A driver experience combined recommendation ranking method and system

Abstract

The application discloses a driver experience combined recommendation sorting method and system, wherein the method comprises the following steps: constructing a basic multi-expert multi-task deep sorting model composed of an input layer network, an expert layer network, a gate layer network and a task layer, which is used for predicting whether a driver will click a candidate cargo source and whether the driver will transact the candidate cargo source; based on the basic multi-expert multi-task deep sorting model, introducing driver experience related features, adding an experience task, and setting an experience penalty loss function; jointly training the business task in the basic model and the added experience task to construct a business-experience multi-expert network, and cooperatively optimizing the business target and the driver experience. The application effectively improves the driver experience and realizes the cooperative optimization of the business target and the driver experience.

Description

Technical Field

[0001] This invention relates to the field of freight platform development technology, and more specifically, to a recommendation ranking method and system that incorporates driver experience. Background Technology

[0002] In the current recommendation and ranking scenarios of freight platforms, the core optimization goal of existing ranking models focuses on improving business metrics, such as the driver click-through rate after a shipper posts a freight request and the transaction rate of cooperation between drivers and shippers. However, these models often neglect the actual user experience of drivers during the order-taking process, leading to drivers frequently encountering various experience problems.

[0003] Specifically, the pain points for drivers are mainly reflected in the following aspects: First, the recommended orders received by drivers may involve excessively long empty driving distances, meaning the distance between the driver's current location and the loading point is too long, which not only increases the driver's time and fuel costs but also reduces the driver's overall operational efficiency. Second, the net freight rates for some recommended orders are too low, failing to match the driver's labor input, which will affect the driver's income level and willingness to accept orders in the long run. Third, the model does not fully consider the actual situation of the driver's current vehicle and may recommend orders that do not match the vehicle's carrying capacity and vehicle type requirements, such as recommending orders for transporting oversized cargo to drivers of small trucks, causing drivers to be unable to accept orders normally and wasting time for both parties.

[0004] To address these issues, most existing solutions involve introducing driver experience-related factors for a secondary ranking during the re-ranking phase after the initial business-goal-oriented ranking. However, this approach tends to go to two extremes in practice: on the one hand, if excessive emphasis is placed on driver experience optimization during the re-ranking phase, while it may improve the driver experience to some extent, it will significantly deviate from the initial business goals, leading to a substantial decline in core business metrics such as click-through rate and conversion rate, thus harming the platform's overall operational efficiency; on the other hand, if the weight of driver experience in the re-ranking phase is weakened to avoid damaging business metrics, it will fail to effectively improve the driver experience, leaving drivers still facing numerous pain points. In the long run, this could lead to driver attrition, hindering the platform's sustainable development.

[0005] Therefore, there is an urgent need for a recommendation ranking model that can improve business metrics while taking into account driver experience, in order to solve the problem of balancing business objectives and driver experience in existing technologies. Summary of the Invention

[0006] To address the aforementioned issues, the present invention aims to provide a recommendation ranking technology that incorporates driver experience, thereby effectively improving driver experience while ensuring improved business metrics, and achieving synergistic optimization between business objectives and driver experience.

[0007] To achieve the above technical objectives, this application provides a recommendation ranking method that incorporates driver experience, comprising the following steps:

[0008] A basic multi-expert, multi-task deep ranking model is constructed, consisting of an input layer network, an expert layer network, a gating layer network, and a task layer, to predict whether a driver will click on a candidate cargo and whether the driver will complete a transaction with a candidate cargo.

[0009] Based on the basic multi-expert multi-task deep ranking model, by introducing driver experience-related features and adding experience tasks, and setting an experience penalty loss function, the basic model can learn behavioral information with low experience.

[0010] By jointly training the business tasks in the basic model with the newly added experience tasks, a business-experience multi-expert network is constructed to collaboratively optimize business objectives and driver experience.

[0011] Preferably, when constructing the basic multi-expert multi-task deep ranking model, the input layer network is used to obtain cargo owner behavior features, driver profile features, cargo owner profile features, driver-cargo owner cross features, and driver behavior features. After processing, these features are concatenated into a matrix and used as the input to the expert layer network.

[0012] Preferably, when constructing a basic multi-expert multi-task deep ranking model, the inputs are concatenated together, and specific task learning is performed through multiple expert networks in the expert layer network, and different information is transferred and shared between tasks.

[0013] Preferably, when constructing the basic multi-expert multi-task deep ranking model, the output of the task expert network, the output of the shared expert network, and the original input layer information are input into the gating layer corresponding to the expert network as the input of the gating layer network, wherein the gating layer network is composed of multiple gating layers;

[0014] By using a gated network, the information of the task expert network and the shared expert network are concatenated into a list. A picker is used to select the input layer, and the output vector is a vector representing the number of expert networks. After adding softmax, the vector is multiplied by the list vector, and the result is output to the task layer. In the task layer, the gated output layer is added to a multi-layer fully connected layer, and cross-entropy loss is applied to the final label.

[0015] Preferably, when introducing experience-related features, the experience-related features include:

[0016] Empty driving distance related features include the straight-line distance between the driver's current location and the loading location, the actual navigation distance, and the proportion of this distance in the driver's historical empty driving distances.

[0017] Net freight-related characteristics include the order's net freight amount, the ratio of net freight to the driver's historical average net freight, and the ratio of net freight to transportation mileage.

[0018] The driver-cargo matching characteristics represent the degree of matching between the driver's current vehicle load weight and the order cargo weight, the degree of matching between the vehicle's cargo dimensions and the cargo dimensions, and the degree of matching between the vehicle's transportation type and the order cargo transportation requirements.

[0019] Preferably, when adding a new experience task, a new experience behavior recognition task is added to identify whether the order is one that will lead to a low driver experience, including: the empty driving distance exceeds 1.5 times the driver's historical maximum empty driving distance, the net freight cost is less than 0.5 times the driver's historical minimum net freight cost, or the vehicle matching degree is less than 0.3.

[0020] Preferably, during joint training, the total loss function is set as the weighted sum of the loss functions of each task and the experience penalty loss function, and the gradient descent optimization algorithm is used to iteratively update the model parameters until the model converges. This ensures that the model learns the optimization directions of both business objectives and experience objectives during training. The total loss function is expressed as follows:

[0021] Ltotal=β1Lclick+β2Ldeal+β3Lexperience+Lpenalty

[0022] In the formula, Ltotal is the total loss of the model, Lclick is the cross-entropy loss of the click prediction task, Ldeal is the cross-entropy loss of the transaction prediction task, Lexperience is the mean squared error loss of the experience rating prediction task, β1, β2, and β3 are the weight coefficients of the loss of each task, and Lpenalty is the experience penalty loss function.

[0023]

[0024] In the formula, Lpenalty represents the experience penalty loss, α is the penalty weight coefficient, and yi is the true label of whether the order is a low-experience order, where yi=1 indicates yes and yi=0 indicates no. The model predicts the label; I(·) is an indicator function that takes the value 1 when the condition in parentheses is true, and 0 otherwise; ∑: sums up the loss terms for all samples; : Represents the squared error between the true label and the predicted label.

[0025] This invention also discloses a recommendation ranking system that incorporates driver experience, used to implement the aforementioned recommendation ranking method that incorporates driver experience, comprising:

[0026] The basic module construction module is used to build a basic multi-expert multi-task deep ranking model consisting of an input layer network, an expert layer network, a gating layer network, and a task layer. This model is used to predict whether a driver will click on a candidate cargo and whether the driver will complete a transaction with a candidate cargo.

[0027] An improved module is used for a foundational multi-expert multi-task deep ranking model. By introducing driver experience-related features and adding experience tasks, and setting an experience penalty loss function, the base model can learn behavioral information with low experience levels.

[0028] The optimization module is used to jointly train the business tasks in the base model with the newly added experience tasks to build a business-experience multi-expert network, and to collaboratively optimize the business objectives and driver experience.

[0029] The present invention discloses the following technical effects:

[0030] This invention effectively improves the driver experience and achieves synergistic optimization between business objectives and driver experience. Attached Figure Description

[0031] 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.

[0032] Figure 1 This is a schematic diagram of the method described in this invention. Detailed Implementation

[0033] 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.

[0034] like Figure 1 As shown, this invention provides a recommendation ranking technology that incorporates driver experience, specifically including the following:

[0035] (a) Building the basic model:

[0036] Based on cargo owner profiles, driver profiles, cross-feed / driver characteristics, and driver / cargo platform behavior information, a basic multi-expert, multi-task deep ranking model is designed. The multi-tasks in this model primarily focus on two core prediction objectives: predicting whether a driver will click on a candidate cargo and predicting whether a driver will actually purchase a candidate cargo. The specific design includes the following:

[0037] 1. Design of the input layer network:

[0038] The features of the cargo owner (mainly statistical features), driver profile features, cargo owner profile features, driver-cargo owner cross features, and driver behavior features (mainly historical click sequences) are processed and concatenated into a matrix as input to the expert layer. The following section mainly introduces the expert layer network.

[0039] 2. Design of the expert layer network:

[0040] After concatenating the inputs, multiple expert networks are constructed, each with different functions.

[0041] Task expert networks are responsible for learning specific tasks and can comprise multiple task expert networks, each consisting of multiple layers of fully connected layers. For example... Figure 1 The expert network A consists of n experts, Ea1 to Ean; the expert network B consists of k experts, Eb1 to Ebk; and the expert network C consists of k experts, Ec1 to Eck.

[0042] A shared expert network is responsible for transferring and sharing different information between tasks. It can contain multiple shared expert networks, each of which consists of multiple layers of fully connected components. Figure 1 The shared expert D network includes Es1 to Esm, with a total of m shared experts.

[0043] 3. Design of the gated layer network:

[0044] 3.1 Gating layer input: Task expert network output, shared expert network output, and original input layer information. It is worth noting that each specific task has its own gating network. The gating network input for a specific task includes the expert network corresponding to that task, the shared expert network, and the input layer information, but does not include the expert networks of other tasks.

[0045] 3.2 Gated Network: The gating layer inputs are concatenated into a list of the corresponding task expert network and shared expert network information; a selector is used to select the input layer, which is actually implemented using a fully connected layer, and the output is a vector representing the number of expert networks; the selected information is then subjected to softmax and multiplied with the concatenated vector in the list to output the result.

[0046] 3.3 Task Layer: Add a multi-layer fully connected layer to the gated output layer and perform cross-entropy loss with the final label.

[0047] (II) Integration of relevant features and experiential tasks:

[0048] Based on the aforementioned basic ranking model, driver experience-related features are introduced, and an experience task is added. At the same time, an experience penalty loss function is set so that the model can learn behavioral information with low experience.

[0049] 1. Introduction of experience-related features:

[0050] Features related to empty driving distance: straight-line distance between the driver's current location and the loading location, actual navigation distance, and the percentage of this distance in the driver's historical empty driving distances, etc.

[0051] Net freight characteristics include: order net freight amount, ratio of net freight to driver's historical average net freight, and ratio of net freight to transportation mileage (i.e., freight per unit mileage).

[0052] Driver-cargo matching characteristics include: the matching degree between the driver's current vehicle load and the order cargo weight, the matching degree between the vehicle's cargo dimensions and the cargo dimensions, and the matching degree between the vehicle's transportation type (such as refrigerated transportation or general transportation) and the order cargo transportation requirements.

[0053] 2. Experience Task Settings

[0054] A new task for predicting experience ratings has been added, which is to predict the driver's experience rating after receiving the order (a tag can be built based on the driver's experience feedback data after accepting orders in the past).

[0055] A new task for identifying experiential behavior has been added, which identifies whether an order is one that would lead to a low driver experience, such as an empty driving distance exceeding 1.5 times the driver's historical maximum empty driving distance, a net freight cost lower than 0.5 times the driver's historical minimum net freight cost, or a vehicle matching degree lower than 0.3. Thresholds are set based on business scenarios to build labels.

[0056] Experience penalty loss function design: In the model's total loss function, a penalty term is set for the experience behavior recognition task. If the model predicts an order with a low experience as an order with a non-low experience, the weight of this part of the loss is increased. An example of the penalty loss function formula is as follows: Where Lpenalty is the experience penalty loss, α is the penalty weight coefficient (which can be adjusted according to the business scenario, initially set to 1.2), and yi is the true label of whether the order is an order with a low experience (1 indicates yes, 0 indicates no). The model predicts the label; I(·) is an indicator function, taking a value of 1 when the condition within the parentheses is true, and 0 otherwise; α (penalty weight coefficient): a hyperparameter used to adjust the penalty intensity, initially set to 1.2. The larger α is, the heavier the penalty for "orders misjudged as having low experience," and the more the model will tend to avoid this type of error; ∑ (summation symbol): sums the loss terms for all samples (from the 1st to the Nth); (yi-ŷi)²: represents the "squared error between the true label and the predicted label." When the true label yi=1 (low experience) and the predicted label... When yi = 1 (correct prediction), (1-1)² = 0, this term has no loss; when the true label yi = 1 and the predicted label yi = 1, the true label yi = 1 and the predicted label yi = 1 is 0. When =0 (misprediction: low experience level is misjudged as not low), (1-0)²=1, and this term will incur a loss; I(yi=1) (indicator function): only applies to samples whose true label is "low experience level" (yi=1) (value is 1), and does not apply to samples whose true label is "not low experience level" (yi=0) (value is 0). This means that only orders whose true label is "low experience level" will be included in the penalty calculation, and orders whose true label is "not low experience level" will not participate in this penalty term.

[0057] (III) Model Training: Multi-expert and multi-task joint learning of business and experience:

[0058] By jointly training the business tasks (click prediction and transaction prediction) in the basic model with the newly added experience tasks (experience behavior recognition), a multi-expert and multi-task network structure is constructed to achieve synergistic optimization of business objectives and driver experience.

[0059] 1. Multi-expert network structure optimization:

[0060] Based on the existing feature expert module, a new experience feature expert module has been added to specifically handle experience-related features;

[0061] Set up a business-experience fusion expert module to integrate the features output by each expert module, while taking into account the importance of both business features and experience features.

[0062] Each task (business task and experience task) has its own dedicated task header to predict the fused features. At the same time, through a parameter sharing mechanism, the tasks can learn from each other.

[0063] 2. Joint training strategy:

[0064] 2.1. A multi-task joint training method is adopted. The total loss function is the weighted sum of the loss functions of each task and the experience penalty loss function, as shown in the following formula: Ltotal = β1Lclick + β2Ldeal + β3Lexperience + Lpenalty.

[0065] Where Ltotal is the total model loss, Lclick is the click prediction task loss (using cross-entropy loss), Ldeal is the transaction prediction task loss (using cross-entropy loss), and Lexperience is the experience rating prediction task loss (using mean squared error loss).

[0066] β1, β2, and β3 are the weighting coefficients for the loss of each task (initially set to β1=0.3, β2=0.4, and β3=0.3, which can be dynamically adjusted according to changes in business and experience metrics during model training).

[0067] 2.2. Use gradient descent optimization algorithms (such as Adam algorithm) to iteratively update the model parameters until the model converges, ensuring that the model learns the optimization direction of both business goals and experience goals during the training process.

[0068] This invention achieves synergistic optimization of business and experience: by integrating driver experience-related features and tasks into the model layer, rather than adjusting them after the reordering stage, it avoids the problem of balancing business indicators and driver experience in the prior art. This allows the model to improve business indicators such as click-through rate and conversion rate while effectively improving driver experience issues such as empty driving distance, net freight cost, and vehicle matching degree.

[0069] This invention accurately identifies and avoids behaviors that result in a poor user experience: by setting an experience penalty loss function, the model can focus on learning the characteristics of orders with poor user experience, reducing the probability of recommending such orders to drivers, and significantly improving the overall order-accepting experience for drivers;

[0070] The model of this invention has strong generalization ability: the model is built based on the multi-dimensional features of cargo owners and drivers, and adopts a multi-expert multi-task learning architecture, which can adapt to the order recommendation needs under different freight scenarios and has strong generalization ability and practicality.

[0071] This invention benefits the platform's sustainable development: by improving the driver experience, it can increase the drivers' willingness to accept orders and their loyalty, reduce driver churn, and at the same time ensure the stable improvement of the platform's business indicators, achieving a win-win situation for the platform, cargo owners, and drivers, and promoting the platform's sustainable development.

[0072] Compared to the original deep ranking model, the experimental data of this invention can achieve an AUC of 10+pp, with the transaction AUC and click AUC remaining the same; online business results: the transaction rate and fulfillment rate within the scenario remain the same, and the average decrease in the experience dissatisfaction index is 20+pp.

[0073] 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 recommendation ranking method that incorporates driver experience, characterized in that, Includes the following steps: A basic multi-expert, multi-task deep ranking model is constructed, consisting of an input layer network, an expert layer network, a gating layer network, and a task layer, to predict whether a driver will click on a candidate cargo and whether the driver will complete a transaction with a candidate cargo. Based on the fundamental multi-expert multi-task deep ranking model, this paper introduces driver experience-related features, adds experience tasks, and sets an experience penalty loss function. By jointly training the business tasks in the basic model with the newly added experience tasks, a business-experience multi-expert network is constructed to collaboratively optimize business objectives and driver experience. When constructing the basic multi-expert multi-task deep ranking model, the input layer network is used to obtain cargo owner behavior features, driver profile features, cargo owner profile features, driver-cargo owner cross features and driver behavior features. After processing, these features are concatenated into a matrix as the input to the expert layer network. When constructing a basic multi-expert multi-task deep ranking model, the inputs are concatenated together, and task learning is performed through multiple expert networks in the expert layer network, and different information is transferred and shared between tasks. When constructing a basic multi-expert multi-task deep ranking model, the output of the task expert network, the output of the shared expert network, and the original input layer information are input into the gating layer corresponding to the expert network as the input of the gating layer network, wherein the gating layer network is composed of multiple gating layers; Through a gated network, the information of the task expert network and the shared expert network are concatenated into a list. A picker is used to select the input layer, and the output vector of the number of expert networks is multiplied by softmax and then multiplied by the list vector. The result is then sent to the task layer. In the task layer, the gated output layer is added to a multi-layer fully connected layer and cross-entropy loss is applied to the final label. When introducing experience-related features, the experience-related features include: Empty driving distance related features include the straight-line distance between the driver's current location and the loading location, the actual navigation distance, and the proportion of this distance in the driver's historical empty driving distances. Net freight-related characteristics include the order's net freight amount, the ratio of net freight to the driver's historical average net freight, and the ratio of net freight to transportation mileage. The driver-cargo matching characteristics represent the degree of matching between the driver's current vehicle load weight and the order cargo weight, the degree of matching between the vehicle's cargo dimensions and the cargo dimensions, and the degree of matching between the vehicle's transportation type and the order cargo transportation requirements. When adding a new experience task, a new experience behavior recognition task is added to identify whether the order is one that will lead to a low driver experience, including: the empty driving distance exceeds 1.5 times the driver's historical maximum empty driving distance, the net freight cost is less than 0.5 times the driver's historical minimum net freight cost, or the vehicle matching degree is less than 0.

3. During joint training, the total loss function is set as the weighted sum of the loss functions of each task and the experience penalty loss function. A gradient descent optimization algorithm is used to iteratively update the model parameters until the model converges, ensuring that the model simultaneously learns the optimization directions of both business and experience objectives during training. The total loss function is expressed as: Ltotal=β1Lclick+β2Ldeal+β3Lexperience+Lpenalty In the formula, Ltotal is the total loss of the model, Lclick is the cross-entropy loss of the click prediction task, Ldeal is the cross-entropy loss of the transaction prediction task, Lexperience is the mean squared error loss of the experience rating prediction task, β1, β2, and β3 are the weight coefficients of the loss of each task, and Lpenalty is the experience penalty loss function. ， In the formula, Lpenalty represents the experience penalty loss, α is the penalty weight coefficient, and yi is the true label of whether the order is a low-experience order, where yi=1 indicates yes and yi=0 indicates no. The model predicts the label; I(·) is an indicator function that takes the value 1 when the condition in parentheses is true, and 0 otherwise; ∑: sums up the loss terms for all samples; : Represents the squared error between the true label and the predicted label.

2. A recommendation ranking system incorporating driver experience, used to implement the recommendation ranking method incorporating driver experience as described in claim 1, characterized in that, include: The basic module construction module is used to build a basic multi-expert multi-task deep ranking model consisting of an input layer network, an expert layer network, a gating layer network, and a task layer. This model is used to predict whether a driver will click on a candidate cargo and whether the driver will complete a transaction with a candidate cargo. An improved module is used for a foundational multi-expert multi-task deep ranking model. This module introduces driver experience-related features, adds experience tasks, and sets an experience penalty loss function. The optimization module is used to jointly train the business tasks in the base model with the newly added experience tasks to build a business-experience multi-expert network, and to collaboratively optimize the business objectives and driver experience.

Images