Internet user behavior accurate analysis method based on artificial intelligence

Abstract

The invention discloses an Internet user behavior accurate analysis method based on artificial intelligence. The method comprises the following steps: obtaining user behavior data of an Internet platform; inputting the user behavior sequence into a Mamba network to generate a behavior state vector; inputting the commodity image set into an improved DeiT model, and setting a hierarchical feature routing controller; executing cross-modal feature interaction to generate a fusion feature tensor; inputting the fusion feature tensor into an IQL network; updating a fusion weight, high-level attention mapping and a routing control vector based on a value gradient resonance feedback mechanism; and outputting a user behavior analysis result and a personalized recommendation result. According to the method, the accuracy of Internet user behavior prediction and personalized recommendation can be remarkably improved, and the method is suitable for scenes needing high-precision user portraits and intelligent recommendation, such as a large-scale e-commerce platform and a content distribution platform.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and user behavior modeling technology, and in particular to a method for accurate analysis of internet user behavior based on artificial intelligence. Background Technology

[0002] In today's internet environment, user behavior analysis and personalized recommendations have become core functions of applications such as e-commerce, content distribution, and social platforms. Traditional user behavior modeling methods mostly rely on statistical rules or shallow machine learning models, such as collaborative filtering and factorization machines. These methods exhibit significant limitations when faced with the complexity of user behavior, multimodal content information, and the need for real-time response. Especially in multi-turn interactions, user behavior sequences exhibit strong temporal dependencies and semantic changes. Modeling based solely on single user behaviors such as clicks and browsing often fails to accurately capture their potential interests and intentions.

[0003] In recent years, with the development of deep learning technology, researchers have begun to use structures such as recurrent neural networks (RNNs) and Transformers to model user behavior sequences and integrate multimodal information such as images and text to improve recommendation accuracy. Visual features, as important semantic information of products, can, to some extent, compensate for the cold start problem and insufficient textual descriptions. However, existing methods suffer from coarse fusion methods and a lack of semantic alignment mechanisms in cross-modal feature fusion, failing to effectively establish a high-quality correspondence between user behavior and visual content. Furthermore, behavior sequence modeling often relies on a uniform structure, lacking path control mechanisms for different levels of semantic features, making it difficult to adapt to the evolution of user interests in different scenarios.

[0004] Existing reinforcement learning-based recommendation strategies mostly employ static policy networks or simple value networks to optimize user behavior feedback, making it difficult to handle value assessment biases caused by high-dimensional state inputs. When user behavior feedback is sparse or delayed, traditional policy update mechanisms are prone to getting stuck in local optima during the early stages of training, resulting in a lack of diversity and generalization ability in recommendation results, thus degrading the user experience. Current technologies still have many shortcomings in multimodal user behavior modeling, feature interaction mechanisms, and policy optimization feedback. There is an urgent need for a closed-loop analysis method that can integrate behavioral states, visual information, and value feedback to achieve more accurate and efficient understanding of internet user behavior and personalized recommendations.

[0005] Therefore, how to provide accurate analysis methods for internet user behavior based on artificial intelligence is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a precise analysis method for internet user behavior based on artificial intelligence. This invention fully integrates the Mamba sequence modeling network, the improved DeiT image coding model, and the bidirectional cross-attention fusion mechanism to construct a deep semantic alignment structure between behavioral state vectors and product visual features. Furthermore, it introduces an IQL reinforcement learning network and a value gradient resonance feedback mechanism to achieve dynamic closed-loop optimization of user behavior analysis and personalized recommendations. This method possesses advantages such as accurate behavior modeling, good cross-modal fusion performance, timely optimization of recommendation results, and a stable feedback mechanism, significantly improving the user understanding and recommendation capabilities of internet platforms in multimodal interaction scenarios.

[0007] The method for precise analysis of internet user behavior based on artificial intelligence according to embodiments of the present invention includes the following steps: Step 1: Obtain user behavior data and product image data from internet platforms, and construct a set of product images associated with user behavior sequences; Step 2: Input the preprocessed user behavior sequence into the Mamba network and generate behavior state vectors based on the selective state space modeling mechanism; Step 3: Input the set of product images into the improved DeiT model, set a hierarchical feature routing controller before the Transformer encoding layer. The hierarchical feature routing controller generates a routing control vector based on the feature statistics output by the previous layer, and performs path selection between the local attention branch and the global attention branch, outputting a visual feature vector and a high-level attention mapping. Step 4: Input the behavioral state vector and visual feature vector into the bidirectional cross-attention network to perform cross-modal feature interaction and generate a fused feature tensor; Step 5: Input the fused feature tensor into the IQL network, generate reward signals based on the user's actual interaction behavior, and output the policy distribution and value gradient distribution; Step 6: Based on the value gradient resonance feedback mechanism, update the fusion weights of the bidirectional cross-attention network, the high-level attention mapping of the improved DeiT model, and the routing control vector of the hierarchical feature routing controller using the value gradient distribution. Step 7: Output user behavior analysis results and personalized recommendation results based on the updated strategy distribution.

[0008] Optionally, step one specifically includes: Read user behavior data generated by the Internet platform, wherein the user behavior data includes user identifier, session identifier, event type, event timestamp, product identifier and page identifier; The user's behavior events within the same session are sorted in ascending order according to the event timestamp, generating a user behavior sequence and creating a sequence number and index table for the user behavior sequence. Read the product image data corresponding to the product identifier; Establish a mapping relationship between user identifier, session identifier, serial number, product identifier, and image identifier; Summarize each product image associated with a user behavior sequence to form a set of product images associated with the user behavior sequence.

[0009] Optionally, the preprocessing specifically includes: Fixed-length category vectors are generated by embedding mapping for event types, product identifiers, and page identifiers; The event timestamp is split into a date field, an hour field, a week sequence number field, and a time difference field between adjacent events, and then encoded into a fixed-time vector. Concatenate the fixed-length category vector and the fixed-length time vector in a preset order to form a fixed-length event vector sequence; The fixed-length event vector sequence is sorted in ascending order based on the event timestamp, and a time window sequence is generated according to a fixed window length and a fixed sliding step. For fixed-length event vector sequences that are shorter than the window length, padding is performed and a valid length marker is recorded.

[0010] Optionally, step two specifically includes: The time window sequence is input into the Mamba network, and the linear mapping from the event vector to the input dimension of the Mamba network is completed in the input mapping layer. In the selective state-space modeling unit, state updates and output mapping are performed sequentially, and inter-layer output alignment is completed in the layer normalization and residual connection unit. Selective state space modeling units are stacked sequentially according to a preset number of layers. The output of the previous selective state space modeling unit is updated and features are transferred. The hidden state vector corresponding to the time window is extracted at the output of the last layer of the Mamba network as the window-level hidden state output. The session-level hidden states are obtained by performing aggregation operations on all window-level hidden states in chronological order, and then the session-level hidden states are transformed into behavior state vectors in the mapping layer.

[0011] Optionally, the hierarchical feature routing controller generates a routing control vector based on the feature statistics output from the previous layer, specifically including: For each product image in the product image set, size normalization and image block processing are performed to obtain a sequence of image blocks arranged in a fixed block order; The image patch sequence is input into the block embedding layer of the improved DeiT model and positional encoding is added to generate an image patch embedding sequence as input to the Transformer coding layer; Set a hierarchical feature routing controller at the input of the Transformer coding layer; The hierarchical feature routing controller reads the statistical values ​​of the output feature vector of the previous Transformer coding layer and generates a routing control vector; The statistical values ​​of the output feature vector include channel mean, variance, and attention response intensity; The feature statistics are used to generate a routing control vector through nonlinear mapping; And determine the feature path of this layer between the local attention branch and the global attention branch based on the routing control vector; The process of determining the feature path includes: performing feature calculations on the outputs of the local attention branch and the global attention branch, respectively. The outputs of the two branches are weighted and fused based on the weight coefficients of the routing control vector. When the weight coefficients are greater than or equal to a preset threshold, the weighting coefficient of the global attention branch in the current Transformer encoding layer output feature vector is greater than the weighting coefficient of the local attention branch. When the weight coefficient is less than the preset threshold, the weight coefficient of the local attention branch in the output feature vector of the current Transformer encoding layer is greater than the weight coefficient of the global attention branch; The weighted fusion result is used as the output feature of this layer.

[0012] Optionally, the step of selecting a path between the local attention branch and the global attention branch, and outputting a visual feature vector and a high-level attention mapping, specifically includes: In the local attention branch, attention is calculated within the range of adjacent embedding vectors of the image patch embedding sequence to obtain the local context weighted result and output the set of local context feature vectors; In the global attention branch, attention calculation is performed on all embedding vectors of the image patch embedding sequence to obtain the global dependency modeling result and output a set of global semantic feature vectors. The local context feature vector set and the global semantic feature vector set are weighted and fused according to the weight coefficient of the routing control vector to obtain the output feature vector of the current Transformer coding layer; The output feature vector is processed sequentially by a feedforward network, residual connection, and layer normalization. The processed result is used as the input to the next Transformer encoding layer until the preset number of encoding layers are completed. The global feature vector is extracted from the output of the final Transformer encoding layer as the visual feature vector, and a high-level attention map is generated based on the attention weights of the final Transformer encoding layer.

[0013] Optionally, step four specifically includes: Based on the cross-attention mechanism, the behavior state vector is used as the query vector, and the visual feature vector is used as the key vector and value vector to calculate the behavior-guided attention weight matrix. Based on the attention weight matrix, a weighted summation is performed on the value vector to generate behavior-guided cross-modal features; The visual feature vector is used as the query vector, and the behavior state vector is used as the key vector and value vector to calculate the visual guidance attention weight matrix; Based on the attention weight matrix, a weighted summation is performed on the value vector to generate visually guided cross-modal features; In the feature fusion layer, the behavior-guided cross-modal features and the vision-guided cross-modal features are concatenated; the concatenated features are linearly mapped according to the fusion weights to generate intermediate fusion features. The fused intermediate features are sequentially subjected to residual connection and layer normalization processing to output the fused feature tensor.

[0014] Optionally, step five specifically includes: The fused feature tensor is received and rearranged and normalized according to a preset dimension. The processed fused feature tensor is written as a state input to the state input terminal of the IQL network. Construct a candidate action set and perform embedding mapping on the candidate action set to generate action vectors. Combine the action vectors with the state input according to a preset concatenation order to form a state-action input. The state input is fed into the policy network and the value network respectively to perform forward inference, and the corresponding policy score and state value are output. The state-action input is fed into the state-action value network for forward inference, and the corresponding state-action value is output. Reward signals are generated based on actual user interaction behavior and are associated with session identifiers, time window identifiers, and action identifiers to form data records for IQL network updates. In the IQL network, parameter update calculations are performed based on reward signals, state values, and state-action values. The gradient weights of the state-action value network are adjusted according to the policy score, and the objective function of the value network is updated based on the difference between the state value and the state-action value. After completing the parameter update, the updated policy network, value network, and state-action value network are obtained. Based on the updated state-action value network output, the partial derivative of the state-action value with respect to the fused feature tensor is calculated. The partial derivative is then used as the gradient tensor and normalized according to a preset mapping rule to generate a value gradient distribution. The updated policy network output is normalized to a policy distribution.

[0015] Optionally, step six specifically includes: The value gradient distribution is converted into weight adjustment coefficients on the bidirectional cross-attention network side according to the preset mapping rules; The behavior-guided attention weight matrix and the visual-guided attention weight matrix are reweighted and normalized according to the weight adjustment coefficients to obtain the updated cross-attention weight parameters. For the linear mapping weights of the feature fusion layer in the bidirectional cross-attention network, perform element-wise update and normalization according to the adjustment coefficients corresponding to the cross-attention weight parameters to obtain the updated fusion weights, and write them into the feature fusion layer. The value gradient distribution is mapped to the high-level attention reweighting coefficients on the improved DeiT model side. The attention weights of the final Transformer encoding layer are reweighted and normalized to generate the updated high-level attention mapping, which is then written into the final Transformer encoding layer. The routing control increment vector is calculated based on the value gradient distribution. The routing control increment vector is combined with the previous round routing control vector by a preset weighting coefficient and normalized to obtain the updated routing control vector, which is then written into the hierarchical feature routing controller.

[0016] Optionally, step seven specifically includes: Receive the updated policy distribution output from the policy network; According to the preset action decoding and filtering rules, the target action set is determined from the candidate action set, and the action identifier and corresponding probability are retained for each action in the target action set; Based on the mapping relationship between action identifiers and product identifiers, the target action set is converted into a recommendation candidate set; The recommended candidate set is sorted according to the probability corresponding to the updated strategy distribution to generate the sorting result; Based on preset quantity parameters, select the preceding items from the sorting results, generate personalized recommendation results, and record the selected product identifier and action identifier; The system combines user identifiers, session identifiers, and time window identifiers to output user behavior analysis results. The personalized recommendation results and the user behavior analysis results are written into the output record, along with a timestamp and sequence number.

[0017] The beneficial effects of this invention are: This invention achieves significant performance improvements in user behavior analysis and recommendation tasks by introducing multiple artificial intelligence models and fusion mechanisms. First, by employing a Mamba network for selective state-space modeling of user behavior sequences, it not only effectively captures the temporal dependencies of user behavior but also enhances the modeling ability for features at different time scales, avoiding the information decay and computational redundancy problems faced by traditional Transformer models in long sequence modeling. Second, this invention inputs product images into an improved DeiT model and introduces a hierarchical feature routing controller during its encoding process. This allows the model to dynamically select local or global attention paths based on the statistical information of features at different levels, thereby improving the diversity and adaptability of visual feature extraction and enhancing the understanding of product image details and semantic information.

[0018] By designing a bidirectional cross-attention network, this invention achieves deep fusion between behavioral state vectors and visual feature vectors, effectively alleviating the information gap between heterogeneous modalities and improving the expressive power of fused features in user behavior modeling. In the policy learning stage, an IQL reinforcement learning network is introduced, and a reward mechanism is constructed based on actual user feedback, significantly improving the adaptability and generalization ability of the policy model to real-world behavior. Furthermore, the proposed value gradient resonance feedback mechanism during the update process can transmit value information from reinforcement learning back to the attention weights and routing control mechanism, achieving a dynamic optimization closed loop from user feedback to feature extraction path. In summary, this invention not only improves the quality and recommendation accuracy of multimodal feature fusion but also enhances the system's responsiveness and interpretability in complex behavioral scenarios. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0020] Figure 1 This is an overall flowchart of the AI-based precise analysis method for internet user behavior proposed in this invention. Figure 2 This is a diagram showing the network parameter update structure of the value gradient resonance feedback method for precise analysis of internet user behavior based on artificial intelligence proposed in this invention. Detailed Implementation

[0021] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0022] refer to Figure 1-2 The method for precise analysis of internet user behavior based on artificial intelligence includes the following steps: Step 1: Obtain user behavior data and product image data from internet platforms, and construct a set of product images associated with user behavior sequences; Step 2: Input the preprocessed user behavior sequence into the Mamba network and generate behavior state vectors based on the selective state space modeling mechanism; Step 3: Input the set of product images into the improved DeiT model, set a hierarchical feature routing controller before the Transformer encoding layer. The hierarchical feature routing controller generates a routing control vector based on the feature statistics output by the previous layer and selects a path between the local attention branch and the global attention branch, and outputs a visual feature vector and a high-level attention mapping. Step 4: Input the behavioral state vector and visual feature vector into the bidirectional cross-attention network to perform cross-modal feature interaction and generate a fused feature tensor; Step 5: Input the fused feature tensor into the IQL network, generate reward signals based on the user's actual interaction behavior, and output the policy distribution and value gradient distribution; Step 6: Based on the value gradient resonance feedback mechanism, the value gradient distribution is mapped to the weight parameters of the bidirectional cross-attention network to update the fusion weights, and the value gradient distribution is mapped to the high-level attention mapping and routing control vector of the improved DeiT model to update the visual attention distribution and feature path selection. Step 7: Output user behavior analysis results and personalized recommendation results based on the updated strategy distribution.

[0023] This invention proposes a method for multimodal fusion modeling and policy feedback optimization based on user behavior and product images. It extracts user behavior state vectors using a Mamba network with selective state space modeling, introduces an improved DeiT model with hierarchical feature path control capabilities to generate visual features and high-level attention mappings, combines a bidirectional cross-attention network to achieve multimodal feature interaction, and performs policy inference and value estimation on the fused features through an IQL network. Furthermore, it utilizes a value gradient resonance feedback mechanism to dynamically optimize visual path selection and fusion weights, ultimately achieving accurate user behavior analysis and personalized recommendation output. This method boasts advantages such as high cross-modal understanding depth, good policy optimization effect, and strong adaptability.

[0024] In this embodiment, step one specifically includes: Read user behavior data generated by the Internet platform, wherein the user behavior data includes user identifier, session identifier, event type, event timestamp, product identifier and page identifier; The user's behavior events within the same session are sorted in ascending order according to the event timestamp, generating a user behavior sequence and creating a sequence number and index table for the user behavior sequence. Read the product image data corresponding to the product identifier, wherein the product image data includes the image identifier, product identifier, acquisition time and version identifier; Based on the product identifier and event timestamp in the user behavior sequence, image acquisition records that match the event timestamp are selected from the product image data according to the preset association rules, and a one-to-one or one-to-many correspondence between the behavior event and the product image is established. Establish a mapping relationship between user identifier, session identifier, serial number, product identifier, and image identifier, and record index information; Summarize each product image associated with a user behavior sequence to form a set of product images associated with the user behavior sequence.

[0025] In the data preprocessing stage, this invention reads user behavior data and product image data generated by the Internet platform, combines event timestamps and product identifiers, constructs a one-to-one or one-to-many correspondence between behavior events and product images, and establishes a complete mapping and indexing system between user identifiers, session identifiers, sequence numbers and image identifiers. Finally, it generates a set of product images associated with user behavior sequences, providing a structured and time-consistent data foundation for subsequent multimodal modeling and personalized recommendations.

[0026] In this embodiment, the preprocessing specifically includes: A fixed-length category vector is generated by embedding mapping for event type, product identifier, and page identifier. The user behavior sequence is input into a Mamba network, and a behavior state vector is generated based on a selective state space modeling mechanism. The corresponding user behavior sequence is read according to the user identifier and session identifier based on the sequence number and index table. For each behavior event in the user behavior sequence, an event feature item containing the user identifier, session identifier, event type, event timestamp, product identifier, and page identifier is assembled. The event timestamp is split into a date field, an hour field, a week sequence number field, and an adjacent event time difference field, and encoded into a fixed-length time vector. The event type, product identifier, page identifier, and event timestamp are fields in the user behavior sequence. The fixed-length category vector and the fixed-length time vector are concatenated in a preset order to form a fixed-length event vector sequence. The fixed-length event vector sequence is sorted in ascending order based on the event timestamp, and a time window sequence is generated according to a fixed window length and a fixed sliding step. For fixed-length event vector sequences that are shorter than the window length, padding is performed and the effective length is recorded. In the preprocessing of user behavior data, this invention generates fixed-length category vectors by performing embedding mapping on event types, product identifiers, and page identifiers, and constructs fixed-length time vectors by combining event timestamp information. The two are then concatenated to form a fixed-length event vector sequence. Subsequently, the sequence is sorted and divided into time window sequences according to the event timestamps. While ensuring temporal continuity, a sliding window mechanism is introduced to improve the context awareness capability of behavior modeling. Finally, the processed sequence is input into a Mamba network, which generates behavior state vectors with user behavior preferences and temporal characteristics based on a selective state space modeling mechanism, providing high-quality input for cross-modal feature fusion and policy learning.

[0027] In this embodiment, step two specifically includes: The time window sequence is input into the Mamba network, and the linear mapping from the event vector to the input dimension of the Mamba network is completed in the input mapping layer; the event vector is a component of a fixed-length event vector sequence; the Mamba network includes four units: the input mapping layer, the selective state-space modeling unit, the layer normalization and residual connection unit, and the mapping layer. In the selective state-space modeling unit, state updates and output mapping are performed sequentially, and inter-layer output alignment is completed in the layer normalization and residual connection unit. Selective state space modeling units are stacked sequentially according to a preset number of layers. The output of the previous selective state space modeling unit is updated and features are transferred. The hidden state vector corresponding to the time window is extracted at the output of the last layer of the Mamba network as the window-level hidden state output. Aggregation operations are performed on all window-level hidden states in chronological order to obtain session-level hidden states, and these session-level hidden states are transformed into behavior state vectors at the mapping layer. An index relationship is established between the user identifier, session identifier, and behavior state vectors, and the behavior state vectors are output for use in subsequent steps.

[0028] This invention introduces the Mamba network to perform deep modeling of the time window structure of user behavior sequences. On the input side, a linear mapping layer converts a fixed-length event vector sequence into an adapted network input format. Subsequently, state update and output mapping operations are performed sequentially in the selective state space modeling unit, and feature stabilization and inter-layer alignment are achieved through layer normalization and residual connection mechanisms. In the multi-layer stacked structure, the hidden state corresponding to each time window is extracted as a window-level representation, and further aggregated into a session-level hidden state in chronological order. Finally, it is converted into a high-dimensional behavior state vector in the mapping layer and an index relationship with user identifier and session identifier is established, providing a structured and time-aware behavior representation for subsequent cross-modal feature fusion and behavior decision-making.

[0029] In this embodiment, the hierarchical feature routing controller generates a routing control vector based on the feature statistics output by the previous layer, specifically including: For each product image in the product image set, size normalization and image block processing are performed to obtain a sequence of image blocks arranged in a fixed block order; The image patch sequence is input into the block embedding layer of the improved DeiT model and positional encoding is added to generate an image patch embedding sequence as input to the Transformer coding layer; Set a hierarchical feature routing controller at the input of the Transformer coding layer; The hierarchical feature routing controller reads the statistical values ​​of the output feature vector of the previous Transformer coding layer and generates a routing control vector; The statistical values ​​of the output feature vector include channel mean, variance, and attention response intensity; The feature statistics are used to generate a routing control vector through nonlinear mapping; And determine the feature path of this layer between the local attention branch and the global attention branch based on the routing control vector; The process of determining the feature path includes: performing feature calculations on the outputs of the local attention branch and the global attention branch, respectively. The outputs of the two branches are weighted and fused based on the weight coefficients of the routing control vector. The weight coefficients of the routing control vector are in Values ​​can be taken within a range; When the weight coefficients are greater than or equal to a preset threshold, the weighting coefficient of the global attention branch in the current Transformer encoding layer output feature vector is greater than the weighting coefficient of the local attention branch. When the weight coefficient is less than the preset threshold, the weight coefficient of the local attention branch in the output feature vector of the current Transformer encoding layer is greater than the weight coefficient of the global attention branch; The weighted fusion result is used as the output feature of this layer and input into the subsequent feedforward network for processing.

[0030] This invention introduces a hierarchical feature routing controller into the improved DeiT model. By performing size normalization and image block processing on images in the product image set, an image block embedding sequence is constructed as input to the Transformer encoding layer. A routing controller is then set up before the encoding layer to dynamically control the feature path. This routing controller generates a routing control vector based on statistical information such as the channel mean, variance, and attention response intensity of the previous layer's output features. Based on this vector, it selects the feature path between local and global attention branches, adjusting the weighted fusion ratio of the branch outputs through nonlinear mapping. When the control weight coefficient is greater than a preset threshold, the global attention response is strengthened; conversely, the local structure modeling capability is enhanced, thereby achieving adaptive path control of the Transformer encoding layer and improving the multi-scale modeling capability and attention distribution accuracy of image features.

[0031] In this embodiment, the step of selecting a path between the local attention branch and the global attention branch, and outputting a visual feature vector and a high-level attention mapping, specifically includes: In the local attention branch, attention is calculated within the range of adjacent embedding vectors of the image patch embedding sequence to obtain the local context weighted result and output the set of local context feature vectors; In the global attention branch, attention calculation is performed on all embedding vectors of the image patch embedding sequence to obtain the global dependency modeling result and output a set of global semantic feature vectors. The local context feature vector set and the global semantic feature vector set are weighted and fused according to the weight coefficient of the routing control vector to obtain the output feature vector of the current Transformer coding layer; The output feature vector is processed sequentially by a feedforward network, residual connection, and layer normalization. The processed result is used as the input to the next Transformer encoding layer until the preset number of encoding layers are completed. The global feature vector is extracted from the output of the final Transformer encoding layer as the visual feature vector; A high-level attention map is generated based on the attention weights of the final Transformer coding layer. The high-level attention map is used to characterize the attention distribution relationship between the embedding vectors of each image patch and the global feature vector.

[0032] This invention sets up local and global attention branches in the improved DeiT model, and jointly outputs visual feature vectors and high-level attention maps through a path selection mechanism. Specifically, the local attention branch focuses on contextual modeling of adjacent regions in the image patch embedding sequence, generating a set of local contextual feature vectors; the global attention branch focuses on semantic dependencies across the entire image, generating a set of global semantic feature vectors. The two branches are weighted and fused according to the weight coefficients of the routing control vectors during each encoding layer, and the output is used as the feature vector of the current encoding layer and input into the next layer, until all encoding is completed. In the final Transformer encoding layer, the global feature vector is extracted as the visual feature vector, and a high-level attention map is generated based on the attention weights to describe the attention contribution relationship of each image patch to the global representation, thereby achieving an organic combination of fine-grained semantic expression and global feature aggregation.

[0033] In this embodiment, step four specifically includes: Based on the cross-attention mechanism, the behavior state vector is used as the query vector, and the visual feature vector is used as the key vector and value vector, which are then input into the cross-attention unit. In the cross-attention unit, the attention weight matrix is ​​calculated based on the similarity between the query vector and the key vector; Based on the attention weight matrix, a weighted summation is performed on the value vector to generate behavior-guided cross-modal features; The visual feature vector is used as the query vector, and the behavior state vector is used as the key vector and value vector, which are then input into the cross-attention unit. In the cross-attention unit, the attention weight matrix is ​​calculated based on the similarity between the query vector and the key vector; Based on the attention weight matrix, a weighted summation is performed on the value vector to generate visually guided cross-modal features; In the feature fusion layer, a concatenation operation is performed on the behavior-guided cross-modal features and the vision-guided cross-modal features; A linear mapping is performed on the concatenated features based on the fusion weights to generate intermediate fusion features; The fusion weights are used to control the weighting ratio of behavioral guidance features and visual guidance features in the fusion intermediate features; The fused intermediate features are sequentially subjected to residual connection and layer normalization processing to output the fused feature tensor.

[0034] This invention introduces a bidirectional cross-attention mechanism into a bidirectional cross-attention network to achieve deep interaction between user behavior states and product visual features. Specifically, firstly, using the behavior state vector as the query and the visual feature vector as the key and value, cross-modal features guided by behavior are calculated; then, using the visual feature vector as the query and the behavior state vector as the key and value, cross-modal features guided by visual information are calculated. Subsequently, the two types of cross-modal features are concatenated in the feature fusion layer, and linearly mapped according to a preset fusion weight ratio to generate fused intermediate features. Finally, the fused feature tensor is output through residual connections and layer normalization operations, providing a high-quality joint feature representation for subsequent policy reasoning and value assessment. This fusion process enhances the coupling relationship between behavioral and visual information, improving the expressive power and discriminative power of cross-modal representation.

[0035] In this embodiment, step five specifically includes: The fused feature tensor is received and rearranged and normalized according to a preset dimension. The processed fused feature tensor is written as a state input to the state input terminal of the IQL network. Construct a candidate action set and perform embedding mapping on the candidate action set to generate action vectors. Combine the action vectors with the state input according to a preset concatenation order to form a state-action input. Constructing a candidate action set includes: extracting an original action set containing product identifiers and behavior types from user historical interaction data and product databases on internet platforms; calculating the semantic similarity of actions based on the behavior state vector and visual feature vector; and selecting actions with semantic similarity greater than a preset threshold as candidate actions. The state input is fed into the policy network and the value network respectively to perform forward inference, and the corresponding policy score and state value are output. The state-action input is fed into the state-action value network for forward inference, and the corresponding state-action value is output. Reward signals are generated based on actual user interaction behavior and are associated with session identifiers, time window identifiers, and action identifiers to form data records for IQL network updates. In the IQL network, parameter update calculations are performed based on reward signals, state values, and state-action values. The gradient weights of the state-action value network are adjusted according to the policy score, and the objective function of the value network is updated based on the difference between the state value and the state-action value. After completing the parameter update, the updated policy network, value network, and state-action value network are obtained. Based on the updated state-action value network output, the partial derivative of the state-action value with respect to the fused feature tensor is calculated. The partial derivative is then used as the gradient tensor and normalized according to a preset mapping rule to generate a value gradient distribution. The updated policy network output is normalized to a policy distribution.

[0036] This invention introduces an IQL (Implicit Q-Learning) reinforcement learning mechanism into the user behavior modeling process. It utilizes a fused feature tensor as state input to construct a candidate action set containing product identifiers and behavior types. Based on the behavior state vector and visual feature vector, it calculates the semantic similarity of actions, filtering out candidate actions highly relevant to the current context. Through joint reasoning of the policy network, value network, and state-action value network, it outputs policy scores, state values, and state-action values. Combined with real user interaction behavior, it generates reward signals and performs multi-network parameter updates. Furthermore, based on the updated state-action value network, it calculates its gradient relative to the fused feature tensor, generating a value gradient distribution for cross-modal guidance. The policy output is then normalized to form a policy distribution, providing an optimization basis for the final recommendation decision and achieving personalized recommendations that better align with users' potential interests.

[0037] In this embodiment, step six specifically includes: Based on the value gradient resonance feedback mechanism, the value gradient distribution is mapped to the weight parameters of the bidirectional cross-attention network to update the fusion weights, and the value gradient distribution is mapped to the high-level attention mapping and routing control vector of the improved DeiT model to update the visual attention distribution and feature path selection. The value gradient distribution is converted into weight adjustment coefficients on the bidirectional cross-attention network side according to the preset mapping rules. The behavior-guided attention weight matrix and the visual-guided attention weight matrix are reweighted and normalized according to the weight adjustment coefficients to obtain the updated cross-attention weight parameters. For the linear mapping weights of the feature fusion layer in the bidirectional cross-attention network, perform element-wise update and normalization according to the adjustment coefficients corresponding to the cross-attention weight parameters to obtain the updated fusion weights, and write them into the feature fusion layer. The value gradient distribution is mapped to the high-level attention reweighting coefficients on the improved DeiT model side. The attention weights of the final Transformer encoding layer are reweighted and normalized to generate the updated high-level attention mapping, which is then written into the final Transformer encoding layer. The routing control increment vector is calculated based on the value gradient distribution. The routing control increment vector is combined with the previous round routing control vector by a preset weighting coefficient and normalized to obtain the updated routing control vector, which is then written into the hierarchical feature routing controller.

[0038] The routing control increment vector is calculated based on the value gradient distribution, specifically including: reading the gradient components corresponding to the visual feature dimensions in the value gradient distribution; performing normalization and threshold truncation on the gradient components of each dimension to obtain a standardized gradient vector; performing element-wise weighted calculation on the standardized gradient vector and the high-level attention mapping of the current Transformer coding layer to generate a gradient-weighted feature vector; calculating the channel importance coefficient based on the mean and variance of the channel responses of the gradient-weighted feature vector; and obtaining the routing control increment vector by performing linear transformation and nonlinear activation mapping on the channel importance coefficient.

[0039] The updated routing control vector is used in subsequent Transformer encoding layers to dynamically adjust the path selection weights of local attention branches and global attention branches, driving the model to adaptively update the visual attention distribution and feature path selection during feature propagation.

[0040] This invention proposes an adaptive tuning method based on a value gradient resonance feedback mechanism. The value gradient distribution generated in the IQL network is mapped to a bidirectional cross-attention network and an improved DeiT model, respectively, enabling dynamic updates of the weights in cross-modal fusion and visual perception modules. Specifically, this includes: adjusting the behavior guidance and visual guidance attention matrices in the bidirectional cross-attention network according to the gradient distribution, and updating the linear mapping weights of the feature fusion layer; simultaneously, applying the gradient distribution to the high-level attention mapping and hierarchical feature routing control vector of the DeiT model, and generating routing control increments by calculating gradient-weighted feature vectors and channel importance coefficients, thereby achieving adaptive updates of the visual path selection strategy and improving the model's responsiveness to potential user preferences and recommendation accuracy.

[0041] In this embodiment, step seven specifically includes: Receive the updated policy distribution output from the policy network; According to the preset action decoding and filtering rules, the target action set is determined from the candidate action set, and the action identifier and corresponding probability are retained for each action in the target action set; The preset action decoding and filtering rules are used to decode the policy distribution output by the policy network into executable user behavior prediction results, including action type parsing rules, action validity filtering rules, action time constraint rules, action conflict elimination rules, action probability normalization and ranking rules, and action mapping and recommendation generation rules.

[0042] Among them, the action type parsing rule is used to map the action index output in the strategy distribution to a predefined behavior type; the action validity filtering rule is used to filter out invalid actions based on the semantic similarity between the behavior state vector and the product feature vector; the action time constraint rule is used to constrain the repeated triggering of the same type of action within the time window; the action conflict elimination rule is used to resolve mutually exclusive behaviors under the same product identifier; the action probability normalization and ranking rule is used to generate a target action set based on the strategy probability; and the action mapping and recommendation generation rule is used to convert the target action set into a recommendation candidate set to output personalized recommendation results.

[0043] The corresponding probabilities are generated by the output layer of the policy network in the IQL network. After receiving the fused feature tensor as state input, the policy network outputs an unnormalized policy score for each action in the candidate action set; the policy scores are then converted into a probability distribution using a normalization function to obtain the policy distribution. Based on the index correspondence between the policy distribution and the candidate action set, a corresponding probability value is assigned to each candidate action.

[0044] When executing action decoding and filtering rules, validity checks and conflict resolution are performed on actions corresponding to probability values. The probability of invalid actions is set to zero, and the probability of valid actions is re-normalized. Finally, in the target action set, each action retains its action identifier and corresponding probability for subsequent recommendation ranking and user behavior analysis output.

[0045] Based on the mapping relationship between action identifiers and product identifiers, the target action set is converted into a recommendation candidate set; The recommended candidate set is sorted according to the probability corresponding to the updated strategy distribution to generate the sorting result; Based on preset quantity parameters, select the preceding items from the sorting results, generate personalized recommendation results, and record the selected product identifier and action identifier; The user behavior analysis results are output by combining the user identifier, session identifier, and time window identifier. The user behavior analysis results include the correspondence between the policy distribution, the target action set and its corresponding probability, and the user identifier, session identifier, and time window identifier. The personalized recommendation results and the user behavior analysis results are written into the output record, along with a timestamp and sequence number.

[0046] This invention proposes a method for action decoding and personalized recommendation generation based on policy distribution. Relying on the updated IQL policy network output, action indices in the candidate action set are mapped to specific behavior types. Multiple rules, including behavior state, product semantics, time constraints, and action conflicts, are combined to filter and generate a target action set and its corresponding probabilities. Further, a recommendation candidate set is constructed based on the mapping relationship between actions and products, and then sorted, filtered, and output based on policy distribution to form the final personalized recommendation result. Simultaneously, contextual information such as user identifier, session identifier, and time window identifier are recorded to generate corresponding user behavior analysis results and write them to the output record, achieving accurate prediction and recommendation response of user behavior.

[0047] Example 1: To verify the feasibility of this invention in practice, it was applied to user behavior modeling and personalized recommendation tasks on a large e-commerce platform. By constructing a visual feature-guided behavior recognition and recommendation system, accurate prediction and optimized recommendations were achieved for active users' behaviors such as browsing, clicking, favoriting, adding to cart, and placing orders. The main problem in this scenario is that traditional recommendation systems have weak ability to fuse multimodal user information (especially visual cues), making it difficult to perceive the deep connection between images and behavioral semantics, thus leading to a decrease in the relevance and accuracy of recommendation results.

[0048] In this embodiment, the platform selects the top 5000 most frequently viewed products in the past 30 days as a candidate set, collects browsing and interaction behavior data from nearly 30,000 active users, and preprocesses the image information related to the products, including size normalization and image block segmentation. The processed image blocks are then input into an improved DeiT network to extract multi-level visual features. For each user session, the system first constructs a behavior state vector, which contains the current user's historical behavior in the session window (e.g., browsing sequence, add-to-cart sequence) and contextual information such as timestamps.

[0049] A bidirectional cross-attention mechanism is employed to separately compute visual attention features guided by behavioral state and behavioral context features guided by visual content, which are then fused to generate a cross-modal representation tensor. This tensor is input into an IQL (Implicit Q-Learning)-based reinforcement learning policy network to predict the most likely interaction type and target product for the current user in the current state. The reinforcement learning network is trained using reward signals constructed from real user feedback, combined with temporal features of behavioral paths to form round samples, ultimately forming multi-branch inference modules such as the policy value function and the state-action value function.

[0050] A value gradient resonance feedback mechanism was incorporated during training, making the recommendation strategy not only sensitive to prediction results but also influencing the attention distribution and feature routing selection within the model structure, thereby improving the model's stability and adaptability. The final recommendation output includes user-level behavior predictions, product-level behavior annotations, behavior probability distributions, and recommendation candidate rankings.

[0051] After deployment, the system ran in a real environment for 7 days, generating over 1.2 million personalized recommendations. Comparative analysis with the platform's existing recommendation system showed significant advantages across several key metrics. Click-through rate increased by approximately 13.7%, conversion rate increased by 9.4%, average user dwell time increased by 18.6 seconds, and the relevance score of recommended products improved from an average of 0.66 to 0.81.

[0052] To further analyze the model's performance under different user groups and behavioral patterns, five user groups (highly active / medium active / lowly active / new users / long-term long-term users) were selected for performance comparison. The recommendation acceptance and interaction response were statistically analyzed for each user group to verify the generalization ability and stability of this invention in multi-user behavior modeling. Specific experimental data are shown in Table 1.

[0053] Table 1 Summary of Comparison Results of Recommendation Systems under Multiple User Behavior Types

[0054] As shown in Table 1, this invention demonstrates significant improvements in recommendation performance across different user types, specifically in multiple dimensions including click-through rate (CTR), conversion rate, dwell time, and recommendation relevance score. Regarding CTR improvement, all user types achieved increases of over 10%, with new users showing the largest increase at 15.6%. In terms of conversion rate improvement, highly active users and long-term existing users saw the most significant improvements, reaching 10.1% and 10.3% respectively. Average dwell time increased significantly across all user types, with the highest increase reaching 21.3 seconds. The recommendation product relevance score was significantly higher than traditional systems, exceeding 0.75 for all user types, reaching a maximum of 0.84. As for recommendation response time, the average response time was kept below 100ms, with new users exhibiting the fastest response time at only 85.3ms.

[0055] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for precise analysis of internet user behavior based on artificial intelligence, characterized in that: Includes the following steps: Step 1: Obtain user behavior data and product image data from internet platforms, and construct a set of product images associated with user behavior sequences; Step 2: Input the preprocessed user behavior sequence into the Mamba network and generate behavior state vectors based on the selective state space modeling mechanism; Step 3: Input the set of product images into the improved DeiT model, set a hierarchical feature routing controller before the Transformer encoding layer. The hierarchical feature routing controller generates a routing control vector based on the feature statistics output by the previous layer, and performs path selection between the local attention branch and the global attention branch, outputting a visual feature vector and a high-level attention mapping. Step 4: Input the behavioral state vector and visual feature vector into the bidirectional cross-attention network to perform cross-modal feature interaction and generate a fused feature tensor; Step 5: Input the fused feature tensor into the IQL network, generate reward signals based on the user's actual interaction behavior, and output the policy distribution and value gradient distribution; Step 6: Based on the value gradient resonance feedback mechanism, update the fusion weights of the bidirectional cross-attention network, the high-level attention mapping of the improved DeiT model, and the routing control vector of the hierarchical feature routing controller using the value gradient distribution. Step 7: Output user behavior analysis results and personalized recommendation results based on the updated strategy distribution.

2. The method for precise analysis of internet user behavior based on artificial intelligence according to claim 1, characterized in that, Step one specifically includes: Read user behavior data generated by the Internet platform, wherein the user behavior data includes user identifier, session identifier, event type, event timestamp, product identifier and page identifier; The user's behavior events within the same session are sorted in ascending order according to the event timestamp, generating a user behavior sequence and creating a sequence number and index table for the user behavior sequence. Read the product image data corresponding to the product identifier; Establish a mapping relationship between user identifier, session identifier, serial number, product identifier, and image identifier; Summarize each product image associated with a user behavior sequence to form a set of product images associated with the user behavior sequence.

3. The method for precise analysis of internet user behavior based on artificial intelligence according to claim 1, characterized in that, The preprocessing specifically includes: Fixed-length category vectors are generated by embedding mapping for event types, product identifiers, and page identifiers; The event timestamp is split into a date field, an hour field, a week sequence number field, and a time difference field between adjacent events, and then encoded into a fixed-time vector. Concatenate the fixed-length category vector and the fixed-length time vector in a preset order to form a fixed-length event vector sequence; The fixed-length event vector sequence is sorted in ascending order based on the event timestamp, and a time window sequence is generated according to a fixed window length and a fixed sliding step. For fixed-length event vector sequences that are shorter than the window length, padding is performed and a valid length marker is recorded.

4. The method for precise analysis of internet user behavior based on artificial intelligence according to claim 1, characterized in that, The generation of behavioral state vectors based on the selective state-space modeling mechanism specifically includes: The time window sequence is input into the Mamba network, and the linear mapping from the event vector to the input dimension of the Mamba network is completed in the input mapping layer. In the selective state-space modeling unit, state updates and output mapping are performed sequentially, and inter-layer output alignment is completed in the layer normalization and residual connection unit. Selective state space modeling units are stacked sequentially according to a preset number of layers. The output of the previous selective state space modeling unit is updated and features are transferred. The hidden state vector corresponding to the time window is extracted at the output of the last layer of the Mamba network as the window-level hidden state output. The session-level hidden states are obtained by performing aggregation operations on all window-level hidden states in chronological order, and then the session-level hidden states are transformed into behavior state vectors in the mapping layer.

5. The method for precise analysis of internet user behavior based on artificial intelligence according to claim 1, characterized in that, The hierarchical feature routing controller generates a routing control vector based on the feature statistics output from the previous layer, specifically including: For each product image in the product image set, size normalization and image block processing are performed to obtain a sequence of image blocks arranged in a fixed block order; The image patch sequence is input into the block embedding layer of the improved DeiT model and positional encoding is added to generate an image patch embedding sequence as input to the Transformer coding layer; Set a hierarchical feature routing controller at the input of the Transformer coding layer; The hierarchical feature routing controller reads the statistical values ​​of the output feature vector of the previous Transformer coding layer and generates a routing control vector; The statistical values ​​of the output feature vector include channel mean, variance, and attention response intensity; The feature statistics are used to generate a routing control vector through nonlinear mapping; And determine the feature path of this layer between the local attention branch and the global attention branch based on the routing control vector; The process of determining the feature path includes: performing feature calculations on the outputs of the local attention branch and the global attention branch, respectively. The outputs of the two branches are weighted and fused based on the weight coefficients of the routing control vector. When the weight coefficients are greater than or equal to a preset threshold, the weighting coefficient of the global attention branch in the current Transformer encoding layer output feature vector is greater than the weighting coefficient of the local attention branch. When the weight coefficient is less than the preset threshold, the weight coefficient of the local attention branch in the output feature vector of the current Transformer encoding layer is greater than the weight coefficient of the global attention branch; The weighted fusion result is used as the output feature of this layer.

6. The method for precise analysis of internet user behavior based on artificial intelligence according to claim 1, characterized in that, The step of selecting a path between the local attention branch and the global attention branch, and outputting a visual feature vector and a high-level attention mapping, specifically includes: In the local attention branch, attention is calculated within the range of adjacent embedding vectors of the image patch embedding sequence to obtain the local context weighted result and output the set of local context feature vectors; In the global attention branch, attention calculation is performed on all embedding vectors of the image patch embedding sequence to obtain the global dependency modeling result and output a set of global semantic feature vectors. The local context feature vector set and the global semantic feature vector set are weighted and fused according to the weight coefficient of the routing control vector to obtain the output feature vector of the current Transformer coding layer; The output feature vector is processed sequentially by a feedforward network, residual connection, and layer normalization. The processed result is used as the input to the next Transformer encoding layer until the preset number of encoding layers are completed. The global feature vector is extracted from the output of the final Transformer encoding layer as the visual feature vector, and a high-level attention map is generated based on the attention weights of the final Transformer encoding layer.

7. The method for precise analysis of internet user behavior based on artificial intelligence according to claim 1, characterized in that, Step four specifically includes: Based on the cross-attention mechanism, the behavior state vector is used as the query vector, and the visual feature vector is used as the key vector and value vector to calculate the behavior-guided attention weight matrix. Based on the attention weight matrix, a weighted summation is performed on the value vector to generate behavior-guided cross-modal features; The visual feature vector is used as the query vector, and the behavior state vector is used as the key vector and value vector to calculate the visual guidance attention weight matrix; Based on the attention weight matrix, a weighted summation is performed on the value vector to generate visually guided cross-modal features; In the feature fusion layer, the behavior-guided cross-modal features and the vision-guided cross-modal features are concatenated; the concatenated features are linearly mapped according to the fusion weights to generate intermediate fusion features. The fused intermediate features are sequentially subjected to residual connection and layer normalization processing to output the fused feature tensor.

8. The method for precise analysis of internet user behavior based on artificial intelligence according to claim 1, characterized in that, Step five specifically includes: The fused feature tensor is received and rearranged and normalized according to a preset dimension. The processed fused feature tensor is written as a state input to the state input terminal of the IQL network. Construct a candidate action set and perform embedding mapping on the candidate action set to generate action vectors. Combine the action vectors with the state input according to a preset concatenation order to form a state-action input. The state input is fed into the policy network and the value network respectively to perform forward inference, and the corresponding policy score and state value are output. The state-action input is fed into the state-action value network for forward inference, and the corresponding state-action value is output. Reward signals are generated based on actual user interaction behavior and are associated with session identifiers, time window identifiers, and action identifiers to form data records for IQL network updates. In the IQL network, parameter update calculations are performed based on reward signals, state values, and state-action values. The gradient weights of the state-action value network are adjusted according to the policy score, and the objective function of the value network is updated based on the difference between the state value and the state-action value. After completing the parameter update, the updated policy network, value network, and state-action value network are obtained. Based on the updated state-action value network output, the partial derivative of the state-action value with respect to the fused feature tensor is calculated. The partial derivative is then used as the gradient tensor and normalized according to a preset mapping rule to generate a value gradient distribution. The updated policy network output is normalized to a policy distribution.

9. The method for precise analysis of internet user behavior based on artificial intelligence according to claim 1, characterized in that, Step six specifically includes: The value gradient distribution is converted into weight adjustment coefficients on the bidirectional cross-attention network side according to the preset mapping rules; The behavior-guided attention weight matrix and the visual-guided attention weight matrix are reweighted and normalized according to the weight adjustment coefficients to obtain the updated cross-attention weight parameters. For the linear mapping weights of the feature fusion layer in the bidirectional cross-attention network, perform element-wise update and normalization according to the adjustment coefficients corresponding to the cross-attention weight parameters to obtain the updated fusion weights, and write them into the feature fusion layer. The value gradient distribution is mapped to the high-level attention reweighting coefficients on the improved DeiT model side. The attention weights of the final Transformer encoding layer are reweighted and normalized to generate the updated high-level attention mapping, which is then written into the final Transformer encoding layer. The routing control increment vector is calculated based on the value gradient distribution. The routing control increment vector is combined with the previous round routing control vector according to the preset weighting coefficient and normalized to obtain the updated routing control vector, which is then written into the hierarchical feature routing controller.

10. The method for precise analysis of internet user behavior based on artificial intelligence according to claim 1, characterized in that, Step seven specifically includes: Receive the updated policy distribution output from the policy network; According to the preset action decoding and filtering rules, the target action set is determined from the candidate action set, and the action identifier and corresponding probability are retained for each action in the target action set; Based on the mapping relationship between action identifiers and product identifiers, the target action set is converted into a recommendation candidate set; The recommended candidate set is sorted according to the probability corresponding to the updated strategy distribution to generate the sorting result; Based on preset quantity parameters, select the preceding items from the sorting results, generate personalized recommendation results, and record the selected product identifier and action identifier; The system combines user identifiers, session identifiers, and time window identifiers to output user behavior analysis results. The personalized recommendation results and the user behavior analysis results are written into the output record, along with a timestamp and sequence number.

Images