Joint marketing ecological intelligent profit sharing system and method based on interface large-scale integration

Abstract

The invention relates to the technical field of automatic processing, and discloses a joint operation ecological intelligent profit sharing system and method based on interface large-scale integration, and the method comprises the steps: firstly, registering a partner interface through a uniform meta-protocol DSL, and automatically generating a sub-adapter, thereby achieving minute-level heterogeneous interface docking; and then synchronously collecting transaction data and a user behavior video stream, extracting a structured behavior event by using technologies such as target detection and posture recognition, aligning the structured behavior event with the transaction stream in a time sequence domain, forming a preprocessed data tensor through feature coding, reasoning the tensor by a multi-modal convolutional neural network, generating a time sequence interaction popularity matrix and a benefit sharing matrix, and finally obtaining a user interaction popularity matrix and a benefit sharing matrix. And real-time display is carried out. According to the technical scheme, minute-level interface adaptation and user behavior interaction strength evaluation are realized, the combined operation and maintenance cost is remarkably reduced, and profit sharing fairness and transparency are improved.

Description

Technical Field

[0001] The present invention relates to the field of automated processing technology, and in particular to a joint venture ecological intelligent profit-sharing system and method based on interface scale integration. Background Art

[0002] With the rise of "platform-based operations" and "ecological joint venture" models, industries such as internet e-commerce, digital content distribution, financial technology, and SaaS have all connected their businesses with third-party partners through open APIs / SDKs, forming a value network that aggregates multiple suppliers and multiple consumers. Joint venture parties often judge popularity based on indicators such as interface call volume, conversion rate, GMV, or ad exposure, which places higher demands on interface adaptation efficiency, real-time data processing capabilities, and profit sharing transparency. Most existing joint venture systems use hard-coded scripts or manual configuration to connect to interfaces. Faced with problems such as heterogeneous interface protocols, inconsistent field naming, and diverse authentication mechanisms, adaptation logic often needs to be developed one by one, resulting in long launch cycles and high operation and maintenance costs. When the number of partners reaches hundreds or thousands, upgrades and iterations caused by interface changes are very frequent, and traditional solutions are difficult to support rapid scale expansion. In terms of process, the mainstream practice is to preset fixed formulas or conditional trigger scripts to complete account splitting: first, the transaction flow is pulled into an offline database, and then the profit sharing calculation is performed in batches on a daily or weekly basis. The interaction intensity does not take into account automated recognition statistics such as the number of clicks, length of stay, and visual attention, resulting in delayed profit sharing results and unable to meet the needs of some financial-level scenarios for T+0 or minute-level settlement; and it is difficult to obtain information such as user interaction intensity based on real-time image and video image processing.

[0003] In recent years, some systems have begun to introduce machine learning models, attempting to dynamically adjust parameters based on user profiles, channel quality, and other factors. However, most are still limited to structured fields and lack in-depth utilization of unstructured behavioral data such as videos and images. They are unable to accurately assess implicit value signals such as "real attention," "dwell time," or "emotional tendencies," and fail to consider the impact of video features on analysis results, particularly the impact of feature graphs on the degree of user behavioral interaction, resulting in inaccurate judgments. Furthermore, revenue results are typically stored in centralized databases, making it difficult to completely eliminate the risk of data tampering and account disputes. The demand for "trusted settlement" in cross-border or cross-industry ecosystems is becoming increasingly prominent. Overall, existing technologies still have significant deficiencies in the following areas: low efficiency in heterogeneous interface adaptation, lack of a unified meta-protocol, and lack of an automatic adapter generation mechanism; coarse and lagging profit distribution granularity, lacking a dynamic account splitting strategy for real-time transactions and multimodal behavioral data; and a lack of visual insight capabilities, making it difficult for operations personnel to intuitively grasp the contribution and risk distribution of each partner. Therefore, there is an urgent need for a new type of joint venture ecological intelligent profit-sharing technology solution that can complete heterogeneous interface docking in minutes, integrate structured transaction data and video image behavior data for multi-dimensional calculations to achieve real-time trusted account splitting and visual operation monitoring, so as to break the existing technical bottleneck, reduce docking and operation and maintenance costs, and improve automation, real-time performance and auditability. Summary of the Invention

[0004] In response to the above-mentioned problems mentioned in the prior art, the present invention discloses a joint venture ecological intelligent profit-sharing system and method based on large-scale interface integration. The method first registers the partner interface through a unified meta-protocol DSL and automatically generates a sub-adapter to achieve minute-level heterogeneous interface docking; then synchronously collects transaction data and user behavior video streams, and uses target detection, posture recognition and other technologies to extract structured behavior events, and aligns them with the transaction flow in the time domain. After feature encoding, a pre-processed data tensor is formed, and a multimodal convolutional neural network is used to infer the tensor to produce a time-series interaction heat matrix and a profit-sharing matrix. This solution realizes minute-level interface adaptation and user behavior interaction intensity evaluation, significantly reduces joint venture operation and maintenance costs, and improves profit-sharing fairness and transparency.

[0005] This application provides a joint ecological intelligent profit-sharing method based on interface scale integration, including the following steps: S1: Register partners and describe their interfaces based on the unified meta-protocol DSL. S2: Automatically generate a sub-adapter based on the DSL and complete interface docking; S3: Collect transaction data and obtain user behavior video stream through image acquisition device; S4: Time-aligning, removing anomalies, and encoding features of the transaction data with the structured user behavior events extracted from the user behavior video stream to obtain a preprocessed data tensor; S5: Input the preprocessed data tensor into a multimodal convolutional neural network, and the multimodal convolutional neural network outputs a temporal interaction heat matrix and a profit matrix; S6: Display the profit matrix and time-series interaction heat matrix in the operation cockpit based on heat maps, cluster maps and KPI dashboards.

[0006] Preferably, the element E(i,t) of the temporal interaction heat matrix represents the weighted interaction intensity of cooperation unit i in time slice t. The interaction intensity is obtained by weighted calculation of the number of clicks, stay duration, and visual attention, and is used to drive the three-dimensional heat map and model online fine-tuning of the operation cockpit.

[0007] Preferably, the structured user behavior events include: using the YOLO deep detection network to identify products, user limbs and interactive objects; using the DeepSORT algorithm to track the target across frames to generate user stay duration and behavior sequence; extracting user posture key points through OpenPose to determine interest; and obtaining structured user behavior events.

[0008] Preferably, the multimodal convolutional neural network includes parallel two-dimensional convolution branches and one-dimensional convolution time series branches, which respectively process the user behavior video stream data features and transaction data sequence features, and perform feature weighted fusion through the attention mechanism in the fusion layer.

[0009] Preferably, the two-dimensional convolution branch adopts an improved activation function after the convolution layer :

[0010] in, is the output of the convolution layer, e is the base of the natural logarithm, is the trainable first scaling factor, where 0.5≤α≤3.0; is the trainable second scaling factor, where 0≤ρ≤1.0; is the trainable third scaling factor, where 0.5≤β≤5.0; is the number of feature maps that are identified as effectively activated in all output feature maps of this convolutional layer in the current training batch. That is, for each output feature map, the average absolute value of the activation values ​​of all samples in the batch and their spatial positions is first calculated. When the average value is higher than the activation threshold set in advance or obtained by online learning, the feature map is determined to be in an effectively activated state. The feature maps that meet this condition are counted and summed one by one to obtain u. The maximum number of network channels that is preset.

[0011] This application also provides a joint ecological intelligent profit-sharing system based on interface scale integration, including: Partner registration and interface description module, which registers partners and describes partner interfaces based on the unified meta-protocol DSL; A sub-adapter generation and interface docking module automatically generates a sub-adapter according to the DSL and completes the interface docking; Transaction data collection and user behavior video capture module, which collects transaction data and obtains user behavior video streams through image acquisition devices; A data alignment and feature coding preprocessing module performs time alignment, anomaly elimination, and feature coding on the transaction data and the structured user behavior events extracted from the user behavior video stream to obtain a preprocessed data tensor; A multimodal convolutional neural network inference module inputs the preprocessed data tensor into the multimodal convolutional neural network, which outputs a temporal interaction heat matrix and a profit matrix; The operation cockpit visualization module displays the profit distribution matrix and time-series interaction heat matrix based on heat maps, cluster maps and KPI dashboards.

[0012] Preferably, the element E(i,t) of the temporal interaction heat matrix represents the weighted interaction intensity of cooperation unit i in time slice t. The interaction intensity is obtained by weighted calculation of the number of clicks, stay duration, and visual attention, and is used to drive the three-dimensional heat map and model online fine-tuning of the operation cockpit.

[0013] Preferably, the structured user behavior events include: using the YOLO deep detection network to identify products, user limbs and interactive objects; using the DeepSORT algorithm to track the target across frames to generate user stay duration and behavior sequence; extracting user posture key points through OpenPose to determine interest; and obtaining structured user behavior events.

[0014] Preferably, the multimodal convolutional neural network includes parallel two-dimensional convolution branches and one-dimensional convolution time series branches, which respectively process the user behavior video stream data features and transaction data sequence features, and perform feature weighted fusion through the attention mechanism in the fusion layer.

[0015] Preferably, the two-dimensional convolution branch adopts an improved activation function after the convolution layer :

[0016] in, is the output of the convolution layer, e is the base of the natural logarithm, is the trainable first scaling factor, where 0.5≤α≤3.0; is the trainable second scaling factor, where 0≤ρ≤1.0; is the trainable third scaling factor, where 0.5≤β≤5.0; is the number of feature maps that are identified as effectively activated in all output feature maps of this convolutional layer in the current training batch. That is, for each output feature map, the average absolute value of the activation values ​​of all samples in the batch and their spatial positions is first calculated. When the average value is higher than the activation threshold set in advance or obtained by online learning, the feature map is determined to be in an effectively activated state. The feature maps that meet this condition are counted and summed one by one to obtain u. The maximum number of network channels that is preset.

[0017] The present invention provides a joint ecological intelligent profit-sharing system and method based on large-scale interface integration, which can achieve the following beneficial technical effects:

[0018] This invention utilizes a unified meta-protocol (DSL), automated sub-adapter generation, and a multimodal convolutional neural network (including improved activation functions) to achieve minute-by-minute interface adaptation. By collaborating with a code generator and DSL descriptions, the integration time for heterogeneous APIs / SDKs has been reduced from the traditional 1-2 weeks to under 30 minutes, significantly reducing the onboarding threshold and labor costs for multi-party collaborations. The refined value assessment of behavior-transaction fusion aligns video / image behavior data (clicks, dwell times, visual attention, gestures / emotions, etc.) with transaction flows in the time domain for the first time, feeding both data into a multimodal CNN and outputting a time-series interaction heat matrix and a profit sharing matrix. This significantly improves the real-time, accuracy, interpretability, and security of interactions, breaking through bottlenecks in existing technologies.

[0019] 2. The present invention adopts an improved activation function to ensure the stability of deep and wide networks. The two-dimensional branch of the present invention adopts an improved activation function. In the current training batch, the number of feature maps identified as effectively activated in all output feature maps of this convolutional layer is added to the convolutional network calculation process, and combined with dynamic effective channel counting, the gradient variance is greatly reduced and the weak texture recall rate is greatly improved, providing numerical stability for training and inference for large-scale parallel reasoning.

[0020] 3. The present invention adopts visual operational insights. The system uses three-dimensional heat maps, cluster maps and KPI dashboards to display the benefit-risk distribution of partners in time-geography dimensions in real time. It supports one-click sandbox simulation and A / B testing, and the operational decision-making efficiency is improved by more than 40%. High scalability and low operation and maintenance, the message queue + containerized architecture supports the concurrent access of thousands of partners and tens of billions of transaction throughput; the activation function and model parameters can be fine-tuned online to reduce the retraining and reconfiguration costs caused by the surge in business. The unified DSL, model library and visualization components can be quickly copied to e-commerce, live broadcast, digital content, financial account sharing and other industry scenarios, providing a standardized and low-cost technical foundation for the expansion of the enterprise ecosystem. It improves the system's real-time performance, accuracy, security and reliability in multiple dimensions, breaks through the existing technical bottlenecks, and has obvious technological advancement. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0022] Figure 1 This is a flowchart of the steps of the joint ecological intelligent profit-sharing method based on interface scale integration of the present invention;

[0023] Figure 2 This is a schematic diagram of the joint ecological intelligent profit-sharing system based on interface large-scale integration of the present invention. DETAILED DESCRIPTION

[0024] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts shall fall within the scope of protection of the present invention.

[0025] Example 1:

[0026] In order to solve the above-mentioned technical problems mentioned in the prior art, the following Figure 1 As shown: This application provides a joint ecological intelligent profit sharing method based on interface scale integration, including the following steps:

[0027] S1: Register partners and describe their interfaces based on the unified meta-protocol DSL.

[0028] S2: Automatically generate a sub-adapter based on the DSL and complete interface docking;

[0029] In one embodiment, the alliance platform first deploys an interface registry and code generation service on a cloud server. Third-party partners open the registry page in a browser, enter basic information such as partner ID, interface root address, authentication method, and traffic threshold, and submit it. The system automatically assigns the partner a pair of access keys and persists all configurations to the database. Simultaneously, a directory named after the partner ID is created in the object storage to store the interface description file. Subsequently, the partner edits a text file written in the unified meta-protocol (DSL) online in the registry. This text describes the interface name, access path, request method, field type, authentication method, retry policy, and other information in a human-readable indented format. After editing, the user clicks "Upload." The registry immediately saves the text file to the object storage and pushes a "New DSL Online" event to the message queue, which contains the partner ID and DSL file address. After subscribing to this event, the code generation service downloads the DSL file and invokes a built-in parser to generate an abstract syntax tree. Based on the syntax tree, it then outputs a set of standardized source code. This source code implements functions such as external request encapsulation, token refresh, flow control, and logging, and is compatible with the alliance platform's internal unified messaging protocol. After the generation is completed, the service automatically compiles the executable file, encapsulates it into a container image, and pushes it to the private image repository. After the continuous delivery pipeline detects the new image tag, it immediately creates a deployment named adapter-partner identity in the Kubernetes cluster, and starts two copies by default. After the new container is started, it automatically reads the access key, applies for a token from the partner interface, and caches it in memory. It then exposes a unified internal endpoint, receives business data forwarded by the platform's internal services, and sends standardized requests to the partner interface in real time. After the monitoring system confirms that the adapter has responded normally through a health check, it updates the partner status to activated. From then on, business events pushed from the partner to the alliance platform can flow into the subsequent data processing and profit-sharing links without obstruction. The entire process, from the partner submitting the DSL to the adapter going online and starting to forward interface traffic, usually takes no more than thirty minutes, effectively solving pain points such as heterogeneous interfaces, slow manual docking, and difficult upgrades.

[0030] S3: Collect transaction data and obtain user behavior video streams through image acquisition devices; in some embodiments, the alliance platform deploys a dual-channel collection mechanism in the client and offline scenarios. In the online part, the platform embeds the tracking SDK in key nodes such as the payment success page, order details page, and advertisement click entrance of the web and mobile applications. Whenever a user completes an order, views an order, or clicks an advertisement, the SDK generates a transaction event containing fields such as order number, product identification, amount, payment channel, terminal type, geographic location, timestamp, etc. at the millisecond level, and immediately sends it to the platform's event gateway via HTTPS. The event gateway uses a backpressure queue to ensure that there is no packet loss in high concurrency, and then writes the original JSON event to the Kafka topic for subsequent cleaning tasks.

[0031] For the offline part, the platform has installed an integrated ultra-high-definition camera above the core experience area. The camera is equipped with an edge computing unit that can run target detection and facial key point models locally in real time. The camera first performs human body detection and multi-target tracking on users entering the screen to generate a unique track number; then, through posture recognition, it estimates the user's viewing angle and duration of stay and captures key frames. If the duration of the track exceeds the set threshold, the edge unit will generate a user behavior segment that contains structured fields such as camera number, track number, entry time, exit time, average gaze direction, and expression confidence. To protect privacy, the edge unit only uploads the extracted numerical features and does not upload the original facial image.

[0032] User behavior snippets collected offline are sent in batches every three seconds to the same event gateway, which writes them to another Kafka topic based on the "offline behavior" tag. When the platform's stream processing engine consumes online transaction events and offline behavior snippets, it uses a unified ISO timestamp field as a benchmark for window alignment and makes a preliminary association based on the geographic location matching strategy between the camera and the order terminal. The aligned data will be written to the real-time feature warehouse to provide input for subsequent feature encoding and multimodal convolutional neural network inference. In this way, the system can simultaneously obtain accurate transaction flows and rich user video behavior data, ensuring that the subsequent profit-sharing model takes into account both economic contribution and real interaction popularity.

[0033] S4: Time-align, anomaly elimination, and feature encoding are performed on the transaction data and the structured user behavior events extracted from the user behavior video stream to obtain a preprocessed data tensor. In one embodiment, the system orchestrates a job named "align-encode-stream" in the stream processing engine Flink to connect to the two Kafka topics obtained in the previous step: the online transaction event topic and the offline behavior fragment topic.

[0034] To align time, the stream job joins two event streams using a five-second sliding window. The window key consists of the user ID, device ID, and the most recent IP-GPS location. When the transaction timestamp and the entry time of the behavior segment differ by no more than two seconds and their geographic locations match, the job merges them into a single "aligned event." If a behavior segment cannot find a corresponding transaction, it is written to the candidate cache and awaits the next window. If no match is found after 30 seconds, it is marked as an "orphan behavior event" for subsequent risk control analysis.

[0035] To eliminate anomalies, the stream job immediately executes three types of rules after merging: First, the value out-of-bounds rule. Transactions with negative amounts or exceeding the daily maximum limit are directly eliminated; behavior segments with a dwell time of more than 30 minutes are considered anomalies. Second, the duplicate detection rule. If the same order number appears twice within one minute with the same amount, it is considered a duplicate push, and only the first one is retained. Third, the probabilistic model rule. The system trains an extreme value density estimation model offline to jointly score dwell time, sentiment confidence, and number of clicks. Events below the threshold are marked as anomalies and written to the bypass Kafka for manual review.

[0036] For feature encoding, aligned events that have passed anomaly filtering are first fed into the encoder. The encoder performs the following field transformations: Continuous fields such as amount and duration are first log-compressed and then normalized to the training set mean-variance. Discrete fields such as payment channel and content category are mapped to fixed-length integer indices, which are then converted into eight-dimensional trainable embedding vectors using a table lookup. The timestamp field is split into two periodic features, representing the hour and the day of the week. The camera pose angle is quantized into three discrete labels and similarly converted into embeddings using a table lookup. All embedding vectors are concatenated with normalized real numbers to form a fixed 128-dimensional feature vector. The stream job uses a rolling step of 3 seconds, stacking feature vectors belonging to the same user and partner in chronological order into a sequence of length 20. This results in a pre-processed data tensor of the form "batch size × 20 × 128" and is written to the Redis stream cache, where it is subsequently read in real time by the multimodal convolutional neural network for inference. The end-to-end latency of the entire alignment, culling, and encoding chain is kept below 300 milliseconds, meeting the real-time requirements of the minute-level account splitting model.

[0037] S5: The preprocessed data tensor is input to a multimodal convolutional neural network, which outputs a temporal interaction heat matrix and a profit matrix. In some embodiments, the system divides the preprocessed data obtained from the previous process into two tensors: one is the video behavior branch input, with a shape of "batch size × 3 frames × 128 × 128," representing the keyframe feature maps extracted from the last three seconds of each user; the other is the transaction-behavior sequence branch input, with a shape of "batch size × 20 × 128," representing 20 feature vectors aligned at the millisecond level within the same time window. In the video behavior 2D convolution branch, the input feature maps of the three frames are first passed through a set of convolutional layers with a kernel size of 3×3 and 64 channels, followed by a batch normalization layer and a BA-Swish-Mish activation. A further 128-channel convolutional layer with the same configuration is then stacked. Dilated convolution is then used to expand the receptive field, and a self-attention gating module is introduced to perform cross-frame fusion of salient regions across the three frames. Finally, global average pooling is used to compress the spatial dimensions to a vector of length 512. In the transaction-behavior one-dimensional convolutional temporal branch, the sequence tensor first enters a one-dimensional convolutional layer with 64 kernels, a stride of 1, and a TG-AdaGELU activation function. This is followed by a 128-kernel convolution of the same structure, followed by a causal residual block to maintain temporal order. The convolutional output undergoes temporal max pooling to produce a vector of length 512. For feature fusion, the visual and temporal vectors are concatenated to form a 1024-dimensional fusion vector, which is then fed into a cross-modal attention layer. The attention weights are determined by content-adaptive gating to ensure that the visual or transaction signal receives the appropriate weight in different scenarios. The fusion result then passes through a fully connected layer with dropout, with the output size remaining at 1024. The dual output head, including the fusion vector, feeds two linear output heads in parallel. The heat header uses softmax to map the 1024-dimensional matrix to the dimension of "number of cooperative units × 20" and rearranges it into a time-series interaction heat matrix. The profit header uses sigmoid to map the 1024-dimensional matrix to the dimension of "number of cooperative units multiplied by number of revenue channels" and rearranges it into a profit matrix. During the inference phase, the model simultaneously generates two matrices for each batch, representing the real-time interaction intensity and the revenue splitting weight.

[0038] The training strategy uses the mean squared error (MSE) of the heat head with the annotated interaction duration and the Huber loss with the true on-chain transaction value for the profit head. The weighted sum of the two is used as the total loss. The AdamW optimizer is selected, with a base learning rate of 0.3% and a weight decay of 0.1%. If the heat error decreases while the profit error stagnates in the validation set, the weights of both factors are dynamically adjusted. Full-network training takes 300 milliseconds per iteration on two 32GB graphics cards, and the latency for batches of 8,000 events in inference is kept below 90 milliseconds.

[0039] The deployment and invocation model weights are saved as ONNX files, compiled with TensorRT and loaded into the containerized inference service. The stream processing engine writes the preprocessed tensors to gRPC, and the inference results are returned in Protocol-buffer format and sent to the subsequent links. The profit matrix is ​​written into the alliance chain smart contract, and the heat matrix is ​​pushed to the operation cockpit in real time for three-dimensional heat map rendering. Through this embodiment, the system is able to stably integrate two types of heterogeneous data, video and transaction, in a deep and wide network, and simultaneously output the time-series interactive heat matrix and profit matrix within a delay of hundreds of milliseconds, providing core computing power support for real-time account splitting and visual operations.

[0040] S6: Display the profit matrix and time series interaction heat matrix in the operations cockpit based on heat maps, clustering diagrams, and KPI dashboards. In some embodiments, after the system completes inference in the stream processing engine, it writes the latest batch of time series interaction heat matrices and profit matrix to the time series database. Next, the front-end visualization service subscribes to the "matrix ready" event via WebSocket, pulls the corresponding batch of data, and performs the following processing:

[0041] The first step is to generate a two-dimensional heat map. The visualization backend expands the temporal interaction heat matrix with collaboration units as rows and time slices as columns. It first calculates the maximum and minimum values ​​of each row, and then uses linear interpolation to normalize all elements to between 0 and 1. The front end uses WebGL to create a canvas and map the matrix into a pixel grid: the horizontal axis is time, the vertical axis is collaboration units, and the pixel color is determined by the normalized value, with a color scale ranging from dark blue to dark red. When the user hovers over the heat map, the front end displays the specific interaction index of the collaboration unit at that point in time and links to the original video clip for review.

[0042] The second step is to generate a profit distribution cluster diagram. The visualization backend first normalizes the profit matrix columns to obtain the proportion vector of each cooperative unit in different income channels. Then use KMeans to cluster the cooperative units into three categories: high contribution, potential, and low activity. The clustering results are returned to the front end through REST. The front end uses ECharts to draw the three types of data into a scatter cloud. The horizontal axis of each point is the first principal component in the profit share, and the vertical axis is the second principal component. The size of the point is proportional to the total profit amount, and the color represents the category. Click on the scatter point to jump to the detailed page of the corresponding partner.

[0043] The third step is to build a KPI dashboard. The visualization backend calculates five metrics in real time according to business definitions: total commission amount, average engagement popularity, average risk adjustment coefficient, number of abnormal intercepts, and commission forecast error. The frontend uses Vue components to render five gauges, along with historical trend lines. If any metric exceeds a threshold, the gauge automatically changes color and an alert pops up.

[0044] The fourth step is linkage operation. When the operator selects a time period in the heat map and selects multiple cooperation units, the front-end immediately initiates a request to the back-end to re-cluster the selected sub-matrix and update the scatter plot; the dashboard synchronously refreshes the KPI of the subset, allowing operations to quickly compare the effects of different profit-sharing strategies. The total delay of the entire visualization link from inference output to front-end rendering is kept within one second, supporting refresh once per second without noticeable lag. This embodiment enables the operation team to have an intuitive insight into "who contributed how much heat and how much revenue when", and can complete strategy simulation and effect evaluation on the same interface, greatly improving revenue management efficiency and decision-making accuracy.

[0045] In some embodiments, the element E(i,t) of the temporal interaction heat matrix represents the weighted interaction intensity of cooperative unit i in time slice t. The interaction intensity is obtained by weighted calculation of the number of clicks, dwell time, and visual attention, and is used to drive the three-dimensional heat map and online fine-tuning of the model in the operation cockpit. In some embodiments, the system organizes the three types of raw indicators of the number of clicks, dwell time, and visual attention within a 5-second sliding window of each cooperative unit into a time series of length 20. The following describes how CNN fuses these three pieces of information into the temporal interaction heat matrix E in end-to-end training. For the same cooperative unit, the click sequence, dwell sequence, and visual sequence collected in the current batch are first normalized according to the historical mean variance of the entire network, and then stacked into a three-channel tensor; the tensor shape is "batch size × 3 channels × 20 time steps". Local temporal convolution, the tensor is fed into a one-dimensional convolution layer with a convolution kernel length of 3, 16 channels, and a stride of 1. This layer is equivalent to performing a weighted sliding calculation of clicks, dwells, and visuals over the last three time steps. However, the weights are automatically learned by the convolution kernel parameters, eliminating the need for manual specification of w1, w2, and w3. The activation function uses TG-AdaGELU to ensure sensitivity to spikes. Cross-metric attention fusion is implemented, and the convolution output enters a lightweight channel attention module. The system first performs global average pooling on the 16 feature maps to produce a vector of length 16. This vector passes through two fully connected layers and a sigmoid function to generate 16 coefficients between 0 and 1, which are then multiplied back to the original feature map to dynamically weight different combinations of "click," "dwell," and "visual" patterns. The dimensionality reduction and normalization of the weighted feature map is then passed through a bottleneck layer with a one-dimensional convolution kernel of one, compressing the 16 channels into a single-channel sequence. A softmax is then applied to ensure that the outputs of different collaborating units within the same time step sum to 1. The single-channel value at each time step is now E(i, t).

[0046] To ensure that E reflects user engagement, the system constructs pseudo-labels: first, the complete viewing time recorded by the player backend, and second, the "hot time period" labels manually assessed in historical marketing experiments. The network uses a combined mean squared error and Focal Loss optimization. When there is a mismatch between predicted popularity and labeled hot spots, the gradient automatically adjusts the convolution kernel and attention weights to achieve the optimal balance between click, dwell, and visual indicators in different scenarios.

[0047] Inference and implementation: During the inference phase, the stream processing engine sends the three-channel tensor into the TensorRT container; the container outputs the E matrix and immediately writes it back to the time series database. After the front-end reads E, it renders each column into a color bar through WebGL. The number of columns corresponds to the time step, the number of rows corresponds to the cooperative unit, and the color depth represents the predicted interaction intensity within the five-second sliding window. Operators can click on any grid to trace back the original click log and video frame to verify the model's interpretability. In offline A / B evaluation, when only click and dwell inputs were used, the accuracy of hotspot detection was 82.7%; after introducing visual attention and using the above-mentioned convolutional-attention structure, the accuracy increased to 85.1%, proving that the weighted fusion of the three indicators can better capture the true interaction heat.

[0048] In some embodiments, the structured user behavior events include: using the YOLO deep detection network to identify products, user body parts, and interactive objects; using the DeepSORT algorithm to track targets across frames to generate user dwell time and behavior sequences; and extracting user pose keypoints using OpenPose to determine interest, thereby obtaining structured user behavior events. In some embodiments, the system sequentially performs YOLO detection, DeepSORT tracking, and OpenPose pose estimation on a video stream captured at frames per second at the edge node. First, the YOLO detection network, loaded with weights fine-tuned using industry data, performs forward inference on each frame, outputting bounding boxes and confidence scores for categories such as products, people, and gestures. Box fields include top-left corner coordinates, width and height, category ID, and probability value. DeepSORT then reads the detection results from the same frame, embeds appearance features into a 128-dimensional vector using a lightweight CNN, and combines this with motion prediction using a Kalman filter to match detections across frames and assign unique track IDs. The system maintains a buffer for each track, accumulating the number of consecutive frames and total duration. A track is considered complete when the detection box with the same ID is interrupted for more than ten frames in the frame sequence. While the trajectory remains active, OpenPose performs pose estimation on the bounding box within the same frame, outputting the 2D coordinates and confidence scores of 18 key points. The system calculates gaze direction based on the coordinates of the nose tip and eyes, and then determines standing or squatting posture based on the ratio of shoulder to hip coordinates. If the user is facing the screen and their gaze is focused, it is considered to be in a high-interest posture. For all frames within the same trajectory, the system calculates the average gaze direction, the percentage of high-interest frames within a second, and the average overall expression confidence. When a trajectory ends, the system invokes an event generator to convert the information summarized in the trajectory buffer into a structured user behavior event. Event fields include: camera number, trajectory ID, entry time, exit time, dwell time, number of dwell frames, average gaze angle, percentage of high-interest frames, a list of detected product categories, and the highest-confidence pose category at the end. The event is written to the local log as a JSON string and uploaded to the central Kafka topic "user.behavior.structured" via MQTT. The upstream streaming engine then aligns the event with the transaction flow by timestamp for subsequent feature encoding and model inference.

[0049] In some embodiments, the multimodal convolutional neural network includes parallel two-dimensional convolutional branches and one-dimensional convolutional time series branches, processing features of user behavior video stream data and transaction data sequence data, respectively. These branches perform weighted feature fusion using an attention mechanism at the fusion layer. In some embodiments, the system splits each batch of data into two tensor streams. The first stream is from the user behavior video stream: three frames are extracted from the last three seconds of video and compressed into three-channel 128×128 feature maps using an edge model. The batch dimension results in a tensor shape of batch size multiplied by 3×128×128. The second stream is from the transaction feature sequence: 20 128-dimensional feature vectors are accumulated within the same time window, resulting in a tensor shape of batch size × 20×128. The video branch uses a two-dimensional convolutional backbone. The input first passes through a convolutional layer with a 3×3 kernel size and 64 output channels, followed by a batch normalization layer and BA-Swish-Mish activation. This is followed by a stacked 3×3 convolution layer with 128 channels.

[0050] To expand the receptive field, a dilated convolution with a dilation rate of 2 and 256 output channels is introduced in the third block. A self-attention gating module is then inserted: the spatial features are convolved one-by-one to obtain the query, key, and value. Attention weights are calculated using scaled dot products, multiplied with the value features, and then added back to the main branch. Global average pooling is then used to compress the spatial dimensions, generating a visual vector of length 512. The transaction branch uses a one-dimensional convolutional temporal architecture. The input sequence first enters a one-dimensional convolutional layer with a kernel length of 3, 64 output channels, and a TG-AdaGELU activation. This is followed by a dilated convolution layer with a dilation rate of 2 and 128 output channels to capture temporal dependencies greater than 10 steps. Both convolutional layers are followed by batch normalization and residual skip connections, and finally by temporal max pooling, compressing the 20-length sequence into a 512-length vector. These two vectors are concatenated into a 1024-dimensional feature vector in a fusion layer and fed into cross-modal attention.

[0051] The implementation method is as follows: the concatenated vector is split in half, with the first 512 dimensions used as the query and the second 512 dimensions copied as both the key and the value. A scaled dot product is used to obtain the attention coefficient, which is then weightedly summed with the value vector. The residual is then added to the original concatenated vector and passed through LayerNorm, resulting in a 1024-dimensional fused vector. The fused vector then enters a fully connected layer with a dropout ratio of 30%, maintaining its 1024-dimensional dimensions, and is then split into two parallel output heads. The heat head is mapped to a dimension of 20 times the number of cooperative units through a linear layer and reshaped into a temporal interaction heat matrix. The profit head is mapped to a dimension of the number of cooperative units times the number of revenue channels through a linear layer and reshaped into a profit matrix. Softmax normalization is used for the heat head, and sigmoid normalization is used for the profit head. During training, the model supervises the heat head with mean squared error and the profit head with Huber loss. The AdamW optimizer has a learning rate of 0.3% and weight decay of 0.1%. During deployment, the model is exported to ONNX and then compiled to an FP16 engine using TensorRT. The latency for inference of 8,000 batches is approximately 90 milliseconds. The entire network can simultaneously obtain a time-series interaction heat matrix and a profit distribution matrix within a hundred-millisecond latency, meeting the performance requirements for real-time account splitting and operational visualization.

[0052] In some embodiments, the 2D convolution branch uses an improved activation function after the convolution layer. :

[0053]

[0054] in, is the output of the convolution layer, e is the base of the natural logarithm, is the trainable first scaling factor, where 0.5≤α≤3.0; is the trainable second scaling factor, where 0≤ρ≤1.0; is the trainable third scaling factor, where 0.5≤β≤5.0; is the number of feature maps that are identified as effectively activated in all output feature maps of this convolutional layer in the current training batch. That is, for each output feature map, the average absolute value of the activation values ​​of all samples in the batch and their spatial positions is first calculated. When the average value is higher than the activation threshold set in advance or obtained by online learning, the feature map is determined to be in an effectively activated state. The feature maps that meet this condition are counted and summed one by one to obtain u. is the maximum number of channels of the network set in advance. In some embodiments, in the PyTorch-1.13 environment, the improved activation function proposed in this article is integrated into the visual convolution branch. First, a BA_SwishMish class is customized in the code, and the trainable parameters alpha, rho, and beta are defined in the constructor. The initial values ​​are set to 0, 0.1, and 0, respectively, and reparameterized through Softplus to ensure that they are always positive. The number of channels in the widest layer of the network, 1024, is written into the constant Cmax. During forward calculation, the activation function traverses the output feature maps of the current batch; for each feature map, the batch dimension and the spatial dimension are first expanded and the average absolute value is calculated. If the average value is higher than the dynamic threshold of 0.05, the feature map is recorded as "valid activation". After counting an entire batch, the total number of valid activation channels u is obtained. Then, according to the formula logic, the convolution layer output, the three trainable scaling factors, and u / Cmax are combined to generate a new activation result. In the training configuration, the AdamW optimizer is used, and the learning rate is 3× , weight decay of 0.1%, and L2 regularization added to alpha, rho, and beta to prevent them from exceeding the specified upper limits. A 48-hour control experiment was conducted on the ImageNet-100 classification subset: the baseline network used conventional SiLU activations, while the improved network replaced only the 2D branch activation with a customized version, while the rest of the structure remained unchanged. The results showed that the improved version exhibited significantly smoother loss curves in the first ten epochs; the final Top-1 accuracy improved from 84.2 to 85.1, and the number of gradient explosion alarms decreased from 8 to 1, demonstrating significant improvement in numerical stability. Grad-CAM heatmaps on the validation set also revealed that the improved network focused on a more complete target region on low-contrast samples, while the baseline network was susceptible to background noise. During inference, the trained weights were exported to ONNX and then compiled to an FP-16 precision engine using TensorRT. At a batch size of 64, inference time increased by approximately 2% compared to the baseline, which was within acceptable operational and maintenance limits. Given the improvements in accuracy and stability, the network was ultimately launched at a 30% grayscale and expanded to full scale two weeks later. Online monitoring statistics show that the average profit sharing error in weak texture scenes is reduced by 0.4%, further proving that this embodiment can bring quantifiable benefits in production.

[0055] In some embodiments, the present application also provides a joint ecological intelligent profit-sharing system based on interface scale integration, such as Figure 2 As shown, the system consists of three layers: cloud core cluster, edge collection nodes and operation terminals. They are interconnected through enterprise-level optical fiber or public network VPN dedicated lines. The overall network topology converges to the cloud core switch in a star shape.

[0056] In terms of the cloud core cluster, three physical server clusters are deployed in the same computer room cabinet: the first cluster is the interface registration and code generation cluster, consisting of three dual-core Xeon servers with 64-core CPUs, 512GB of memory, and two 10G network cards. It is used to run the DSL registration center, CICD pipeline, and container image repository. The second cluster is the inference computing cluster, equipped with 8 single-node A100 GPU servers, each containing 8 40GB GPUs, dual-core AMD EPYC processors, and a default of 1TB of memory, carrying multimodal convolutional neural network inference services. The third cluster is the data and blockchain cluster, including 3 NVMe AllFlash storage servers running the time series database and feature cache, and 2 independent high-performance servers arranging the Fabric blockchain alliance node and CA service. All three clusters of servers are connected to the intranet through the core 10G switch.

[0057] As for edge collection nodes, each offline partner store or display area is equipped with an integrated smart camera and a miniPC. The camera, equipped with a Sony 4K sensor and Cambricon edge AI chip, is responsible for real-time inference of YOLO detection and OpenPose pose estimation. The extracted structured features are then transmitted to the local miniPC via a PoE switch. The miniPC uses an 8-core ARM processor, 16GB of RAM, and a 256GB SSD, running a Dockerized edge event aggregation program. After aggregation, data is directly uploaded to the cloud event gateway via an enterprise dedicated line / MPLS VPN.

[0058] Regarding the operational cockpit terminals, the headquarters dispatch center features a hyperbolic, ultra-wide 8K LED wall, with two RTXA6000 workstations behind it serving as visualization rendering nodes. Six i7-based office PCs serve as consoles for operators, communicating with the LED wall workstations via a switch within a Layer 2 network. The cockpit terminals subscribe to cloud-based inference output via WebSocket, with latency under 100 milliseconds.

[0059] Regarding communication and security, all edge nodes access the cloud via a public VPN. The VPN terminal gateway is located above the core switch and supports active-active redundancy. The core switch connects to the firewall and then to the internet border router. Blockchain nodes are interconnected through independent TLS tunnels to avoid interoperability with production data networks. Inference clusters and data clusters are interconnected via RDMA, improving the efficiency of GPU access to the time series database. Only gRPC and Docker Registry ports are open between the interface registration cluster and the inference cluster, complying with the zero-trust access policy.

[0060] Regarding hardware interaction during the operational process, after the local AI chip detects video clips captured by the edge camera, extracted elements such as product frames, face frames, and gesture key points are sent via Ethernet to the miniPC. The miniPC then packages and uploads these data once per second to the inference cluster, which hosts the cloud-based Kafka cluster. The inference cluster server reads the transaction flow and behavior tensors from the same batch, performs multimodal CNN inference on the GPU, writes the resulting profit matrix into the blockchain cluster's smart contract, and pushes the interaction heat matrix to the LED wall rendering node via the time series database. Operations personnel can view the 3D heat map in real time via a browser on their desktop PC and issue policy adjustment commands. These commands are routed to the DSL service through the interface registration cluster's load balancer, which then invokes CICD to automatically regenerate or hot-update the adapter image before returning to the inference cluster for a new round of online model fine-tuning. Through this hardware and connectivity design, the system ensures throughput of tens of thousands of concurrent video streams and hundreds of millions of transactions while achieving an end-to-end, second-level data loop, trusted on-chain ledger resolution, and visual decision support in the operations cockpit.

[0061] The system includes: a partner registration and interface description module, which registers partners and describes partner interfaces based on the unified meta-protocol DSL;

[0062] A sub-adapter generation and interface docking module automatically generates a sub-adapter according to the DSL and completes the interface docking;

[0063] Transaction data collection and user behavior video capture module, which collects transaction data and obtains user behavior video streams through image acquisition devices;

[0064] A data alignment and feature coding preprocessing module performs time alignment, anomaly elimination, and feature coding on the transaction data and the structured user behavior events extracted from the user behavior video stream to obtain a preprocessed data tensor;

[0065] A multimodal convolutional neural network inference module inputs the preprocessed data tensor into the multimodal convolutional neural network, which outputs a temporal interaction heat matrix and a profit matrix;

[0066] The operation cockpit visualization module displays the profit distribution matrix and time-series interaction heat matrix based on heat maps, cluster maps and KPI dashboards.

[0067] In some embodiments, the element E(i,t) of the temporal interaction heat matrix represents the weighted interaction intensity of the cooperation unit i in time slice t. The interaction intensity is obtained by weighted calculation of the number of clicks, stay duration, and visual attention, and is used to drive the three-dimensional heat map and model online fine-tuning of the operation cockpit.

[0068] In some embodiments, the structured user behavior events include: using the YOLO deep detection network to identify products, user limbs and interactive objects; using the DeepSORT algorithm to track the target across frames to generate user stay duration and behavior sequence; extracting user posture key points through OpenPose to determine interest; and obtaining structured user behavior events.

[0069] In some embodiments, the multimodal convolutional neural network includes parallel two-dimensional convolution branches and one-dimensional convolution time series branches, which respectively process user behavior video stream data features and transaction data sequence features, and perform feature weighted fusion through an attention mechanism in the fusion layer.

[0070] In some embodiments, the 2D convolution branch uses an improved activation function after the convolution layer. :

[0071]

[0072] in, is the output of the convolution layer, e is the base of the natural logarithm, is the trainable first scaling factor, where 0.5≤α≤3.0; is the trainable second scaling factor, where 0≤ρ≤1.0; is the trainable third scaling factor, where 0.5≤β≤5.0; is the number of feature maps that are identified as effectively activated in all output feature maps of this convolutional layer in the current training batch. That is, for each output feature map, the average absolute value of the activation values ​​of all samples in the batch and their spatial positions is first calculated. When the average value is higher than the activation threshold set in advance or obtained by online learning, the feature map is determined to be in an effectively activated state. The feature maps that meet this condition are counted and summed one by one to obtain u. The maximum number of network channels that is preset.

[0073] The present invention provides a joint ecological intelligent profit-sharing system and method based on large-scale interface integration, which can achieve the following beneficial technical effects:

[0074] This invention utilizes a unified meta-protocol (DSL), automated sub-adapter generation, and a multimodal convolutional neural network (including improved activation functions) to achieve minute-by-minute interface adaptation. By collaborating with a code generator and DSL descriptions, the integration time for heterogeneous APIs / SDKs has been reduced from the traditional 1-2 weeks to under 30 minutes, significantly reducing the onboarding threshold and labor costs for multi-party collaborations. The refined value assessment of behavior-transaction fusion aligns video / image behavior data (clicks, dwell times, visual attention, gestures / emotions, etc.) with transaction flows in the time domain for the first time, feeding both data into a multimodal CNN and outputting a time-series interaction heat matrix and a profit sharing matrix. This significantly improves the real-time, accuracy, interpretability, and security of interactions, breaking through bottlenecks in existing technologies.

[0075] 2. The present invention adopts an improved activation function to ensure the stability of deep and wide networks. The two-dimensional branch of the present invention adopts an improved activation function. In the current training batch, the number of feature maps identified as effectively activated in all output feature maps of this convolutional layer is added to the convolutional network calculation process, and combined with dynamic effective channel counting, the gradient variance is greatly reduced and the weak texture recall rate is greatly improved, providing numerical stability for training and inference for large-scale parallel reasoning.

[0076] 3. The present invention adopts visual operational insights. The system uses three-dimensional heat maps, cluster maps and KPI dashboards to display the benefit-risk distribution of partners in time-geography dimensions in real time. It supports one-click sandbox simulation and A / B testing, and the operational decision-making efficiency is improved by more than 40%. High scalability and low operation and maintenance, the message queue + containerized architecture supports the concurrent access of thousands of partners and tens of billions of transaction throughput; the activation function and model parameters can be fine-tuned online to reduce the retraining and reconfiguration costs caused by the surge in business. The unified DSL, model library and visualization components can be quickly copied to e-commerce, live broadcast, digital content, financial account sharing and other industry scenarios, providing a standardized and low-cost technical foundation for the expansion of the enterprise ecosystem. It improves the system's real-time performance, accuracy, security and reliability in multiple dimensions, breaks through the existing technical bottlenecks, and has obvious technological advancement.

[0077] The above is a detailed introduction to a joint ecological intelligent profit-sharing system and method based on large-scale interface integration. This article uses specific examples to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the core idea of ​​the present invention; at the same time, for general technical personnel in this field, according to the ideas and methods of the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as a limitation on the present invention.

Claims

1. The intelligent profit-sharing method for joint ecological operation based on large-scale interface integration is characterized by: Including steps: S1: Register partners and describe their interfaces based on the unified meta-protocol DSL. S2: Automatically generate a sub-adapter based on the DSL and complete interface docking; S3: Collect transaction data and obtain user behavior video stream through image acquisition device; S4: Time-aligning, removing anomalies, and encoding features of the transaction data with the structured user behavior events extracted from the user behavior video stream to obtain a preprocessed data tensor; S5: Input the preprocessed data tensor into a multimodal convolutional neural network, and the multimodal convolutional neural network outputs a temporal interaction heat matrix and a profit matrix; S6: Display the profit matrix and time-series interaction heat matrix in the operation cockpit based on heat maps, cluster maps and KPI dashboards.

2. The intelligent profit-sharing method for joint ecological operation based on interface scale integration according to claim 1 is characterized in that: The element E(i,t) of the temporal interaction heat matrix represents the weighted interaction intensity of cooperation unit i in time slice t. The interaction intensity is obtained by weighted calculation of the number of clicks, stay duration, and visual attention, and is used to drive the three-dimensional heat map and model online fine-tuning of the operation cockpit.

3. The intelligent profit-sharing method for joint ecological operation based on interface scale integration according to claim 1 is characterized in that: The structured user behavior events include: using the YOLO deep detection network to identify products, user limbs and interactive objects; using the DeepSORT algorithm to track the target across frames to generate user stay duration and behavior sequences; extracting user posture key points through OpenPose to determine interest; and obtaining structured user behavior events.

4. The intelligent profit-sharing method for joint ecological operation based on interface scale integration according to claim 1 is characterized in that: The multimodal convolutional neural network includes parallel two-dimensional convolution branches and one-dimensional convolution time series branches, which respectively process the user behavior video stream data features and transaction data sequence features, and perform feature weighted fusion through the attention mechanism in the fusion layer.

5. The intelligent profit-sharing method for joint ecological operation based on interface scale integration according to claim 4 is characterized in that: The 2D convolution branch uses an improved activation function after the convolution layer. : , in, is the output of the convolution layer, e is the base of the natural logarithm, is the trainable first scaling factor, where 0.5≤α≤3.0; is the trainable second scaling factor, where 0≤ρ≤1.0; is the trainable third scaling factor, where 0.5≤β≤5.0; is the number of feature maps that are identified as effectively activated in all output feature maps of this convolutional layer in the current training batch. That is, for each output feature map, the average absolute value of the activation values ​​of all samples in the batch and their spatial positions is first calculated. When the average value is higher than the activation threshold set in advance or obtained by online learning, the feature map is determined to be in an effectively activated state. The feature maps that meet this condition are counted and summed one by one to obtain u. The maximum number of network channels that is preset.

6. The intelligent profit-sharing system for joint venture ecology based on large-scale interface integration is characterized by: include: Partner registration and interface description module, which registers partners and describes partner interfaces based on the unified meta-protocol DSL; A sub-adapter generation and interface docking module automatically generates a sub-adapter according to the DSL and completes the interface docking; Transaction data collection and user behavior video capture module, which collects transaction data and obtains user behavior video streams through image acquisition devices; A data alignment and feature coding preprocessing module performs time alignment, anomaly elimination, and feature coding on the transaction data and the structured user behavior events extracted from the user behavior video stream to obtain a preprocessed data tensor; A multimodal convolutional neural network inference module inputs the preprocessed data tensor into the multimodal convolutional neural network, which outputs a temporal interaction heat matrix and a profit matrix; The operation cockpit visualization module displays the profit distribution matrix and time-series interaction heat matrix based on heat maps, cluster maps and KPI dashboards.

7. The intelligent profit-sharing system for joint venture ecology based on interface scale integration as claimed in claim 6 is characterized in that: The element E(i,t) of the temporal interaction heat matrix represents the weighted interaction intensity of cooperation unit i in time slice t. The interaction intensity is obtained by weighted calculation of the number of clicks, stay duration, and visual attention, and is used to drive the three-dimensional heat map and model online fine-tuning of the operation cockpit.

8. The intelligent profit-sharing system for joint venture ecology based on interface scale integration as claimed in claim 6 is characterized in that: The structured user behavior events include: using the YOLO deep detection network to identify products, user limbs and interactive objects; using the DeepSORT algorithm to track the target across frames to generate user stay duration and behavior sequences; extracting user posture key points through OpenPose to determine interest; and obtaining structured user behavior events.

9. The intelligent profit-sharing system for joint venture ecology based on interface scale integration as claimed in claim 6 is characterized in that: The multimodal convolutional neural network includes parallel two-dimensional convolution branches and one-dimensional convolution time series branches, which respectively process the user behavior video stream data features and transaction data sequence features, and perform feature weighted fusion through the attention mechanism in the fusion layer.

10. The intelligent profit-sharing system for joint venture ecology based on interface scale integration as claimed in claim 9, characterized in that: The 2D convolution branch uses an improved activation function after the convolution layer. : , in, is the output of the convolution layer, e is the base of the natural logarithm, is the trainable first scaling factor, where 0.5≤α≤3.0; is the trainable second scaling factor, where 0≤ρ≤1.0; is the trainable third scaling factor, where 0.5≤β≤5.0; is the number of feature maps that are identified as effectively activated in all output feature maps of this convolutional layer in the current training batch. That is, for each output feature map, the average absolute value of the activation values ​​of all samples in the batch and their spatial positions is first calculated. When the average value is higher than the activation threshold set in advance or obtained by online learning, the feature map is determined to be in an effectively activated state. The feature maps that meet this condition are counted and summed one by one to obtain u. The maximum number of network channels that is preset.

Images