An intelligent product arrangement system based on live conversion data

Abstract

This invention discloses an intelligent product ranking system based on live streaming conversion data, specifically relating to the field of data processing technology. It includes a real-time event acquisition module that connects to the live streaming platform's message channel, extracts user behavior events, and encapsulates them into an event log stream; a behavior sequence encoding module that aggregates user behavior sequences according to session windows, fusing exponential decay weights and positional encoding to form a spatiotemporal behavior tensor; a multi-task conversion prediction module that outputs add-to-cart and order probabilities based on a deep interest network, obtaining a comprehensive conversion potential score after Wilson interval correction; a constrained product ranking queue generation module that combines return rate reduction and triple constraints, outputting the optimal display queue through heuristic beam search; and a closed-loop feedback adaptive module that monitors real-time conversion deviations and uses an elastic weight consolidation algorithm to incrementally update the deep interest network online. This invention dynamically adapts to the real-time conversion environment of live streaming, improving product ranking accuracy and sales volume.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and more specifically, to an intelligent ranking system based on live streaming conversion data. Background Technology

[0002] With the rapid development of live-streaming e-commerce, the rational arrangement of product display order in live-streaming rooms has become a key factor affecting users' purchasing decisions and the total transaction amount. How to effectively predict the conversion potential of products based on users' real-time behavior and generate the optimal product display sequence under multiple constraints such as inventory and product variety is a technical problem that urgently needs to be solved in the current live-streaming e-commerce field.

[0003] Current technical solutions obtain user interests and live stream state vectors by parsing the live stream context, calculate the matching degree in a cross-modal semantic space, predict the role of the product recommendation function, and generate a recommendation list based on multi-objective optimization and dynamic weight allocation mechanisms. Alternatively, they obtain feedback by pushing live stream preview information to user terminals, start the live stream and interact with user terminals to obtain interaction data, filter product-related feature terms from the interaction data to obtain user preference information, and then prompt the broadcaster with this preference information to adjust the product push order. Other solutions collect user historical data, product information, etc., extract user preferences and needs, and push products after feature parameter association and matching.

[0004] However, in practical use, these solutions still have some drawbacks. For example, most rely solely on users' static preferences or historical behavior data for product recommendations, failing to fully explore the dynamic interest evolution information in the real-time behavioral sequences generated during live streaming. They also lack differentiated modeling of shallow interaction behavior and deep conversion behavior, resulting in insufficient accuracy in predicting conversion potential. Secondly, existing solutions generally do not consider supply chain constraints such as product inventory quantity and return rate when generating recommendation lists, nor do they consider ranking constraints such as category diversity and the continuous exposure limit for a single product. This can easily lead to overselling risks, increased return losses, and user visual fatigue. Furthermore, existing solutions mostly generate recommendation lists offline and statically, lacking closed-loop monitoring and dynamic re-ranking mechanisms for real-time conversion effects, making it impossible to adjust subsequent ranking strategies in a timely manner based on deviations in actual conversion data during live streaming. Finally, when faced with new samples, the recommendation models in existing solutions typically employ either full retraining or simple fine-tuning. The former is computationally expensive and difficult to meet real-time requirements, while the latter can easily lead to catastrophic forgetting of historical knowledge, making it difficult to maintain stable predictive performance in continuous learning scenarios. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide an intelligent product ranking system based on live streaming conversion data, which solves the problems mentioned in the background art through the following solutions.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an intelligent product ranking system based on live streaming conversion data, comprising a real-time event acquisition module: accessing the real-time message channel of the live streaming platform, extracting user identifiers, product identifiers, event types and timestamps, and encapsulating the events into an event log stream;

[0007] Behavior sequence encoding module: Receives event log stream, aggregates and generates user behavior sequence according to user session window, calculates shallow behavior frequency vector and deep conversion indicator vector for each user-product pair, applies exponential decay weight to each event in the sequence according to the occurrence time, and forms a spatiotemporal behavior tensor after fusing location encoding.

[0008] Multi-task conversion prediction module: Input the spatiotemporal behavior tensor into the deep interest network, and obtain the product-level comprehensive conversion potential score by Wilson interval correction based on the current real-time exposure of the product.

[0009] Constrained product queue generation module: Reads the comprehensive conversion potential score, product inventory quantity, return rate parameters and category tags, reduces the comprehensive conversion potential score to the net conversion index according to the return rate, and determines the product arrangement subsequence through heuristic beam search under the preset triple constraints and outputs it as the next round of display queue. The product arrangement subsequence satisfies the maximization of the total expected transaction amount of all products.

[0010] The closed-loop feedback adaptive module monitors the real-time order quantity of products in the display queue, calculates the real-time order conversion rate, converts the comprehensive conversion potential score into the expected order conversion rate using a preset calibration function, and calculates the cumulative deviation between the real-time order conversion rate and the expected order conversion rate. It also acquires payment events as incremental samples, uses an elastic weight consolidation algorithm to calculate the importance weights of each parameter in the deep interest network to historical tasks, and updates the parameters of the deep interest network online based on the incremental samples and importance weights.

[0011] Preferably, the event types include at least browsing, clicking, adding to cart, placing an order, and payment.

[0012] Preferably, the shallow behavior frequency vector is composed of the weighted cumulative value of the number of views, clicks, add-to-cart times, and total interactions, wherein the weighted cumulative value is obtained by weighted aggregation based on the exponential decay weight assigned to the occurrence time of each event; the deep conversion indicator vector is a three-dimensional vector including add-to-cart indicator, order indicator, and payment indicator, and the value of each dimension is determined according to whether the most recent occurrence time of the corresponding type of event is higher than a preset decay threshold.

[0013] Preferably, the spatiotemporal behavior tensor is obtained by converting each event in the sequence into an embedding vector according to the event type, multiplying it by the time decay weight, adding it to the sinusoidal position code, and finally concatenating the features of the shallow behavior frequency vector and the deep transformation indicator vector after linear mapping.

[0014] Preferably, the deep interest network includes a shared bottom layer, an add-to-cart task tower, and an order placement task tower. The shared bottom layer extracts user interest representations, the add-to-cart task tower outputs add-to-cart probability predictions, and the order placement task tower outputs order placement probability predictions.

[0015] Preferably, the preset triple constraints are the minimum interval constraint between product categories, the maximum number of consecutive exposures for a single product, and the lower limit of inventory.

[0016] Preferably, the heuristic beam search limits the beam width, gradually expands the queue, traverses candidate products at each step and checks whether they violate the preset triple constraints, and retains a limited number of optimal states based on the cumulative expected transaction amount until the queue reaches the preset length.

[0017] Preferably, the cumulative deviation is calculated using a cumulative control chart. When the cumulative deviation of any product exceeds a preset rearrangement trigger threshold, a rearrangement instruction is sent, the remaining undisplayed parts of the queue are cleared, and the heuristic beam search is re-executed.

[0018] Preferably, the elastic weight consolidation algorithm calculates the diagonal values ​​of the Fisher information matrix of each parameter in the deep interest network based on the historical training set as importance weights, and constructs a total loss function including incremental sample prediction loss and regularization term for gradient update during online update.

[0019] The technical effects and advantages of this invention are as follows:

[0020] 1. This invention utilizes a behavior sequence encoding module to deeply mine real-time user behavior sequences. For each user-product pair, it calculates a shallow behavior frequency vector and a deep conversion indicator vector, and combines exponentially decaying weights and sinusoidal positional encoding to generate a spatiotemporal behavior tensor. This fully captures the dynamic evolution of user interests and the differentiated information between different levels of conversion behavior. Compared to existing technologies that rely solely on static preferences or simple behavioral statistics, this invention can more accurately characterize a user's real-time purchase intention in the current live stream session, providing a richer and more accurate feature base for subsequent conversion potential prediction, and significantly improving the accuracy of product conversion potential estimation.

[0021] 2. This invention, through a constrained product queue generation module, strictly satisfies three constraints: minimum interval between product categories, maximum consecutive exposure times for a single product, and minimum inventory level, based on the net conversion index calculated by reducing the comprehensive conversion potential score according to the return rate. Furthermore, it employs a heuristic beam search algorithm to optimize the total expected transaction amount for all products globally. Compared to existing recommendation schemes that do not consider supply chain constraints and display diversity constraints, this invention effectively avoids overselling risks and return losses, while ensuring the diversity and uniformity of product category display, preventing user visual fatigue, and maximizing overall transaction efficiency while meeting actual business constraints.

[0022] 3. This invention establishes a cumulative deviation monitoring mechanism between the real-time order conversion rate and the expected order conversion rate through a closed-loop feedback adaptive module. When the deviation exceeds a preset threshold, a reordering instruction is automatically triggered, and beam search is re-executed for the queues not yet displayed, achieving dynamic adaptive adjustment of the product ranking strategy to the real-time conversion environment. Simultaneously, this invention employs an elastic weight consolidation algorithm to calculate the importance weights of each parameter of the deep interest network to historical tasks, updating regularized parameters only based on incremental samples. This avoids the high computational cost of full retraining and effectively overcomes the catastrophic forgetting of historical knowledge caused by simple fine-tuning, ensuring the model maintains stable predictive performance and rapid response to new patterns throughout continuous learning. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the overall structure of the present invention;

[0024] Figure 2 This is a schematic diagram of the behavior sequence encoding module of the present invention;

[0025] Figure 3 This is a schematic diagram of the multi-task conversion prediction module of the present invention;

[0026] Figure 4 This is a schematic diagram of the constrained product queue generation module of the present invention;

[0027] Figure 5 This is a schematic diagram of the closed-loop feedback adaptive module of the present invention. Detailed Implementation

[0028] 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.

[0029] like Figure 1As shown, an intelligent product ranking system based on live streaming conversion data includes a real-time event acquisition module: it accesses the real-time message channel of the live streaming platform, extracts user identifiers, product identifiers, event types and timestamps, and encapsulates the events into an event log stream.

[0030] It should be noted that the event types include at least browsing, clicking, adding to cart, placing an order, and payment.

[0031] It should be further explained that the live streaming platform provides all user interaction events within the live stream to the outside world via message queues, such as high-throughput message pipelines based on Kafka. The real-time event acquisition module subscribes to live stream event topics and continuously receives event messages in JSON format. Each event message includes at least a user identifier, a product identifier, an event type, and a timestamp. The event types cover five basic behaviors: browsing, clicking, adding to cart, placing an order, and payment.

[0032] This module extracts the above fields from the message. The live session identifier is obtained by directly parsing the message or by querying the session mapping table based on the live room identifier in the message. The session identifier is dynamically assigned by the module according to the user and the preset timeout threshold. If the interval between adjacent events of the same user exceeds the threshold, a new session identifier is assigned. Otherwise, the current identifier is used. The above identifiers and the extracted fields are encapsulated together to form a standard event log stream.

[0033] To ensure the accuracy of time semantics, all timestamps are uniformly converted to UTC millisecond precision and written to the internal streaming queue in strict order according to the event occurrence time. The event log stream also includes a snapshot of the product's basic attributes, such as category tags, price, and current inventory quantity, which is obtained by periodically refreshing the bypass cache.

[0034] like Figure 2 As shown, the behavior sequence encoding module receives the event log stream, aggregates it according to the user session window to generate a user behavior sequence, calculates the shallow behavior frequency vector and the deep conversion indicator vector for each user-product pair, and applies an exponentially decaying weight to each event in the sequence according to the occurrence time, and forms a spatiotemporal behavior tensor after fusing the position encoding.

[0035] It should be specifically noted that the shallow behavior frequency vector is composed of the weighted cumulative value of the number of views, clicks, add-to-cart times, and total interactions. The weighted cumulative value is obtained by weighting and aggregating the exponential decay weights assigned to the occurrence time of each event. The deep conversion indicator vector is a three-dimensional vector including add-to-cart indicators, order indicators, and payment indicators. The value of each dimension is determined based on whether the most recent occurrence time of the corresponding type of event is higher than a preset decay threshold.

[0036] The spatiotemporal behavior tensor is obtained by converting each event in the sequence into an embedding vector according to the event type, multiplying it with the time decay weight, adding it with the sinusoidal position code, and finally concatenating the features of the shallow behavior frequency vector and the deep transformation indicator vector after linear mapping.

[0037] It should be further explained that a session window is defined as the continuous interactive operations of the same user in the live broadcast room. In this embodiment, if the time interval between two adjacent events exceeds 30 minutes, the session is considered to have ended, and a new event will start the next session. For each session, the behavior sequence encoding module first sorts the products interacted with by the user by time, using the user as the key, to generate a user-product granular behavior sequence, and calculates two types of complementary behavioral features for each user-product pair: a shallow behavior frequency vector and a deep conversion indicator vector.

[0038] The shallow behavior frequency vector characterizes the density of shallow user interactions with products, and consists of the following statistical components: number of pageviews. Number of clicks Number of times add to cart and total number of interactions .

[0039] To differentiate the importance of recent behaviors, each interaction event is assigned an exponentially decaying weight based on its occurrence time when being counted for frequency. Specifically, for the current time T, the time decay weight of the i-th past event is defined as follows: ,in It is the attenuation factor. It is the timestamp of the event. The value range is (0,1], and the more recent the event, the higher the value. The closer it is to 1. Based on this, weighted aggregation is performed on different event types to obtain the weighted cumulative value of each behavior: , , , That is, the weighted cumulative value of pageviews. The sum of the decay weights of all browsing events, plus the weighted cumulative value of clicks. This is the sum of the decay weights for all click events, and so on. This forms a multi-dimensional shallow behavior frequency vector: . The larger the value of each component, the higher the frequency and the more recent the occurrence of the corresponding behavior, thus giving higher feature representation to recent behaviors.

[0040] The deep conversion indicator vector focuses on reflecting whether a user has engaged in high-intent conversion behavior and the timeliness of these behaviors. This embodiment constructs the deep conversion indicator vector as a binary or multi-valued vector of length 3. .in, This indicates whether the user has added the product to their cart. If the time decay weight of the most recent addition to the cart is higher than a preset threshold, it is set to 1; otherwise, it is set to 0. , These indicate whether the user has placed an order or made payment for the product. , and The same logic is used to determine whether an order has been placed and payment has been made. In this way, the deep conversion indicator vector not only reflects the type of behavior, but also implicitly assesses the timeliness of the behavior; old conversion behaviors that have become ineffective will be considered invalid as they decay over time.

[0041] To further preserve the temporal and locational information of the behaviors, this module performs embedding and location encoding on each event in the user-product behavior sequence, and fuses it with the aforementioned frequency and transformation vectors to generate a spatiotemporal behavior tensor. Specifically, the nth event in the event sequence is first converted into a d-dimensional embedding vector according to its event type. Simultaneously, time decay weights are calculated for events. ,Will and Element-wise multiplication yields a time-aware representation of events.

[0042] Specifically, this embodiment introduces a sinusoidal position code PE(pos), where pos is the absolute position of the event in the sequence:

[0043]

[0044]

[0045] In the formula, Let d be the dimension index and d be the embedding layer dimension. The time-aware event representation is added to the location encoding to obtain a sequence representation with spatiotemporal information. Then, the sequence representations of all events are stacked into a matrix, and a shallow behavior frequency vector is appended to the first or last row. and deep transformation indicator vector The features, after linear mapping, ultimately form a fixed-size spatiotemporal behavior tensor. This design enables the model to capture users' dynamic interests from sequence evolution, while also perceiving behavioral intensity and phased transitions from statistical features.

[0046] like Figure 3 As shown, the multi-task conversion prediction module inputs the spatiotemporal behavior tensor into the deep interest network, and obtains the product-level comprehensive conversion potential score by using Wilson interval correction based on the current real-time exposure of the product.

[0047] It should be specifically noted that the deep interest network includes a shared bottom layer, an add-to-cart task tower, and an order placement task tower. The shared bottom layer extracts user interest representations, the add-to-cart task tower outputs add-to-cart probability predictions, and the order placement task tower outputs order placement probability predictions.

[0048] It should be further explained that the multi-task conversion prediction module incorporates a deep interest network, which consists of a shared underlying layer, an add-to-cart task tower, and an order placement task tower. The shared underlying layer employs multi-layer fully connected layers and a self-attention mechanism, drawing from the input spatiotemporal behavior tensor... Extracting the user's deep interest representation vector Both the add-to-cart task tower and the order task tower receive corresponding interest representations, which are then processed through independent fully connected layers and Sigmoid activation, respectively outputting the predicted add-to-cart probability Pa and the predicted order probability Po.

[0049] After obtaining the predicted probability, this module introduces the Wilson interval correction method. Let the real-time exposure of product m at the current moment be denoted as . This refers to the cumulative number of times a product is shown to users during a live stream session, which predicts the probability of placing an order. For a confidence level of 1-α, select the quantiles of the standard normal distribution. If z = 1.96 when α = 0.05, then the formula for calculating the lower bound of the Wilson interval for this commodity is:

[0050]

[0051] The Wilson lower bound adjusted score automatically decreases when exposure is low, reflecting a penalty for prediction uncertainty and preventing the inclusion of a large number of unverified products in the queue, thus avoiding risks to total sales. Meanwhile, the overall conversion potential score can incorporate add-to-cart conversion probability as a supplementary factor, for example, by... The product's overall conversion potential score is obtained by weighting the Pa score by 0.1. .

[0052] like Figure 4 As shown, the constrained product queue generation module reads the comprehensive conversion potential score, product inventory quantity, return rate parameters, and category tags. It then reduces the comprehensive conversion potential score to a net conversion index based on the return rate. Under the condition of satisfying the preset triple constraints, it determines the product arrangement subsequence through heuristic beam search and outputs it as the next round of display queue. The product arrangement subsequence satisfies the maximization of the total expected transaction amount of all products.

[0053] It should be specifically noted that the preset triple constraints are the minimum interval constraint between product categories, the maximum number of consecutive exposures for a single product, and the lower limit of inventory.

[0054] The heuristic beam search limits the beam width, gradually expands the queue, traverses candidate products at each step and checks whether they violate the preset triple constraints, and retains a limited number of optimal states based on the cumulative expected transaction amount until the queue reaches the preset length.

[0055] It should be further explained that the constrained product ranking queue generation module first reduces the overall conversion potential score to a net conversion index based on the return rate. Specifically, for product m, the historical return rate of m is known. , It can be dynamically updated, net conversion index Read the real-time inventory and category tags of the products.

[0056] The product placement process strictly adheres to three pre-defined constraints: a minimum interval constraint between product categories, requiring that within the same display queue, the distance between any two products belonging to the same category at their display position indices is no less than the minimum interval. For example, in this embodiment The reason for doing this in the specific implementation is to ensure product variety and prevent user visual fatigue; the maximum number of consecutive exposures for a single product is limited, restricting the same product from being displayed continuously for more than [number missing]. For example, in this embodiment The reason for doing this in the specific implementation process is to avoid over-harvesting of a single product; the lower limit of inventory constraint stipulates that the total number of times a product is arranged in the queue shall not exceed the current inventory quantity, and when the inventory is less than 2 units, it shall not be queued again to prevent overselling.

[0057] This embodiment designs a heuristic beam search algorithm. Let the length of the display queue be M, and the beam width be B; in this embodiment, B=5. The search process proceeds by progressively expanding the queue. Initially, the beam contains only an empty queue state, with an expected total transaction amount of 0. In step t (t ranges from 1 to M), for each state in the beam, considering the currently displayed queue of length t-1 and the consumed inventory, all candidate products are traversed, and the following checks are performed sequentially: if the product is added, does it violate the minimum interval constraint for the category (verified by recording the most recent position of each category), does it violate the continuous exposure limit (checking the number of consecutive identical products at the tail of the queue), and is the inventory sufficient? For products that pass the constraint checks, the cumulative expected total transaction amount is calculated if it is placed at position t. The expected total transaction amount is calculated based on the product price and the expected order probability after calibration mapping. Specifically, the calibration function g(·) trained using historical data is used to calculate the net conversion index. The expected order conversion rate is mapped to the product price, which gives the expected transaction amount per exposure. This amount is then added to the total amount in the current state to form a new state.

[0058] After completing all feasible expansions, sort the queues by cumulative expected transaction amount from largest to smallest, retain only the top B optimal states to proceed to the next search step, and when the queue length reaches M, select the complete queue with the largest cumulative expected transaction amount as the next round of display queue output.

[0059] like Figure 5 As shown, the closed-loop feedback adaptive module monitors the real-time order quantity of products in the display queue, calculates the real-time order conversion rate, converts the comprehensive conversion potential score into the expected order conversion rate using a preset calibration function, and calculates the cumulative deviation between the real-time order conversion rate and the expected order conversion rate; it acquires payment events as incremental samples, uses an elastic weight consolidation algorithm to calculate the importance weights of each parameter in the deep interest network to historical tasks, and updates the parameters of the deep interest network online based on the incremental samples and importance weights.

[0060] It should be specifically noted that the cumulative deviation is calculated through the cumulative sum control chart. When the cumulative deviation of any product exceeds the preset rearrangement trigger threshold, a rearrangement instruction is sent, the remaining part of the queue that is not displayed is cleared, and the heuristic beam search is re-executed.

[0061] The elastic weight consolidation algorithm calculates the diagonal values ​​of the Fisher information matrix of each parameter in the deep interest network based on the historical training set as importance weights. During online updates, it constructs a total loss function including incremental sample prediction loss and regularization term for gradient updates.

[0062] It should be further explained that the closed-loop feedback adaptive module calculates the real-time order conversion rate for each displayed product. :

[0063]

[0064] At the same time, a pre-defined calibration function is used to score the overall conversion potential of the product. Mapped to expected order conversion rate The calibration function can be pre-trained using ordinal-preserving regression or Platt scaling. For each item, the instantaneous deviation is calculated. The cumulative deviation is calculated using a Cumulative Sum (CUSUM) control chart. Specifically, the cumulative deviation statistic... The update method is to approximate ,in This is the allowable slight drift amount. The cumulative deviation of any given item... When the preset reordering trigger threshold is exceeded, the module determines that there is a significant mismatch between the current product queue and the real-time conversion environment, and then sends a reordering instruction to the constrained product queue generation module. After receiving the reordering instruction, the product queue module will clear the remaining products in the queue that have not yet been displayed (the part that has been locked in the queue but has not been exposed), re-execute the heuristic beam search, and dynamically adjust the subsequent display order.

[0065] Simultaneously, this module undertakes the online update task of the deep interest network. Whenever a payment event is received, the system associates the payment event with the previous behavior sequence to form a complete conversion sample. To avoid catastrophic forgetting caused by directly fine-tuning with new samples, this embodiment employs the Elastic Weight Consolidation (EWC) algorithm for incremental learning. First, after offline pre-training of the deep interest network, each parameter is calculated using the historical training set. Weight of importance for historical tasks The following can be obtained by approximating the diagonal of the Fisher information matrix:

[0066]

[0067] in Represents mathematical expectation, For historical task losses, during the online update phase, a total loss function is constructed for newly arriving incremental sample batches. for:

[0068]

[0069] in This is the prediction loss for incremental samples under the current model (e.g., binary cross-entropy). This represents the total number of parameters in a deep interest network. These are the model parameter values ​​corresponding to historical tasks. To balance the importance of new and old tasks, the hyperparameters are minimized. Gradient updates impose strong regularization constraints on parameters that are important for historical tasks, allowing only minor adjustments in directions that benefit the new task, while less important parameters can flexibly adapt to the new pattern.

[0070] Secondly: The accompanying drawings of the embodiments disclosed in this invention only involve the structures involved in the embodiments of this disclosure. Other structures can refer to the general design. In the absence of conflict, the same embodiment and different embodiments of this invention can be combined with each other.

[0071] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An intelligent product ranking system based on live streaming conversion data, characterized in that, include: Real-time event acquisition module: Connects to the live streaming platform's real-time message channel, extracts user identifiers, product identifiers, event types, and timestamps, and encapsulates the events into an event log stream; Behavior sequence encoding module: Receives event log stream, aggregates and generates user behavior sequence according to user session window, calculates shallow behavior frequency vector and deep conversion indicator vector for each user-product pair, applies exponential decay weight to each event in the sequence according to the occurrence time, and forms a spatiotemporal behavior tensor after fusing location encoding. Multi-task conversion prediction module: Input the spatiotemporal behavior tensor into the deep interest network, and obtain the product-level comprehensive conversion potential score by Wilson interval correction based on the current real-time exposure of the product. Constrained product queue generation module: Reads the comprehensive conversion potential score, product inventory quantity, return rate parameters and category tags, reduces the comprehensive conversion potential score to the net conversion index according to the return rate, and determines the product arrangement subsequence through heuristic beam search under the preset triple constraints and outputs it as the next round of display queue. The product arrangement subsequence satisfies the maximization of the total expected transaction amount of all products. The closed-loop feedback adaptive module monitors the real-time order quantity of products in the display queue, calculates the real-time order conversion rate, converts the comprehensive conversion potential score into the expected order conversion rate using a preset calibration function, and calculates the cumulative deviation between the real-time order conversion rate and the expected order conversion rate. It also acquires payment events as incremental samples, uses an elastic weight consolidation algorithm to calculate the importance weights of each parameter in the deep interest network to historical tasks, and updates the parameters of the deep interest network online based on the incremental samples and importance weights.

2. The intelligent product ranking system based on live streaming conversion data according to claim 1, characterized in that: The event types include at least browsing, clicking, adding to cart, placing an order, and payment.

3. The intelligent product ranking system based on live streaming conversion data according to claim 1, characterized in that: The shallow behavior frequency vector is composed of the weighted cumulative value of the number of views, clicks, add-to-cart times, and total interactions. The weighted cumulative value is obtained by weighting and aggregating the exponential decay weights assigned to the occurrence time of each event. The deep conversion indicator vector is a three-dimensional vector including add-to-cart indicators, order indicators, and payment indicators. The value of each dimension is determined based on whether the most recent occurrence time of the corresponding type of event is higher than a preset decay threshold.

4. The intelligent product ranking system based on live streaming conversion data according to claim 1, characterized in that: The spatiotemporal behavior tensor is obtained by converting each event in the sequence into an embedding vector according to the event type, multiplying it with the time decay weight, adding it with the sinusoidal position code, and finally concatenating the features of the shallow behavior frequency vector and the deep transformation indicator vector after linear mapping.

5. The intelligent product ranking system based on live streaming conversion data according to claim 1, characterized in that: The deep interest network includes a shared bottom layer, a add-to-cart task tower, and an order task tower. The shared bottom layer extracts user interest representations, the add-to-cart task tower outputs add-to-cart probability predictions, and the order task tower outputs order placement probability predictions.

6. The intelligent product ranking system based on live streaming conversion data according to claim 1, characterized in that: The preset triple constraints are the minimum interval constraint between product categories, the maximum number of consecutive exposures for a single product, and the lower limit of inventory.

7. The intelligent product ranking system based on live streaming conversion data according to claim 1, characterized in that: The heuristic beam search limits the beam width, gradually expands the queue, traverses candidate products at each step and checks whether they violate the preset triple constraints, and retains a limited number of optimal states based on the cumulative expected transaction amount until the queue reaches the preset length.

8. The intelligent product ranking system based on live streaming conversion data according to claim 1, characterized in that: The cumulative deviation is calculated using a cumulative sum control chart. When the cumulative deviation of any product exceeds a preset rearrangement trigger threshold, a rearrangement command is sent, the remaining undisplayed parts of the queue are cleared, and the heuristic beam search is re-executed.

9. The intelligent product ranking system based on live streaming conversion data according to claim 1, characterized in that: The elastic weight consolidation algorithm calculates the diagonal values ​​of the Fisher information matrix of each parameter in the deep interest network based on the historical training set as importance weights, and constructs a total loss function including incremental sample prediction loss and regularization term for gradient update during online update.

Images