A Joint Venture Ecosystem Intelligent Profit-Sharing System and Method Based on Interface-Based Large-Scale Integration

Abstract

This invention relates to the field of automated processing technology and discloses a joint-venture ecosystem intelligent profit-sharing system and method based on large-scale interface integration. The method first registers partner interfaces through a unified meta-protocol (DSL) and automatically generates sub-adaptors, achieving minute-level heterogeneous interface integration. Subsequently, it synchronously collects transaction data and user behavior video streams, extracts structured behavioral events using target detection and pose recognition technologies, aligns them with transaction flows in the time domain, forms a preprocessed data tensor through feature encoding, and uses a multimodal convolutional neural network to infer the tensor, producing a time-series interaction heat matrix and a profit-sharing matrix, which are then displayed in real time. This technical solution achieves minute-level interface adaptation and user behavior interaction intensity evaluation, significantly reducing joint-venture operation and maintenance costs and improving the fairness and transparency of profit-sharing.

Description

Technical Field

[0001] This invention relates to the field of automated processing technology, specifically to a joint-venture ecosystem intelligent profit-sharing system and method based on interface-based large-scale integration. Background Technology

[0002] With the rise of "platform-based operation" and "ecosystem alliance" models, industries such as e-commerce, digital content distribution, fintech, and SaaS are increasingly using open APIs / SDKs to connect with third-party partners, forming a value network aggregating multiple suppliers and consumers. Alliance partners often judge popularity based on metrics such as API call volume, conversion rate, GMV, or ad exposure, which places higher demands on API adaptation efficiency, real-time data processing capabilities, and revenue sharing transparency. Most existing alliance systems use hard-coded scripts or manual configuration to connect APIs. Faced with issues such as heterogeneous API protocols, inconsistent field naming, and diverse authentication mechanisms, it is often necessary to develop adaptation logic for each API, resulting in long deployment cycles and high maintenance costs. When the number of partners reaches hundreds or thousands, API changes lead to frequent upgrades and iterations, making it difficult for traditional solutions to support rapid scalability. Furthermore, in terms of process, the mainstream approach is to pre-set fixed formulas or conditions to trigger scripts to complete the profit sharing: first, the transaction flow is pulled to an offline database, and then the profit sharing calculation is performed in batches on a daily or weekly basis. Moreover, the interaction intensity does not take into account the number of clicks, dwell time, visual attention, and other automated identification and statistics, which leads to delays in profit sharing results and cannot meet the needs of some financial-grade scenarios for T+0 or minute-level settlement. In addition, it is difficult to obtain information such as the intensity of user interaction 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 and channel quality. However, most remain limited to structured fields, lacking in-depth utilization of unstructured behavioral data such as videos and images. This makes it difficult to accurately assess implicit value signals such as "genuine attention," "stay time," or "emotional tendency," and fails to consider the impact of video features on analysis results, especially the influence of feature maps on user interaction levels, leading to inaccurate judgments. Furthermore, revenue results are typically stored in centralized databases, making it difficult to completely eliminate the risk of data tampering and accounting disputes. The demand for "trustworthy settlement" is increasingly prominent across cross-border and cross-industry ecosystems. In summary, existing technologies still have significant shortcomings in the following aspects: low efficiency in adapting to heterogeneous interfaces, lack of a unified meta-protocol and automatic adapter generation mechanism; coarse and lagging granularity of profit sharing, lacking dynamic accounting strategies for real-time transactions and multimodal behavioral data; and insufficient visualization and 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 intelligent profit-sharing technology solution for joint venture ecosystems that can complete heterogeneous interface docking within minutes, integrate structured transaction data and video image behavior data for multi-dimensional calculations to achieve real-time reliable accounting and visualized operation monitoring, so as to break through existing technical bottlenecks, reduce docking and maintenance costs, and improve automation, real-time performance and auditability. Summary of the Invention

[0004] To address the aforementioned problems in existing technologies, this invention discloses a joint-venture ecosystem intelligent profit-sharing system and method based on large-scale interface integration. This method first registers partner interfaces via a unified meta-protocol (DSL) and automatically generates sub-adaptors, achieving minute-level heterogeneous interface integration. Subsequently, it synchronously collects transaction data and user behavior video streams, extracts structured behavioral events using target detection and pose recognition technologies, aligns them with transaction flows in the temporal domain, and forms a preprocessed data tensor through feature encoding. A multimodal convolutional neural network infers from the tensor, producing a temporal interaction heat matrix and a profit-sharing matrix. This solution achieves minute-level interface adaptation and user behavior interaction intensity evaluation, significantly reducing joint-venture operation and maintenance costs and improving the fairness and transparency of profit-sharing.

[0005] This application provides a method for intelligent profit-sharing within a collaborative ecosystem based on large-scale interface integration, including the following steps:

[0006] S1: Register the partner and describe the partner interface based on the Unified Meta Protocol DSL;

[0007] S2: Automatically generate a sub-adaptor based on the DSL and complete the interface connection;

[0008] S3: Collect transaction data and acquire user behavior video streams through image acquisition devices;

[0009] S4: Perform time alignment, anomaly removal, and feature encoding on the transaction data and the structured user behavior events extracted from the user behavior video stream to obtain a preprocessed data tensor;

[0010] S5: Input the preprocessed data tensor into a multimodal convolutional neural network, which outputs a temporal interaction heat matrix and a distribution matrix;

[0011] S6: Display the profit-sharing matrix and time-series interactive heat matrix in the operations cockpit based on heatmaps, cluster maps, and KPI dashboards.

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

[0013] Preferably, the structured user behavior event includes: using the YOLO deep detection network to identify products, user limbs, and interactive objects; using the DeepSORT algorithm to perform cross-frame tracking of the target to generate user dwell time and behavior sequence; extracting user pose key points through OpenPose to determine interest level; and obtaining the structured user behavior event.

[0014] Preferably, the multimodal convolutional neural network includes parallel two-dimensional convolutional branches and one-dimensional convolutional temporal 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.

[0015] Preferably, the two-dimensional convolutional branch employs an improved activation function after the convolutional layer. :

[0016]

[0017] in, This is the output of the convolutional layer, where e is the base of the natural logarithm. Let be the first trainable scaling factor, where 0.5 ≤ α ≤ 3.0; Let be the trainable second scaling factor, where 0 ≤ ρ ≤ 1.0; Let β be the trainable third scaling factor, where 0.5 ≤ β ≤ 5.0; To determine the number of validly activated feature maps in the current training batch across all output feature maps of this convolutional layer, we first calculate the absolute average of the activation values ​​of all samples and their spatial locations within each output feature map. If this average value exceeds a pre-set or online-learned activation threshold, the feature map is considered validly activated. We then count and sum the activation values ​​of all feature maps that meet this condition to obtain u. The maximum number of network channels is preset.

[0018] This application also provides a joint venture ecosystem intelligent profit-sharing system based on interface-based large-scale integration, including:

[0019] The partner registration and interface description module registers partners and describes their interfaces based on the Unified Meta Protocol (DSL).

[0020] The sub-adaptor generation and interface docking module automatically generates sub-adaptors and completes interface docking based on the DSL.

[0021] The transaction data acquisition and user behavior video capture module collects transaction data and acquires user behavior video streams through an image acquisition device.

[0022] The data alignment and feature encoding preprocessing module performs time alignment, anomaly removal, and feature encoding on the transaction data and the structured user behavior events extracted from the user behavior video stream to obtain a preprocessed data tensor.

[0023] The multimodal convolutional neural network inference module inputs preprocessed data tensors into the multimodal convolutional neural network, which outputs a temporal interaction heat matrix and a distribution matrix.

[0024] The operation cockpit visualization module displays the profit-sharing matrix and time-series interactive heat matrix based on heat maps, cluster maps, and KPI dashboards.

[0025] Preferably, the element E(i,t) of the time-series interaction heat matrix represents the weighted interaction intensity of cooperative unit i in time slice t. The interaction intensity is obtained by weighting the number of clicks, dwell time, and visual attention, and is used to drive the online fine-tuning of the three-dimensional heat map and model of the operation cockpit.

[0026] Preferably, the structured user behavior event includes: using the YOLO deep detection network to identify products, user limbs, and interactive objects; using the DeepSORT algorithm to perform cross-frame tracking of the target to generate user dwell time and behavior sequence; extracting user pose key points through OpenPose to determine interest level; and obtaining the structured user behavior event.

[0027] Preferably, the multimodal convolutional neural network includes parallel two-dimensional convolutional branches and one-dimensional convolutional temporal 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.

[0028] Preferably, the two-dimensional convolutional branch employs an improved activation function after the convolutional layer. :

[0029]

[0030] in, This is the output of the convolutional layer, where e is the base of the natural logarithm. Let be the first trainable scaling factor, where 0.5 ≤ α ≤ 3.0; Let be the trainable second scaling factor, where 0 ≤ ρ ≤ 1.0; Let β be the trainable third scaling factor, where 0.5 ≤ β ≤ 5.0; To determine the number of validly activated feature maps in the current training batch across all output feature maps of this convolutional layer, we first calculate the absolute average of the activation values ​​of all samples and their spatial locations within each output feature map. If this average value exceeds a pre-set or online-learned activation threshold, the feature map is considered validly activated. We then count and sum the activation values ​​of all feature maps that meet this condition to obtain u. The maximum number of network channels is preset.

[0031] This invention provides a joint-venture ecosystem intelligent profit-sharing system and method based on interface-based large-scale integration, which can achieve the following beneficial technical effects:

[0032] 1. This invention achieves minute-level interface adaptation by employing technologies such as a unified meta-protocol (DSL), automated sub-adaptor generation, and multimodal convolutional neural networks (including improved activation functions). Through the collaboration of DSL description and code generator, the integration time for heterogeneous APIs / SDKs is shortened from the traditional 1-2 weeks to less than 30 minutes, significantly reducing the launch threshold and manpower costs for multi-party joint ventures. The refined value assessment of behavior-transaction fusion is the first to align video / image behavior data (clicks, dwell time, visual attention, posture / emotion, etc.) with transaction flow in the time domain and input them together into a multimodal CNN, outputting a time-series interaction heat matrix and profit-sharing matrix. This greatly improves the real-time performance, accuracy, interpretability, and security of interactions, breaking through existing technological bottlenecks in multiple dimensions.

[0033] 2. This invention employs an improved activation function to ensure the stability of deep and wide networks. The two-dimensional branch of this invention uses an improved activation function. In the current training batch, the number of feature maps that are recognized as effectively activated in all output feature maps of this convolutional layer is added to the calculation process of the convolutional network. Combined with dynamic effective channel counting, this significantly reduces gradient variance and greatly improves weak texture recall, providing numerical stability for training and inference for large-scale parallel inference.

[0034] 3. This invention employs visualized operational insights. The system displays the revenue-risk distribution of partners across time and geography in real time using 3D heatmaps, clustering diagrams, and KPI dashboards. It supports one-click sandbox simulation and A / B testing, improving operational decision-making efficiency by over 40%. High scalability and low maintenance are achieved through a message queue + containerized architecture supporting concurrent access from thousands of partners and transaction throughput of tens of billions. Activation functions and model parameters can be fine-tuned online, reducing the costs of retraining and reconfiguration due to business surges. A unified DSL, model library, and visualization components can be quickly replicated to multiple industry scenarios such as e-commerce, live streaming, digital content, and financial revenue sharing, providing a standardized and low-cost technical foundation for enterprise ecosystem expansion. It overcomes existing technical bottlenecks in multiple dimensions, including improved system real-time performance, accuracy, and security, demonstrating significant technological advancement. Attached Figure Description

[0035] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0036] Figure 1 This is a flowchart of the steps of a joint venture ecosystem intelligent profit-sharing method based on interface-scale integration according to the present invention;

[0037] Figure 2 This is a schematic diagram of a joint-venture ecosystem intelligent profit-sharing system based on interface-scale integration according to the present invention. Detailed Implementation

[0038] 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.

[0039] Example 1:

[0040] In view of the aforementioned problems mentioned in the prior art, and in order to solve the above technical problems, as shown in the appendix. Figure 1 As shown: This application provides a method for intelligent profit sharing in a joint venture ecosystem based on large-scale interface integration, including the following steps:

[0041] S1: Register the partner and describe the partner interface based on the Unified Meta Protocol DSL;

[0042] S2: Automatically generate a sub-adaptor based on the DSL and complete the interface connection;

[0043] In one embodiment, the alliance platform first deploys an interface registration center and code generation service on a cloud server. Third-party partners open the registration center page in a browser, fill in basic information such as partner identifier, interface root address, authentication method, and traffic threshold, and submit. The system automatically assigns an access key pair to the partner and persists all configurations to the database. Simultaneously, a directory named after the partner identifier is created in object storage to store the interface description file. Subsequently, the partner edits a text document written in the Unified Meta-Protocol (DSL) online in the registration center. This text describes the interface name, access path, request method, field type, authentication method, retry strategy, etc., in a human-readable indented format. After editing, the user clicks upload, and the registration center immediately saves the text to object storage and pushes a "New DSL Online" event to the message queue. The event includes the partner identifier and the DSL file address. After subscribing to this event, the code generation service downloads the DSL file and calls its built-in parser to generate an abstract syntax tree (AST), then outputs a standardized set of source code based on the AST. The source code implements functions such as request encapsulation, token refresh, flow control, and logging, and is compatible with the unified message protocol within the alliance platform. After generation, the service automatically compiles the executable file, packages it into a container image, and pushes it to a private image repository. Upon detecting a new image tag, the continuous delivery pipeline immediately creates a deployment named `adapter-partner identifier` in the Kubernetes cluster, starting two replicas by default. After the new container starts, it automatically reads the access key, requests a token from the partner interface, and caches it in memory. It then exposes a unified internal endpoint, receives business data forwarded by services within the platform, and sends standardized requests to the partner interface in real time. Once the monitoring system confirms the adapter has responded normally through health checks, it updates the partner status to "activated." From then on, business events pushed from the partner to the alliance platform can flow seamlessly into subsequent data processing and revenue sharing. The entire process, from the partner submitting the DSL to the adapter going live and starting to forward interface traffic, typically takes no more than thirty minutes, effectively solving pain points such as heterogeneous interfaces, slow manual integration, and difficult upgrades.

[0044] S3: Collect transaction data and acquire user behavior video streams via image acquisition devices; in some embodiments, the alliance platform deploys a dual-channel acquisition mechanism in both client-side and offline scenarios. Online, the platform embeds a tracking SDK at key nodes such as payment success pages, order details pages, and ad click entry points on the web and mobile applications. Whenever a user completes an order, views an order, or clicks an ad, the SDK generates a transaction event within milliseconds, containing fields such as order number, product identifier, amount, payment channel, terminal type, geographic location, and timestamp, and immediately sends it to the platform's event gateway via HTTPS. The event gateway uses a backpressure queue to ensure no packet loss under high concurrency, and then writes the raw JSON event to a Kafka topic for subsequent cleaning tasks.

[0045] In the offline portion, the platform has installed an integrated ultra-high-definition camera above the core experience area. The camera incorporates an edge computing unit, enabling real-time local object detection and facial landmark modeling. The camera first performs human detection and multi-object tracking on users entering the frame, generating a unique trajectory number. Next, it estimates the user's viewing angle and dwell time through pose recognition and captures keyframes. If the trajectory duration exceeds a set threshold, the edge unit generates a user behavior segment containing structured fields such as camera ID, trajectory number, entry time, exit time, average gaze direction, and expression confidence level. To protect privacy, the edge unit only uploads the extracted numerical features, not the original facial image.

[0046] User behavior segments collected offline are sent in batches to the same event gateway every three seconds. The gateway then writes them to another Kafka topic tagged with "offline behavior." When the platform's stream processing engine consumes online transaction events and offline behavior segments, it performs window alignment based on a unified ISO timestamp field and makes initial associations based on the geographic location matching strategy between the camera and the order terminal. The aligned data is then written to a real-time feature warehouse, providing input for subsequent feature encoding and multimodal convolutional neural network inference. In this way, the system can simultaneously obtain accurate transaction logs and rich user video behavior data, ensuring that the subsequent profit-sharing model considers both economic contribution and genuine interaction activity.

[0047] S4: Perform time alignment, anomaly removal, and feature encoding 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.

[0048] For time alignment, the stream job joins two event streams using a five-second sliding window. The window key includes the user identifier, device identifier, and the most recent IP-GPS aggregation location. When the transaction timestamp and the entry time of the behavior segment are within two seconds and their geographical locations match, the job merges them into an "aligned event." If no corresponding transaction is found for the behavior segment, it is written to a waitlist cache to wait for the next window; if no match is found after thirty seconds, it is marked as an "isolated behavior event" for subsequent risk control analysis.

[0049] Regarding anomaly removal, the workflow immediately executes three types of rules after merging: First, the numerical out-of-bounds rule: Transactions with negative amounts or exceeding the daily maximum limit are directly removed; behavior segments with a dwell time exceeding thirty 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 has trained an extreme value density estimation model offline, and performs joint scoring on dwell time, sentiment confidence, and number of clicks. Events below the threshold are marked as anomalies and written to a bypass Kafka for manual review.

[0050] In terms of feature encoding, events that have been aligned and passed anomaly filtering first enter the encoder. The encoder performs the following transformations on the fields: continuous fields such as amount and dwell time are first log-compressed and then normalized according to the mean and variance of the training set; discrete fields such as payment channel and content category are mapped to fixed-length integer indices, and then an eight-dimensional trainable embedding vector is obtained by looking up a table; the timestamp field is split into two periodic features, representing the hour and the day of the week respectively; the camera pose angle is quantized into three discrete labels and also converted into embeddings by looking up a table; all embedding vectors are concatenated with normalized real numbers to form a fixed 128-dimensional feature vector. The streaming job rolls in 3-second increments, stacking feature vectors belonging to the same user and the same partner in chronological order into a sequence of length 20, finally obtaining a preprocessed data tensor in the form of "batch size × 20 × 128", which is written to the Redis streaming cache, waiting for the subsequent multimodal convolutional neural network to read it in real time for inference. The end-to-end latency of the entire alignment, elimination, and encoding chain is kept within 300 milliseconds, meeting the real-time requirements of the minute-level accounting model.

[0051] S5: The preprocessed data tensor is input into a multimodal convolutional neural network, which outputs a temporal interaction heat matrix and a distribution matrix. In some embodiments, the system divides the preprocessed data obtained from the preceding 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 each user within the last three seconds; 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. Regarding the two-dimensional convolutional branch for video behavior, the input three-frame feature maps first pass through a set of convolutional layers with a kernel size of 3 × 3 and 64 channels, followed by a batch normalization layer and BA-Swish-Mish activation; then, a 128-channel convolutional layer with the same configuration is stacked. Subsequently, dilated convolution is used to expand the receptive field, and a self-attention gating module is introduced to perform cross-frame fusion on salient regions between the three frames. Finally, global average pooling is used to compress the spatial dimension into a vector of length 512. The transaction-behavior one-dimensional convolutional temporal branch first feeds the sequence tensor into a one-dimensional convolutional layer with 64 kernels, a stride of 1, and TG-AdaGELU activation. This is followed by a 128-kernel convolution with the same structure, and then an additional causal residual block to preserve temporal order. The convolutional output is temporally max-pooled to obtain a vector of length 512. For feature fusion, the visual and temporal vectors are concatenated to form a 1024-dimensional fused vector, which is then fed into a cross-modal attention layer. The attention weights are determined by content-adaptive gating to ensure that visual or transaction signals occupy appropriate proportions in different scenarios. The fused result is then passed through a fully connected layer with dropout, with the output size remaining 1024. The dual output heads include the fused vector, which is fed in parallel into two linear output heads. The heat index head uses softmax to map the 1024-dimensional data to a "number of cooperating units × 20" dimension and rearranges it into a time-series interaction heat index matrix. The profit sharing head uses sigmoid to map the 1024-dimensional data to a "number of cooperating units × number of revenue channels" dimension and rearranges it into a profit sharing matrix. During the inference phase, the model generates two matrices for each batch, representing the real-time interaction intensity and the revenue sharing weight, respectively.

[0052] The training strategy uses the mean squared error of interaction duration with labeled data as the heat index header and the Huber loss with the actual on-chain transaction splitting as the profit sharing header. The weighted sum of the two is used as the total loss. The optimizer is AdamW, with a base learning rate of 0.3% and a weight decay of 0.1%. If the heat index error decreases but the profit sharing error stagnates on the validation set, the weights of both are dynamically adjusted. The entire network training takes 300 milliseconds per iteration on two 32GB GPUs, and the latency for batches of 8000 events during the inference phase remains within 90 milliseconds.

[0053] The deployed and invoked model weights are saved as ONNX files, compiled using TensorRT, and loaded into the containerized inference service. The stream processing engine writes the preprocessed tensors into gRPC, and the inference results are returned in Protocol-buffer format and sent to subsequent links. The profit-sharing matrix is ​​written into the consortium blockchain smart contract, and the heat map matrix is ​​pushed to the operations dashboard in real time for 3D heat map rendering. Through this embodiment, the system can stably integrate two types of heterogeneous data, video and transaction data, in a deep and wide network, and simultaneously output the time-series interaction heat map matrix and profit-sharing matrix within a latency of hundreds of milliseconds, providing core computing power support for real-time accounting and visualized operations.

[0054] S6: In the operations dashboard, the revenue sharing matrix and time-series interaction heat map are displayed based on heatmaps, clustering graphs, 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 maps and revenue sharing matrices into the time-series database. Then, the front-end visualization service subscribes to the "matrix ready" event via WebSocket, retrieves the corresponding batch of data, and performs the following processing:

[0055] The first step is to generate a 2D heatmap. The visualization backend expands the time-series interaction heatmap matrix by rows of collaboration units and columns of time slices. It first calculates the maximum and minimum values ​​for each row, and then normalizes all elements to between 0 and 1 using linear interpolation. The frontend uses WebGL to create a canvas, mapping the matrix sequentially into a pixel grid: the horizontal axis represents time, the vertical axis represents collaboration units, and the pixel color is determined by the normalized value, ranging from dark blue to dark red. When the user hovers over the heatmap, the frontend displays the specific interaction index of that collaboration unit at that time point and links to the original video clip for verification.

[0056] The second step is to generate a revenue-sharing distribution clustering chart. The visualization backend first performs column normalization on the revenue-sharing matrix to obtain the proportion vector of each partner unit in different revenue channels. Then, KMeans is used to cluster the partners into three categories: high contribution, high potential, and low activity. The clustering results are returned to the frontend via REST, where ECharts is used to plot the three categories of data as a scatter cloud. The x-axis of each point represents the first principal component of the revenue-sharing proportion, and the y-axis represents the second principal component. The point size is proportional to the total revenue-sharing amount, and the color represents the category. Clicking on a scatter point will take you to the corresponding partner's details page.

[0057] The third step is to build a KPI dashboard. The visualization backend calculates five metrics in real time according to business definitions: total revenue sharing, average interaction activity, average risk adjustment factor, number of blocked exceptions, and revenue sharing prediction error. The frontend uses Vue components to render the five dashboards, along with historical line trends. If any metric exceeds a threshold, the dashboard automatically changes color and displays an alert.

[0058] The fourth step is coordinated operation. When operations personnel select a time period and multiple cooperative units on the heatmap, the front end immediately sends a request to the back end to re-cluster the selected sub-matrix and update the scatter plot; the dashboard synchronously refreshes the KPIs of that subset, allowing operations to quickly compare the effects of different revenue-sharing strategies. The total latency of the entire visualization chain, from inference output to front-end rendering, remains within one second, supporting refreshes every second without noticeable lag. This implementation allows the operations team to intuitively understand "who contributed how much heat and how much revenue, when," and can complete strategy simulation and effect evaluation on the same interface, significantly improving revenue management efficiency and decision-making accuracy.

[0059] In some embodiments, the element E(i,t) of the temporal interaction heatmap represents the weighted interaction intensity of cooperative unit i in time slice t. The interaction intensity is obtained by weighting the number of clicks, dwell time, and visual attention, and is used to drive the online fine-tuning of the 3D heatmap and model in the operation cockpit. In some embodiments, the system organizes the three types of raw indicators—clicks, dwell time, and visual attention—within a 5-second sliding window for each cooperative unit into a time series of length 20. The following describes how the CNN fuses these three types of information into a temporal interaction heatmap 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 and 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 is performed, and the tensor is fed into a one-dimensional convolutional layer with a kernel length of 3, 16 channels, and a stride of 1. This layer is equivalent to performing a weighted sliding calculation on clicks, pauses, and views over the three most recent time steps, but the weights are automatically learned by the convolutional 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, with the convolutional output entering a lightweight channel attention module. The system first performs global average pooling on the 16 feature maps to obtain a vector of length 16. This vector is then passed through two fully connected layers and a sigmoid function to generate 16 coefficients between 0 and 1, which are then multiplied back into the original feature map, achieving dynamic weighting for different combinations of "clicks," "pauses," and "views." The dimensionality-reduced and normalized weighted feature maps are then passed through a bottleneck layer with a single one-dimensional convolutional kernel, compressing the 16 channels into a single-channel sequence. Softmax is then applied to ensure that the outputs of different cooperating units within the same time step sum to 1. At this point, the single-channel value at each time step is E(i,t).

[0060] To ensure that E reflects user engagement, the system constructs pseudo-labels: one is the complete viewing duration recorded by the player's backend, and the other is the "hot time period" label manually evaluated in historical marketing experiments. The network uses a combination of mean squared error and FocalLoss for optimization. When the predicted popularity mismatches with the labeled hotspots, the gradient automatically adjusts the convolution kernel and attention weights to achieve optimal weighting for clicks, dwell time, and visual metrics in different scenarios.

[0061] In the inference and implementation phase, the stream processing engine feeds the three-channel tensor into the TensorRT container; the container outputs the E matrix and immediately writes it back to the time-series database. The front-end reads the E matrix and renders each column as a color bar using WebGL. The number of columns corresponds to the time step, the number of rows corresponds to the cooperative unit, and the color intensity represents the predicted interaction intensity within a five-second sliding window. Operators can click on any cell to retrieve the original click logs and video frames, verifying the model's interpretability. In offline A / B testing, the accuracy of hotspot detection is 82.7% when only click and dwell inputs are used; after introducing visual attention and using the aforementioned convolutional-attention structure, the accuracy improves to 85.1%, proving that the weighted fusion of the three metrics can better capture the true interaction intensity.

[0062] In some embodiments, the structured user behavior events include: identifying products, user limbs, and interactive objects using a YOLO deep detection network; performing cross-frame tracking of targets using the DeepSORT algorithm to generate user dwell time and behavior sequences; extracting user pose key points using OpenPose to determine interest levels; and obtaining structured user behavior events. In some embodiments, the system sequentially performs YOLO detection, DeepSORT tracking, and OpenPose pose estimation on a video stream sampled per second at edge nodes. First, the YOLO detection network loads weights finely tuned with industry data, performs forward inference on each frame, and outputs bounding boxes and confidence scores for categories such as products, human bodies, and gestures; the box fields include the top-left corner coordinates, width and height, category number, and probability value. Subsequently, DeepSORT reads the detection results of the same frame, embeds the appearance features into a 128-dimensional vector using a lightweight CNN, and combines it with motion prediction using a Kalman filter to match cross-frame detections and assign unique trajectory IDs. The system maintains a buffer for each trajectory, accumulating the number of consecutively existing frames and the total duration; when the detection box with the same ID is interrupted in the frame sequence for more than ten frames, the trajectory is considered to have ended. While the trajectory remains active, OpenPose estimates the pose of the human bounding box within the same frame, outputting the 2D coordinates and confidence scores of 18 key points. The system calculates the gaze direction based on the coordinates of the nose tip and eyes, and then determines the standing or squatting posture based on the ratio of the shoulder to hip coordinates; if the user is facing the screen and focusing their gaze, it is judged as a high-interest posture. For all frames within the same trajectory, the system calculates the average gaze direction, the proportion of high-interest frames within one second, and the overall average confidence score of facial expressions. When the trajectory ends, the system calls an event generator to transform the information aggregated 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, proportion of interest frames, list of detected product categories, and the maximum confidence pose category at the end time. The event is written to the local log in JSON string format and uploaded to the central Kafka topic "user.behavior.structured" via MQTT. The upstream streaming engine can then align the event with the transaction flow according to the timestamp to complete subsequent feature encoding and model inference.

[0063] In some embodiments, the multimodal convolutional neural network includes parallel two-dimensional convolutional branches and one-dimensional convolutional temporal 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. In some embodiments, the system splits each batch of data into two tensor streams. The first stream comes from the user behavior video stream: three frames are extracted from the most recent three seconds of video, compressed into a three-channel 128×128 feature map by an edge model; the tensor shape after batch dimension is batch size multiplied by 3×128×128. The second stream comes from the transaction feature sequence: 20 128-dimensional feature vectors are accumulated within the same time window, and the tensor shape is batch size × 20×128. The video branch adopts a two-dimensional convolutional skeleton. The input first passes through a convolutional layer with a kernel size of 3×3 and an output channel of 64, followed by a batch normalization layer and BA-Swish-Mish activation; then another 3×3, 128-channel convolution is stacked.

[0064] To expand the receptive field, a dilated convolution with a dilation rate of 2 and 256 output channels is introduced in the third block. Next, a self-attention gating module is inserted: the spatial features are subjected to a 1-to-1 convolution to obtain the query, key, and value; the attention weights are calculated using a scaled dot product and multiplied by the value features, then added back to the main branch. Global average pooling is then used to compress the spatial dimension, generating a visual vector of length 512. The transaction branch uses a one-dimensional convolutional temporal structure. The input sequence first enters a one-dimensional convolutional layer with a kernel length of 3 and 64 output channels, activated by TG-AdaGELU; next is a dilated convolution with a dilation rate of 2 and 128 output channels to capture temporal dependencies of 10 steps or more. Both convolutional layers are followed by batch normalization and residual skip connections, and finally, temporal max pooling is performed to compress the 20-length sequence into a 512-length vector. The two vectors are concatenated in a fusion layer to form a 1024-dimensional feature and then fed into the cross-modal attention layer.

[0065] The implementation is as follows: The concatenated vector is divided in half. The first 512 dimensions are used as the query, and the last 512 dimensions are copied simultaneously as the key and value. Attention coefficients are obtained using a scaled dot product, then weighted and summed with the value vector. This sum is then added to the original concatenated vector as a residual and passed through LayerNorm to output a 1024-dimensional fused vector. This fused vector enters a fully connected layer with a 30% dropout ratio, maintaining its 1024 dimensions, and is then split into two parallel output heads. The popularity head is mapped to the number of collaborating units multiplied by 20 dimensions via a linear layer and rearranged into a temporal interaction popularity matrix. The profit-sharing head is mapped to the number of collaborating units multiplied by the number of revenue channels via a linear layer and rearranged into a profit-sharing matrix. The popularity head uses softmax normalization, and the profit-sharing head uses sigmoid normalization. During training, the model uses mean squared error to supervise the popularity head and Huber loss to supervise the profit-sharing head. The AdamW optimizer has a learning rate of 0.3% and a weight decay of 0.1%. During deployment, the model is exported as ONNX and then compiled into the FP16 engine using TensorRT, with a batch inference latency of approximately 90 milliseconds for 8000 models. The entire network can simultaneously obtain the time-series interaction heat matrix and profit-sharing matrix within a latency of hundreds of milliseconds, meeting the performance requirements for real-time accounting and operational visualization.

[0066] In some embodiments, the two-dimensional convolutional branch employs an improved activation function after the convolutional layer. :

[0067]

[0068] in, This is the output of the convolutional layer, where e is the base of the natural logarithm. Let be the first trainable scaling factor, where 0.5 ≤ α ≤ 3.0; Let be the trainable second scaling factor, where 0 ≤ ρ ≤ 1.0; Let β be the trainable third scaling factor, where 0.5 ≤ β ≤ 5.0; To determine the number of validly activated feature maps in the current training batch across all output feature maps of this convolutional layer, we first calculate the absolute average of the activation values ​​of all samples and their spatial locations within each output feature map. If this average value exceeds a pre-set or online-learned activation threshold, the feature map is considered validly activated. We then count and sum the activation values ​​of all feature maps that meet this condition to obtain u. The maximum number of channels in the network is preset. In some embodiments, under the PyTorch-1.13 environment, the improved activation function proposed in this paper is integrated into the visual convolution branch. First, a custom class BA_SwishMish is defined in the code, and the trainable parameters alpha, rho, and beta are defined in the constructor, with initial values ​​set to 0, 0.1, and 0 respectively, and reparameterized using Softplus to ensure they are always positive. The maximum number of channels in the widest layer of the network, 1024, is written into the constant Cmax. During forward computation, the activation function traverses the output feature maps of the current batch; for each feature map, the batch dimension and spatial dimension are fully expanded and the absolute value of the average is calculated. If the average value is higher than the dynamic threshold of 0.05, the feature map is recorded as "effectively activated". After counting an entire batch, the total number of effective activated channels u is obtained. Then, according to the formula logic, the convolutional layer output, the three trainable scaling factors, and u / Cmax are combined to generate a new activation result. In terms of training configuration, the AdamW optimizer is used with a learning rate of 3× The weights were decayed by 0.1%, and L2 regularization was added to alpha, rho, and beta to prevent exceeding the set limits. A 48-hour control experiment was conducted on a subset of ImageNet-100 classification: the baseline network used conventional SiLU activation, while the improved network only replaced the 2D branch activation with a custom version, keeping the rest of the structure the same. The results showed that the loss curve of the improved version was significantly smoother 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, indicating a significant improvement in numerical stability. The Grad-CAM heatmap on the validation set also showed that the improved network could focus on a more complete target region on low-contrast samples, while the baseline network was more susceptible to background noise. During the inference phase, the trained weights were exported to ONNX and then compiled into an FP-16 accuracy engine using TensorRT. When the batch size was 64, the inference time increased by about 2% compared to the baseline, which was within an acceptable range for operation and maintenance; given the improvement in accuracy and stability, 30% grayscale was finally deployed, and then expanded to full scale after two weeks. Online monitoring statistics show that the average profit sharing error in weak texture scenes decreased by 0.4%, further proving that this embodiment can bring quantifiable benefits in production.

[0069] In some embodiments, this application also provides a joint venture ecosystem intelligent profit-sharing system based on interface-scale integration, such as... Figure 2 As shown, the system consists of three layers: a cloud core cluster, edge acquisition nodes, and operation terminals. These layers are interconnected through enterprise-grade fiber optic cables or public VPN leased lines, and the overall network topology converges in a star shape to the cloud core switch.

[0070] For the core cloud cluster, three physical server clusters are deployed in the same data center rack: The first cluster is the interface registration and code generation cluster, consisting of three dual-socket Xeon servers equipped with 64-core CPUs, 512GB of memory, and two 10 Gigabit Ethernet cards, used to run the DSL registration center, CICD pipeline, and container image repository. The second cluster is the inference computing cluster, configured with eight single-node A100 GPU servers, each containing eight 40GB GPUs, dual-socket AMD EPYC processors, and a default 1TB of memory, supporting multimodal convolutional neural network inference services. The third cluster is the data and blockchain cluster, including three NVMe AllFlash storage servers running time-series databases and feature caching, and two independent high-performance servers deploying Fabric blockchain consortium nodes and CA services. All three clusters of servers are connected to the internal network through a core 10 Gigabit switch.

[0071] For edge acquisition nodes, each offline partner store or showroom is equipped with an integrated smart camera and a miniPC. The camera has a built-in Sony 4K sensor and Cambricon edge AI chip, responsible for real-time inference YOLO detection and OpenPose pose estimation, and sends the extracted structured features to the local miniPC via a PoE switch. The miniPC uses an 8-core ARM processor, 16GB of memory and 256GB of SSD, and runs a Dockerized edge event aggregation program; after aggregation, it is directly uploaded to the cloud event gateway via enterprise leased line / MPLSVPN.

[0072] For the operations cockpit terminals, a hyperbolic ultra-wide 8K LED wall is placed in the headquarters dispatch center, with two RTX A6000 workstations behind it serving as visualization rendering nodes. The operations staff's desktops use six i7 office PCs as control consoles, communicating with the LED wall workstations within a Layer 2 network via a switch. The cockpit terminals subscribe to cloud inference output via WebSocket, with a latency of no more than 100 milliseconds.

[0073] For communication and security connectivity, all edge nodes access the cloud via a public VPN. The VPN terminal gateway is located above the core switch and supports dual-active redundancy. The core switch connects to the firewall, which in turn connects to the internet border router. Blockchain nodes network with each other through independent TLS tunnels to avoid sharing with the production data network. The inference cluster and data cluster are interconnected via an RDMA network to improve the efficiency of GPUs reading time-series databases. Only gRPC and Docker Registry ports are open between the interface registration cluster and the inference cluster, in conjunction with a zero-trust access policy.

[0074] In terms of hardware interaction during operation, video clips captured by edge cameras are detected by the local AI chip, and the extracted product frames, face frames, pose key points, and other elements are sent to the miniPC via Ethernet. The miniPC packages and uploads the data once per second to the inference cluster where the Kafka cluster resides in the cloud. The inference cluster server reads the transaction flow and behavior tensor of the same batch, performs multimodal CNN inference on the GPU, writes the generated profit-sharing matrix into the smart contract of the blockchain cluster, and then pushes the interaction heat matrix to the LED wall rendering node through the time-series database. Operators can view the 3D heat map in real time on their desktop PCs through a browser and issue strategy adjustment commands. The commands are passed through the load balancer of the interface registration cluster to the DSL service, and then the CICD is called to automatically regenerate or hot-update the adapter image, finally flowing back to the inference cluster to complete a new round of online model fine-tuning. Through the above hardware and connectivity design, the system can achieve end-to-end second-level data closure, on-chain trusted splitting, and visual decision support for the operation dashboard while ensuring tens of thousands of concurrent video streams and hundreds of millions of transaction throughput.

[0075] The system includes: a partner registration and interface description module, which registers partners and describes their interfaces based on the Unified Meta Protocol (DSL);

[0076] The sub-adaptor generation and interface docking module automatically generates sub-adaptors and completes interface docking based on the DSL.

[0077] The transaction data acquisition and user behavior video capture module collects transaction data and acquires user behavior video streams through an image acquisition device.

[0078] The data alignment and feature encoding preprocessing module performs time alignment, anomaly removal, and feature encoding on the transaction data and the structured user behavior events extracted from the user behavior video stream to obtain a preprocessed data tensor.

[0079] The multimodal convolutional neural network inference module inputs preprocessed data tensors into the multimodal convolutional neural network, which outputs a temporal interaction heat matrix and a distribution matrix.

[0080] The operation cockpit visualization module displays the profit-sharing matrix and time-series interactive heat matrix based on heat maps, cluster maps, and KPI dashboards.

[0081] In some embodiments, the element E(i,t) of the time-series interaction heat matrix represents the weighted interaction intensity of cooperative unit i in time slice t. The interaction intensity is obtained by weighting the number of clicks, dwell time, and visual attention, and is used to drive the online fine-tuning of the three-dimensional heat map and model of the operation cockpit.

[0082] In some embodiments, the structured user behavior event includes: using the YOLO deep detection network to identify products, user limbs, and interactive objects; using the DeepSORT algorithm to perform cross-frame tracking of the target to generate user dwell time and behavior sequence; extracting user pose key points through OpenPose to determine interest level; and obtaining the structured user behavior event.

[0083] In some embodiments, the multimodal convolutional neural network includes parallel two-dimensional convolutional branches and one-dimensional convolutional temporal 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.

[0084] In some embodiments, the two-dimensional convolutional branch employs an improved activation function after the convolutional layer. :

[0085]

[0086] in, This is the output of the convolutional layer, where e is the base of the natural logarithm. Let be the first trainable scaling factor, where 0.5 ≤ α ≤ 3.0; Let be the trainable second scaling factor, where 0 ≤ ρ ≤ 1.0; Let β be the trainable third scaling factor, where 0.5 ≤ β ≤ 5.0; To determine the number of validly activated feature maps in the current training batch across all output feature maps of this convolutional layer, we first calculate the absolute average of the activation values ​​of all samples and their spatial locations within each output feature map. If this average value exceeds a pre-set or online-learned activation threshold, the feature map is considered validly activated. We then count and sum the activation values ​​of all feature maps that meet this condition to obtain u. The maximum number of network channels is preset.

[0087] This invention provides a joint-venture ecosystem intelligent profit-sharing system and method based on interface-based large-scale integration, which can achieve the following beneficial technical effects:

[0088] 1. This invention achieves minute-level interface adaptation by employing technologies such as a unified meta-protocol (DSL), automated sub-adaptor generation, and multimodal convolutional neural networks (including improved activation functions). Through the collaboration of DSL description and code generator, the integration time for heterogeneous APIs / SDKs is shortened from the traditional 1-2 weeks to less than 30 minutes, significantly reducing the launch threshold and manpower costs for multi-party joint ventures. The refined value assessment of behavior-transaction fusion is the first to align video / image behavior data (clicks, dwell time, visual attention, posture / emotion, etc.) with transaction flow in the time domain and input them together into a multimodal CNN, outputting a time-series interaction heat matrix and profit-sharing matrix. This greatly improves the real-time performance, accuracy, interpretability, and security of interactions, breaking through existing technological bottlenecks in multiple dimensions.

[0089] 2. This invention employs an improved activation function to ensure the stability of deep and wide networks. The two-dimensional branch of this invention uses an improved activation function. In the current training batch, the number of feature maps that are recognized as effectively activated in all output feature maps of this convolutional layer is added to the calculation process of the convolutional network. Combined with dynamic effective channel counting, this significantly reduces gradient variance and greatly improves weak texture recall, providing numerical stability for training and inference for large-scale parallel inference.

[0090] 3. This invention employs visualized operational insights. The system displays the revenue-risk distribution of partners across time and geography in real time using 3D heatmaps, clustering diagrams, and KPI dashboards. It supports one-click sandbox simulation and A / B testing, improving operational decision-making efficiency by over 40%. High scalability and low maintenance are achieved through a message queue + containerized architecture supporting concurrent access from thousands of partners and transaction throughput of tens of billions. Activation functions and model parameters can be fine-tuned online, reducing the costs of retraining and reconfiguration due to business surges. A unified DSL, model library, and visualization components can be quickly replicated to multiple industry scenarios such as e-commerce, live streaming, digital content, and financial revenue sharing, providing a standardized and low-cost technical foundation for enterprise ecosystem expansion. It overcomes existing technical bottlenecks in multiple dimensions, including improved system real-time performance, accuracy, and security, demonstrating significant technological advancement.

[0091] The above provides a detailed description of a joint-venture ecosystem intelligent profit-sharing system and method based on interface-based large-scale integration. Specific examples have been used to illustrate the principles and implementation methods of this invention. The descriptions of the above embodiments are only for the purpose of helping to understand the core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas and methods of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. A joint-venture ecosystem intelligent profit-sharing method based on large-scale interface integration, characterized in that: Including the following steps: S1: Register the partner and describe the partner interface based on the Unified Meta Protocol DSL; S2: Automatically generate a sub-adaptor based on the DSL and complete the interface connection; S3: Collect transaction data and acquire user behavior video streams through image acquisition devices; S4: Perform time alignment, anomaly removal, and feature encoding on the transaction data and 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, which outputs a temporal interaction heat matrix and a distribution matrix; S6: Display the profit-sharing matrix and time-series interactive heat matrix in the operations dashboard based on heatmaps, clustering graphs, and KPI dashboards; 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 perform cross-frame tracking of the target to generate user dwell time and behavior sequences; extracting user pose key points through OpenPose to determine interest level; and obtaining structured user behavior events. The multimodal convolutional neural network includes parallel two-dimensional convolutional branches and one-dimensional convolutional temporal 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; The two-dimensional convolutional branch employs an improved activation function after the convolutional layer. : ， in, This is the output of the convolutional layer, where e is the base of the natural logarithm. Let be the first trainable scaling factor, where 0.5 ≤ α ≤ 3.0; Let be the trainable second scaling factor, where 0 ≤ ρ ≤ 1.0; Let β be the trainable third scaling factor, where 0.5 ≤ β ≤ 5.0; To determine the number of validly activated feature maps in the current training batch across all output feature maps of this convolutional layer, we first calculate the average absolute value of the activation values ​​of all samples and their spatial locations within each output feature map. If this average absolute value exceeds a pre-set or online-learned activation threshold, the feature map is considered validly activated. We then count and sum the activation values ​​of all feature maps that meet this condition to obtain u. The maximum number of network channels is preset.

2. The joint venture ecosystem intelligent profit-sharing method based on interface-based large-scale integration as described in claim 1, characterized in that, The time-series interaction heat matrix has an element E(i,t) that represents the weighted interaction intensity of cooperative unit i in time slice t. The interaction intensity is obtained by weighting the number of clicks, dwell time, and visual attention, and is used to drive the online fine-tuning of the three-dimensional heat map and model of the operation cockpit.

3. A joint-venture ecosystem intelligent profit-sharing system based on large-scale interface integration, characterized in that: include: The partner registration and interface description module registers partners and describes their interfaces based on the Unified Meta Protocol (DSL). The sub-adaptor generation and interface docking module automatically generates sub-adaptors and completes interface docking based on the DSL. The transaction data acquisition and user behavior video capture module collects transaction data and acquires user behavior video streams through an image acquisition device. The data alignment and feature encoding preprocessing module performs time alignment, anomaly removal, and feature encoding on the transaction data and the structured user behavior events extracted from the user behavior video stream to obtain a preprocessed data tensor. The multimodal convolutional neural network inference module inputs preprocessed data tensors into the multimodal convolutional neural network, which outputs a temporal interaction heat matrix and a distribution matrix. The operation cockpit visualization module displays the profit-sharing matrix and time-series interactive heat matrix based on heat maps, cluster maps, and KPI dashboards; 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 perform cross-frame tracking of the target to generate user dwell time and behavior sequences; extracting user pose key points through OpenPose to determine interest level; and obtaining structured user behavior events. The multimodal convolutional neural network includes parallel two-dimensional convolutional branches and one-dimensional convolutional temporal 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; The two-dimensional convolutional branch employs an improved activation function after the convolutional layer. : ， in, This is the output of the convolutional layer, where e is the base of the natural logarithm. Let be the first trainable scaling factor, where 0.5 ≤ α ≤ 3.0; Let be the trainable second scaling factor, where 0 ≤ ρ ≤ 1.0; Let β be the trainable third scaling factor, where 0.5 ≤ β ≤ 5.0; To determine the number of validly activated feature maps in the current training batch across all output feature maps of this convolutional layer, we first calculate the average absolute value of the activation values ​​of all samples and their spatial locations within each output feature map. If this average absolute value exceeds a pre-set or online-learned activation threshold, the feature map is considered validly activated. We then count and sum the activation values ​​of all feature maps that meet this condition to obtain u. The maximum number of network channels is preset.

4. The joint-venture ecosystem intelligent profit-sharing system based on interface-based large-scale integration as described in claim 3, characterized in that, The time-series interaction heat matrix has an element E(i,t) that represents the weighted interaction intensity of cooperative unit i in time slice t. The interaction intensity is obtained by weighting the number of clicks, dwell time, and visual attention, and is used to drive the online fine-tuning of the three-dimensional heat map and model of the operation cockpit.

Images