A data processing method and system based on multi-time scale dynamic behavior

Abstract

The application relates to a data processing method and system based on multi-time scale dynamic behavior, belonging to the field of customer behavior data processing, and the method comprises the following steps: acquiring high-level summary information and numerical characteristic representation for multi-time scale historical behavior data including category characteristics and numerical characteristics; acquiring a time token embedding vector sequence of each time scale based on original category characteristics, high-level summary information and numerical characteristic representation; acquiring parameters of the time token embedding vector sequence for describing the stability and fluctuation degree of user behavior and performing fusion processing to obtain an enhanced time sequence; performing compression processing on the enhanced time sequence to acquire a feature vector with long-term overall performance and capable of focusing on the latest state of a user; and acquiring a loss probability by processing the feature vector of each time scale based on a dynamic gating weight and an activation function. The application avoids information omission of a single time perspective and improves the timeliness and accuracy of prediction.

Description

Technical Field

[0001] This invention relates to the field of customer behavior data processing technology, and in particular to a data processing method and system based on dynamic behavior over multiple time scales. Background Technology

[0002] In the digital content service industry, as market competition intensifies and user scale gradually reaches its peak, companies generally face the challenge of continuously rising marginal customer acquisition costs. Against this backdrop, reducing customer churn and extending user lifetime value have become key to improving operational efficiency and profitability.

[0003] Existing customer relationship management (CRM) solutions often rely on static data for customer churn prediction, failing to fully capture the dynamic evolution of customer behavior over time. Current methods cannot incorporate long-term dependencies within user behavior sequences or information on the influence of external environmental factors. Consequently, churn risk identification often lags behind actual behavioral changes, causing businesses to miss the optimal intervention window for customer retention.

[0004] Furthermore, as user behavior data becomes increasingly dimensional and spans longer, traditional machine learning models have significant limitations in feature extraction capabilities, temporal modeling depth, and multi-scale pattern recognition, making it difficult to effectively extract predictive dynamic features from massive amounts of high-dimensional data.

[0005] Therefore, how to achieve accurate and interpretable customer status prediction in complex time-series scenarios, and generate personalized intervention strategies accordingly, is a technical bottleneck that urgently needs to be overcome in the field of intelligent customer relationship management. Summary of the Invention

[0006] (a) Technical problems to be solved

[0007] In view of the above-mentioned shortcomings and deficiencies of the prior art, the present invention provides a data processing method and system based on dynamic behavior at multiple time scales, which solves the technical problems of the prior art, which relies on static data and is difficult to capture the dynamic evolution of customer behavior, has recognition lag, and has insufficient prediction accuracy.

[0008] (II) Technical Solution

[0009] To achieve the above objectives, the main technical solutions adopted by the present invention include:

[0010] In a first aspect, embodiments of the present invention provide a data processing method based on dynamic behavior across multiple time scales. This method is used for personalized customer relationship management in the digital content service industry. The method includes:

[0011] For historical behavioral data with multiple time scales including categorical and numerical features, we obtain high-level summary information of the label vector sequence to which the categorical features belong, and obtain numerical feature representations of numerical features that include the interaction strength on the hidden dimensions.

[0012] Based on the original category features, high-level summary information, and numerical feature representation, the temporal word embedding vector sequence for each time scale is obtained.

[0013] The parameters describing the stability and volatility of user behavior are obtained from the temporal lexical embedding vector sequence and fused to obtain an enhanced time series sequence containing global dynamic evolution information from a complete time series perspective.

[0014] The enhanced time series is compressed to obtain feature vectors that have long-term overall performance and can focus on the latest state of the user.

[0015] Based on the feature vector at each time scale, dynamic gating weights and activation functions are used to process the data and obtain the churn probability.

[0016] Optionally, for historical behavioral data across multiple time scales, including categorical and numerical features, high-level summary information of the label vector sequence to which the categorical features belong is obtained, and numerical feature representations of the numerical features, including the interaction strength on the hidden dimensions, are obtained, including:

[0017] Based on the established time scale division rules, the historical behavior data of target customers is divided into multiple time scales. The historical behavior data is time-series data in customer relationship management.

[0018] Extract the categorical and numerical features at each time step from the historical behavioral data belonging to each time scale;

[0019] The categorical features at each time step under the same time scale are converted into fixed-dimensional categorical vectors and labeled. Then, the relationships between features are extracted based on the self-attention mechanism to obtain high-level summary information of the categorical features at the current time scale.

[0020] Based on the latent vector of numerical features, the interaction intensity vector between numerical features at the same time scale is extracted, and the interaction intensity vector is fused with the corresponding original numerical features to generate a numerical feature representation containing the interaction intensity of the hidden dimension.

[0021] The time scale division rules include: a medium-term time scale of 12 months, a short-term time scale of 3 months, and a current time scale of 1 month.

[0022] Optionally, the categorical features at each time step within the same time scale are converted into fixed-dimensional categorical vectors and labeled. Then, relationships between features are extracted based on a self-attention mechanism to obtain high-level summary information of the categorical features at the current time scale, including:

[0023] The categorical features at each time step within the same time scale are mapped into fixed-dimensional categorical vectors through a specified embedding layer.

[0024] All category vectors are concatenated, and a preset label vector is added to the resulting category vector sequence to obtain the label vector sequence of category features at the current time scale;

[0025] The labeled vector sequence is input into a pre-defined two-layer Transformer encoder for context learning, and the relationships between features are extracted based on the self-attention mechanism to obtain high-level summary information of the category features at the current time scale.

[0026] Optionally, based on the latent vectors of numerical features, the interaction strength vectors between numerical features at the same time scale are extracted, and the interaction strength vectors are fused with the corresponding original numerical features to generate a numerical feature representation that includes the interaction strength of the latent dimension, including:

[0027] The latent vectors of numerical features are used to perform pairwise interaction calculations by the feature interaction module of the pre-defined factorization machine, and the interaction strength between corresponding numerical features is quantified by calculating the inner product of the latent vectors to generate an interaction strength vector.

[0028] The interaction intensity vector is fused with the corresponding original numerical features, and the resulting high-dimensional feature vector is projected onto the original numerical feature dimension to obtain the initial numerical feature representation.

[0029] The initial numerical feature representation is subjected to layer normalization to obtain a numerical feature representation that includes the interaction intensity of the hidden dimensions.

[0030] Optionally, the parameters describing the stability and volatility of user behavior in the temporal lexical embedding vector sequence are obtained and fused to obtain an enhanced time-series sequence containing global dynamic evolution information from a complete time-series perspective, including:

[0031] The sequence of time word embedding vectors is convolved on a specified time dimension, and the short-term dynamic evolution trend is captured based on the difference information of the time word embedding vectors between adjacent time steps, generating a temporal feature evolution feature map of short-term change trend.

[0032] The mean, standard deviation, and extreme value difference are calculated for each feature dimension of the temporal lexical embedding vector sequence to obtain statistical aggregate information for describing the stability and volatility of user behavior.

[0033] The evolutionary feature map of temporal features is fused with statistical aggregation information, and the fused features are projected back to the original time dimension through a preset linear layer to obtain enhanced temporal lexical features that contain global statistical characteristics and local dynamic change information.

[0034] The enhanced temporal lexical features are fused into the original temporal lexical embedding vector sequence using a residual connection method to obtain the enhanced temporal sequence of the temporal lexical embedding vector sequence.

[0035] Optionally, the enhanced time-series sequence is compressed to obtain a feature vector that has long-term overall performance and can focus on the user's latest state, including:

[0036] Learnable dynamic weights are assigned to each feature in the augmented time series, with the weight values ​​derived from the feature's time decay factor and importance score.

[0037] The dynamic weights of each feature are iterated and multiplied element-wise with the enhanced time series sequence to obtain the global user time series.

[0038] Based on the pre-set global churn feature query vector, the global user time series is queried. The attention distribution of each time step in the sequence is calculated through a multi-head attention mechanism. The global user time series is then weighted and summed according to the attention distribution to obtain a weighted average global context vector.

[0039] The global context vector is concatenated with the features of the last time step in the global user time series to obtain the feature vector of the current time scale.

[0040] Optionally, based on the feature vector at each time scale, dynamic gating weights and activation functions are used to obtain the churn probability, including:

[0041] By concatenating feature vectors representing users' long-term trends, recent anomalies, and current status across multiple time scales, a panoramic vector of user behavior containing global information is obtained.

[0042] The user behavior panoramic vector is input into a pre-defined gated multilayer perceptron network, which dynamically learns and outputs the initial importance scores of feature vectors at each time scale. The initial importance scores are then normalized using the Softmax function to generate dynamic gate weights corresponding to each time scale.

[0043] Based on gating weights, the feature vectors of each time scale are weighted and fused to generate a fused feature that can simultaneously reflect the user's long-term trend, recent changes and current state.

[0044] The fused features are mapped to scalar values ​​using a multilayer perceptron classification head model based on parameterized modified linear units, and the sigmoid activation function is combined to map the scalar values ​​to user churn probability values ​​between 0 and 1.

[0045] Optionally, after processing the feature vectors at each time scale using dynamic gating weights and activation functions to obtain the churn probability, the method further includes:

[0046] Attribution analysis is performed on the probability of user churn to obtain the average contribution of each user behavior feature to the prediction of user churn.

[0047] The importance of global user behavior features is ranked based on their contribution, and customer relationship maintenance strategies are adaptively adjusted according to the ranking results.

[0048] Users are segmented based on the analysis results of feature contribution, and differentiated customer retention strategies are generated for different groups.

[0049] Optionally, after processing the feature vectors at each time scale using dynamic gating weights and activation functions to obtain the churn probability, the method further includes:

[0050] Based on the set churn risk threshold, users with a churn probability higher than the churn risk threshold are selected from all users, and a warning user list is generated.

[0051] For users on the warning user list, one or more numerical features in the user's historical behavior data are adjusted to simulate one or more intervention measures.

[0052] Based on the adjusted historical behavior data, the churn probability of the corresponding users is re-predicted to obtain the user churn probability after simulated intervention.

[0053] The simulated churn probability is compared with the original churn probability, and the effectiveness of each simulated intervention is analyzed based on the comparison results to determine the optimal intervention strategy to be adopted in actual customer relationship management.

[0054] Secondly, embodiments of the present invention provide a data processing system based on dynamic behavior across multiple time scales. This system is used for personalized customer relationship management in the digital content service industry. The system includes:

[0055] The feature extraction module is used to obtain high-level summary information of the label vector sequence to which the categorical features belong, and to obtain the numerical feature representation of the numerical features, which includes the interaction intensity on the hidden dimension, for historical behavioral data with multiple time scales including categorical features and numerical features.

[0056] The temporal embedding module is used to obtain the temporal word embedding vector sequence for each time scale based on the original category features, high-level summary information, and numerical feature representation.

[0057] The temporal enhancement module is used to obtain and fuse the parameters of the temporal lexical embedding vector sequence that describe the stability and volatility of user behavior, thereby obtaining an enhanced temporal sequence containing global dynamic evolution information from a complete time series perspective.

[0058] The sequence compression module is used to compress the enhanced time series sequence to obtain a feature vector that has long-term overall performance and can focus on the latest state of the user.

[0059] The prediction module is used to obtain the churn probability by processing the feature vector at each time scale using dynamic gating weights and activation functions.

[0060] (III) Beneficial Effects

[0061] This invention analyzes user historical behavior data by breaking it down into multiple independent branches. First, it enhances the temporal dynamic evolution of the time-series embedding vector sequences at each time scale using a complete time-series perspective, and compresses the sequences to obtain long-term overall performance while focusing on the user's latest state. The resulting feature vectors automatically extract deep behavioral patterns and evolutionary trends of users at different time scales. Then, through gating weights and activation function feature vectors, high-precision churn probability prediction is performed. Compared with existing technologies, this invention achieves churn probability prediction across multiple time scales and reference modes, avoiding the omission of important information that might be overlooked at a single time scale. This significantly improves the timeliness and accuracy of customer churn warnings, providing crucial data support for personalized customer relationship management. Attached Figure Description

[0062] Figure 1 A flowchart illustrating a data processing method based on dynamic behavior at multiple time scales, provided in an embodiment of the present invention;

[0063] Figure 2 This is a skeleton diagram of an artificial intelligence model backbone network provided in an embodiment of the present invention;

[0064] Figure 3 This is a flowchart of multi-timescale data processing based on an artificial intelligence model, provided as an embodiment of the present invention. Detailed Implementation

[0065] To better explain and facilitate understanding of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0066] refer to Figure 1As shown in the figure, this invention proposes a data processing method based on dynamic behavior across multiple time scales. This method is used for personalized customer relationship management in the digital content service industry. The entity executing the data processing method can be, for example, […]. Figure 2 , 3 The processor of the AI ​​model described includes the following methods: for historical behavioral data at multiple time scales, including categorical features and numerical features, obtaining high-level summary information of the label vector sequence to which the categorical features belong, and obtaining numerical feature representations of numerical features containing interaction strengths on hidden dimensions; based on the original categorical features, high-level summary information, and numerical feature representations, obtaining a temporal word embedding vector sequence for each time scale; obtaining and fusing parameters of the temporal word embedding vector sequence to describe the stability and volatility of user behavior, obtaining an enhanced time-series sequence containing global dynamic evolution information from a complete time-series perspective; compressing the enhanced time-series sequence to obtain feature vectors with long-term overall performance that can focus on the latest state of the user; and processing the feature vectors at each time scale using dynamic gating weights and activation functions to obtain the churn probability.

[0067] This example analyzes user historical behavior data by breaking it down into multiple independent branches. First, it enhances the temporal embedding vector sequences at each time scale through global dynamic evolution from a complete time-series perspective, and compresses the sequences to capture long-term overall performance while focusing on the user's latest state. The resulting feature vectors automatically extract deep behavioral patterns and evolutionary trends at different time scales. Then, through gating weights and activation function feature vectors, high-precision churn probability prediction is performed. Compared to existing technologies, this example achieves churn probability prediction across multiple time scales and reference patterns, avoiding the omission of important information that might be overlooked at a single time scale. This significantly improves the timeliness and accuracy of customer churn warnings, providing crucial data support for personalized customer relationship management.

[0068] To better understand the above technical solutions, exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that the present invention can be understood more clearly and thoroughly, and that the scope of the present invention can be fully conveyed to those skilled in the art.

[0069] Specifically, the execution subject of the data processing method based on multi-timescale dynamic behavior proposed in this embodiment includes, for example, the following: Figure 2 , 3The processor of the AI ​​model shown is divided into three time-scale data processing channels to simultaneously capture long-term trends, recent anomalies, and the current state of user behavior. Each channel has three data processing modules: a table feature fusion module based on self-attention feature fusion and factorization machine (Module 1), a time evolution modeling module based on convolutional neural networks and feature enhancement (Module 2), and a global feature importance aggregation module based on hierarchical attention (Module 3). The feature vectors output from the three data processing channels are finally used by a gating module based on multi-time-scale fusion (Module 4) to predict user churn and output the user churn probability. (Reference) Figure 1 As shown, the data processing method based on multi-timescale dynamic behavior may include the following steps S100 to S500:

[0070] S100. For historical behavioral data with multiple time scales including categorical features and numerical features, obtain high-level summary information of the label vector sequence to which the categorical features belong, and obtain numerical feature representations of numerical features including the interaction intensity on the hidden dimensions.

[0071] In this embodiment, user historical behavior data is divided into multiple time scales, enabling subsequent data processing to simultaneously capture long-term trends, recent anomalies, and current status information of user behavior, thereby improving the accuracy of user churn probability prediction. Furthermore, data at each time scale is split into categorical features and numerical features for independent analysis. For categorical features, high-level summary information of their respective tag vector sequences is extracted using embedding representation and sequence modeling techniques to capture the semantic relationships and evolutionary patterns between categorical behaviors. For numerical features, a numerical feature representation that reflects the intensity of interactions between hidden dimensions is constructed to characterize the trends and intensity of continuous behavioral changes.

[0072] In this embodiment, step S100 may include the following sub-steps S110 to S140:

[0073] S110. Based on the established time scale division rules, the historical behavior data of the target customers is divided into multiple time scales. The historical behavior data is time-series data in customer relationship management. The time scale division rules include: a medium-term time scale of 12 months, a short-term time scale of 3 months, and a current time scale of 1 month.

[0074] S120. Extract the categorical and numerical features of each time step in the historical behavioral data of each time scale.

[0075] S130. Convert the category features of each time step at the same time scale into fixed-dimensional category vectors and label them. Then, extract the relationship between features based on the self-attention mechanism to obtain high-level summary information of the category features at the current time scale.

[0076] Further, step S130 may include the following sub-steps S131 to S133:

[0077] S131. The categorical features at each time step under the same time scale are mapped into fixed-dimensional categorical vectors through a specified embedding layer.

[0078] To further explain, at the same time scale, the categorical features at each time step (such as gender, package name, etc.) are transformed into fixed-dimensional vector representations through a specific embedding layer:

[0079] e i =Emb i (x) cav i) (1)

[0080] In equation (1), x cav i represents the categorical feature, and Emb... i For the embedding layer, e i This is a category vector.

[0081] S132. Concatenate all category vectors and add a preset label vector to the resulting category vector sequence to obtain the label vector sequence of category features at the current time scale.

[0082] To further explain, the category vectors of all categorical features are concatenated together, and according to the BERT pattern, a label vector is added before all category vectors to generate a sequence of label vectors:

[0083] E in =[e cls ;e1; e2;…;e N (2)

[0084] In equation (2), E in For the sequence of labeled vectors, e cls Let N be the label vector, and N be the number of class features.

[0085] S133. Input the labeled vector sequence into the preset two-layer Transformer encoder for context learning, and extract the relationship between features based on the self-attention mechanism to obtain high-level summary information of the category features at the current time scale.

[0086] To further explain, the labeled vector sequence is input into a pre-defined two-layer Transformer encoder for context learning, capturing complex relationships between features based on a self-attention mechanism. For example, learning the combination of the category features "package" and "high-star user" might represent a high-value user. The expression for extracting the relationships between features is as follows:

[0087] Eout =TransformerEncoder(E in (3)

[0088] In equation (3), e cls The output vector h of the label cls =E out Considered as a high-level summary of features across all categories, TransformerEncode(·) is a two-layer Transformer encoder.

[0089] S140. Based on the latent vector of numerical features, extract the interaction intensity vector between numerical features at the same time scale, and fuse the interaction intensity vector with the corresponding original numerical features to generate a numerical feature representation containing the interaction intensity of the hidden dimension.

[0090] Furthermore, this example, by introducing a latent vector interaction mechanism, effectively captures the potential complex relationships between numerical features, thereby enhancing the model's ability to model the synergistic effects of multiple features. In time series models, directly using normalized numerical features may overlook the latent relationships between them. Therefore, this embodiment introduces a feature interaction module based on a factorization machine to interact with the original numerical features x. num Interaction intensity vector extraction is performed to enhance the expressive power of feature interactions. Unlike traditional factorization machines that directly output scalars for final prediction, in this embodiment, the factorization machine does not directly output prediction results, but rather acts as an interaction feature extractor, extracting interaction intensity vectors for each numerical feature x at the same time scale. i Learning a latent vector v i =[v i1 v i2 , ..., v ik ], and calculate the interaction strength vector between all numerical features through the vector inner product. The calculation process is based on formula (4). This design enables the model to extract structured interaction information from the interaction process, rather than a single interaction score, thereby more fully exploring the second-order interaction relationship between features and enhancing the representation ability of numerical features.

[0091] (4)

[0092] In equation (4), M is the total number of original numerical features, and k is the dimension of the numerical variable factorization machine.

[0093] Further, step S140 may include the following sub-steps S141 to S143:

[0094] S141. The latent vectors of numerical features are used to perform pairwise interaction calculations using the feature interaction module of the preset factorization machine, and the interaction strength between corresponding numerical features is quantified by calculating the inner product of the latent vectors to generate an interaction strength vector.

[0095] To further explain, each dimension in the interaction intensity vector no longer directly corresponds to any original numerical feature. Instead, each dimension represents an abstract latent factor that is automatically learned by the factorization machine model by aggregating all numerical features, and then the total interaction intensity of all numerical features on all latent dimensions is calculated.

[0096] S142. The interaction intensity vector is fused with the corresponding original numerical features, and the high-dimensional feature vector obtained after fusion is projected onto the original numerical feature dimension to obtain the initial numerical feature representation.

[0097] S143. Perform layer normalization on the initial numerical feature representation to obtain a numerical feature representation that includes the interaction intensity of the hidden dimensions.

[0098] To further explain, the calculation expression for the numerical feature representation is as follows:

[0099] h num =LayerNorm(W(Concat[x]) num, h f ])) (5)

[0100] In equation (5), h num For numerical feature representation, LayerNorm(·) is the layer normalization process, W is the projection layer, and Concat[·] represents concatenation according to the column dimensions. num As the initial numerical feature, h f This is the interaction strength vector between numerical features.

[0101] S200: Based on the original category features, high-level summary information, and numerical feature representation, obtain the temporal word embedding vector sequence for each time scale.

[0102] In this embodiment, the high-level summary information of the category features at each time step under each time scale, the original category feature embedding, and the numerical feature representation are concatenated to generate the complete temporal lexical embedding representation for that time step:

[0103] S t =Concat[h cls ;e1;e2;…;e N h num (6)

[0104] In equation (6), St This represents a complete temporal lexical embedding.

[0105] Then, iterate through all complete temporal word embeddings at the current time scale, and generate the temporal word embedding vector sequence S for the current time scale according to the time step order. seq =[S1,S2,…,S t ].

[0106] S300: Obtain the parameters of the time lexical embedding vector sequence used to describe the stability and volatility of user behavior, and fuse them to obtain an enhanced time series sequence containing global dynamic evolution information from a complete time series perspective.

[0107] In this embodiment, by extracting and fusing stability and volatility parameters, the evolutionary characteristics of user behavior over time can be characterized from multiple dimensions, thereby capturing its dynamic evolutionary patterns more comprehensively and improving the accuracy and robustness of subsequent behavior predictions. This embodiment employs convolutional neural networks and feature enhancement for temporal evolution modeling, enabling the temporal enhancement process to perceive the entire evolutionary process of the temporal lexical embedding vector sequence, thus obtaining global dynamic evolutionary information from a complete time series perspective. Specifically, by introducing convolutional operations to extract features from the sequence and combining them with feature enhancement mechanisms to strengthen key evolutionary patterns, the ability to perceive long-term dependencies and local dynamics is effectively improved, leading to a more accurate understanding of user behavior trends and providing more reliable temporal feature representations for applications such as personalized recommendations and abnormal behavior detection.

[0108] In this embodiment, step S300 may include the following sub-steps S310 to S340:

[0109] S310. Perform convolution operation on the time word embedding vector sequence in the specified time dimension, and capture the short-term dynamic evolution trend based on the difference information of the time word embedding vectors between adjacent time steps, and generate a time-series feature evolution feature map of short-term change trend.

[0110] Furthermore, for the temporal lexical embedding vector sequence S seq In the time dimension, a one-dimensional convolution module (Conv1d) with a kernel size of 2 is used to calculate the difference information of temporal word embedding vectors between adjacent time steps, thereby capturing the short-term dynamic evolution trend of feature values, defined as F. trend , i.e., time sequence feature evolution feature map.

[0111] S320. Calculate the mean, standard deviation, and extreme value difference for each feature dimension of the time term embedding vector sequence to obtain statistical aggregate information for describing the stability and volatility of user behavior.

[0112] Furthermore, the expression for calculating the mean is:

[0113] (7)

[0114] In equation (7), μ s Let T be the mean, and T be the total number of sequences in the time-term embedding vector sequence.

[0115] The expression for calculating standard deviation is:

[0116] (8)

[0117] In equation (8), σ s The standard deviation is denoted as .

[0118] The expression for calculating the extreme value difference is:

[0119] R s =max(S seq ) - min(S seq (9)

[0120] In equation (9), R s This represents the extreme value difference.

[0121] S330. The temporal feature evolution feature map is fused with statistical aggregation information, and the fused features are projected back to the original time dimension through a preset linear layer to obtain enhanced temporal lexical features containing global statistical characteristics and local dynamic change information.

[0122] Furthermore, the short-term dynamic evolution trends captured at each time step are fused with statistical aggregation information. Then, the fused features are projected back onto the original time dimension to obtain enhanced temporal lexical features that contain both global statistical characteristics and local dynamic change information.

[0123] (10)

[0124] In equation (10), To enhance temporal lexical features, Linear(·) is used as a linear layer.

[0125] S340. The enhanced temporal lexical features are fused into the original temporal lexical embedding vector sequence using a residual connection method to obtain the enhanced temporal sequence of the temporal lexical embedding vector sequence.

[0126] Furthermore, to inject contextual and global statistical information captured by the convolutional evolution model into the original temporal lexical embedding vector sequence, the enhanced temporal lexical features after temporal enhancement are fused to the corresponding time steps in the original temporal lexical embedding vector sequence via residual connections, outputting the final enhanced temporal sequence of the temporal lexical embedding vector sequence:

[0127] (11)

[0128] In equation (11), S' seq To enhance the time series, α is a scaling factor used for stable training and adaptive learning in the early stages of training.

[0129] S400: Compress the enhanced time series sequence to obtain a feature vector that has long-term overall performance and can focus on the latest state of the user.

[0130] In this embodiment, a hierarchical attention-based global feature importance aggregation module is employed to compress the enhanced time-series sequence into a fixed-length feature vector. This not only extracts a stable overall performance (global vector) from long-period sequences but also adaptively strengthens attention to recent key behaviors (the final step), overcoming the problems of information smoothing or dilution of recent signals in traditional methods during long-sequence modeling. This implementation introduces a feature-level attention weight allocation mechanism to obtain the global user time series and a time-step-level attention weighted summation to obtain the global context vector. This enables the autonomous identification of the differences in the importance of each feature to the current prediction task at each time step, enhancing the discriminative power of feature representation.

[0131] In this embodiment, step S400 may include the following sub-steps S410 to S440:

[0132] S410. Configure learnable dynamic weights for each feature in the enhanced time series, with the weight values ​​obtained based on the feature's time decay factor and importance score.

[0133] Furthermore, a learnable dynamic weight is generated for each feature in the enhanced time series using a two-layer fully connected neural network and a GELU layer:

[0134] (12)

[0135] In equation (12), For linear layers, , W f This is the weight matrix. Sigmoid is the activation function.

[0136] S420. Iterate through the dynamic weights of each feature and multiply them element-wise with the enhanced time series to obtain the global user time series.

[0137] Furthermore, the expression for the global user time series is:

[0138] (13)

[0139] In equation (13), S weighted This is a global user time series.

[0140] S430. Based on the pre-set global churn feature query vector, query the global user time series, calculate the attention distribution of each time step in the sequence through a multi-head attention mechanism, and perform weighted summation on the global user time series according to the attention distribution to obtain a weighted average global context vector.

[0141] Furthermore, a learnable global churn feature query vector is pre-defined for all users. The system queries the global user time series, calculates the attention distribution at each time step in the series, and then performs a weighted summation of the global user time series based on the attention distribution to obtain a weighted average global context vector.

[0142] (14)

[0143] In equation (4), h global For global context vectors, MultiHeadAttention(·) is the multi-head attention function, Q is the global churn factor vector, K is the key index obtained by linear transformation of the user sequence, and V is the value index obtained by linear transformation of the user time series, which are used for training the multi-head attention mechanism model.

[0144] S440. Concatenate the global context vector with the features of the last time step in the global user time series to obtain the feature vector of the current time scale.

[0145] Furthermore, the expression for calculating the eigenvector is:

[0146] (15)

[0147] (16)

[0148] In equation (16), h final These are the eigenvectors.

[0149] S500 uses dynamic gating weights and activation functions to process the feature vectors at each time scale to obtain the churn probability.

[0150] In this embodiment, the following is adopted: Figure 3The gating module shown is based on multi-timescale fusion and performs user churn prediction on the input across multiple timescales. Specifically, based on all three input feature vectors, a gating weight is dynamically calculated for each branch (mid, short, and current). Then, through an intelligent fusion mechanism, information from all timescales is coordinated to achieve churn probability prediction across multiple timescales and reference modes. This avoids overlooking important information that might be missed at a single timescale, thus greatly improving the timeliness and accuracy of customer churn warnings and providing crucial data support for personalized customer relationship management.

[0151] In this embodiment, step S500 may include the following sub-steps S510 to S540:

[0152] S510. The feature vectors representing the user's long-term trends, recent changes, and current state at multiple time scales are concatenated to obtain a panoramic vector of user behavior containing global information.

[0153] Furthermore, the expression for calculating the user behavior panoramic vector is as follows:

[0154] (17)

[0155] In equation (17), h all h is a panoramic vector of user behavior. mid h is the eigenvector of the intermediate time scale. short h is the feature vector with a short time scale. current This is the feature vector at the current time scale.

[0156] S520. Input the panoramic vector of user behavior into a pre-defined gated multilayer perceptron network, dynamically learn and output the initial importance score of the feature vector at each time scale, and normalize the initial importance score through the Softmax function to generate dynamic gate weights corresponding to each time scale.

[0157] Furthermore, based on all three input feature vectors, an independent gated multilayer perceptron network (Gates) is used to dynamically calculate the initial importance score of the feature vectors at each time scale for each branch. All scores are normalized using Softmax to generate the final gating weights. This ensures that the sum of the weights of the corresponding feature dimensions at each time scale is 1. The above process can be formalized as follows:

[0158] (18)

[0159] (19)

[0160] In equation (19), Stack(·) is used to stack according to the dimension of the row.

[0161] S530: Based on gating weights, the feature vectors of each time scale are weighted and fused to generate a fused feature that can simultaneously reflect the user's long-term trend, recent changes and current state.

[0162] Furthermore, the calculation expression for the fused features is as follows:

[0163] (20)

[0164] In equation (20), h fused For feature fusion, Linear(·) is a linear transformation layer. , , These are feature vectors at three time scales.

[0165] For example, if a user's data shows that their behavior is extremely stable in the medium term, but fluctuates drastically in the short term (e.g., a sharp increase in traffic, a sudden decrease in costs), the gating module might assign higher weight to the "short term" branch. Conversely, for users whose behavior remains relatively stable, the model might place more emphasis on the long-term trends shown by the "medium term" branch. This adaptive weighted fusion mechanism makes the model more intelligent than simply concatenating or averaging features.

[0166] S540: The multilayer perceptron classification head model based on parameterized modified linear units is used to map the fused features into scalar values, and the sigmoid activation function is combined to map the scalar values ​​into user churn probability values ​​between 0 and 1.

[0167] Furthermore, the formula for calculating the user churn probability value is as follows:

[0168] (twenty one)

[0169] In equation (21), p is the user churn probability value, and PRelu is the parameterized modified linear unit.

[0170] In this embodiment, the data processing method described in steps S100 to S500 above is used for personalized customer relationship management in the digital content service industry. This management primarily operates at two levels: the strategy layer and the execution layer. The decision-making layer uses the user churn probability values ​​obtained from data processing to gain insights into feature importance and segment customer groups, thereby configuring corresponding customer relationship management strategies or adaptively adjusting existing strategies. The execution layer uses the user churn probability values ​​obtained from data processing to conduct individual churn simulation interventions and provide guidance on responses, implementing optimal intervention measures for customers based on the intervention results to achieve the best customer retention effect.

[0171] Furthermore, the customer management methodological steps implemented at the strategy level include G100 to G300:

[0172] G100 performs attribution analysis on user churn probability to obtain the average contribution of each user behavior feature to predicting user churn.

[0173] G200 ranks the importance of global user behavior features based on feature contribution and adaptively adjusts customer relationship maintenance strategies according to the ranking results.

[0174] G300 segments users based on the analysis results of feature contribution and generates differentiated customer retention strategies for different groups.

[0175] In one specific implementation, the SHAP (SHapley Additive exPlanations) model interpretability tool is used to perform attribution analysis on user churn probability, obtaining the average contribution of each feature in user behavior to predicting user churn, thereby obtaining a global feature importance ranking. This allows for a data-driven approach to reveal the key factors influencing user churn, enabling business teams to adaptively adjust customer relationship maintenance strategies and focus resources on the most important influencing factors. Then, by analyzing SHAP values, a clear high-risk user profile is created, and users are precisely segmented (e.g., young users with high data consumption but recently decreased activity), and differentiated retention strategies are generated for different groups.

[0176] Furthermore, the customer management method steps executed at the execution level include F100 to F400:

[0177] F100: Based on the set churn risk threshold, filter out users whose churn probability is higher than the churn risk threshold from all users and generate a warning user list.

[0178] F200: For users on the warning user list, simulate one or more intervention measures by adjusting one or more numerical features in the user's historical behavior data.

[0179] F300: Based on adjusted historical behavior data, re-predict the churn probability of the corresponding user and obtain the user churn probability after simulated intervention.

[0180] F400: Compare the simulated churn probability with the original churn probability, and analyze the effect of each simulated intervention based on the comparison results to determine the optimal intervention strategy to be adopted in actual customer relationship management.

[0181] In one specific embodiment, the core of the data processing steps S100 to S500 is predicting the churn probability of each user in the future. Thus, by setting a churn risk threshold, the execution layer can filter out a list of users requiring high attention. Based on this list, customer relationship management shifts from passive response to proactive intervention, allowing for early intervention before users actually churn and the implementation of personalized customer retention strategies, thereby significantly improving the success rate of user retention. Furthermore, for high-risk users, this embodiment can adjust the numerical characteristics in the user's historical behavioral data (e.g., simulating a certain amount of free data or a reduction in their plan fees), and then re-perform the user churn probability prediction in steps S100 to S500. Finally, by observing the changes in the predicted churn probability, the potential effects of different retention measures are evaluated, thereby selecting the optimal retention measure or optimal intervention strategy. This provides data support for the selection of customer retention response measures and further enables precise resource allocation.

[0182] Furthermore, this embodiment also provides a data processing system based on dynamic behavior across multiple time scales. This system is used for personalized customer relationship management in the digital content service industry. The system includes:

[0183] The feature extraction module is used to obtain high-level summary information of the label vector sequence to which the categorical features belong, and to obtain the numerical feature representation of the numerical features, which includes the interaction intensity on the hidden dimension, for historical behavioral data with multiple time scales including categorical features and numerical features.

[0184] The temporal embedding module is used to obtain the temporal word embedding vector sequence for each time scale based on the original category features, high-level summary information, and numerical feature representation.

[0185] The temporal enhancement module is used to acquire and fuse parameters from the temporal lexical embedding vector sequence that describe the stability and volatility of user behavior, thereby obtaining an enhanced temporal sequence containing global dynamic evolution information from a complete time series perspective.

[0186] The sequence compression module is used to compress the enhanced time series sequence to obtain feature vectors that have long-term overall performance and can focus on the latest state of the user.

[0187] The prediction module is used to obtain the churn probability by processing the feature vector at each time scale using dynamic gating weights and activation functions.

[0188] In this embodiment, before performing historical probability prediction for customers, the data processing system needs to train the configured artificial intelligence model through a preset training optimization module. This artificial intelligence model function is implemented through the system's feature extraction module, temporal embedding module, temporal enhancement module, sequence compression module, and prediction module. Specifically, model training involves: First, acquiring time-series data containing churn and non-churn rates. Next, due to class imbalance in the time-series data (taking telecommunications data as an example, where churned users account for only 2% of the total data), this embodiment introduces an optimized loss function based on Focal Loss for imbalanced datasets. This allows setting different penalty factors for positive samples (churned) and negative samples (non-churned), thereby balancing the time-series data to obtain the final training dataset. Finally, the training dataset is used for model training, and the area under the precision and recall curves is used as an evaluation metric to validate and early stop the trained model to obtain the optimal artificial intelligence model.

[0189] In summary, this invention proposes a data processing method and system based on dynamic behavior across multiple time scales, applicable to personalized customer relationship management in the digital content service industry. The core of this method lies in parallel analysis of historical user behavior data across multiple time scales, employing a modular artificial intelligence model for deep feature extraction and intelligent fusion, ultimately outputting a high-precision user churn probability. Specifically, the method first divides the time-series behavioral data into different time scales and refines both categorical and numerical features: categorical features are processed using embedding representations and a two-layer Transformer encoder to obtain high-level semantic summary information, while numerical features are captured using a factorization machine model to capture the interaction strength between hidden dimensions. Subsequently, the processed features are concatenated to form a temporal word embedding sequence, which is further enhanced by fusing convolutional neural networks with statistical features to capture the dynamic evolution of user behavior from a global perspective. Next, a hierarchical attention mechanism is used to compress the enhanced sequence into feature vectors that reflect both long-term overall performance and focus on the latest state. Finally, a gating fusion module dynamically weighs the importance of feature vectors at each time scale, and a classifier outputs the churn probability.

[0190] The significant advantage of this invention lies in its multi-timescale, global analytical capabilities, avoiding information omissions from a single time perspective, thereby greatly improving the timeliness and accuracy of predictions. Furthermore, this invention not only excels in prediction but its outputs can also be directly applied to business decisions, supporting the entire customer relationship management process from feature attribution and customer segmentation at the strategy level to proactive early warning and measure simulation at the execution level. This provides strong data support for achieving precise resource allocation and personalized customer retention.

[0191] Since the systems / devices described in the above embodiments of the present invention are systems / devices used to implement the methods of the above embodiments of the present invention, those skilled in the art can understand the specific structure and modifications of the systems / devices based on the methods described in the above embodiments of the present invention, and therefore will not be repeated here. All systems / devices used in the methods of the above embodiments of the present invention fall within the scope of protection of the present invention.

[0192] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0193] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions.

[0194] It should be noted that in the description of this invention, the word "a" or "an" preceding a component does not exclude the existence of multiple such components. This invention can be implemented by means of hardware comprising several different components and by means of a suitably programmed computer. The use of terms such as first, second, third, etc., is merely for convenience and does not indicate any order. These terms can be understood as part of the component names.

[0195] Furthermore, it should be noted that in the description of this specification, the terms "one embodiment," "some embodiments," "embodiment," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Furthermore, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0196] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning of the basic inventive concept, can make other changes and modifications to these embodiments.

[0197] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from the spirit and scope of the invention.

Claims

1. A data processing method based on dynamic behavior at multiple time scales, characterized in that, The method is used for personalized customer relationship management in the digital content service industry, and the method includes: For historical behavioral data across multiple time scales, including categorical and numerical features, this method obtains high-level summary information of the labeled vector sequence to which the categorical features belong, and numerical feature representations of the numerical features containing the interaction strength on the hidden dimensions. This includes: dividing the historical behavioral data of target customers into multiple time scales according to a set time scale division rule; the historical behavioral data is time-series data in customer relationship management; extracting categorical and numerical features at each time step in the historical behavioral data for each time scale; converting the categorical features at each time step within the same time scale into fixed-dimensional categorical vectors and labeling them; then extracting the relationships between features based on a self-attention mechanism to obtain high-level summary information of the categorical features at the current time scale; and extracting the interaction strength vectors between numerical features at the same time scale based on the latent vectors of the numerical features, and fusing the interaction strength vectors with the corresponding original numerical features to generate numerical feature representations containing the interaction strength on the hidden dimensions. Based on the original category features, high-level summary information, and numerical feature representation, the temporal word embedding vector sequence for each time scale is obtained. The parameters describing the stability and volatility of user behavior are obtained from the temporal lexical embedding vector sequence and fused to obtain an enhanced time series sequence containing global dynamic evolution information from a complete time series perspective. The enhanced time series is compressed to obtain feature vectors that have long-term overall performance and can focus on the latest state of the user. Based on the feature vector at each time scale, dynamic gating weights and activation functions are used to process the data and obtain the churn probability.

2. The method as described in claim 1, characterized in that, The time scale division rules include: a medium-term time scale of 12 months, a short-term time scale of 3 months, and a current time scale of 1 month.

3. The method as described in claim 1, characterized in that, The categorical features at each time step within the same time scale are converted into fixed-dimensional categorical vectors and labeled. Then, relationships between features are extracted based on a self-attention mechanism to obtain high-level summary information of the categorical features at the current time scale, including: The categorical features at each time step within the same time scale are mapped into fixed-dimensional categorical vectors through a specified embedding layer. All category vectors are concatenated, and a preset label vector is added to the resulting category vector sequence to obtain the label vector sequence of category features at the current time scale; The labeled vector sequence is input into a pre-defined two-layer Transformer encoder for context learning, and the relationships between features are extracted based on the self-attention mechanism to obtain high-level summary information of the category features at the current time scale.

4. The method as described in claim 1, characterized in that, Based on the latent vectors of numerical features, the interaction strength vectors between numerical features at the same time scale are extracted, and the interaction strength vectors are fused with the corresponding original numerical features to generate a numerical feature representation that includes the interaction strength of the latent dimension, including: The latent vectors of numerical features are used to perform pairwise interaction calculations by the feature interaction module of the pre-defined factorization machine, and the interaction strength between corresponding numerical features is quantified by calculating the inner product of the latent vectors to generate an interaction strength vector. The interaction intensity vector is fused with the corresponding original numerical features, and the resulting high-dimensional feature vector is projected onto the original numerical feature dimension to obtain the initial numerical feature representation. The initial numerical feature representation is subjected to layer normalization to obtain a numerical feature representation that includes the interaction intensity of the hidden dimensions.

5. The method as described in claim 1, characterized in that, The parameters describing the stability and volatility of user behavior are obtained from the temporal lexical embedding vector sequence and fused to obtain an enhanced time series sequence containing global dynamic evolution information from a complete time series perspective, including: The sequence of time word embedding vectors is convolved on a specified time dimension, and the short-term dynamic evolution trend is captured based on the difference information of the time word embedding vectors between adjacent time steps, generating a temporal feature evolution feature map of short-term change trend. The mean, standard deviation, and extreme value difference are calculated for each feature dimension of the temporal lexical embedding vector sequence to obtain statistical aggregate information for describing the stability and volatility of user behavior. The evolutionary feature map of temporal features is fused with statistical aggregation information, and the fused features are projected back to the original time dimension through a preset linear layer to obtain enhanced temporal lexical features that contain global statistical characteristics and local dynamic change information. The enhanced temporal lexical features are fused into the original temporal lexical embedding vector sequence using a residual connection method to obtain the enhanced temporal sequence of the temporal lexical embedding vector sequence.

6. The method as described in claim 1, characterized in that, The enhanced time-series sequence is compressed to obtain feature vectors that have long-term overall performance and can focus on the latest state of the user, including: Learnable dynamic weights are assigned to each feature in the augmented time series, with the weight values ​​derived from the feature's time decay factor and importance score. The dynamic weights of each feature are iterated and multiplied element-wise with the enhanced time series sequence to obtain the global user time series. Based on the pre-set global churn feature query vector, the global user time series is queried. The attention distribution of each time step in the sequence is calculated through a multi-head attention mechanism. The global user time series is then weighted and summed according to the attention distribution to obtain a weighted average global context vector. The global context vector is concatenated with the features of the last time step in the global user time series to obtain the feature vector of the current time scale.

7. The method as described in claim 1, characterized in that, Based on the feature vector at each time scale, dynamic gating weights and activation functions are used to obtain the churn probability, including: By concatenating feature vectors representing users' long-term trends, recent anomalies, and current status across multiple time scales, a panoramic vector of user behavior containing global information is obtained. The user behavior panoramic vector is input into a pre-defined gated multilayer perceptron network, which dynamically learns and outputs the initial importance scores of feature vectors at each time scale. The initial importance scores are then normalized using the Softmax function to generate dynamic gate weights corresponding to each time scale. Based on gating weights, the feature vectors of each time scale are weighted and fused to generate a fused feature that can simultaneously reflect the user's long-term trend, recent changes and current state. The fused features are mapped to scalar values ​​using a multilayer perceptron classification head model based on parameterized modified linear units, and the sigmoid activation function is combined to map the scalar values ​​to user churn probability values ​​between 0 and 1.

8. The method as described in claim 1, characterized in that, After processing the feature vectors at each time scale using dynamic gating weights and activation functions to obtain the churn probability, the process also includes: Attribution analysis is performed on the probability of user churn to obtain the average contribution of each user behavior feature to the prediction of user churn. The importance of global user behavior features is ranked based on their contribution, and customer relationship maintenance strategies are adaptively adjusted according to the ranking results. Users are segmented based on the analysis results of feature contribution, and differentiated customer retention strategies are generated for different groups.

9. The method as described in claim 1, characterized in that, After processing the feature vectors at each time scale using dynamic gating weights and activation functions to obtain the churn probability, the process also includes: Based on the set churn risk threshold, users with a churn probability higher than the churn risk threshold are selected from all users, and a warning user list is generated. For users on the warning user list, one or more numerical features in the user's historical behavior data are adjusted to simulate one or more intervention measures. Based on the adjusted historical behavior data, the churn probability of the corresponding users is re-predicted to obtain the user churn probability after simulated intervention. The simulated churn probability is compared with the original churn probability, and the effectiveness of each simulated intervention is analyzed based on the comparison results to determine the optimal intervention strategy to be adopted in actual customer relationship management.

10. A data processing system based on dynamic behavior at multiple time scales, characterized in that, The system is used for personalized customer relationship management in the digital content service industry. The system includes: The feature extraction module is used to obtain high-level summary information of the labeled vector sequence to which the categorical features belong, and to obtain the numerical feature representation of the numerical features containing the interaction strength on the hidden dimension, for historical behavioral data with multiple time scales including categorical features and numerical features. This includes: dividing the historical behavioral data of the target customer into multiple time scales according to a set time scale division rule; extracting categorical and numerical features at each time step in the historical behavioral data of each time scale; converting the categorical features of each time step at the same time scale into fixed-dimensional categorical vectors and labeling them; then extracting the relationships between features based on a self-attention mechanism to obtain high-level summary information of the categorical features at the current time scale; and extracting the interaction strength vector between numerical features at the same time scale based on the latent vector of the numerical features, and fusing the interaction strength vector with the corresponding original numerical features to generate a numerical feature representation containing the interaction strength on the hidden dimension. The temporal embedding module is used to obtain the temporal word embedding vector sequence for each time scale based on the original category features, high-level summary information, and numerical feature representation. The temporal enhancement module is used to obtain and fuse the parameters of the temporal lexical embedding vector sequence that describe the stability and volatility of user behavior, and obtain an enhanced temporal sequence containing global dynamic evolution information from a complete time series perspective. The sequence compression module is used to compress the enhanced time series sequence to obtain a feature vector that has long-term overall performance and can focus on the latest state of the user. The prediction module is used to obtain the churn probability by processing the feature vector at each time scale using dynamic gating weights and activation functions.

Images