A method and system for regulating cargo flow based on cancellation rate prediction

Abstract

The application discloses a kind of based on cancellation rate prediction's cargo source flow regulation method and system, it is related to the development technical field of material flow platform, including: by CTR tower, CVR tower and CCR tower, construct three tower architecture, wherein, CTR tower is used to predict exposure click rate, CVR tower is used to predict click transaction rate, CCR tower is used to predict cancellation rate;Based on three tower architecture, obtain the shared features between each tower and unique feature, carry out weighted fusion processing, and introduce loss function when training, by executing gradient balance strategy, to predict cancellation rate, regulate cargo source flow.The application avoids invalid cargo exposure, saves driver time cost, improves driver platform use experience and trust degree, realizes that freight cargo source recommendation is from " pursues transaction quantity " to " guarantee real performance effect " core upgrade.

Description

Technical Field

[0001] This invention relates to the field of freight logistics platform development technology, and more specifically, to a method and system for regulating cargo flow based on cancellation rate prediction. Background Technology

[0002] In the freight matching business of freight logistics platforms, the core objective of the recommendation system is to assist drivers in quickly matching suitable freight and completing transportation fulfillment. The driver's operation behavior forms a four-stage conversion funnel of exposure, click, transaction, and fulfillment. Currently, the mainstream industry adopts the ESMM dual-tower multi-task model to realize freight recommendation ranking. This model predicts the exposure click-through rate through the CTR tower and the click-to-transaction rate through the CVR tower. The exposure-to-transaction rate is obtained by multiplying the two and used as the ranking basis. This effectively solves the sample selection bias problem of traditional conversion rate models in exposure space training and can optimize the conversion link from exposure to transaction. Therefore, it is widely used in this scenario. However, this model has obvious technical defects in actual business implementation and cannot adapt to the full-link business needs. On the one hand, existing ESMM models only focus on the initial conversion stage from exposure to transaction, failing to model the fulfillment and order cancellation stages after the transaction. In the freight business, over 20% of completed orders are cancelled due to reasons such as inaccurate cargo information, driver no-shows, shipper cancellations, and price disputes. This leads the model to prioritize cargo with high completion rates but high cancellation risks, resulting in inflated transaction volume and stagnant actual fulfillment volume, deviating from the platform's core fulfillment optimization goals. On the other hand, the model's lack of cancellation risk identification capabilities causes it to prioritize high-cancellation-risk cargo, wasting drivers' time and energy and reducing their trust in the platform and their willingness to use it. Furthermore, the traditional dual-tower architecture cannot adapt to full-link modeling of the four-stage conversion funnel. Fulfillment and cancellation behaviors only exist in completed transaction samples; directly training the model will lead to severe sample selection bias, and the number of completed transaction samples is far less than the number of exposure samples, resulting in data sparsity and significantly reducing the model's predictive accuracy and stability. In summary, existing technologies suffer from several technical pain points, including failure to cover the fulfillment cancellation process, inability to identify high-risk sources of goods, large biases in end-to-end modeling, data sparsity, and a disconnect between recommendation effects and fulfillment goals. There is an urgent need to design a new multi-task model architecture and supporting traffic control strategies to achieve unified modeling of the four-stage conversion funnel, suppress the recommendation of high-cancellation-risk sources of goods, and improve the platform's exposure fulfillment rate and user experience. Summary of the Invention

[0003] To address the aforementioned issues, the present invention aims to provide a method and system for regulating cargo flow based on cancellation rate prediction, thereby overcoming the technical shortcomings of existing ESMM cargo recommendation models that only optimize the transaction process, do not consider the risk of contract cancellation, and cannot model the entire supply chain.

[0004] To achieve the above technical objectives, this application provides a method for regulating cargo flow based on cancellation rate prediction, comprising the following steps: A three-tower architecture is constructed using the CTR tower, CVR tower, and CCR tower. The CTR tower is used to predict the click-through rate, the CVR tower is used to predict the click-to-conversion rate, and the CCR tower is used to predict the cancellation rate. Based on a three-tower architecture, shared and unique features between each tower are obtained, weighted and fused, and a loss function is introduced during training. By implementing a gradient balancing strategy, the cancellation rate is predicted and the flow of goods is regulated.

[0005] Preferably, when acquiring shared features, a portion of common features shared among the three towers are used as shared features, wherein the shared features include: driver profile, cargo attributes, context, and driver-cargo matching features.

[0006] Preferably, when acquiring unique features, the features that affect whether a click occurs are taken as the unique features of the CTR tower. The unique features of the CTR tower include: exposure position, display style, historical and real-time exposure and click volume. The unique features of CVR towers are the characteristics that affect whether a transaction is completed. These unique features include: the matching degree of the source driver tag information and the matching degree of loading and unloading time. The unique features of CCR Tower include: the fulfillment history of shippers and drivers, the completeness of cargo information, the cancellation rate of cooperation, and the sequence of cargo cancelled by drivers in their history.

[0007] Preferably, during weighted fusion processing, a common representation is learned through a shared expert network to obtain shared features; By leveraging the expert networks dedicated to each tower to learn differentiated representations, unique features can be obtained. A gating mechanism is used to perform weighted fusion of shared and unique features, which preserves shared knowledge while preventing unique information from being interfered with by other towers.

[0008] Preferably, when designing the loss function, a loss function is constructed based on the supervised CTR tower loss, the supervised CTR×CVR joint prediction loss, the supervised CTR×CVR×(1-cancellation rate) joint prediction loss, and the supervised CCR tower loss.

[0009] Preferably, when implementing the gradient balancing strategy, if the number of transaction samples is much smaller than the number of exposure samples, the transaction samples are lightly upsampled and a higher initial weight is assigned to the supervised CCR tower loss.

[0010] Preferably, when implementing the gradient balancing strategy, the gradient of the CCR tower is propagated twice through the supervised joint prediction loss of CTR×CVR×(1-cancellation rate), and its value is proportional to the exposure click-through rate × click-to-conversion rate.

[0011] Preferably, when implementing the gradient balancing strategy, the loss of the supervised CCR tower is multiplied by a compensation coefficient = 1 / (average exposure click-through rate × average click-through conversion rate) to compensate for the chain decay effect.

[0012] Preferably, when implementing cargo flow control, a maximum threshold and a higher threshold for cancellation rate are set, and the predicted cancellation rate is compared with the maximum threshold and the higher threshold respectively, and cargo flow control is performed based on the comparison results.

[0013] Based on the same inventive concept, this invention discloses a cargo flow control system based on cancellation rate prediction, comprising: The model building module is used to construct a three-tower architecture using the CTR tower, CVR tower, and CCR tower. The CTR tower is used to predict the click-through rate, the CVR tower is used to predict the click-to-conversion rate, and the CCR tower is used to predict the cancellation rate. The flow control module is used to obtain the shared and unique features between each tower based on the three-tower architecture, perform weighted fusion processing, introduce a loss function during training, and regulate the flow of goods by executing a gradient balancing strategy.

[0014] The present invention discloses the following technical effects: This invention avoids exposing invalid cargo listings, saves drivers' time costs, improves the user experience and trustworthiness of the driver platform, and achieves a core upgrade in freight listing recommendations from "pursuing the number of transactions" to "ensuring genuine fulfillment results". Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments 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.

[0016] Figure 1 This is a schematic diagram of the method described in this invention.

[0017] Figure 2 This is a schematic diagram illustrating the execution process of the method described in this invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely represents selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0019] like Figure 1 As shown, this invention provides a method for regulating the flow of goods based on cancellation rate prediction. By upgrading the model architecture, designing differentiated features, implementing joint supervision of multiple losses and gradient balance optimization, and combining an interpretable and configurable grouping threshold regulation mechanism, it achieves accurate full-space modeling of the three-stage conversion funnel of exposure, click, transaction, and fulfillment, effectively suppressing the recommendation of goods with high cancellation risk and improving the overall fulfillment conversion effect of the platform.

[0020] In one embodiment, the present invention adds a cancellation rate prediction tower (CCR tower) to the traditional ESMM dual-tower model, forming a three-tower architecture. The entire recommendation model consists of three "prediction towers", each tower is responsible for predicting the probability of one link in the conversion funnel.

[0021] For example, the first tower (CTR tower) is used to predict whether "drivers click after seeing the cargo", i.e., the exposure click-through rate, where the tower is trained and supervised on all exposure samples.

[0022] For example, the second tower (CVR tower) is used to predict whether "the driver will accept the order and complete the transaction after clicking on the cargo source", i.e., the click-to-conversion rate. The feature learning of this tower occurs on the click samples, but is jointly supervised in the exposure space by multiplying with the first tower.

[0023] For example, the third tower (CCR tower) is used to predict whether "drivers cancel after accepting and completing an order", i.e., the cancellation rate. The feature learning of this tower occurs on the completed sample, but is jointly supervised in the exposure space through chained product with the first two towers. The fulfillment rate = 1 - cancellation rate.

[0024] For example, the predicted values ​​from the three towers are multiplied by a chain to obtain the final exposure fulfillment rate: Exposure fulfillment rate = Exposure click-through rate × Click-through conversion rate × (1 - Cancellation rate).

[0025] For example, although the cancellation rate is only defined on the completed samples (uncompleted orders are not considered cancelled), the supervision signal of the cancellation rate tower can be fed back to the exposure space through chained product, avoiding the bias and sparsity problems caused by training only on a small subset of completed samples; this inherits the core idea of ​​"full space supervision" of the ESMM model and extends it from a two-funnel to a three-funnel.

[0026] In one implementation, feature design is performed based on the three-tower architecture: the three towers share some common features such as driver profiles, freight attributes, contextual environment, and driver-freight matching features. These shared features enable more thorough training of sparse features through parameter sharing. Simultaneously, each tower also has its own unique features, including: CTR Tower's unique features include: exposure ranking, display style, historical and real-time exposure and click-through rates, and other features that influence whether or not a click is made. CVR Tower's unique features include: matching degree of freight driver tag information, matching degree of loading and unloading time, and other features that affect "whether a deal is made"; CCR Tower's unique features include: the fulfillment history of shippers and drivers, such as historical cancellation rates and default records; the completeness of cargo information, such as address accuracy and contact person validity; and cooperation cancellation rates, which affect "whether or not the contract is fulfilled." In addition, it also includes the sequence features of cargo cancelled by drivers in the past, which is used to capture drivers' cancellation preference patterns.

[0027] For example, feature sharing and uniqueness are achieved through a structure of "shared experts + unique experts + gating fusion": shared features learn a general representation through a set of shared expert networks, unique features learn a differentiated representation through the dedicated expert networks of each tower, and the gating mechanism weights and fuses the two, which preserves shared knowledge and prevents unique information from being interfered with by other towers, thus avoiding negative transfer.

[0028] In one implementation, loss function design is performed based on feature design.

[0029] For example, a four-way supervised loss is used during model training: Loss 1: Monitor the CTR tower in the exposure space, labeled "Click or not"; Loss 2: Monitor the joint prediction of CTR×CVR in the exposure space, labeled "whether a transaction was completed"; Loss 3: Monitor the joint prediction of CTR×CVR×(1-cancellation rate) in the exposure space, labeled as "fulfillment", to directly optimize the ultimate business goal of exposure fulfillment rate; Loss 4: Directly monitor the CCR tower in the transaction space, labeled "Cancel or Not"; among which, the auxiliary loss provides a more direct gradient signal for the CCR tower, avoiding it from obtaining too weak an indirect gradient only through chain multiplication.

[0030] For example, the total loss = weight1 × loss1 + weight2 × loss2 + weight3 × loss3 + weight4 × loss4, where the weights are determined by a dynamic adjustment strategy.

[0031] In one implementation, based on the loss function design, a gradient balancing strategy is executed. Compared to the dual-tower model, the gradient imbalance is more severe in the three-tower model. The present invention designs the following balancing strategy: For example, sample compensation: the number of transaction samples is much smaller than the number of exposure samples, resulting in sparse CCR tower supervision signals. Transaction samples are lightly upsampled (2-3 times), and the auxiliary loss is given a higher initial weight to ensure that the CCR tower obtains sufficient training signals.

[0032] For example, chained gradient compensation: the gradient of the cancellation rate tower of the main target loss is propagated through two product passes, and its value is proportional to the exposure click-through rate × click-through conversion rate; when these two values ​​are small (unpopular cargo source - driver matching), the cancellation rate tower receives almost no gradient.

[0033] For example, in this embodiment, the auxiliary loss is multiplied by a compensation coefficient, which is 1 / (average exposure click-through rate × average click-through conversion rate), to compensate for the chain decay effect, and the compensation coefficient is clipped at upper and lower limits to prevent numerical instability.

[0034] In one implementation, a traffic control strategy based on cancellation rate prediction is designed based on the designed three-tower model.

[0035] For example, after the model predicts the cancellation rate, this embodiment designs a traffic control strategy based on statistical thresholds, transforming the model prediction results into adjustment logic for list sorting. This strategy is decoupled from the model prediction; the model is only responsible for scoring, while the control strategy is independently configurable.

[0036] For example, threshold calculation: group according to two dimensions, origin province and vehicle length (vehicle length), and calculate the model prediction score of cancellation rate for all source-driver matching pairs in the past period (such as the last 7 days).

[0037] For example, the percentiles of the scores are calculated within each group: θ1: 98th quantile, representing the threshold at which the cancellation rate is predicted to be extremely high in this group; θ2: 65th quantile, representing the threshold for a higher predicted cancellation rate in this group.

[0038] For example, the sorting adjustment rule is as follows: When sorting the list, after each shipment is scored by the cancellation rate model for the current driver, the corresponding θ1 and θ2 thresholds are queried based on the departure province and vehicle length: 1) If the cancellation rate score is greater than θ1: it is considered that the probability of the cargo being cancelled by the driver is extremely high, and the cargo will be pushed to the bottom of the list, its ranking score will be greatly reduced, and the cargo will be pushed to the end of the list; 2) If the cancellation rate score is greater than θ2 and less than or equal to θ1: it is considered that the probability of cancellation for this driver is high. The driver will be demoted in the list, the ranking score will be appropriately reduced, and the cargo will be placed at the end of the list but not at the end. 3) If the cancellation rate score is ≤ θ2: the goods will be displayed according to the normal sorting logic.

[0039] For example, the advantages of this strategy are: the threshold is dynamically updated based on historical statistical data and can be flexibly adjusted according to different regions and vehicle models; the control logic is decoupled from the model prediction, making it easier for business users to understand and adjust; and the threshold is set by quantiles rather than absolute values, adapting to the differences in cancellation rate distribution in different groups.

[0040] Example: This example is applied to the cargo recommendation scenario of trunk freight logistics. It is aimed at the real-time matching and ranking of massive drivers and cargo on the platform. It fully deploys the three-tower multi-task recommendation model and grouped cancellation risk traffic control strategy proposed in this invention, realizing the upgrade and transformation of cargo recommendation from "optimizing transactions" to "optimizing actual fulfillment". The specific implementation process includes six parts: model construction, feature configuration, model training, gradient balance optimization, traffic control deployment and online application.

[0041] In one embodiment, this example builds a three-tower multi-task recommendation model based on a deep learning framework. It retains the CTR (Click-Through Rate) and CVR (Click-to-Convert Rate) prediction towers from the traditional ESMM model, and adds a CCR (Order Cancellation Rate) prediction tower, forming a three-segment end-to-end modeling architecture. The model's final ranking score uses the exposure fulfillment rate calculation formula: Exposure Fulfillment Rate = Exposure Click-Through Rate × Click-to-Convert Rate × (1 - Cancellation Rate). This uses the actual transportation fulfillment target as the core basis for cargo ranking, replacing the traditional ranking logic based solely on exposure fulfillment rate.

[0042] In one implementation, this embodiment employs a "shared expert + unique expert + gating fusion" structure to complete feature modeling. The shared features across the three towers include common features such as driver basic profiles, cargo category and weight attributes, time and location context, and driver-cargo matching similarity. A shared expert network is used to mine common patterns across various conversion tasks. Each tower is configured with unique features: the CTR tower introduces click preference features such as exposure location, display style, and real-time exposure popularity; the CVR tower introduces transaction preference features such as loading / unloading timeliness matching, freight matching, and route suitability; and the newly added CCR tower focuses on introducing performance risk features such as cargo owner historical cancellation rate, driver historical default records, cargo address accuracy, contact validity, cancellation rate of bilateral cooperation in the past 7 days, and driver historical cargo cancellation sequence characteristics. After the unique features of each tower are learned and differentiated by the dedicated expert network, they are adaptively weighted and fused with the shared representations through a gating mechanism, effectively avoiding the multi-task negative transfer problem.

[0043] In one embodiment, this example employs four-way loss joint supervised training to ensure a full fit of the entire task: First, the CTR click prediction task is supervised on the full exposure sample; second, the CTR×CVR joint transaction prediction task is supervised in the exposure space; third, the CTR×CVR×(1-CCR) joint fulfillment prediction main task is supervised in the exposure space using order fulfillment status as a label; fourth, a CCR cancellation prediction auxiliary loss is added in the transaction sample space to provide a direct training gradient for the cancellation rate prediction branch. During training, a dynamic weight mechanism balances the four-way losses, prioritizing the optimization effect of the fulfillment main task while also considering the prediction accuracy of each sub-task, including clicks, transactions, and cancellations.

[0044] In one embodiment, to address the issues of gradient imbalance, sparse CCR tower samples, and chain gradient decay in the three-tower model, this embodiment simultaneously employs a dual gradient balancing strategy: First, the scarce transaction fulfillment samples are lightly upsampled by 2 times, and the initial weight of the CCR auxiliary loss is increased to alleviate the underfitting problem caused by sample sparsity; Second, a chain gradient compensation coefficient is introduced, and gradient decay is compensated by "1 / (platform average exposure click-through rate × platform average click-through conversion rate)", while upper and lower limits are set for the compensation coefficient to avoid numerical overflow and gradient oscillation during training, ensuring stable convergence of the three-tower model training.

[0045] In one implementation, this embodiment employs an interpretable traffic control mechanism decoupled from the model. It does not interfere with the model training process, but only performs post-processing corrections on the model's ranking results. The specific grouping rules are as follows: using the province of origin and vehicle length as grouping dimensions, historical data from the past 7 days is updated daily. The predicted cancellation rate (CCR) scores for all freight-driver matching pairs within each group are calculated, and the 65th quantile threshold θ2 and the 98th quantile threshold θ1 for each group are calculated. The actual ranking control logic is as follows: when the predicted cancellation rate score > θ1, the freight is judged as having extremely high cancellation risk and is placed at the bottom of the list; when θ2 < cancellation rate score ≤ θ1, the freight is judged as having relatively high cancellation risk and is subject to ranking reduction; when the cancellation rate score ≤ θ2, the freight is displayed normally according to the model's original fulfillment rate score. This strategy achieves differentiated risk management by region and vehicle type, and the control rules are transparent and configurable.

[0046] In one implementation, after the online A / B test of this embodiment was carried out, compared with the baseline of the traditional ESMM dual-tower model, the platform's core indicators were significantly optimized: the overall fulfillment volume of the platform increased by 0.51%, and after the full rollout, the average daily increase in valid fulfillment orders was 136; the overall order cancellation rate decreased by 1.50%, an absolute decrease of 0.35 percentage points. At the same time, the invalid exposure of high-risk cargo sources was greatly reduced, the negative experience of drivers running empty after accepting orders and order cancellations was significantly improved, and the accuracy of platform cargo recommendation and driver retention experience were improved in tandem, verifying the practical effectiveness and business advancement of the three-tower modeling architecture and traffic control strategy of this invention.

[0047] In summary, this invention addresses the shortcomings of existing freight recommendation technologies by constructing a three-tower, multi-task modeling architecture and implementing a quantifiable traffic control strategy, achieving significant technical and business results across multiple dimensions. First, this invention effectively improves platform exposure fulfillment rates and actual fulfilled order volume. By adding a CCR cancellation rate prediction tower, it optimizes the entire link from exposure to fulfillment through modeling, abandoning the one-sidedness of traditional models that only optimize transaction metrics. Simultaneously, combined with traffic control measures such as de-weighting and relegating high-cancellation-risk freight, it shifts the platform's freight recommendation logic from simply pursuing transaction volume to ensuring genuine transportation fulfillment, effectively driving growth in the platform's effective fulfilled order volume. Second, this invention significantly reduces order cancellation rates. Relying on the CCR tower to accurately identify high-cancellation-risk freight, and through a controllable traffic control strategy to avoid pre-exposing high-risk freight, it greatly reduces order cancellations after drivers accept orders, effectively saving drivers' time and effort, and improving driver experience and trust in the platform. Furthermore, this invention significantly improves the overall robustness and prediction accuracy of the model. Through a complete spatial supervision mechanism, it thoroughly solves the sample selection bias problem inherent in traditional subsequent task training. Simultaneously, combined with a gradient balancing strategy, it effectively avoids underfitting issues caused by sparse transaction samples and chain-like gradient decay in the newly added CCR tower, ensuring that the click, transaction, and cancellation tasks can all be fully trained and collaboratively optimized. In addition, the traffic control strategy of this invention is based on statistical quantile thresholds. The control rules are clear and transparent, the logic is interpretable, and the parameters can be flexibly adjusted, overcoming the black-box defects of pure model ranking and facilitating business implementation, iteration, and refined operation. Online A / B testing verified that after the implementation of this invention, the platform's overall fulfillment volume increased by 0.51%, 136 new valid fulfilled orders were added after full implementation, and the overall order cancellation rate decreased by 1.50%, a decrease of 0.35 percentage points. This fully demonstrates that this invention can effectively improve driver fulfillment conversion and reduce order cancellation risk, possessing significant business benefits and technological advancement.

[0048] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0049] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for regulating cargo flow based on cancellation rate prediction, characterized in that, Includes the following steps: A three-tower architecture is constructed using the CTR tower, CVR tower, and CCR tower, where the CTR tower is used to predict the click-through rate, the CVR tower is used to predict the click-to-conversion rate, and the CCR tower is used to predict the cancellation rate. Based on the aforementioned three-tower architecture, shared and unique features between each tower are obtained, weighted fusion processing is performed, and a loss function is introduced during training. By executing a gradient balancing strategy, the cancellation rate is predicted, and the flow of goods is regulated.

2. The method for regulating cargo flow based on cancellation rate prediction according to claim 1, characterized in that: When acquiring shared features, a portion of common features shared among the three towers are used as shared features, which include: driver profile, cargo attributes, context, and driver-cargo matching features.

3. The method for regulating cargo flow based on cancellation rate prediction according to claim 1, characterized in that: When acquiring unique features, the features that affect whether a click occurs are taken as the unique features of the CTR tower. The unique features of the CTR tower include: exposure position, display style, historical and real-time exposure and click volume. The unique features of the CVR tower are the characteristics that affect whether a transaction is completed. These unique features include: matching degree of the source driver tag information and matching degree of loading and unloading time. The CCR Tower uses features that affect contract fulfillment as its unique features. These unique features include: the fulfillment history of cargo owners and drivers, the completeness of cargo information, the cancellation rate of cooperation, and the sequence of cargo cancelled by drivers in their history.

4. The method for regulating cargo flow based on cancellation rate prediction according to claim 1, characterized in that: During weighted fusion processing, a common representation is learned through a shared expert network to obtain the shared features; By utilizing the expert network dedicated to each tower to learn differentiated representations, the unique features are obtained; A gating mechanism is used to perform weighted fusion of the shared features and the unique features, in order to retain shared knowledge while preventing unique information from being interfered with by other towers.

5. The method for regulating cargo flow based on cancellation rate prediction according to claim 1, characterized in that: When designing the loss function, the loss function is constructed based on the supervised CTR tower loss, the supervised CTR×CVR joint prediction loss, the supervised CTR×CVR×(1-cancellation rate) joint prediction loss, and the supervised CCR tower loss.

6. The method for regulating cargo flow based on cancellation rate prediction according to claim 5, characterized in that: When implementing the gradient balancing strategy, if the number of transaction samples is much smaller than the number of exposure samples, the transaction samples are lightly upsampled and a higher initial weight is assigned to the supervised CCR tower loss.

7. The method for regulating cargo flow based on cancellation rate prediction according to claim 5, characterized in that: When implementing the gradient balancing strategy, the gradient of the CCR tower is propagated twice through the supervised joint prediction loss of CTR×CVR×(1-cancellation rate), and its value is proportional to the exposure click-through rate × click-to-conversion rate.

8. The method for regulating cargo flow based on cancellation rate prediction according to claim 7, characterized in that: When implementing the gradient balancing strategy, the loss of the supervised CCR tower is multiplied by a compensation coefficient = 1 / (average exposure click rate × average click-through rate) to compensate for the chain decay effect.

9. The method for regulating cargo flow based on cancellation rate prediction according to claim 1, characterized in that: When implementing cargo flow control, a maximum threshold and a higher threshold for cancellation rate are set. The predicted cancellation rate is compared with the maximum threshold and the higher threshold respectively, and cargo flow control is performed based on the comparison results.

10. A cargo flow control system based on cancellation rate prediction, characterized in that, include: The model building module is used to construct a three-tower architecture using CTR tower, CVR tower, and CCR tower, wherein the CTR tower is used to predict the click-through rate, the CVR tower is used to predict the click-to-conversion rate, and the CCR tower is used to predict the cancellation rate. The flow control module is used to obtain the shared features and unique features between each tower based on the three-tower architecture, perform weighted fusion processing, introduce a loss function during training, and regulate the flow of goods by executing a gradient balancing strategy.

Images