A cross-border e-commerce pushing optimization system and method based on big data

Abstract

The application relates to the technical field of cross-border e-commerce intelligent pushing, and discloses a cross-border e-commerce pushing optimization system and method based on big data. The method comprises the following steps: acquiring a user multi-modal behavior track, decoupling a behavior intention, separating exploratory browsing from target search to construct a differentiation graph, aligning asynchronous behavior sequences by using dynamic time warping, generating a normalized interest evolution context, and reconstructing static commodity association into a dynamic interest flow connection network according to the interest evolution context; simulating the diffusion process of information in the network, calculating the penetration strength and the decay rate, recombining a pushing information segment so that the pushing information segment is embedded into the interest flow evolution direction; capturing user feedback in real time and correcting the intention judgment boundary in a backflow manner to form a closed-loop optimization. The method can deeply analyze user intentions and make predictive pushing based on a dynamic interest network, and the accuracy and timeliness of pushing are improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent push technology for cross-border e-commerce, specifically to a cross-border e-commerce push optimization system and method based on big data. Background Technology

[0002] In current cross-border e-commerce platforms, the core of the recommendation system relies on the collection and analysis of users' historical behavior data. Existing technologies generally employ methods based on collaborative filtering or content tagging, aggregating user clicks, browsing time, purchase records, and other behaviors to construct static user interest profiles and product association graphs. These methods treat user interactions as a homogeneous set of events, performing recommendation calculations based on fixed rules or statistical models. The relationships between products are usually established on a pre-set category tree or fixed co-occurrence relationships, forming a structurally stable recommendation network.

[0003] These conventional technical solutions have shortcomings. They fail to effectively analyze the heterogeneous intentions behind users' continuous interactive behaviors, resulting in noise in the interest features mined from the original behavioral trajectories. The user profiles are coarse and cannot accurately reflect users' instantaneous decision-making psychology and potential interest shifts. Static association networks and fixed time windows are difficult to adapt to the asynchronous nature of user behavior and the dynamic nature of interest evolution in real-world scenarios. Different users' behavioral sequences within the same product category differ in rhythm and pattern, and static models cannot effectively align and compare them. The spread and rise and fall of product information among user groups is a dynamic process, and static associations based on historical snapshots cannot capture the real-time changes and diffusion patterns of interest streams, leading to inaccurate push timing and content that is out of touch with the current interest stream. Summary of the Invention

[0004] The purpose of this invention is to provide a cross-border e-commerce push optimization system and method based on big data to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, this invention provides a cross-border e-commerce push optimization method based on big data, the method comprising:

[0006] The system acquires multimodal behavior trajectories generated by users during platform interactions, including serialized records of visual dwell time, interface swiping, and transaction decisions.

[0007] The multimodal behavioral trajectories are decoupled from behavioral intent, separating tentative browsing and targeted search from complex interactive actions to form a behavioral intent differentiation map;

[0008] A dynamic time warping mechanism is embedded in the behavioral intent differentiation map to align asynchronous behavioral sequences of different users under the same product category and generate a normalized interest evolution path.

[0009] Based on the normalized interest evolution path, the topology of product information in the push network is reconstructed, and the association between product nodes is transformed from static category affiliation to dynamic interest flow connection.

[0010] Through the dynamic interest flow connection, the diffusion process of product information in the user interest network is simulated, and the penetration intensity and decay rate of information on potential interest paths are calculated.

[0011] Based on the penetration intensity and attenuation rate, information fragments are reorganized in an orderly manner in the push information pool, and information units with high penetration intensity are embedded in the evolution direction of the user's current interest stream.

[0012] Real-time capture of user feedback signals triggered by the recombined information unit, the user feedback signals including micro changes in the interface operation trajectory and abnormal fluctuations in decision duration;

[0013] The user feedback signal is fed back to the behavior intent decoupling stage to dynamically correct the judgment boundary of the behavior intent differentiation map, forming a closed-loop optimization flow from intent parsing to feedback correction.

[0014] Preferably, the step of decoupling the multimodal behavior trajectory from behavioral intent, separating tentative browsing and targeted search from the composite interactive actions, specifically includes:

[0015] Analyze the continuous operations within a single user session, identify the logical discontinuities and target continuity between operations, and mark operation sequences that are clearly guided by search terms and point to a single category as target search trajectories;

[0016] Extract operation sequences from the conversation that do not contain explicit search terms and involve skipping between different product categories, calculate the frequency and span of the skips, and mark high-frequency and large-span sequences as exploratory browsing trajectories;

[0017] In the targeted search trajectory and the tentative browsing trajectory, the operation intensity features are further extracted, which are quantitatively characterized by sliding speed, click intensity and page scrolling mode.

[0018] Based on the quantitative differences in the operational strength characteristics, sub-intents are divided within the marked trajectory to distinguish between in-depth exploration with a strong purchase tendency and general interest exploration with a weak purchase tendency.

[0019] By fusing trajectory labeling results with sub-intent segmentation results, a multi-layer network graph is constructed with users as nodes, intent types as edges, and operation intensity as weights, namely the behavioral intent differentiation graph.

[0020] Preferably, the step of embedding a dynamic time warping mechanism into the behavioral intent differentiation map to align asynchronous behavioral sequences of different users under the same product category specifically includes:

[0021] In the behavioral intent differentiation map, different user behavior sequences with the same terminating product category are selected as the normalization objects;

[0022] A non-linear alignment algorithm is used to stretch or compress the time axis of the behavior sequence so that the key intent transition points in the sequence are aligned in the time dimension. The key intent transition points include the moment when the user turns from tentative browsing to targeted search.

[0023] In the timeline aligned sequences, the shared intent transformation pattern is extracted, and each aligned sequence is transformed into a regular sequence consisting of intent state and dwell time.

[0024] By overlaying the normalized sequences of all users targeting the same product category, high-frequency paths of intent state transitions are identified in the overlay area. These high-frequency paths constitute the core framework of the normalized interest evolution network.

[0025] Preferably, the step of reconstructing the topology of product information in the push network based on the normalized interest evolution path specifically includes:

[0026] Each product in the platform's product library is treated as an independent node in the network, and the initial connection is established based on the fixed category hierarchy of the product.

[0027] The high-frequency intent transfer paths identified in the normalized interest evolution context are mapped to product nodes, and temporary directed connections based on interest transfer are established between product nodes that originally had no direct category association.

[0028] Each temporary directed connection established based on interest transfer is assigned a weight, the weight value of which is determined by the frequency and intensity of the high-frequency intention transfer path appearing in the normalized interest evolution context.

[0029] Based on the temporary directed connections and their weights, the topology connection matrix of the product network is dynamically updated, so that the association structure between products evolves from a tree-like static category tree into a graph-like dynamic interest flow network that can change over time.

[0030] Preferably, the process of simulating the diffusion of product information in the user interest network through the dynamic interest stream connection specifically includes:

[0031] A single product push event is abstracted as a diffusion simulation of information particles in the dynamic interest flow network;

[0032] Using the product node currently interacting with by the user as the diffusion source point of the information particles, and based on the network topology and connection weight of the dynamic interest flow connection, the probability of the information particles spreading to adjacent nodes along different directed connections is calculated.

[0033] In the diffusion simulation process, an attenuation factor is introduced. The attenuation factor is negatively correlated with the length of the connection path and the historical exposure of the nodes on the path. The penetration intensity of the information particle is attenuated according to the attenuation factor each time it passes through a node.

[0034] Traverse the potential nodes that can be reached from the diffusion source point through several hop connections, calculate the final penetration intensity of the information particles to reach each potential node within a set number of simulation steps, and record the number of steps required to reach the node and the decay rate.

[0035] Preferably, the step of performing ordered recombination of information fragments in the push information pool based on the penetration intensity and attenuation rate specifically includes:

[0036] Set a penetration intensity threshold and a decay rate threshold, and filter out potential product nodes whose final penetration intensity is higher than the penetration intensity threshold and whose decay rate is lower than the decay rate threshold, as high-potential target push nodes.

[0037] From the product information corresponding to the target push node, key information units are extracted. The key information units include the core visual feature description of the product, differentiated attribute tags, and scenario-based usage fragments.

[0038] Based on the specific path of information particles from the diffusion source to each target push node, the interest transfer logic represented by the path is determined, and the extracted key information units are narratively sorted according to this logic.

[0039] The sorted key information units are merged and recombined with the product information of the diffusion source to generate a coherent information push example that conforms to the evolution direction of user interest flow, thus completing the orderly recombination of information fragments.

[0040] Preferably, the user feedback signal triggered by the real-time capture and recombination information unit specifically includes:

[0041] When displaying the reorganized information push example to the user, high-precision interaction log recording is simultaneously enabled. The high-precision interaction log recording captures the user's gaze thermal change trajectory in the information area with millisecond-level precision.

[0042] By analyzing the thermal change trajectory of the gaze, abnormal dwell time, rapid skip sequence, and repeated retracement pattern of the gaze on specific information units are identified. The abnormal dwell time is marked as a deep interest signal, and the rapid skip sequence is marked as an ignore signal.

[0043] Simultaneously monitor the user's touch or cursor movement trajectory on the information display page, and extract the smoothness of the trajectory, the hesitation trajectory before clicking, and the acceleration characteristics when closing the page;

[0044] The analysis results of the gaze thermal change trajectory are coupled with the characteristics of the touch cursor movement trajectory to cross-verify the user's true feedback intensity to the recombined information unit, generating the user feedback signal containing multi-dimensional evidence.

[0045] Preferably, the step of feeding the user feedback signal back to the behavioral intent decoupling stage to dynamically correct the determination boundary of the behavioral intent differentiation map specifically includes:

[0046] The information unit corresponding to the deep attention signal in the user feedback signal is reverse-mapped to the interest transfer path on which the information unit was generated.

[0047] By reinforcing the interest transfer path into the construction process of the behavioral intent differentiation map, the weight of the intent conversion pattern corresponding to the interest transfer path is increased, so that this intent pattern can be identified earlier and more definitively in similar behavioral sequences.

[0048] The information units corresponding to the ignored signals in the user feedback signals are also reverse-mapped to the interest transfer paths they generate.

[0049] In the behavioral intent differentiation map, suppression weights are applied to the mapped interest transfer paths to reduce the priority of the interest transfer paths in intent determination, thereby dynamically narrowing the determination boundary of the intent pattern.

[0050] By cyclically executing the reinforcement and inhibition process, the judgment rules of the behavioral intention differentiation map are continuously evolved, forming a closed-loop optimization flow of intention analysis and feedback correction.

[0051] Preferably, the method further includes:

[0052] Establish cross-session user interest state continuity tracking, seal the interest stream state at the end of the previous session, and load it as the initial diffusion source point at the beginning of the next session to realize long-term interest coherence push simulation.

[0053] Preferably, the present invention also includes a cross-border e-commerce push optimization system based on big data. The system includes a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the steps of the cross-border e-commerce push optimization method based on big data as described above.

[0054] Compared with the prior art, the beneficial effects of the present invention are:

[0055] By decoupling user multimodal behavioral trajectories at the intent level, different psychologically driven behavioral patterns, such as tentative browsing and targeted searching, are separated from complex interactive actions, and a behavioral intent differentiation map is constructed. This transforms raw behavioral sequence data into structured intent metadata, enabling refined modeling of users' complex psychological activities. This allows for the differentiation between users' casual attention and genuine needs in push notification decisions, improving the accuracy and depth of interest recognition and fundamentally improving the signal quality at the push source.

[0056] A dynamic time warping mechanism is introduced to align asynchronous behavior sequences, and based on a normalized interest evolution framework, the relationships between product nodes are reconstructed from static category affiliation to dynamic interest flow connections. This enables the push system to understand and simulate the dynamic diffusion process of interests among user groups. By calculating the penetration intensity and decay rate of information along potential paths, the system can predict the evolution direction of interest flows and accordingly reorganize push information in an orderly manner. Push notifications no longer rely solely on the static matching of products and user history, but rather on the position and flow state of products in the current dynamic interest network. This realizes a shift in push strategy from responding to history to predicting and guiding real-time interest flows, enhancing the timeliness and contextual adaptability of push notifications. Attached Figure Description

[0057] Figure 1 This is a schematic diagram illustrating the working principle of the cross-border e-commerce push optimization method based on big data as described in this invention.

[0058] Figure 2 A flowchart for generating dynamic time warping and normalized interest evolution context;

[0059] Figure 3 A flowchart for reconstructing the network topology for product information push;

[0060] Figure 4 Optimize the user feedback distribution heatmap at each iteration stage of the closed-loop push for cross-border e-commerce;

[0061] Figure 5 This is a graph showing the relationship between the diffusion and penetration intensity of product information and its attenuation rate. Detailed Implementation

[0062] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0063] Please see Figure 1This invention provides a big data-based method for optimizing cross-border e-commerce push notifications. The method includes: acquiring multimodal behavioral trajectories generated by users during interactions on a cross-border e-commerce platform, including temporal records of visual dwell time, interface swiping, and transaction decisions; decoupling behavioral intent from the multimodal behavioral trajectories, separating tentative browsing from targeted searches in complex interactive actions, forming a structured behavioral intent differentiation map; embedding a dynamic time warping mechanism into this map to align asynchronous behavioral sequences of different users within the same product category, thereby generating a normalized interest evolution path; reconstructing the topology of product information in the push network based on this interest evolution path, transforming the association between product nodes from static category affiliation to connections based on dynamic interest flows; simulating the diffusion process of product information in the user's interest network through these dynamic interest flow connections, calculating the penetration intensity and decay rate of information along potential interest paths; and performing ordered recombination of information fragments in the push information pool based on the calculated penetration intensity and decay rate, embedding information units with high penetration intensity into the evolution direction of the user's current interest flow. Real-time capture of user feedback signals triggered by recombined information units includes subtle changes in interface operation trajectories and abnormal fluctuations in decision-making time. These user feedback signals are then fed back to the behavioral intent decoupling stage, dynamically correcting the judgment boundaries of the behavioral intent differentiation map, thus forming a closed-loop optimization flow from intent parsing to feedback correction.

[0064] Example 1: See Figure 2In the process of decoupling behavioral intent, the continuous operations within a single user session are analyzed to identify logical discontinuities and target continuity between operations. Sequences guided by explicit search terms and pointing to a single category are marked as targeted search trajectories. Operation sequences without explicit search terms and involving skipping between different product categories are extracted from the session. The frequency and span of such skips are calculated, and high-frequency, large-span sequences are marked as exploratory browsing trajectories. Within the marked targeted search trajectories and exploratory browsing trajectories, operation intensity features are further extracted. Operation intensity features are quantified using swipe speed, click intensity, and page scrolling patterns. Based on the quantified differences in operation intensity features, sub-intents are segmented within the marked trajectories to distinguish between in-depth exploration with strong purchase inclination and general interest exploration with weak purchase inclination. The trajectory marking results and sub-intent segmentation results are integrated to construct a multi-layered network graph with users as nodes, intent types as edges, and operation intensity as weights, i.e., a behavioral intent differentiation graph. A dynamic time warping mechanism is embedded in the behavioral intent differentiation graph, selecting different user behavior sequences with the same terminating product category as warping objects. A non-linear alignment algorithm is employed to stretch or compress the timelines of these behavioral sequences, aligning key intent transition points along the temporal dimension. These key intent transition points include the moment when browsing shifts to a targeted search. Shared intent transition patterns are extracted from the timeline-aligned sequences, transforming each aligned sequence into a regularized sequence composed of intent state and dwell time. All regularized sequences targeting the same product category are overlaid, and high-frequency paths of intent state transitions are identified within the overlaid area. These high-frequency paths form the core framework of the normalized interest evolution network.

[0065] In practical implementation, the behavioral intent decoupling process analyzes the continuous operations within a single user session, identifying logical discontinuities and goal continuity between operations. For example, a user session might contain the following sequence of operations: entering the search term "summer dress," clicking on the first search result product, lingering to view the product details page for thirty seconds, returning to the search results list, clicking on the third search result product, and quickly swiping through the product details page. In this session, the subsequence explicitly guided by the search term "summer dress" and where all subsequent operations revolve around the "women's clothing - dress" category is marked as a targeted search trajectory. Another sequence of operations from the same session is extracted: the user does not use the search function but clicks on a vase in the "home decor" category, an earphone in the "electronics" category, and a water bottle in the "outdoor sports" category. The jump frequency is calculated to be an average interval of five seconds between each click, and the jump span involves three unrelated primary categories. This high-frequency, large-span sequence is marked as an exploratory browsing trajectory. Within the marked targeted search trajectory and tentative browsing trajectory, further analysis of operational intensity characteristics is conducted. These characteristics are quantified using swipe speed, click intensity, and page scrolling pattern. For example, client-side data tracking records swipe speed as the number of pixels per second, click intensity as the pressure value of the touch event, and page scrolling pattern as the acceleration sequence of the scroll trajectory. Based on the quantified differences in operational intensity characteristics, sub-intents are segmented within the targeted search trajectory: segments with swipe speeds below 100 pixels per second, click intensity values ​​greater than 0.5, and page scrolling exhibiting multiple up-and-down repetitive patterns are classified as deep exploration with strong purchase intent; segments with swipe speeds above 300 pixels per second, click intensity values ​​less than 0.3, and page scrolling exhibiting a unidirectional rapid downward movement are classified as general interest exploration with weak purchase intent. By integrating trajectory labeling results with sub-intent segmentation results, a multi-layer network graph is constructed with users as nodes, intent types as edges, and operation intensity as weights, namely, a behavioral intent differentiation graph. For example, for user identifier UID_123, the behavioral intent differentiation graph includes edges: from node "tentative browsing" to "targeted search" with a weight of 0.75, and from node "general interest exploration" to "deep exploration" with a weight of 0.6.

[0066] In some embodiments, a dynamic time warping mechanism is embedded in the behavioral intent differentiation map. Different user behavior sequences with the same terminating product category are selected as warping objects. For example, the behavior sequences of users A, B, and C all end with viewing a product details page under the "Smartphone" category. A non-linear alignment algorithm is used to stretch or compress the time axis of these behavior sequences, aligning the key intent transition points in the sequence in the time dimension. Key intent transition points include the moment of transition from tentative browsing to targeted search. For example, user A transitions from tentative browsing to targeted search at the 150th second after the session starts, and user B makes the same transition at the 90th second after the session starts. The dynamic time warping algorithm aligns these two transition points to the same reference moment on the virtual time axis. In the specific implementation of the dynamic time warping algorithm, the behavioral sequences of user A and user B are first selected as warping objects. These sequences all terminate at the same product category. The algorithm adopts a non-linear alignment method, dynamically adjusting the time axis scale by calculating the local distance between sequence elements and minimizing the cumulative alignment cost. For example, it stretches the time axis of user B's sequence or compresses the time axis of user A's sequence, so that the transition point from tentative browsing to targeted search that occurs in user A at the 150th second after the start of the session is aligned to a common reference time on the virtual time axis with the same transition point that occurs in user B at the 90th second. During the alignment process, the algorithm identifies key intent transition points in the sequence as alignment anchors and implements flexible mapping of the time dimension based on the minimum cost path, thereby eliminating the asynchronicity caused by differences in the rhythm of individual user behaviors. In the sequences that have completed time axis alignment, the shared intent transition pattern is extracted, and each aligned sequence is transformed into a warped sequence composed of intent state and dwell time. For example, the warped sequence is expressed as an ordered list: ["Tentative browsing", 80 seconds], ["Targeted search", 120 seconds].

[0067] By overlaying all regularized sequences of users targeting the same product category, high-frequency paths of intent state transitions are identified in the overlay area. These high-frequency paths constitute the core framework of the normalized interest evolution pattern. For example, the overlay statistics of one thousand regularized sequences show that the path from "tentative browsing" directly to "targeted search" occurs eight hundred times, and the path from "general interest exploration" through "deep exploration" and then to "targeted search" occurs five hundred and fifty times.

[0068] Optionally, the dynamic time warping mechanism uses the following formula to calculate the minimum cumulative cost of sequence alignment:

[0069]

[0070] in: Indicates length is source sequence With length target sequence Align to position and Minimum cumulative cost at that time Represents the source sequence The element With the target sequence The element Local distance between elements and All include two dimensions: intent state type and dwell time, and local distance. The calculation is a Boolean value representing the difference in intent state type plus the normalized absolute value of the difference in dwell time, for the sequence. and sequence This represents the regularized sequences of behavior from two different users. and It is a sequence index variable, with values ​​ranging from one to the length of the respective sequence.

[0071] Example 2: See Figure 3 When reconstructing the product information topology based on the normalized interest evolution framework, each product in the platform's product library is treated as an independent node in the network. Initial connections are established based on the fixed category hierarchy of the products. High-frequency intent transfer paths identified in the normalized interest evolution framework are mapped to the product nodes, thus establishing temporary directed connections based on interest transfers between product nodes that originally had no direct category association. Each temporary directed connection based on interest transfers is assigned a weight, the weight value of which is determined by the frequency and intensity of the high-frequency intent transfer path appearing in the normalized interest evolution framework. Based on these temporary directed connections and their weights, the topology connection matrix of the product network is dynamically updated, transforming the association structure between products from a tree-like static category tree into a graph-like, time-varying dynamic interest flow network.

[0072] In practical implementation, when reconstructing the topology of product information in the push network based on the normalized interest evolution path, each product in the platform's product library is regarded as an independent node in the network. For example, the product library contains product nodes "Smartphone A", "Bluetooth Headset B", "Smartwatch C", and "Coffee Machine D". The initial connection is established based on the fixed category hierarchy of the products. For example, "Smartphone A" belongs to the category "Electronic Digital > Mobile Communication", and "Bluetooth Headset B" belongs to the category "Electronic Digital > Audio-Visual Entertainment". According to the fixed category hierarchy, these two nodes share the parent category "Electronic Digital", so an initial undirected connection is established between these two product nodes. The high-frequency intent transfer paths identified in the normalized interest evolution path are mapped to the product nodes. For example, the normalized interest evolution path shows a high-frequency path: users frequently transfer their browsing behavior from the "smartphone" category to the "smartwatch" category. However, "smartphone A" and "smartwatch C" belong to different subcategories in the fixed category tree and have no direct parent category relationship. At this time, a temporary directed connection based on interest transfer is established between the product nodes "smartphone A" and "smartwatch C", with the direction from "smartphone A" to "smartwatch C".

[0073] In some embodiments, each temporary directed connection established based on interest transfer is assigned a weight. The weight value is determined by the frequency and intensity of high-frequency intention transfer paths appearing in the normalized interest evolution context. For example, within a one-day time window, it is detected that user interest evolution paths starting from the category interest corresponding to product node "Smartphone A" and ultimately leading to the category corresponding to product node "Smartwatch C" occur 2,000 times. Among these, 1,500 times represent deep exploration transfers with strong purchase inclination, and 500 times represent general interest exploration transfers with weak purchase inclination. The frequency of occurrence is quantified by the total number of paths, and the path intensity is quantified by the proportion of deep exploration transfers in the path. The weight of the temporary directed connection is calculated by combining the frequency and intensity. Based on the temporary directed connections and their weights, the topology connection matrix of the product network is dynamically updated. Each element of the connection matrix... Indicates from product node To the product node The connection weights, originally based on a fixed category, are set to a basic constant value. The temporary directed connection weights based on interest transfer are updated in real time, which makes the association structure between products evolve from a tree-like static category tree to a graph-like dynamic interest flow network that can change over time. In the dynamic interest flow network, the product node "Smartphone A" is connected to both "Bluetooth Headset B" and "Smartwatch C", but the weights and properties of the connections are different.

[0074] Optionally, the calculation of weights for temporary directed connections based on interest transfer follows the formula:

[0075]

[0076] in: Indicates from the source product node Point to the target product node Temporary directed connection weights, This indicates that within the statistical time window, data is collected from the source product node. Starting from the corresponding category interests and ultimately leading to the target product node. The original occurrence count of high-frequency intent transfer paths for the corresponding category. Indicates in Within this secondary path, the proportion of the deep exploration transfer path marked as having a strong purchasing tendency. It is a positive coefficient used to adjust the degree of influence of path intensity. It is a set The index variable in the table represents each product node connected to node p. This represents the original number of occurrences of the high-frequency intent transfer path from the source product node p to the target product node r. This indicates that in a dynamic interest flow network, all nodes related to the source product node... The set of target product nodes associated by temporary directed connections, the denominator pairs All temporary connections are normalized. Product node and product nodes It refers to any two product nodes in the network. It is the preset intensity adjustment coefficient.

[0077] Example 3: Simulating the diffusion of product information through dynamic interest flow connections, a product push event is abstracted as a diffusion simulation of information particles in a dynamic interest flow network. Taking the product node currently interacting with by the user as the diffusion source point of the information particle, the probability of the information particle diffusing to adjacent nodes along different directed connections is calculated based on the network topology and connection weights of the dynamic interest flow connections. During the diffusion simulation, an attenuation factor is introduced, which is negatively correlated with the length of the connection path and the historical exposure of nodes on the path. The penetration intensity of the information particle decreases according to the attenuation factor each time it passes through a node. The potential nodes reachable from the diffusion source point via several hop connections are traversed, and the final penetration intensity of the information particle reaching each potential node within a set number of simulation steps is calculated, recording the required number of steps, i.e., the attenuation rate. Based on the penetration intensity and attenuation rate, ordered recombination of information fragments is performed. A penetration intensity threshold and an attenuation rate threshold are set, and potential product nodes with a final penetration intensity higher than the penetration intensity threshold and an attenuation rate lower than the attenuation rate threshold are selected as high-potential target push nodes. From the product information corresponding to the target push nodes, key information units are extracted. These key information units include descriptions of the product's core visual features, differentiated attribute tags, and contextual usage snippets. Based on the specific paths taken by information particles from the diffusion source to each target push node, the interest transfer logic represented by the paths is determined, and the extracted key information units are narratively ordered according to this logic. The ordered key information units are then merged and recombined with the product information at the diffusion source to generate a coherent information push example that conforms to the evolution direction of the user's interest flow, thus completing the orderly reorganization of information fragments.

[0078] In practical implementation, the diffusion process of product information in the user's interest network is simulated through dynamic interest flow connections. A single product push event is abstracted as a diffusion simulation of information particles in the dynamic interest flow network. Taking the product node currently being interacted with by the user as the diffusion source point of the information particles, for example, if user ID UID_456 is detected browsing the details page of product node "Smartphone X", the dynamic interest flow network shows three temporary directed connections originating from product node "Smartphone X", pointing to product nodes "Smartwatch Y", "Bluetooth Speaker Z", and "Tablet M", respectively. Based on the network topology and connection weights of the dynamic interest flow connections, the probability of information particles diffusing to adjacent nodes along different directed connections is calculated. For example, with connection weights of 0.8, 0.5, and 0.3, after normalization, the probability of information particles diffusing to product node "Smartwatch Y" is 50%, to product node "Bluetooth Speaker Z" is 31.25%, and to product node "Tablet M" is 18.75%. In the diffusion simulation, an attenuation factor is introduced. This attenuation factor is negatively correlated with the length of the connection path and the historical exposure of nodes along the path. Each time an information particle passes through a node, its penetration intensity decreases according to the attenuation factor. For example, if an information particle diffuses from the product node "Smartphone X" to the product node "Smartwatch Y," and "Smartwatch Y" has been exposed to user UID_456 three times in the past hour, its historical exposure is high. Therefore, the attenuation factor for this path is set to 0.7. If the initial penetration intensity of the information particle is one, then the penetration intensity at the product node "Smartwatch Y" will decrease to 0.7. The simulation iterates through potential nodes reachable from the diffusion source point via several hops, calculating the final penetration intensity of the information particle reaching each potential node within a set number of simulation steps, and recording the required number of steps, i.e., the attenuation rate. For example, if the maximum number of simulation steps is set to three hops, the simulation shows that the product node "Sports Headphones N" requires two hops from the product node "Smartphone X" to reach, with a final penetration intensity of 0.42 and an attenuation rate of two.

[0079] In some embodiments, information fragments are ordered and recombined in the push information pool based on penetration intensity and attenuation rate. Penetration intensity thresholds and attenuation rate thresholds are set, for example, a penetration intensity threshold of 0.3 and an attenuation rate threshold of three. Potential product nodes with a final penetration intensity higher than the penetration intensity threshold and an attenuation rate lower than the attenuation rate threshold are selected as high-potential target push nodes. For example, the final penetration intensity of product node "Smartwatch Y" is 0.7 and the attenuation rate is one, while the final penetration intensity of product node "Sports Headphones N" is 0.42 and the attenuation rate is two; both nodes meet the selection criteria. Key information units are extracted from the product information corresponding to the target push nodes. These key information units include descriptions of the product's core visual features, differentiated attribute tags, and scenario-based usage fragments. For example, key information units extracted from the information of product node "Smartwatch Y" include: "1.3-inch AMOLED circular display," "Supports continuous blood oxygen and heart rate monitoring," and "Can be worn while swimming to record sports data." Based on the specific paths taken by information particles from the diffusion source to each target push node, the interest transfer logic represented by the paths is determined. The extracted key information units are then narratively ordered according to this logic. For example, the path from an information particle to the product node "Smartwatch Y" is "Smartphone X -> Smartwatch Y," representing the "phone accessory extension" interest logic. The path from an information particle to the product node "Sports Headphones N" is "Smartphone X -> Sports and Health App K -> Sports Headphones N," representing the "healthy lifestyle extension" interest logic. Key information units are ordered according to the narrative order of "accessories first, extension later." The ordered key information units are then merged and recombined with the product information from the diffusion source to generate a coherent information push example that aligns with the user's interest flow evolution, completing the orderly reorganization of information fragments. For example, the generated push example text is: "The smartphone X you are currently interested in boasts excellent performance. Pairing it with the same brand's smartwatch Y enables seamless interconnection. Smartwatch Y features a 1.3-inch AMOLED circular display and supports continuous blood oxygen and heart rate monitoring. Furthermore, to start a healthy lifestyle, you also need a waterproof sports headphone N, which can be worn while swimming to record exercise data."

[0080] Optionally, the attenuation factor is calculated during the diffusion simulation according to the following formula:

[0081]

[0082] in: Information particles originate from the product node. Directly spread to adjacent product nodes The decay factor at time, It is a fundamental decay rate constant between zero and one. Indicates from product node To the product node The length of the connection path in the current single-hop connection. The value is one. Indicates the target product node The total number of historical exposures to the current user within a recent specific time window. It is the base of the natural logarithm, used to prevent the denominator from being zero. (Product node) and product nodes These are two product nodes with a direct temporary directed connection in a dynamic interest flow network.

[0083] It is understandable that the information particle diffusion simulation process performs probabilistic calculations based on the topology of a dynamic interest flow network. The introduction of an attenuation factor allows the calculation of penetration intensity to take into account the influence of path length and the historical exposure of nodes. Traversing potential nodes and calculating the final penetration intensity and attenuation rate provides a quantitative basis for subsequent screening. Setting a threshold to screen target push nodes is a transformation step from simulation calculation results to actual push content. Extracting and sorting key information units according to interest transfer logic ensures the coherence and directionality of the recombined push information. The information push examples generated by fusion realize the orderly recombination of information fragments.

[0084] Example 4: Real-time capture of user feedback signals triggered by recombined information units. While displaying examples of recombined information to the user, high-precision interaction log recording is simultaneously initiated. This high-precision log recording captures the user's gaze thermal trajectory in the information area with millisecond-level accuracy. Analysis of the gaze thermal trajectory identifies abnormal dwell time, rapid skip sequences, and repeated retracement patterns on specific information units. Abnormal dwell time is marked as a deep attention signal, and rapid skip sequences are marked as ignored signals. Simultaneously, the user's touch or cursor movement trajectory on the information display page is monitored, extracting the smoothness of the trajectory, the hesitation trajectory before clicking, and the acceleration features when closing the page. The analysis results of the gaze thermal trajectory are coupled with the features of the touch cursor movement trajectory to cross-validate the user's true feedback intensity to the recombined information unit, generating a user feedback signal containing multi-dimensional evidence. This user feedback signal is fed back to the behavioral intent decoupling stage, dynamically correcting the judgment boundary of the behavioral intent differentiation map. The information unit corresponding to the deep attention signal in the user feedback signal is back-mapped to the interest transfer path on which the information unit was generated. By reinforcing the input of interest transfer paths into the construction of the behavioral intent differentiation map, the weight of the intent conversion patterns corresponding to these paths is increased, enabling earlier and more definitive identification of these intent patterns in similar behavioral sequences. Similarly, the information units corresponding to ignored signals in user feedback signals are back-mapped to their generated interest transfer paths. In the behavioral intent differentiation map, suppression weights are applied to the mapped interest transfer paths, reducing their priority in intent determination and dynamically narrowing the determination boundary of intent patterns. Through iterative execution of the reinforcement and suppression processes, the determination rules of the behavioral intent differentiation map continuously evolve, forming a closed-loop optimization flow of intent parsing and feedback correction.

[0085] In practical implementation, user feedback signals triggered by the recombined information units are captured in real time, and high-precision interaction log recording is simultaneously activated when displaying examples of the recombined information push to the user. High-precision interaction log recording captures the user's gaze thermal changes in the information area with millisecond-level accuracy. For example, when displaying a recombined push interface containing two key information units, "Smartwatch Battery Life" and "Sports Headphone Waterproof Rating," to user UID_789, the interface tracking module records the movement sequence of the user's gaze focus on the screen pixel coordinates at a frequency of 120 times per second. Analysis of the gaze thermal change trajectory identifies abnormal dwell time, rapid skip sequences, and repeated retracement patterns on specific information units. For example, analysis reveals that the user's gaze lingered for a cumulative 3,500 milliseconds in the screen area corresponding to the "Smartwatch Battery Life" information unit, exceeding twice the standard deviation of the historical average dwell time in that area; this abnormal dwell time is marked as a deep attention signal. Simultaneously, it identifies that the gaze only stayed in the "Sports Headphone Waterproof Rating" information unit area for 200 milliseconds before quickly moving away, forming a rapid skip sequence; this rapid skip sequence is marked as an ignored signal. Simultaneously, the system monitors user touch or cursor movement trajectories on the information display page, extracting the smoothness of the trajectory, the hesitation trajectory before clicking, and the acceleration features when closing the page. For example, it records that the user's finger made multiple minor directional corrections in the last 500 milliseconds when approaching the "Smartwatch Battery Life" information unit area, forming a hesitation trajectory; while when swiping to close the entire push page, it records that the initial upward swipe velocity of the finger exceeded 1,500 pixels per second, forming a high-acceleration closing feature. The analysis results of the gaze thermal change trajectory are coupled with the features of the touch cursor movement trajectory to cross-validate the user's true feedback intensity to the reconstructed information unit, generating user feedback signals containing multi-dimensional evidence. For example, for the "Smartwatch Battery Life" information unit, coupling the deep attention signal and the hesitation trajectory generates a strong positive feedback signal; for the "Sports Headphone Waterproof Rating" information unit, coupling the ignored signal and the high-acceleration closing feature generates a strong negative feedback signal.

[0086] In some embodiments, the cross-validation process integrates multidimensional evidence through a mapping table, as shown in Table 1, which displays the preliminary feedback strength determinations corresponding to different combinations of trajectory features:

[0087] Table 1: Multidimensional Feature Coupling Mapping Table for User Feedback Signals

[0088] Line-of-sight thermal characteristic pattern Touch / cursor trajectory feature patterns Type of feedback signal generated by coupling Feedback Intensity Level The dwell time of a specific unit is abnormally long. A hesitation trajectory appears before clicking the corresponding area. Pay close attention to confirmation signals High-intensity positive The dwell time of a specific unit is abnormally long. No operation was performed in the corresponding area or the area was simply swiped over. Pay close attention to unresolved signals Medium intensity positive Quickly skip sequence High acceleration when closing the page Explicitly ignore signal High intensity negative Quickly skip sequence Low acceleration when closing the page Potential ignored signals Medium intensity negative Repeated retrace mode The trajectory is smooth, accompanying the final click. Comparison Decision Signals Medium intensity positive Repeated retrace mode Smooth trajectory, no final click Interest assessment signals Low intensity positive

[0089] Optionally, the analysis for identifying anomalies in dwell time within the line-of-sight thermal change trajectory follows the quantitative judgment based on the following formula:

[0090]

[0091] in: Representation of information unit Abnormal dwell time values ​​for corresponding screen areas This indicates that the user's gaze is on the information unit in the current session. The actual total dwell time for the corresponding screen area. Representation of information unit The average dwell time of the corresponding screen area within the historical statistical period. Representation of information unit Standard deviation of dwell time for the corresponding screen area within the historical statistical period. Information unit Represents any key information unit in the reorganization promotion example, when Greater than the preset threshold If the stay duration is deemed abnormal, then... Then it is marked as a deep interest signal, if Then, combined with other modes, it is marked as an ignored signal. It is a preset significance threshold parameter.

[0092] Understandably, high-precision interaction logs provide raw gaze and operation trajectory data. Analyzing gaze thermal changes identifies basic attention and neglect patterns, monitoring touch cursor trajectories extracts supplementary features of operation force and intent, and coupling analysis integrates evidence from both visual gaze and physical interaction to generate more reliable user feedback signals containing intensity levels. These user feedback signals are then fed back to the behavioral intent decoupling stage to dynamically correct the judgment boundaries of the behavioral intent differentiation map. Information units corresponding to deep attention signals in the user feedback signals are mapped inversely to the interest transfer path upon which the information unit was generated. For example, a high-intensity positive "deep attention confirmation signal" corresponds to the "smartwatch battery life" information unit. This information unit originates from the product node "smartwatch Y," which was selected through the interest transfer path "smartphone X -> smartwatch Y." By reinforcing the input of interest transfer paths into the construction of the behavioral intent differentiation map, the weight of intent conversion patterns corresponding to these paths is increased. This allows for earlier and more definitive identification of these intent patterns in similar behavioral sequences. For example, the weight of the intent conversion path "from smartphone browsing to smartwatch browsing" is increased by 10%. Information units corresponding to ignored signals in user feedback signals are also back-mapped to their generated interest transfer paths. For example, a high-intensity negative "explicitly ignored signal" corresponds to the "waterproof rating of sports headphones" information unit, originating from the path "smartphone X -> sports health app K -> sports headphones N". In the behavioral intent differentiation map, a suppression weight is applied to the mapped interest transfer paths, reducing their priority in intent determination and dynamically narrowing the determination boundary of intent patterns. For example, the weight of the path "from smartphone browsing to sports health app browsing" is reduced by 15%, and the threshold for triggering this path's determination in subsequent behavioral sequences is correspondingly increased.

[0093] See Figure 4In stage 4 of the closed-loop optimization of cross-border e-commerce push notifications, the distribution characteristics of user feedback under different iterative optimization stages are presented in the form of stacked bar charts. The feedback types cover four categories: high-intensity positive, medium-intensity positive, medium-intensity negative, and high-intensity negative. Specifically, in the initial stage, medium-intensity negative (20) and high-intensity negative (15) feedback are the main types, while the proportion of positive feedback is relatively low. As the first to fourth iterations progress, the high-intensity positive feedback increases from the initial 5 to 30, the medium-intensity positive feedback increases from 10 to 25, while the high-intensity negative feedback gradually decreases from 15 to 3, and the medium-intensity negative feedback decreases from 20 to 8. This trend directly reflects the effect of the closed-loop optimization flow: by correcting the behavioral intent differentiation map through the feedback signal feedback, the strengthening and suppression mechanism of interest transfer path effectively improves the matching degree of push information. The proportion of positive feedback increases significantly with the number of iterations, while the proportion of negative feedback continues to be compressed, verifying the effectiveness of the dynamic interest flow network and information reorganization strategy.

[0094] Example 5: Establish cross-session user interest state continuity tracking by preserving the interest stream state at the end of the previous session. The preserving content includes the normalized interest evolution path node positions and dynamic interest stream connection states exhibited by the user at the end of the current session. At the start of the user's next session, the preserving interest stream state is loaded as the initial diffusion source point to initialize a new round of information diffusion simulation and push reorganization process. This mechanism enables long-term interest coherence push simulation, allowing the evolution of user interests to be seamlessly connected across different sessions.

[0095] In practice, a cross-session user interest state continuity tracking system is established. The interest stream state at the end of the previous session is archived. The archived content includes the normalized interest evolution network node positions and dynamic interest stream connection states exhibited by the user at the end of the current session. For example, in an afternoon session, user ID UID_1001's last interaction was a deep browsing of the product node "Wireless Noise Cancelling Headphones Pro". The dynamic interest stream network shows multiple directed connections extending from the "Wireless Noise Cancelling Headphones Pro" node. The connection weight pointing to the "Portable Headphone Case" node is 0.85, and the connection weight pointing to the "High-Resolution Audio Player" node is 0.60. After the user session ends or idle timeout, the system takes the "Wireless Noise Cancelling Headphones Pro" node as the termination node of the current interest network, and serializes it together with the aforementioned dynamic interest stream connections and their weights into an interest state snapshot, which is then bound and stored with user ID UID_1001. At the start of the next user session, the sealed interest stream state is loaded as the initial diffusion source point to initialize a new round of information diffusion simulation and push reorganization process. For example, if user ID UID_1001 starts a new session the next day, the system retrieves and loads its stored interest state snapshot, directly sets the "Wireless Noise Cancelling Headphones Pro" product node as the diffusion source point for the first information diffusion simulation in the new session, and initializes the network topology of this simulation based on the dynamic interest stream connection state sealed in the snapshot. Thus, nodes such as "Portable Headphone Case" and "High-Resolution Audio Player" become high-probability potential diffusion directions at the beginning of the session.

[0096] In some embodiments, the sealed interest stream state is not loaded exactly as is. The system fine-tunes the state based on a time decay factor to reflect the natural decay or change of interest over time. For example, if twenty hours have passed since the end of the previous session, the system applies a time decay factor to the connection weights of the loaded dynamic interest stream. All connection weights are multiplied by a decay coefficient calculated based on the time interval, thus reducing the direct impact of the old state. This mechanism realizes the simulation of long-term interest coherence push, allowing the evolution of user interests to be connected between different sessions. For example, in three independent sessions within a week, the interest evolution path of user ID UID_1001 is "smartphone -> wireless headphones", "wireless headphones -> audio player", and "audio player -> record player". The cross-session tracking mechanism ensures that the push start point of each subsequent session can connect to the end point of the previous session, thereby simulating a coherent interest evolution path across discrete sessions: "smartphone -> wireless headphones -> audio player -> record player".

[0097] Optionally, the calculation of time decay fine-tuning for the loaded archived state follows the following formula:

[0098]

[0099] in: This represents the interest stream state vector used for initialization at the start of this session, after time decay correction. This represents the original interest stream state vector loaded from storage and archived in the previous session. This vector contains the normalized interest evolution context node position encoding and related dynamic interest stream connection weight information. It is a positive-zero interest decay coefficient used to control the rate at which the interest state weakens over time. This indicates the time interval elapsed from the end of the previous session to the start of the current session. It is the base of the natural logarithm. Interest decay coefficient. It is a preset system parameter, time interval It is the actual value obtained through calculation.

[0100] It is understandable that sealing the interest stream state at the end of the previous session captures the user's immediate interest "section" when the discrete session is interrupted. Loading this state at the beginning of the next session essentially provides the user's interest evolution process with memory and continuity across inactive time periods. Using the loaded state as the initial diffusion source point allows the new round of push simulation to be built on the continuation point of the user's own historical interests without having to start from scratch or rely on a general cold start strategy.

[0101] See Figure 5 In the quantitative analysis of the product information diffusion process, the diffusion steps (positively correlated with the decay rate) were used as the horizontal axis and the penetration intensity as the vertical axis to construct the diffusion characteristic distribution of different product nodes in the interest flow network. Specifically, each data point in the figure corresponds to a different product (such as smartphones, wireless noise-canceling headphones Pro, etc.), and their horizontal and vertical axes respectively represent the decay rate correlation dimension and penetration intensity of product information diffusion in the interest network. The red curve fitted to the overall trend shows a negative correlation of "increasing diffusion steps and decreasing penetration intensity," which is consistent with the decay mechanism of information diffusion in the dynamic interest flow network: as the diffusion path lengthens (increasing the number of steps), the decay effect of information on the interest path is enhanced, leading to a decrease in penetration intensity. From the product dimension, products such as smartphones and record players are in the region with low diffusion steps and high penetration intensity, indicating that they have the characteristics of "short path and high penetration" in the interest flow network and belong to the core nodes in the interest flow; while products such as headphone cleaning kits and audio cables are in the region with high diffusion steps and low penetration intensity, corresponding to the diffusion characteristics of edge nodes in the interest flow network. This distribution result can directly support the information reorganization strategy of the push information pool: product information units with high penetration intensity and low decay rate can be prioritized to be embedded in the evolution direction of the user's current interest stream in order to improve the accuracy of push.

[0102] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0103] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for optimizing push notifications in cross-border e-commerce based on big data, characterized in that, include: The system acquires multimodal behavior trajectories generated by users during platform interactions, including serialized records of visual dwell time, interface swiping, and transaction decisions. The multimodal behavioral trajectories are decoupled from behavioral intent, separating tentative browsing and targeted search from complex interactive actions to form a behavioral intent differentiation map; A dynamic time warping mechanism is embedded in the behavioral intent differentiation map to align asynchronous behavioral sequences of different users under the same product category and generate a normalized interest evolution path. Based on the normalized interest evolution path, the topology of product information in the push network is reconstructed, and the association between product nodes is transformed from static category affiliation to dynamic interest flow connection. Through the dynamic interest flow connection, the diffusion process of product information in the user interest network is simulated, and the penetration intensity and decay rate of information on potential interest paths are calculated. Based on the penetration intensity and attenuation rate, information fragments are reorganized in an orderly manner in the push information pool, and information units with high penetration intensity are embedded in the evolution direction of the user's current interest stream. Real-time capture of user feedback signals triggered by the recombined information unit, the user feedback signals including micro changes in the interface operation trajectory and abnormal fluctuations in decision duration; The user feedback signal is fed back to the behavior intent decoupling stage to dynamically correct the judgment boundary of the behavior intent differentiation map, forming a closed-loop optimization flow from intent parsing to feedback correction. Based on the normalized interest evolution path, the topology of product information in the push network is reconstructed, and the association between product nodes is transformed from static category affiliation to dynamic interest flow connection, specifically including: Each product in the platform's product library is treated as an independent node in the network, and the initial connection is established based on the fixed category hierarchy of the product. The high-frequency intent transfer paths identified in the normalized interest evolution context are mapped to product nodes, and temporary directed connections based on interest transfer are established between product nodes that originally had no direct category association. Each temporary directed connection established based on interest transfer is assigned a weight, the weight value of which is determined by the frequency and intensity of the high-frequency intention transfer path appearing in the normalized interest evolution context. Based on the temporary directed connections and their weights, the topology connection matrix of the product network is dynamically updated, so that the association structure between products evolves from a tree-like static category tree into a graph-like dynamic interest flow network that can change over time.

2. The method for optimizing cross-border e-commerce push notifications based on big data as described in claim 1, characterized in that, The process of decoupling the multimodal behavioral trajectories from behavioral intent, separating tentative browsing from targeted searching in complex interactive actions, and forming a behavioral intent differentiation map specifically includes: Analyze the continuous operations within a single user session, identify the logical discontinuities and target continuity between operations, and mark operation sequences that are clearly guided by search terms and point to a single category as target search trajectories; Extract operation sequences from the conversation that do not contain explicit search terms and involve skipping between different product categories, calculate the frequency and span of the skips, and mark high-frequency and large-span sequences as exploratory browsing trajectories; In the targeted search trajectory and the tentative browsing trajectory, the operation intensity features are further extracted, which are quantitatively characterized by sliding speed, click intensity and page scrolling mode. Based on the quantitative differences in the operational strength characteristics, sub-intents are divided within the marked trajectory to distinguish between in-depth exploration with a strong purchase tendency and general interest exploration with a weak purchase tendency. By fusing trajectory labeling results with sub-intent segmentation results, a multi-layer network graph is constructed with users as nodes, intent types as edges, and operation intensity as weights, namely the behavioral intent differentiation graph.

3. The method for optimizing cross-border e-commerce push notifications based on big data according to claim 2, characterized in that, The process of embedding a dynamic time warping mechanism into the behavioral intent differentiation map to align asynchronous behavioral sequences of different users within the same product category and generate a normalized interest evolution profile includes: In the behavioral intent differentiation map, different user behavior sequences with the same terminating product category are selected as the normalization objects; A non-linear alignment algorithm is used to stretch or compress the time axis of the behavior sequence so that the key intent transition points in the sequence are aligned in the time dimension. The key intent transition points include the moment when the user turns from tentative browsing to targeted search. In the timeline aligned sequences, the shared intent transformation pattern is extracted, and each aligned sequence is transformed into a regular sequence consisting of intent state and dwell time. By overlaying the normalized sequences of all users targeting the same product category, high-frequency paths of intent state transitions are identified in the overlay area. These high-frequency paths constitute the core framework of the normalized interest evolution network.

4. The cross-border e-commerce push optimization method based on big data according to claim 3, characterized in that, The process of simulating the diffusion of product information in the user's interest network through the dynamic interest stream connection, and calculating the penetration intensity and attenuation rate of information on potential interest paths, specifically includes: A single product push event is abstracted as a diffusion simulation of information particles in the dynamic interest flow network; Using the product node currently interacting with by the user as the diffusion source point of the information particles, and based on the network topology and connection weight of the dynamic interest flow connection, the probability of the information particles spreading to adjacent nodes along different directed connections is calculated. In the diffusion simulation process, an attenuation factor is introduced. The attenuation factor is negatively correlated with the length of the connection path and the historical exposure of the nodes on the path. The penetration intensity of the information particle is attenuated according to the attenuation factor each time it passes through a node. Traverse the potential nodes that can be reached from the diffusion source point through several hop connections, calculate the final penetration intensity of the information particles to reach each potential node within a set number of simulation steps, and record the number of steps required to reach the node and the decay rate.

5. The cross-border e-commerce push optimization method based on big data according to claim 4, characterized in that, The step of performing ordered recombination of information fragments in the push information pool based on the penetration intensity and attenuation rate, embedding high-penetration-intensity information units into the evolution direction of the user's current interest stream, specifically includes: Set a penetration intensity threshold and a decay rate threshold, and filter out potential product nodes whose final penetration intensity is higher than the penetration intensity threshold and whose decay rate is lower than the decay rate threshold, as high-potential target push nodes. From the product information corresponding to the target push node, key information units are extracted. The key information units include the core visual feature description of the product, differentiated attribute tags, and scenario-based usage fragments. Based on the specific path of information particles from the diffusion source to each target push node, the interest transfer logic represented by the path is determined, and the extracted key information units are narratively sorted according to this logic. The sorted key information units are merged and recombined with the product information of the diffusion source to generate a coherent information push example that conforms to the evolution direction of user interest flow, thus completing the orderly recombination of information fragments.

6. The method for optimizing cross-border e-commerce push notifications based on big data according to claim 1, characterized in that, The real-time capture and reorganization of user feedback signals triggered by the information unit includes micro-changes in the interface operation trajectory and abnormal fluctuations in decision-making time, specifically including: When displaying the reorganized information push example to the user, high-precision interaction log recording is simultaneously enabled. The high-precision interaction log recording captures the user's gaze thermal change trajectory in the information area with millisecond-level precision. By analyzing the thermal change trajectory of the gaze, abnormal dwell time, rapid skip sequence, and repeated retracement pattern of the gaze on specific information units are identified. The abnormal dwell time is marked as a deep interest signal, and the rapid skip sequence is marked as an ignore signal. Simultaneously monitor the user's touch or cursor movement trajectory on the information display page, and extract the smoothness of the trajectory, the hesitation trajectory before clicking, and the acceleration characteristics when closing the page; The analysis results of the gaze thermal change trajectory are coupled with the characteristics of the touch cursor movement trajectory to cross-verify the user's true feedback intensity to the recombined information unit, generating the user feedback signal containing multi-dimensional evidence.

7. The method for optimizing cross-border e-commerce push notifications based on big data according to claim 6, characterized in that, The step of feeding the user feedback signal back to the behavior intent decoupling stage, dynamically correcting the determination boundary of the behavior intent differentiation map, and forming a closed-loop optimization flow from intent parsing to feedback correction specifically includes: The information unit corresponding to the deep attention signal in the user feedback signal is reverse-mapped to the interest transfer path on which the information unit was generated. By reinforcing the interest transfer path into the construction process of the behavioral intent differentiation map, the weight of the intent conversion pattern corresponding to the interest transfer path is increased, so that this intent pattern can be identified earlier and more definitively in similar behavioral sequences. The information units corresponding to the ignored signals in the user feedback signals are also reverse-mapped to the interest transfer paths they generate. In the behavioral intent differentiation map, suppression weights are applied to the mapped interest transfer paths to reduce the priority of the interest transfer paths in intent determination, thereby dynamically narrowing the determination boundary of the intent pattern. By cyclically executing the reinforcement and inhibition process, the judgment rules of the behavioral intention differentiation map are continuously evolved, forming a closed-loop optimization flow of intention analysis and feedback correction.

8. The method for optimizing cross-border e-commerce push notifications based on big data according to claim 1, characterized in that, The method further includes: Establish cross-session user interest state continuity tracking, seal the interest stream state at the end of the previous session, and load it as the initial diffusion source point at the beginning of the next session to realize long-term interest coherence push simulation.

9. A cross-border e-commerce push optimization system based on big data, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the big data-based cross-border e-commerce push optimization method as described in any one of claims 1 to 8.

Images