 The flowchart shown is a process for obtaining a multi-dimensional time series matrix in a communication user churn warning method based on multimodal fusion, according to an embodiment of this application. Figure 2 As shown, in one implementation of this embodiment, processing the time series data to obtain a multidimensional time series matrix includes the following steps S111 to S114.

[0053] Step S111: Extract the user's daily communication behavior data over the past T days. The communication behavior data includes multiple types of traffic data, voice data, communication quality data, and communication anomaly data. Wherein, T is a preset positive integer, T≥1, and preferably T=30.

[0054] Step S112: Select D key business dimensions from the communication behavior data. The D key business dimensions include multiple indicators such as traffic indicators, voice indicators, quality indicators, and anomaly indicators. D is a preset positive integer, D≥1, for example, D=16.

[0055] In this embodiment, the D key business dimensions specifically include multiples of the following: daily cumulative traffic, uplink rate, downlink rate, call duration, call frequency, number of outgoing calls, number of incoming calls, average RSRP signal strength, average SINR value, TCP retransmission rate, latency, number of network outages, call drop rate, number of handover failures, number of access failures, and service interruption duration.

[0056] The traffic-related metrics include at least one of daily cumulative traffic and uplink / downlink speed; the voice-related metrics include at least one of call duration and call frequency; the quality-related metrics include at least one of average RSRP signal strength and TCP retransmission rate; and the anomaly-related metrics include at least one of network outages and call drop rate.

[0057] Step S113: Perform daily aggregation and normalization processing on the filtered D key business dimensions data to obtain daily aggregated and normalized data of D-dimensional business indicators for users.

[0058] Specifically, the raw data of the selected D key business dimensions are aggregated daily to obtain the daily D-dimensional business indicator aggregated data for each target user; then, the daily aggregated D-dimensional business indicator aggregated data is subjected to Min-Max normalization to eliminate the influence of different units on the data.

[0059] Specifically, the daily aggregation process involves: for each target user and each key business dimension, statistically aggregating all the original data under that dimension on that day. The statistical aggregation method includes at least one of summation, averaging, finding the maximum value, and finding the minimum value.

[0060] The formula for calculating Min-Max normalization is:

[0061]

[0062] in, Here, x represents the normalized data, and x represents the original aggregated data before normalization. This is the minimum value among the original aggregated data for all users and all days under this key business dimension. This refers to the maximum value in the original aggregated data for all users and all days under this key business dimension.

[0063] Step S114: Based on the aggregated and normalized data of the D-dimensional business indicators, a multi-dimensional time series matrix with dimension T×D is constructed.

[0064] Each row of the multidimensional time series matrix corresponds to the D-dimensional full behavioral state of the target user on a certain day, and the multidimensional time series matrix is ​​used as the input of the subsequent LSTM time-series coding tower.

[0065] Figure 3 The flowchart shown is a process for obtaining a text token sequence in a communication user churn warning method based on multimodal fusion, according to an embodiment of this application. Figure 3 As shown, in one implementation of this embodiment, processing the text data to obtain a text token sequence includes the following steps S121 to S124.

[0066] Step S121: Obtain the text data of the user complaint, and use a word segmenter to perform character-level segmentation on the text data of the user complaint, breaking down the text data of the user complaint into the smallest character units.

[0067] A word segmenter matched with a pre-trained language model is used to segment the unstructured user complaint text at the character level. The word segmenter is matched with the pre-trained language model specifically: the word segmenter's word list and segmentation rules are consistent with the underlying word list and semantic encoding rules of the pre-trained language model, ensuring that the segmented character units can be recognized by the pre-trained language model.

[0068] In this embodiment, the character granularity segmentation specifically involves treating each Chinese character and punctuation mark in the user complaint text as an independent smallest character unit, without performing word-level merging processing.

[0069] In step S122, a classification marker CLS for carrying global information is inserted at the beginning of the smallest character unit; and a separator marker SEP for defining sentence boundaries is inserted at the end of the valid text of the character unit sequence to obtain the initial token sequence.

[0070] Step S123: Perform a fixed-length transformation on the initial token sequence based on a preset maximum length threshold to obtain a fixed-length token sequence. That is, perform fixed-length processing on the obtained initial token sequence based on the preset maximum length threshold to meet the tensor computation requirements of the neural network with fixed input dimensions.

[0071] The preset maximum length threshold is a preset positive integer, set according to the input dimension requirements of the subsequent semantic coding tower, and its value ranges from 128 to 1024.

[0072] The fixed-length processing specifically involves: if the total length of the initial token sequence exceeds the preset maximum length threshold, the excess portion at the end of the sequence is truncated and discarded, while the content carrying the core message at the beginning of the sequence is retained; if the total length of the initial token sequence is less than the preset maximum length threshold, meaningless padding markers are continuously appended to the end of the sequence until the total length reaches the preset maximum length threshold, resulting in a fixed-length token sequence. For example, extracting customer service complaint tickets from the same period, after word segmentation and cleaning, results in a token sequence of length 128.

[0073] The meaningless padding mark is a preset fixed character. The padding mark does not carry any textual semantic information and does not repeat the CLS mark, SEP mark, or segmented character unit.

[0074] Step S124: Map the fixed-length token sequence to a discrete integer identifier sequence, and determine the discrete integer identifier sequence as the final text token sequence.

[0075] Specifically, the obtained fixed-length token sequence is mapped to a discrete integer identifier sequence corresponding to the underlying vocabulary of the model. The discrete integer identifier sequence is the final fixed-length token sequence, which serves as the standard input feature source for the subsequent semantic coding tower.

[0076] In one embodiment, the mapping process specifically involves: querying the underlying vocabulary of the model, and mapping each token (CLS token, SEP token, character unit, padding token) in the fixed-length token sequence to a unique discrete integer identifier, thereby forming a discrete integer identifier sequence.

[0077] Therefore, in this embodiment, user complaint text is segmented using a word segmentation tool and converted into a fixed-length token sequence. For unstructured user complaint text, a word segmenter matching the pre-trained language model is used for character-level segmentation, breaking the text down into the smallest character units. After basic segmentation, the system forcibly inserts a classification tag (CLS) to carry global information at the beginning of the sequence and a separator tag (SEP) to define sentence boundaries at the end of the effective text. To meet the tensor computation requirements of the fixed input dimension of the neural network, the system performs a fixed-length transformation operation on the sequence: when the total length of the sequence exceeds a preset maximum length threshold, the excess portion at the end is directly truncated and discarded to retain the core message at the beginning; when the total length of the sequence is less than the preset threshold, meaningless padding tags are continuously added to the end of the sequence until the total length reaches the threshold. After the above standardization process, the original natural language text is finally accurately mapped to a discrete integer identifier sequence corresponding to the underlying vocabulary of the model, i.e., a fixed-length token sequence, which serves as the standard input feature source for the subsequent semantic coding tower.

[0078] Step S200: Input the multidimensional temporal matrix into a preset temporal coding tower, and output a hidden layer sequence vector containing temporal context information through the temporal coding tower; input the text token sequence into a preset semantic coding tower, and output a global semantic vector representing the user's complaint intent through the semantic coding tower.

[0079] In this embodiment, the temporal coding tower is used to capture the temporal dependence of user behavior. The temporal coding tower uses a Long Short-Term Memory (LSTM) network model as its core component. The LSTM network model includes an input gate, a forget gate, and an output gate, which can effectively solve the gradient vanishing problem in long sequence training.

[0080] Figure 4 The flowchart shown is a process for outputting hidden layer sequence vectors through a timing coding tower in a communication user churn warning method based on multimodal fusion according to an embodiment of this application. Figure 4 As shown, in one implementation of this embodiment, the step of inputting the multidimensional time series matrix into a preset time series coding tower and outputting a hidden layer sequence vector containing time context information through the time series coding tower includes the following steps S211 to S213.

[0081] Step S211: Input the multidimensional time series matrix into the LSTM network model. The LSTM network model reads and processes the data of the multidimensional time series matrix row by row in the order of time steps.

[0082] Step S212: For any given day, the LSTM network model extracts the user communication behavior feature vector for that day and the hidden layer state of the previous day from the data in the multi-dimensional time series matrix. Through the collaborative processing of the input gate, the forget gate, and the output gate, the hidden layer vector for that day is obtained. The forget gate uses the Sigmoid activation function to calculate the information retention ratio and remove invalid historical memories from the core cell state. The input gate, in conjunction with the Tanh activation function, extracts new features with high information entropy from the user communication behavior feature vector for that day and linearly superimposes them onto the core cell state, completing the dynamic update of the memory bank. The output gate performs non-linear filtering based on the updated core cell state, calculates and outputs the hidden layer vector for that day.

[0083] Step S213: Capture the hidden layer vector generated by the LSTM network model for each day, and concatenate the hidden layer vectors of all days to output a hidden layer sequence vector composed of features from multiple days.

[0084] Figure 5 This diagram illustrates the data processing principles of the input gate, forget gate, and output gate in an LSTM network model within a multimodal fusion-based communication user churn warning method according to an embodiment of this application. Figure 5 As shown, the forget gate uses the Sigmoid activation function to calculate the information retention ratio and remove invalid historical memories from the core cell state; the input gate, in conjunction with the Tanh activation function, extracts new features with high information entropy from the user communication behavior feature vector of that day and linearly superimposes them onto the core cell state to complete the dynamic update of the memory bank; the output gate performs nonlinear filtering based on the updated core cell state, calculates and outputs the hidden layer vector of that day, including:

[0085] 1) The forget gate maps the user's communication behavior feature vector for the day and the hidden layer state of the previous day to the 0-1 interval through the Sigmoid activation function to obtain the information retention ratio in order to strip out invalid historical memories;

[0086] 2) The input gate first filters the effective information of the current day from the user communication behavior feature vector of the day through the Sigmoid activation function, and then coordinates with the Tanh activation function to map the effective information to the -1-1 interval to generate a high information entropy feature vector. The newly added feature vector with high information entropy is linearly superimposed on the core cell state of the LSTM network model to realize the dynamic update of the memory bank.

[0087] 3) The output gate performs nonlinear filtering on the updated core cell state, calculates and outputs the hidden layer vector for the day. The updated state refers to the core cell state updated by the LSTM network after the input gate completes the linear superposition of new features and the forget gate completes the stripping of invalid memories (i.e., the final state after the dynamic update of the memory bank, which integrates historical valid memories and newly added behavioral features for the day).

[0088] The specific implementation of the nonlinear filtering is as follows: the output gate does not directly output the core cell state, but first receives three types of inputs: the refreshed core cell state, the communication behavior feature vector of the current time step, and the hidden layer state of the previous time step; then, the Sigmoid activation function is used to generate filtering coefficients (mapped to the 0-1 interval), and the Tanh activation function is used to map the refreshed core cell state to the -1-1 interval. The two are multiplied element by element to complete the nonlinear screening of the core cell state, filtering out irrelevant and redundant memory information and retaining the key features related to the communication behavior of the day.

[0089] After the above nonlinear filtering, the final vector is the hidden layer vector of the current time step (day). This vector carries the core features of the user's communication behavior on the day and incorporates historical context information. It will be captured by the system and used for the concatenation of subsequent hidden layer sequence vectors as a feature retrieval source for the cross-attention mechanism.

[0090] Therefore, in this embodiment, for the input time-series matrix, the Long Short-Term Memory (LSTM) network reads and processes the data row by row in chronological order of time steps. On any given monitoring day, the LSTM network synchronously receives the communication behavior feature vector of that day and the hidden layer state of the previous day. The forget gate inside the LSTM first uses the Sigmoid activation function to calculate an information retention ratio, determining how much invalid historical behavior memory to remove from the core cell state; then, the input gate, in conjunction with the Tanh activation function, extracts features with high information entropy from the new behavior data of the day and linearly superimposes them onto the core cell state to complete the dynamic update of the memory bank; finally, the output gate performs nonlinear filtering based on the refreshed state, calculates and releases the hidden layer vector of the day. In this process, the system is configured to fully capture the hidden layer vectors generated by the network at each independent time step, and finally concatenates and outputs a high-dimensional matrix composed of features from multiple days. This hidden layer sequence vector meticulously preserves a snapshot of the micro-evolution of the user's daily behavior fluctuations, providing a feature retrieval source accurate to the day for the subsequent cross-attention mechanism. After the input time series matrix is ​​processed by the LSTM unit, the output is a hidden layer sequence vector containing time context information. This sequence preserves the state of each time step of each day, rather than just outputting the result of the last time step.

[0091] In this embodiment, the semantic encoding tower is used to extract deep semantics from the text. The semantic encoding tower uses a pre-trained BERT model to encode the complaint text sequence. The semantic encoding tower uses a BERT pre-trained model with fixed weights as a static feature extractor. The BERT model is based on a Transformer bidirectional encoder architecture, which can capture the polysemy of words in context.

[0092] Figure 6 The flowchart shown is a process for outputting a global semantic vector through a semantic coding tower in a communication user churn warning method based on multimodal fusion according to an embodiment of this application. Figure 6 As shown, in one implementation of this embodiment, the semantic encoding tower adopts a pre-trained BERT model; the step of inputting the text token sequence into the preset semantic encoding tower and outputting a global semantic vector representing the user's complaint intent through the semantic encoding tower includes the following steps S221 to S223.

[0093] Step S221: Input the text token sequence into the BERT model, and the BERT model maps the text token sequence into a composite embedding vector matrix that integrates the word features and absolute position features of each character in the token sequence.

[0094] Step S222: The composite embedding vector matrix is ​​fed into the multi-layered Transformer bidirectional encoder of the BERT model. In each layer of the network structure, the forward and backward dependency weights between any character nodes in the composite embedding vector matrix are calculated in parallel through a multi-head self-attention mechanism, so that the feature representations of each character are deeply integrated with the global context of the entire complaint text.

[0095] Step S223: After processing by multiple layers of Transformer bidirectional encoders, the hidden state vector corresponding to the classification label CLS at the starting position is extracted from the sequence output at the top layer of the BERT model. This hidden state vector is used as a global semantic vector representing the user's complaint intent and is directly output by the semantic encoding tower.

[0096] Since the classification marker CLS has completed full feature aggregation with all valid characters in the Token sequence during the layer-by-layer information transmission and calculation, its corresponding hidden state vector highly condenses the macro sentiment and core business demands of the user complaint text. This hidden state vector is used as a global semantic vector representing the user's complaint intent and is directly output by the semantic coding tower.

[0097] In this embodiment, the BERT pre-trained model is a static feature extractor with fixed weights, and the training parameters are not dynamically updated. The multi-head self-attention mechanism realizes deep interaction of global context information within the token sequence by parallel computing the forward and backward dependency weights between each character node, ensuring that the features of each character can fully integrate the semantic information of the whole sentence. The global semantic vector is a high-dimensional hidden state vector corresponding to the classification label CLS, which can accurately represent the core intent of the user's complaint, including but not limited to business-related demands such as dissatisfaction with tariffs and poor network signal.

[0098] Specifically, the process of outputting a global semantic vector through the semantic encoding tower is as follows: The global semantic vector output by the BERT model represents the core intent of the user's complaint, such as dissatisfaction with the tariff or poor network signal. In this embodiment, the semantic encoding tower directly loads a pre-trained BERT model with fixed weights as a static feature extractor. When processing the input text, a fixed-length token sequence with the first classification label CLS is first mapped into a composite embedding vector matrix that integrates the features of the words themselves and the absolute position features. Subsequently, this matrix is ​​fed into a multi-layered Transformer bidirectional encoder. Within each layer of the network structure, the multi-head self-attention mechanism promotes the deep integration of the feature representations of each character with the global context of the entire sentence by parallelly calculating the forward and backward dependency weights between any character nodes in the sequence. After multi-layer self-attention interaction and nonlinear transformation, the system extracts the hidden state vector corresponding to the classification label CLS located at the starting position in the top-level output sequence of the model. Because the classification label has completed full feature aggregation with all valid characters in the text sequence during the layer-by-layer information transmission and calculation, the high-dimensional state vector highly condenses the macro sentiment and core business demands of the entire text, and is ultimately directly output by the system as a global semantic vector representing the user's complaint intent.

[0099] Step S300: Perform dimensionality reduction projection on the global semantic vector based on the fully connected layer, so that the dimension of the global semantic vector is aligned with the dimension of the hidden layer sequence vector.

[0100] Since semantic vectors typically have high dimensionality (e.g., 768 dimensions) while temporal feature vectors have low dimensionality (e.g., 64 dimensions), they reside in different feature spaces and cannot be directly subjected to mathematical operations. This embodiment introduces a learnable linear mapping layer (fully connected layer) to perform dimensionality reduction projection on the semantic vectors.

[0101] In one implementation of this embodiment, the formula used to perform dimensionality reduction projection on the global semantic vector based on the fully connected layer, so that the dimension of the global semantic vector is aligned with the dimension of the hidden layer sequence vector, is as follows:

[0102]

[0103] in, This is the global semantic vector after dimensionality reduction. This is the global semantic vector before dimensionality reduction. It is the weight matrix, a two-dimensional real matrix whose dimension is equal to the product of the input semantic vector dimension and the target temporal feature dimension. It is the bias term, which is a one-dimensional parameter vector corresponding to the dimension of the target feature space. ReLU is the modified linear unit activation function.

[0104] After performing dimensionality reduction projection on the global semantic vector, the dimension of the semantic vector is compressed to be consistent with that of the temporal vector (e.g., 64 dimensions), thereby solving the problem of heterogeneous data interaction.

[0105] In this embodiment, the weight matrix is ​​a two-dimensional real matrix with dimensions equal to the product of the input semantic vector dimension and the target temporal feature dimension, and the bias term is a one-dimensional parameter vector with dimensions corresponding to the target feature space. Regarding the parameter setting and update mechanism, the internal elements of both do not rely on pre-set human experience, but rather participate in the entire lifecycle calculation as learnable parameters. During the end-to-end joint training phase of the model, adaptive iteration is performed based on the gradient backpropagated from the prediction error.

[0106] The principle of dimensionality reduction alignment is as follows: First, matrix multiplication is performed on the high-dimensional initial semantic vector and the weight matrix. The original 768 text features are re-weighted and information converged using the 64 independently distributed hidden layer nodes in the matrix, realizing linear projection compression of the feature space from high-dimensional sparse to low-dimensional dense. Then, the system adds the linear projection result to the bias term, assigning an independent basic activation offset to each compressed feature dimension. Finally, the system applies a modified linear unit activation function to the whole, directly filtering out negative response noise in the operation matrix and introducing non-linear expressive power, thereby smoothly purifying the complex text intent into a 64-dimensional low-dimensional semantic representation that is completely aligned with the temporal vector.

[0107] Step S400: Use the dimension-aligned global semantic vector as the query vector and the hidden layer sequence vector as the key vector. Obtain the attention weights in the time dimension based on the query vector and the key vector, and use the attention weights to perform a weighted summation on the hidden layer sequence vector to generate a fused feature vector.

[0108] This embodiment designs a semantically guided cross-attention mechanism to find corresponding abnormal time points using complaint content. The aligned semantic vector is used as the query vector, and the temporal feature sequence is used as the key vector; the dot product correlation between the two is calculated. The larger the dot product result, the higher the correlation between the behavioral features at that moment and the complaint content.

[0109] In one implementation of this embodiment, obtaining the attention weights in the time dimension based on the query vector and the key vector includes: calculating the dot product correlation between the query vector and the key vector, wherein the calculation method is as follows:

[0110]

[0111] in, This represents the correlation value of the dot product. For query vector, The key vector;

[0112] The dot product correlation values ​​are normalized using the Softmax function to obtain the attention weights in the time dimension. Then, these attention weights are used to perform a weighted summation of the hidden layer sequence vectors to generate a fused feature vector.

[0113]

[0114] This process achieves a denoising effect: the weight of normal time periods unrelated to the complaint semantics is suppressed, while the weight of abnormal time periods leading to complaints is amplified. This embodiment, through a semantic guidance mechanism, enables the model to capture instantaneous anomalies masked by averages, solving the average trap problem, and significantly improving the prediction accuracy for network outages caused by sudden network failures. Even if two users' time-series data are completely identical, if user A complains about high costs and user B complains about slow internet, the model will focus on the date of the charge and the date of signal fluctuation respectively, based on the different semantic vectors, generating differentiated risk assessment results.

[0115] Step S500: Input the fused feature vector into the fully connected classification layer and output the probability value of the user leaving the network.

[0116] A fully connected classification layer is a type of neural network layer that contains a one-dimensional learnable weight matrix and bias terms. It performs linear transformations and nonlinear mappings on the high-dimensional abstract features of the input, and finally outputs a prediction result that meets the task requirements (the probability value of user leaving the network in this embodiment).

[0117] The fused feature vector is fed into a fully connected classification layer. First, the feature vector is input into the fully connected classification layer, where a dot product is performed between the input vector and the one-dimensional learnable weight matrix within the layer, and a bias term is added. This linearly compresses the multi-dimensional abstract features into a single-dimensional scalar score, which directly reflects the target user's overall churn tendency. The probability value of the user churning (between 0 and 1) is then output using the Sigmoid function.

[0118] In this embodiment, the fully connected classification layer also outputs an attention weight sequence. Operations personnel can visualize this weight distribution: the date with the highest weight is the key time point leading to user churn, achieving a shift from black-box prediction to white-box attribution. Therefore, by outputting an attention weight distribution map, this embodiment allows business personnel to intuitively see the specific dates and reasons for user churn without manually sifting through massive amounts of logs.

[0119] After the fused feature vector is fed into the fully connected classification layer, the fully connected classification layer completes the off-network probability output according to the following process:

[0120] 1) The fully connected classification layer performs a dot product summation operation with the input fusion feature vector through a pre-set one-dimensional learnable weight matrix within the layer, and simultaneously superimposes a pre-set bias term. Through this linear transformation, the multi-dimensional abstract fusion features are compressed into a single-dimensional scalar score. The magnitude of this scalar score directly corresponds to the target user's comprehensive churn tendency. The higher the score, the greater the likelihood of the user churning.

[0121] 2) Input the obtained scalar score into the Sigmoid activation function. Through the non-linear mapping effect of the Sigmoid function, the scalar score is transformed into a value between 0 and 1. This value is the user's churn probability value. The closer the probability value is to 1, the higher the risk of the user churning; the closer it is to 0, the lower the risk of the user churning.

[0122] Based on the preset parameters in this embodiment, supplementary examples are provided to illustrate the input, processing, and related parameters of the fully connected classification layer, clarifying the relationship between each parameter and the probability output:

[0123] 1. Input relevant parameters: The dimension of the fused feature vector is consistent with the dimension of the temporal feature and the dimension of the aligned semantic vector, that is, 64 dimensions (based on the parameters of this embodiment: the dimension of the LSTM hidden layer is set to 64 dimensions, and the projection layer compresses the 768-dimensional original semantic vector output by BERT to 64 dimensions to ensure that the dimensions are consistent after the dual-modal feature fusion).

[0124] 2. Time window parameter: The time window T is set to 30 days, which means that the fused feature vector contains the time-series feature information of the user in the past 30 days. The probability value output by the fully connected classification layer is based on the user's communication behavior in the past 30 days and the semantic judgment of complaints during the period.

[0125] 3. Training parameters: The model uses the binary cross-entropy loss function for end-to-end training. By continuously iterating and optimizing the learnable weight matrix and bias terms of the fully connected classification layer, the accuracy of off-network probability prediction is improved.

[0126] 4. Output example: If the scalar score obtained after the fused feature vector is linearly transformed by the fully connected classification layer is 2.3, and the output churn probability value after mapping by the Sigmoid function is 0.91, it indicates that the user has an extremely high risk of churn. At the same time, the visualization of the attention weight sequence shows that the weight is the highest on the 12th day, which means that the communication anomaly on the 12th day (such as dropped connection, poor signal, etc.) is the key time node that leads to the user churn.

[0127] Simultaneously, while outputting the churn probability value, the model also outputs the attention weight sequence. Operations personnel can visualize this weight distribution, with the date with the highest weight indicating the critical time point leading to user churn. This breaks through the limitations of traditional black-box prediction models, achieving white-box attribution of churn risk, facilitating operations personnel to accurately pinpoint the causes of churn and develop targeted intervention strategies. Furthermore, during model training, a binary cross-entropy loss function is used for end-to-end training, ensuring that the churn probability values ​​output by the classification layer are accurate and reliable, closely reflecting actual business scenarios.

[0128] Figure 7 The diagram shown illustrates the implementation architecture of a communication user churn warning method based on multimodal fusion, according to an embodiment of this application. Figure 7 As shown in the figure, the implementation process of the communication user churn early warning method based on multimodal fusion in this embodiment is divided into data construction, dual-tower coding, feature fusion and prediction. The specific implementation process is as follows:

[0129] 1) Data Construction: Acquire and process two types of input data to provide standardized input for subsequent coding. On the one hand, construct time-series input: Backtrack the target user's daily communication behavior data over the past 30 days. The data includes key indicators such as average daily data traffic, call duration, and number of dropped calls. Construct a time-series matrix based on this daily data, preserving the original fluctuation characteristics of the daily data throughout, without monthly average aggregation, to ensure that the time-series data can truly reflect the user's daily behavioral changes. On the other hand, construct text input: Extract customer service complaint tickets for the user and the same period mentioned above. Perform word segmentation and data cleaning on the complaint ticket text, remove invalid and redundant information, and then truncate it into a token sequence of fixed length of 128 to complete the standardized processing of text input.

[0130] 2) Dual-Tower Encoding: Core features of two types of data are extracted using temporal and semantic towers respectively. Temporal Tower Processing: The constructed temporal matrix is ​​input into a pre-defined LSTM network with a hidden layer dimension of 64. The network processes the temporal matrix data row by row according to the chronological order of time steps, fully preserving the state information of 30 time steps, and finally outputting a sequence vector containing the states of 30 time steps, achieving deep extraction of temporal features. Semantic Tower Processing: The previously obtained 128-character text token sequence is input into the BERT pre-trained model. Through bidirectional encoding and feature aggregation, a global semantic vector with a dimension of 768 is extracted. This vector can accurately represent the core demands and sentiment tendencies in user complaint tickets.

[0131] 3) Feature fusion and prediction: Effective fusion of dual-modal features to output churn warning results. First, dimensional alignment is performed: the 768-dimensional global semantic vector output by the semantic tower is input into a fully connected layer. Through the mapping effect of the fully connected layer, it is converted into a 64-dimensional semantic vector, ensuring that its dimension is consistent with the sequence vector output by the temporal tower, laying the foundation for subsequent fusion. Next, attention calculation is performed: the 64-dimensional semantic vector after dimensional alignment is used as the query vector, and the sequence vector containing 30 time steps output by the temporal tower is used as the key vector. The attention weight of the temporal dimension is obtained by calculating the correlation between the two. For example, when the text token sequence contains churn-related semantics, the model will automatically assign a higher attention weight to the time step corresponding to the network outage day (e.g., day 5) in the temporal vector, achieving precise guidance of temporal features by semantic intent. Finally, prediction output is performed: the attention-weighted fused feature vector is input into a Sigmoid classification layer. The Sigmoid activation function maps the output result to the 0-1 interval. This output result is the user's churn probability, completing the entire implementation process of user churn warning.

[0132] The scope of protection of the communication user churn warning method based on multimodal fusion described in this application is not limited to the execution order of the steps listed in this embodiment. Any solution implemented by adding, deleting, or replacing steps in the prior art based on the principles of this application is included within the scope of protection of this application.

[0133] This application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the communication user churn warning method based on multimodal fusion provided in any embodiment of this application.

[0134] In this embodiment, any combination of one or more storage media can be used. The storage medium can be a computer-readable signal medium or a computer-readable storage medium. A computer-readable storage medium can be, for example—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable computer disk, a hard disk, RAM, ROM, erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this embodiment, the computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0135] This application also provides an electronic device. Figure 8 The diagram shown is a structural schematic of the electronic device 100 provided in an embodiment of this application. In some embodiments, the electronic device may be a terminal device such as an in-vehicle device, a seat control device, or a smart screen.

[0136] like Figure 8 As shown, the electronic device 100 provided in this application embodiment includes a memory 101 and a processor 102.

[0137] The memory 101 is used to store computer programs; preferably, the memory 101 includes various media that can store program code, such as ROM, RAM, magnetic disk, USB flash drive, memory card or optical disk.

[0138] Specifically, memory 101 may include computer system readable media in the form of volatile memory, such as random access memory (RAM) and / or cache memory. Electronic device 100 may further include other removable / non-removable, volatile / non-volatile computer system storage media. Memory 101 may include at least one program product having a set (e.g., at least one) of program modules configured to perform the functions of the embodiments of this application.

[0139] The processor 102 is connected to the memory 101 and is used to execute the computer program stored in the memory 101 so that the electronic device 100 executes the communication user churn warning method based on multimodal fusion provided in any embodiment of this application.

[0140] Optionally, the processor 102 may be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it may also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0141] Optionally, in this embodiment, the electronic device 100 may further include a display 103. The display 103 is communicatively connected to the memory 101 and the processor 102, and is used to display the relevant GUI interactive interface of the communication user churn warning method based on multimodal fusion.

[0142] The above embodiments are merely illustrative of the principles and effects of this application and are not intended to limit this application. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of this application. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in this application should still be covered by the claims of this application.

Claims

1. A method for early warning of user churn based on multimodal fusion, characterized in that, include: Acquire time-series data and text data of user communication, and process the time-series data and text data respectively to obtain a multi-dimensional time-series matrix and a text token sequence; The multidimensional temporal matrix is ​​input into a preset temporal encoding tower, and the hidden layer sequence vector containing temporal context information is output through the temporal encoding tower; the text token sequence is input into a preset semantic encoding tower, and the global semantic vector representing the user's complaint intent is output through the semantic encoding tower. The global semantic vector is dimensionality-reduced and projected using a fully connected layer, so that the dimension of the global semantic vector is aligned with the dimension of the hidden layer sequence vector. The dimension-aligned global semantic vector is used as the query vector, and the hidden layer sequence vector is used as the key vector. Attention weights in the time dimension are obtained based on the query vector and the key vector, and the attention weights are used to perform a weighted summation on the hidden layer sequence vector to generate a fused feature vector. The fused feature vector is input into a fully connected classification layer, which outputs the probability value of the user leaving the network.

2. The communication user churn early warning method based on multimodal fusion according to claim 1, characterized in that, Processing the time-series data to obtain a multi-dimensional time-series matrix includes: Extract daily communication behavior data of users over the past T days, including multiple types of traffic data, voice data, communication quality data, and communication anomaly data; D key business dimensions are selected from the communication behavior data. The D key business dimensions include multiple indicators such as traffic indicators, voice indicators, quality indicators, and anomaly indicators. The data of the selected D key business dimensions are aggregated and normalized on a daily basis to obtain the daily aggregated and normalized data of the D-dimensional business indicators of the user. Based on the aggregated and normalized data of the D-dimensional business indicators, a multi-dimensional time series matrix with dimension T×D is constructed.

3. The communication user churn early warning method based on multimodal fusion according to claim 1, characterized in that, Processing the text data to obtain a text token sequence includes: The text data of user complaints is obtained, and a word segmenter is used to segment the text data of user complaints into character-level segments, breaking down the text data of user complaints into the smallest character units; At the beginning of the smallest character unit, a classification marker CLS for carrying global information is inserted; at the end of the valid text of the character unit sequence, a separator marker SEP for defining sentence boundaries is inserted to obtain the initial token sequence. The initial token sequence is converted to a fixed length based on a preset maximum length threshold to obtain a fixed-length token sequence. The fixed-length token sequence is mapped to a discrete integer identifier sequence, and the discrete integer identifier sequence is determined as the final text token sequence.

4. The communication user churn early warning method based on multimodal fusion according to claim 1 or 2, characterized in that, The time-series coding tower employs an LSTM network model, which includes an input gate, a forget gate, and an output gate. The process of inputting the multi-dimensional time series matrix into a preset time-series coding tower and outputting a hidden layer sequence vector containing temporal context information through the time-series coding tower includes: The multidimensional time series matrix is ​​input into the LSTM network model, and the LSTM network model reads and processes the data of the multidimensional time series matrix row by row in the order of time steps. For any given day, the LSTM network model extracts the user communication behavior feature vector and the hidden layer state of the previous day from the data in the multi-dimensional time series matrix. Through the collaborative processing of the input gate, the forget gate, and the output gate, it obtains the hidden layer vector for that day: the forget gate uses the Sigmoid activation function to calculate the information retention ratio and remove invalid historical memories from the core cell state; the input gate, in conjunction with the Tanh activation function, extracts new features with high information entropy from the user communication behavior feature vector for that day and linearly superimposes them onto the core cell state, completing the dynamic update of the memory bank; the output gate performs non-linear filtering based on the updated core cell state, calculates and outputs the hidden layer vector for that day. Capture the hidden layer vectors generated by the LSTM network model for each day, and concatenate the hidden layer vectors of all days to output a hidden layer sequence vector composed of features from multiple days.

5. The communication user churn early warning method based on multimodal fusion according to claim 4, characterized in that, The forget gate uses the Sigmoid activation function to calculate the information retention ratio and remove invalid historical memories from the core cell state; the input gate, in conjunction with the Tanh activation function, extracts new features with high information entropy from the user communication behavior feature vector of the day and linearly superimposes them onto the core cell state to complete the dynamic update of the memory bank; the output gate performs nonlinear filtering based on the updated core cell state and calculates and outputs the hidden layer vector of the day, including: The forget gate maps the user's communication behavior feature vector for the day and the hidden layer state of the previous day to the 0-1 interval through the Sigmoid activation function to obtain the information retention ratio in order to strip away invalid historical memories. The input gate first filters the valid information of the current day from the user communication behavior feature vector of the day through the Sigmoid activation function, and then coordinates with the Tanh activation function to map the valid information to the -1-1 interval to generate a high information entropy feature vector. The newly added high information entropy feature vector is linearly superimposed on the core cell state of the LSTM network model to realize the dynamic update of the memory bank. The output gate performs nonlinear filtering on the updated core cell state, calculates and outputs the hidden layer vector for the day.

6. The communication user churn early warning method based on multimodal fusion according to claim 3, characterized in that, The semantic encoding tower employs a pre-trained BERT model; the process of inputting the text token sequence into the preset semantic encoding tower and outputting a global semantic vector representing the user's complaint intent through the semantic encoding tower includes: The text token sequence is input into the BERT model, and the BERT model maps the text token sequence into a composite embedding vector matrix that integrates the word features and absolute position features of each character in the token sequence. The composite embedding vector matrix is ​​fed into the multi-layered Transformer bidirectional encoder of the BERT model. Within each network layer, the forward and backward dependency weights between any character nodes in the composite embedding vector matrix are calculated in parallel through a multi-head self-attention mechanism, so that the feature representations of each character are deeply integrated with the global context of the entire complaint text. After processing by multiple layers of Transformer bidirectional encoders, the hidden state vector corresponding to the classification label CLS at the starting position is extracted from the sequence output at the top layer of the BERT model. This hidden state vector is used as a global semantic vector representing the user's complaint intent and is directly output by the semantic encoding tower.

7. The communication user churn early warning method based on multimodal fusion according to claim 1, characterized in that, The formula used to perform dimensionality reduction projection on the global semantic vector based on the fully connected layer, so that the dimension of the global semantic vector is aligned with the dimension of the hidden layer sequence vector, is as follows: in, This is the global semantic vector after dimensionality reduction. This is the global semantic vector before dimensionality reduction. It is the weight matrix, a two-dimensional real matrix whose dimension is equal to the product of the input semantic vector dimension and the target temporal feature dimension. It is the bias term, which is a one-dimensional parameter vector corresponding to the dimension of the target feature space. ReLU is the modified linear unit activation function.

8. The communication user churn early warning method based on multimodal fusion according to claim 1, characterized in that, The attention weights obtained based on the query vector and the key vector in the time dimension include: The dot product correlation between the query vector and the key vector is calculated as follows: in, This represents the correlation value of the dot product. For query vector, The key vector; The attention weights in the time dimension are obtained by normalizing the dot product correlation values ​​using the Softmax function.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the communication user churn warning method based on multimodal fusion as described in any one of claims 1 to 8.

10. An electronic device, characterized in that, The electronic device includes: Processor and memory; The memory stores program instructions; The processor is configured to run the program instructions to execute the communication user churn warning method based on multimodal fusion as described in any one of claims 1 to 8.

Images