Target object recruitment number prediction method, model training method and equipment

Abstract

The embodiment of the invention provides a target object recruitment quantity prediction method, a model training method, equipment, a storage medium and a program product. The input cost of recruiting the target object and at least one influence factor capable of influencing the recruiting number of the target object except the input cost are obtained, the input cost can be input into a pre-trained basic sub-model to predict the basic recruiting number of the target object, and the recruiting number of the target object is predicted for each influence factor. And inputting the influence factor into the exclusive adjustment sub-model of the influence factor to output an adjustment coefficient corresponding to the influence factor, and sequentially adjusting the basic recruitment number by using the adjustment coefficient corresponding to the at least one influence factor to obtain a predicted recruitment number corresponding to the input cost. Through the scheme, the precision of the recruitment number predicted by the model can be improved, the interpretability of the model can be improved, and a basis is provided for customization of the recruitment scheme.

Description

Technical Field

[0001] The embodiments in this specification relate to the field of artificial intelligence technology, and in particular to a method for predicting the number of target recruitments, a model training method, a device, a storage medium, and a program product. Background Technology

[0002] In sectors such as local services, internet-based transportation, and the sharing economy, the number of recruits from target groups (such as delivery personnel, ride-hailing drivers, and platform merchants) directly determines the platform's service supply capacity. The forecast results serve as the core basis for platforms to formulate recruitment budgets, optimize resource allocation, and balance supply and demand. Recruitment costs are the primary driver of recruitment volume, while recruitment numbers are also influenced by a combination of external factors, including order trends, seasonality, and the level of special subsidies.

[0003] In related technologies, when training a recruitment quantity prediction model to predict the number of recruits for a target group, either only the input cost is considered, resulting in poor accuracy of the prediction results, or the input cost and other multiple influencing factors are used as inputs to the model. This method produces more accurate prediction results, but the interpretability is poor, and the impact of each influencing factor on the recruitment quantity cannot be quantified and accurately measured, resulting in a lack of data support and basis when formulating recruitment strategies.

[0004] It is evident that the prediction solutions in related technologies are insufficient to meet users' needs for accurate prediction of recruitment numbers and refined strategy formulation. Summary of the Invention

[0005] To overcome the problems existing in related technologies, this specification provides a method for predicting the number of recruited targets, a model training method, an apparatus, a storage medium, and a program product.

[0006] According to a first aspect of the embodiments of this specification, a method for predicting the number of recruits for a target group is provided, the method comprising: Obtain the input cost of recruiting target individuals, and at least one other influencing factor besides the input cost that can affect the number of target individuals recruited; The input cost is fed into the base sub-model of the pre-trained recruitment quantity prediction model to predict the base recruitment quantity of the target object. The base sub-model is trained with historical input cost as input and historical actual recruitment quantity corresponding to historical input cost as label. For each influencing factor, the influencing factor is input into the dedicated adjustment sub-model of the recruitment quantity prediction model to output the adjustment coefficient corresponding to the influencing factor; wherein, the dedicated adjustment sub-model of each influencing factor is trained with the historical data of the influencing factor as input and the ratio of the historical actual recruitment quantity under the influence of the influencing factor to the historical predicted recruitment quantity without considering the influencing factor as the label. The basic recruitment quantity is adjusted sequentially using the adjustment coefficients corresponding to each of the at least one influencing factor to obtain the predicted recruitment quantity corresponding to the input cost.

[0007] In some embodiments, the base sub-model is trained prior to the dedicated adjustment sub-models for each influencing factor, and the historical predicted recruitment number without considering the influencing factor is determined at least based on the prediction results of the trained base sub-model.

[0008] In some embodiments, the dedicated moderating sub-models of each of the at least one influencing factor are trained sequentially in a chronological order, wherein the dedicated moderating sub-models of the influencing factor that have a greater impact on the recruitment quantity of the target object are trained earlier, and the historical predicted recruitment quantity without considering the influencing factor is determined by combining the prediction results of the trained base sub-model and the prediction results of the dedicated moderating sub-models of other influencing factors.

[0009] In some embodiments, the dedicated regulation sub-model for each influencing factor is trained in the following manner: Acquire multiple sets of historical recruitment data, each set of historical recruitment data corresponding to one historical recruitment event. Each set of historical recruitment data includes at least the historical investment cost of that historical recruitment event, the historical data of that influencing factor, and the historical actual number of the target objects recruited for that historical recruitment event. The historical input cost of this recruitment event is input into the trained base sub-model to obtain the prediction result; At least based on the prediction results, determine the historical predicted recruitment numbers without considering this influencing factor; Determine the ratio of the historical actual recruitment number to the historical predicted recruitment number; The historical data of the influencing factor is used as the input to the specific regulation sub-model, and the ratio is used as the label to train the specific regulation sub-model.

[0010] In some embodiments, the historical recruitment data further includes historical data of other influence factors besides the at least one influence factor, wherein the specific moderating sub-models for the other influence factors are trained prior to the specific moderating sub-model for the influence factor, and determining the historical predicted recruitment quantity without considering the influence factor is based at least on the prediction results, including: For each other influencing factor, the historical data of that other influencing factor is input into the trained dedicated adjustment sub-model of that other historical influencing factor to output the corresponding historical adjustment coefficient; The prediction results are adjusted sequentially using the historical adjustment coefficients corresponding to each of the other historical influencing factors to obtain the historical predicted recruitment number.

[0011] In some embodiments, the degree of influence of each influencing factor on the recruitment quantity of the target object is characterized by a goodness-of-fit score, which is determined based on the following: Obtain multiple historical recruitment data points, each corresponding to a historical recruitment event. Each historical recruitment data point includes the historical investment cost of that recruitment event, the historical data of that influencing factor, and the historical actual recruitment quantity of that recruitment event. For each historical recruitment data point, the historical investment cost in that historical recruitment data point is input into the trained base sub-model to output a prediction result, and the difference between the prediction result and the actual historical recruitment number is determined. The difference sequence formed by the differences corresponding to each of the multiple historical recruitment data is used as the dependent variable, and the historical data sequence formed by the historical data of the influencing factor in the multiple historical recruitment data is used as the independent variable; The dependent and independent variables are fitted, and the goodness of fit between the independent and dependent variables is calculated, wherein the goodness of fit is positively correlated with the degree of influence.

[0012] In some embodiments, if the impact factor is a sequential impact factor, then the specific moderating sub-model of the impact factor is a time series model. If the impact factor is a continuous numerical impact factor, then the specific moderating sub-model of the impact factor shall be selected from one of the following: autolinear regression model, multinomial regression model, or logarithmic regression model. If the impact factor is a categorical impact factor, then the specific moderating sub-model for the impact factor shall be one of the following: logistic regression model, LightGBM classification-derived regression model, or categorical embedding regression model.

[0013] In some embodiments, the target object includes delivery capacity, and the at least one influencing factor includes one or more of the following: The order trend factor describes the changing trend of delivery orders, the seasonal factor describes seasonal attributes, and the subsidy cost subsidizes delivery capacity.

[0014] In some embodiments, the order trend factor is characterized by an order volume ratio sequence, which is determined based on the following: Obtain the historical order volume sequence, which includes the historical order volume corresponding to each of multiple historical time periods; Get the current order volume for the current time period; The ratio of each historical order volume in the historical order volume sequence to the current order volume is determined to obtain the order volume ratio sequence.

[0015] In some embodiments, the target object includes delivery capacity capable of handling specific types of waybills, the input costs include multiple factors, and the method further includes: Determine the predicted recruitment quantity of the delivery capacity capable of handling specific types of waybills under multiple input costs, so as to obtain the mapping relationship between input costs and predicted recruitment quantity; Obtain historical waybill data for this specific type of waybill, as well as the predicted waybill volume for this specific type of waybill over a future period. Based on the waybill data and the predicted waybill volume, determine the quantity of delivery capacity gap that can handle the specific type of waybill; The actual input cost in the recruitment scheme for recruiting delivery capacity capable of handling this specific type of waybill is determined based on the number of gaps and the mapping relationship.

[0016] In some embodiments, the method further includes: For each influencing factor, after determining the adjustment coefficient corresponding to that influencing factor, the basic recruitment quantity is adjusted using the adjustment coefficient corresponding to that influencing factor to obtain the adjusted recruitment quantity; The change in the number of recruits for the target group caused by the influencing factor is determined based on the difference between the adjusted recruitment quantity and the basic recruitment quantity.

[0017] In some embodiments, the input costs include multiple input costs corresponding to each of multiple recruitment channels used to recruit the target objects, and the method further includes: For each recruitment channel, determine the predicted recruitment quantity for that channel under multiple input costs, so as to obtain the mapping relationship between the input cost and the predicted recruitment quantity for that recruitment channel; An optimization model is constructed based on the mapping relationship corresponding to each of the multiple recruitment channels. The optimization model uses the actual investment cost corresponding to each recruitment channel as the optimization variable, minimizes the total investment cost of the multiple recruitment channels as the optimization objective, and uses the mapping relationship and the sum of the predicted recruitment quantity of each of the multiple recruitment channels as constraints that the total number of recruitments is not less than the target quantity. Solve the optimization model to obtain the actual investment cost corresponding to each recruitment channel.

[0018] According to a second aspect of the embodiments of this specification, a method for training a recruitment quantity prediction model is provided. The recruitment quantity prediction model includes a base sub-model and at least one dedicated adjustment sub-model. Each dedicated adjustment sub-model corresponds to an influence factor, which is different from the input cost and can influence the recruitment quantity of the target object. The method includes: Obtain first sample data, which includes the historical investment cost of the first historical recruitment event and the historical actual recruitment quantity corresponding to the first historical recruitment event; The basic sub-model is trained using the historical investment cost of the first historical recruitment event as input and the historical actual recruitment quantity corresponding to the first historical recruitment event as label. Obtain second sample data, which includes the historical investment cost of the second historical recruitment event, the historical data of at least one of the influencing factors, and the historical actual recruitment quantity corresponding to the second historical recruitment event. The historical input cost of the second historical recruitment event is input into the pre-trained base sub-model to obtain the prediction result; For each dedicated adjustment sub-model, at least based on the prediction results, the historical predicted recruitment quantity of the second historical recruitment event is determined, and the ratio of the historical actual recruitment quantity corresponding to the second historical recruitment event to the historical predicted recruitment quantity is determined. The historical data of the influence factor corresponding to the dedicated adjustment sub-model is used as the input of the dedicated adjustment sub-model, and the ratio is used as the label to train the dedicated adjustment sub-model. The historical predicted recruitment quantity is the predicted recruitment quantity of the second historical recruitment event without considering the influence factor corresponding to the dedicated adjustment sub-model.

[0019] In some embodiments, the dedicated moderating sub-models of each of the at least one influencing factor are trained sequentially, wherein the dedicated moderating sub-models of the influencing factor that have a greater impact on the recruitment quantity of the target object are trained earlier, and the step of determining the historical predicted recruitment quantity of the second historical recruitment event based at least on the prediction results includes: For each other dedicated moderating sub-model trained prior to this dedicated moderating sub-model, the historical data of the influencing factors corresponding to the other dedicated moderating sub-model are input into the other dedicated moderating sub-model to output the corresponding historical moderating coefficients; The prediction results are adjusted sequentially using the historical adjustment coefficients output by each of the other dedicated adjustment sub-models to obtain the historical predicted recruitment quantity.

[0020] According to a third aspect of the embodiments of this specification, a computer program product is provided, including a computer program that, when executed by a processor, implements the methods mentioned in the first and / or second aspects above.

[0021] According to a fourth aspect of the embodiments of this specification, an electronic device is provided, the electronic device including a processor, a memory, and a computer program stored in the memory and executable by the processor, wherein the computer program, when executed, implements the methods mentioned in the first and / or second aspects above.

[0022] According to a fifth aspect of the embodiments of this specification, a computer storage medium is provided, on which a computer program is stored, which, when executed by a processor, implements the methods mentioned in the first and / or second aspects above.

[0023] The beneficial effects of the embodiments in this specification are as follows: A recruitment quantity prediction model can be pre-trained. This model includes a basic sub-model and at least one specific adjustment sub-model. The basic sub-model is used to predict the impact of input costs on recruitment quantity. Each specific adjustment sub-model corresponds to an influencing factor different from the input cost, used to predict the impact of the corresponding influencing factor on recruitment quantity. Specifically, the basic sub-model can be trained using historical input costs as input and the historical actual recruitment quantity corresponding to the historical input costs as labels. The specific adjustment sub-model for each influencing factor is trained using historical data of that influencing factor as input and the ratio of the historical actual recruitment quantity under the influence of that influencing factor to the historical predicted recruitment quantity without considering that influencing factor as labels. When predicting the recruitment quantity of a target object, the input cost of recruiting the target object and at least one other influencing factor besides the input cost that can affect the recruitment quantity of the target object can be obtained. The input cost is input into the basic sub-model to predict the basic recruitment quantity of the target object. For each influencing factor, the influencing factor is input into the specific adjustment sub-model of that influencing factor to output the adjustment coefficient corresponding to that influencing factor. Then, the basic recruitment quantity is adjusted sequentially using the adjustment coefficients corresponding to each of the at least one influencing factor to obtain the predicted recruitment quantity corresponding to the input cost.

[0024] By constructing a hierarchical forecasting architecture of "basic sub-model + dedicated adjustment sub-model", the basic sub-model firmly anchors the core driving role of input costs on recruitment quantity, ensuring the consistency of business logic in forecasting. Furthermore, by configuring an independent dedicated adjustment sub-model for each non-cost-related influencing factor, combined with a label design of "the ratio of actual recruitment quantity to the predicted value without considering this factor", the adjustment effect of individual influencing factors is accurately isolated and quantified. The basic recruitment quantity is then adjusted sequentially by various adjustment coefficients, effectively integrating the core cost-driven effect with the synergistic influence of multiple factors. Compared to the shortcomings of traditional single-factor forecasting that ignores the effects of multiple factors and the confusion of factor influences in mixed models, this approach improves the accuracy of recruitment quantity forecasting while making the adjustment effect of each factor clear and traceable. It provides reliable support for the platform's refined recruitment budget planning, balancing forecasting accuracy with business decision-making value.

[0025] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit the embodiments of this specification. Attached Figure Description

[0026] The accompanying drawings, which are incorporated herein by reference and form part of the embodiments thereof, illustrate embodiments consistent with those described herein and, together with the description, serve to explain the principles of those embodiments.

[0027] Figure 1 This is a schematic diagram illustrating an application scenario as an exemplary embodiment of this specification; Figure 2 This is a schematic diagram illustrating the structure of a recruitment quantity prediction model as an exemplary embodiment of this specification; Figure 3 A flowchart illustrating a method for predicting the number of target recruits, as shown in an exemplary embodiment of this specification; Figure 4 This is a schematic diagram illustrating the training process of a recruitment quantity prediction model as an exemplary embodiment of this specification; Figure 5 A flowchart illustrating a training method for a recruitment quantity prediction model, as shown in an exemplary embodiment of this specification; Figure 6 This is a logic block diagram of an electronic device illustrated in an exemplary embodiment of this specification. Detailed Implementation

[0028] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with those described in this specification. Rather, they are merely examples of apparatuses and methods consistent with some aspects of the embodiments described in this specification as detailed in the appended claims.

[0029] The terminology used in the embodiments of this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the embodiments of this specification. The singular forms “a,” “described,” and “the” as used in the embodiments of this specification and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0030] It should be understood that although the terms first, second, third, etc., may be used to describe various information in the embodiments of this specification, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, first information may also be referred to as second information without departing from the scope of the embodiments of this specification, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to a determination."

[0031] In sectors such as local services, internet-based transportation, and the sharing economy, the number of recruits from target groups (such as delivery personnel, ride-hailing drivers, and platform merchants) directly determines the platform's service supply capacity. The forecast results serve as the core basis for platforms to formulate recruitment budgets, optimize resource allocation, and balance supply and demand. Recruitment costs are the primary driver of recruitment volume, while recruitment numbers are also influenced by a combination of external factors, including order trends, seasonality, and the level of special subsidies.

[0032] In related technologies, methods for predicting the number of recruits for a target group are mainly divided into two categories: One type is single-factor prediction technology. This type of method uses only the input cost as the sole input variable to train the recruitment quantity prediction model. This method ignores the impact of other influencing factors such as order trends, seasonality, and special subsidies on the recruitment quantity. For example, in actual business, the recruitment quantity during periods of rapid order growth (when potential recruits have a higher willingness to work) is often significantly higher than during periods of stable order growth, and the recruitment conversion efficiency in peak seasons also differs significantly from that in off-seasons. This leads to a large deviation between the prediction results of the single-factor model and the actual values, making it difficult to meet the accuracy requirements of business operations.

[0033] Another type is the multi-factor hybrid prediction technique. This method uses recruitment costs along with other influencing factors as input to train a recruitment quantity prediction model. However, this type of method also has significant drawbacks. First, because this method indiscriminately includes all influencing factors in the same model, the dominant role of cost is weakened, and the interaction interference between different factors is difficult to isolate, making it impossible to trace the source of prediction errors. Second, the models trained by this method have poor interpretability, cannot quantify the independent impact of individual influencing factors on recruitment quantity, and cannot support refined recruitment strategy decisions.

[0034] Therefore, there is an urgent need for a solution that can balance the accuracy of recruitment quantity prediction results with the interpretability of the model, so as to meet users' needs for accurate recruitment quantity prediction and refined recruitment strategy formulation.

[0035] Based on this, embodiments of this specification provide a method for predicting the recruitment quantity of a target object. A recruitment quantity prediction model can be pre-trained, comprising a base sub-model and at least one dedicated adjustment sub-model. The base sub-model predicts the impact of input costs on the recruitment quantity, and each dedicated adjustment sub-model corresponds to an influencing factor different from the input cost, used to predict the impact of the corresponding influencing factor on the recruitment quantity. The base sub-model can be trained using historical input costs as input and the historical actual recruitment quantity corresponding to the historical input costs as labels. The dedicated adjustment sub-model for each influencing factor is trained using historical data of that influencing factor as input and the ratio of the historical actual recruitment quantity under the influence of that influencing factor to the historical predicted recruitment quantity without considering that influencing factor as labels. When predicting the recruitment quantity of a target object, the input cost for recruiting the target object and at least one other influencing factor besides the input cost that can affect the recruitment quantity of the target object can be obtained. The input cost is input into the base sub-model to predict the base recruitment quantity of the target object. For each influencing factor, the influencing factor is input into the dedicated adjustment sub-model for that influencing factor to output the adjustment coefficient corresponding to that influencing factor. Then, the basic recruitment quantity is adjusted sequentially using the adjustment coefficients corresponding to each of the at least one influencing factor to obtain the predicted recruitment quantity corresponding to the input cost.

[0036] By constructing a hierarchical prediction architecture consisting of a "basic sub-model + multiple dedicated adjustment sub-models," the basic sub-model firmly grasps the core driving role of input costs in recruitment quantity, ensuring the alignment of prediction logic with the essence of business. Furthermore, by configuring dedicated adjustment sub-models for each other influencing factor, the impact of various factors is accurately isolated and independently quantified. Specifically, the dedicated adjustment sub-models, trained using the "ratio of the actual value to the predicted value without considering the factor," directly output targeted adjustment coefficients. Combined with the prediction logic of "sequentially adjusting the basic recruitment quantity," this not only effectively avoids the accuracy defects of traditional single-factor models that ignore multi-factor adjustment, but also solves the problems of factor confusion and difficulty in error traceability in multi-factor mixed modeling. This makes the specific adjustment effect of each influencing factor on recruitment quantity clearly verifiable, improving the accuracy of recruitment quantity prediction and increasing the model's interpretability, providing reliable data support for recruitment plan planning.

[0037] The target recruitment quantity prediction method provided in the embodiments of this specification can be executed by various electronic devices that have deployed recruitment quantity prediction models, such as mobile phones, tablets, computers, physical servers, cloud servers, and server clusters.

[0038] The target objects in the embodiments of this specification can be various types of objects to be recruited, such as delivery capacity, platform merchants, ride-hailing drivers, etc.

[0039] like Figure 1The diagram illustrates an application scenario of an embodiment of this specification. Recruitment quantity prediction software can be installed on a user terminal device (e.g., a mobile phone or computer). This software provides an interactive interface through which users can input the type of target objects to be recruited (e.g., delivery capacity, platform merchants, etc.). Simultaneously, users can input various configuration information in the recruitment plan configuration interface provided by the prediction software. This configuration information can include various influencing factors related to the recruitment quantity, such as input costs, subsidy plans for the target objects, and recent order volume. After configuring the above information, the user can click the submit button on the interactive interface, and the prediction software can send the user-input configuration information to a cloud server. The cloud server deploys a pre-trained recruitment quantity prediction model, which can then be used to determine the predicted recruitment quantity of the target objects based on the user-input configuration information and send it to the prediction software. Furthermore, the prediction software can generate and display corresponding reports based on the prediction results. For example, it can generate a "input cost - recruitment quantity" mapping curve and output quantitative results of the influence of each influencing factor on the recruitment quantity. Alternatively, a customized recruitment plan can be created based on the user's recruitment needs, the aforementioned "cost-recruitment quantity" mapping curve, and the quantitative results of the impact of each influencing factor on the recruitment quantity. This plan can then be presented to the user. The recruitment plan may include details of cost input, subsidy activities, etc.

[0040] like Figure 2 The diagram shows the structure of the recruitment quantity prediction model in this embodiment. This model includes a basic sub-model and at least one dedicated adjustment sub-model (dedicated adjustment sub-models 1-n in the diagram). To facilitate backtracking on the specific impact of each influencing factor on the recruitment quantity, both the basic sub-model and each dedicated adjustment sub-model are single-factor models, meaning the model input is a single influencing factor. Considering that input cost is the most direct and crucial factor in recruiting target groups, directly impacting the recruitment quantity, the basic sub-model can be used to learn the intrinsic relationship between input cost and recruitment quantity. Simultaneously, for other influencing factors besides input cost that affect the recruitment quantity (influence factors 1-n in the diagram), a dedicated adjustment sub-model can be set for each influencing factor to learn its impact on the recruitment quantity. The number of dedicated adjustment sub-models can be adapted to the types of influencing factors. For example, taking delivery capacity as the target group, in addition to input cost, four other influencing factors need to be considered: order trend factor, seasonal factor, subsidy cost, and the current delivery capacity gap. Therefore, four dedicated adjustment sub-models can be set.

[0041] like Figure 3As shown, the target recruitment quantity prediction method provided in the embodiments of this specification may include the following steps: S302. Obtain the input cost of recruiting target objects, and at least one other influencing factor besides the input cost that can affect the number of target objects recruited; In step S302, the input costs for recruiting target candidates can be obtained. These input costs may include advertising expenses, recruitment channel service fees, etc. Simultaneously, at least one other influencing factor, besides these input costs, that can affect the number of target candidates recruited can also be obtained. The types and number of influencing factors can be specifically set based on the type of target candidates to be recruited. For example, if the target candidates are delivery capacity, the at least one influencing factor could be an order trend factor, a seasonal factor, subsidy costs, the current delivery capacity gap, etc.

[0042] S304. Input the input cost into the base sub-model of the pre-trained recruitment quantity prediction model to predict the base recruitment quantity of the target object, wherein the base sub-model is trained with historical input cost as input and historical real recruitment quantity corresponding to historical input cost as label. A recruitment quantity prediction model can be pre-trained, wherein the base sub-model and at least one dedicated adjustment sub-model in the recruitment quantity prediction model can be trained in stages. In some embodiments, when training the base sub-model, historical investment costs can be used as input, and the historical actual recruitment quantity corresponding to the historical investment costs can be used as the label to train the base sub-model. For example, recruitment data corresponding to historical recruitment events can be obtained, which may include the investment cost of the historical recruitment event (i.e., historical investment cost) and the number of target objects actually recruited in the historical recruitment event (i.e., historical actual recruitment quantity). Then, the historical investment cost is used as the input of the base sub-model, and the predicted recruitment quantity is output by the base sub-model. Based on the difference between the predicted recruitment quantity and the historical actual recruitment quantity, the model parameters of the base sub-model are adjusted to train the base sub-model.

[0043] The core design of the dedicated moderating sub-model is to quantify the moderating effect of a single factor on recruitment volume by isolating the independent influence of that factor. Therefore, in some embodiments, when training the dedicated moderating sub-model for each influencing factor, historical data of that factor can be used as input, and the ratio of the historical actual recruitment volume under the influence of that factor to the historical predicted recruitment volume without considering that factor can be used as the label to train the dedicated moderating sub-model. For example, recruitment data corresponding to historical recruitment events can be obtained, which may include historical data of the influencing factor. For instance, if the influencing factor is an order trend factor, the value corresponding to the order trend factor of that historical recruitment event can be obtained as the model input. Simultaneously, since the dedicated moderating sub-model learns the moderating effect of the influencing factor on recruitment volume, the historical predicted recruitment volume without considering that factor can be determined, and the ratio of the historical actual recruitment volume under the influence of that factor to the historical predicted recruitment volume can be used as the label to train the dedicated moderating sub-model.

[0044] The historical predicted recruitment quantity for a recruitment event without considering the influencing factor can be determined in various ways. In some embodiments, a trained base sub-model can be used to output the historical predicted recruitment quantity without considering the influencing factor based on the historical input cost of the recruitment event. In some embodiments, after the base sub-model outputs the historical predicted recruitment quantity, the adjustment coefficient output by the dedicated adjustment sub-model for other influencing factors can also be used to adjust the historical predicted recruitment quantity to obtain the historical predicted recruitment quantity. In some embodiments, the historical predicted recruitment quantity without considering the influencing factor can also be determined by analyzing and fitting a large amount of historical recruitment data. For example, taking the seasonal factor as an example, peak seasons may usually cause an increase in recruitment quantity. The increase in recruitment quantity caused by the peak season factor can be determined by combining multiple historical recruitment data. Subtracting this increase from the historical actual recruitment quantity gives the historical predicted recruitment quantity without considering the influencing factor.

[0045] In step S304, the input cost can be fed into the base sub-model of the pre-trained recruitment quantity prediction model to predict the base recruitment quantity of the target object through the base sub-model.

[0046] S306. For each influencing factor, the influencing factor is input into the dedicated adjustment sub-model of the recruitment quantity prediction model to output the adjustment coefficient corresponding to the influencing factor. The dedicated adjustment sub-model of each influencing factor is trained with the historical data of the influencing factor as input and the ratio of the historical actual recruitment quantity under the influence of the influencing factor to the historical predicted recruitment quantity without considering the influencing factor as the label. In step S306, for each influencing factor, it can be input into the dedicated adjustment sub-model of that influencing factor in the recruitment quantity prediction model to output the corresponding adjustment coefficient. For example, if the influencing factors include order change trend factor, seasonal factor, and subsidy cost factor, then these three can be input into the dedicated adjustment sub-models for order change trend factor, seasonal factor, and subsidy cost factor, respectively, to obtain their respective adjustment coefficients. A adjustment coefficient greater than 1 indicates that the influencing factor promotes recruitment quantity growth; a adjustment coefficient less than 1 indicates that the influencing factor inhibits recruitment quantity growth; and a adjustment coefficient close to 1 indicates that the influencing factor has a weak impact on recruitment quantity.

[0047] S308. The basic recruitment quantity is adjusted sequentially using the adjustment coefficients corresponding to each of the at least one influencing factor to obtain the predicted recruitment quantity corresponding to the input cost.

[0048] In step S308, after obtaining the adjustment coefficients corresponding to each influencing factor, these adjustment coefficients can be used to adjust the basic recruitment quantity output by the basic sub-model in sequence, thereby obtaining the predicted recruitment quantity corresponding to the input cost.

[0049] For example, assuming the basic recruitment quantity predicted by the basic sub-model is N0, the adjustment coefficient corresponding to the order trend factor is 0.7, the adjustment coefficient corresponding to the seasonal factor is 0.9, and the adjustment coefficient corresponding to the subsidy cost factor is 1.1, then the final predicted recruitment quantity N = N0 × 0.7 × 0.9 × 1.1.

[0050] In some embodiments, the training order of the base sub-model precedes the training order of the specific regulation sub-models for each influencing factor. For example, the base sub-model can be trained first, and then the specific regulation sub-models can be trained, thereby utilizing the prediction results of the trained base sub-model to assist in the training of the specific regulation sub-models. The "historical predicted recruitment quantity without considering the influencing factor" used in the training of each specific regulation sub-model can be determined based on the prediction results of the base sub-model.

[0051] In some embodiments, the dedicated moderating sub-models of each of the at least one influencing factor are trained sequentially in a chronological order, wherein the dedicated moderating sub-models of the influencing factor that have a greater impact on the recruitment quantity of the target object are trained earlier. For each dedicated moderating sub-model of an influencing factor, the "historical predicted recruitment quantity without considering the influencing factor" used in the training process can be determined by combining the prediction results of the trained base sub-model and the prediction results of the dedicated moderating sub-models of other influencing factors.

[0052] For example, the degree of independent influence of each influencing factor on the recruitment quantity can be determined. Then, following the principle of "the greater the influence, the earlier the training order," the training process of each dedicated regulation sub-model is initiated sequentially. That is, the dedicated regulation sub-model corresponding to the factor with the most significant effect on regulating the recruitment quantity is trained first. After its training is completed and validated, the dedicated regulation sub-model corresponding to the next most influential factor is trained, and so on, until all sub-models are trained. For example, taking delivery capacity as the target, other influencing factors may include order trend factors, seasonal factors, subsidy costs, etc. The influence of these factors on the delivery capacity recruitment quantity, from largest to smallest, is: order trend factors, seasonal factors, subsidy cost factors. Therefore, after training the basic sub-model, the dedicated regulation sub-model for the order trend factor can be trained first, then the dedicated regulation sub-model for the seasonal factor, and finally the dedicated regulation sub-model for the subsidy cost factor.

[0053] For each dedicated regulation sub-model to be trained, the "historical predicted recruitment quantity without considering the influencing factor" required to calculate its training label (i.e. the ratio mentioned above) can be obtained by combining the historical basic predicted value output by the already trained basic sub-model and the historical regulation coefficient output by the dedicated regulation sub-models corresponding to other high-influence factors that have been trained with priority. By using these high-priority regulation coefficients to sequentially correct the historical basic predicted value, a comprehensive predicted value that excludes the interference of the current influencing factor to be trained and integrates the core cost-driven and high-priority factor regulation effects can be obtained.

[0054] By linking the training order of dedicated adjustment sub-models to the influence of influencing factors on recruitment numbers, it ensures that key factors that play a decisive role in recruitment numbers are optimized first, guaranteeing the model's core adjustment capabilities from the outset. By comprehensively determining historical predicted recruitment numbers considering a given influencing factor using both a trained base sub-model and a trained dedicated adjustment sub-model for a high-importance influencing factor, this approach is more closely aligned with historical business scenarios compared to predictions based solely on a single base sub-model. This significantly reduces the baseline error in label calculation and allows for more accurate quantification of the independent adjustment effect of the current factor. Furthermore, by including only the adjustment results of trained high-importance influencing factors, interference from currently untrained factors and low-importance factors is eliminated, avoiding training bias caused by multi-factor confusion. This ensures that each dedicated adjustment sub-model accurately learns the adjustment patterns of its corresponding influencing factor, effectively accumulating the training results of high-importance influencing factors, achieving synergistic optimization of multi-factor adjustment, and ultimately significantly improving the accuracy and stability of the entire prediction model.

[0055] The ranking of the influence of at least one influencing factor on the recruitment quantity of the target group can be determined based on experience, or the influence of the influencing factor on the recruitment quantity of the target group can be quantified for easier ranking and comparison. For example, to facilitate the ranking of the influence of the recruitment quantity of the target group and to accurately determine the training order of the dedicated regulation sub-model, some parameters can be used to measure the magnitude of the influence. For example, in some embodiments, linear or non-linear correlation coefficients between each influencing factor and the recruitment quantity can be calculated, and these linear or non-linear correlation coefficients can be used to characterize the magnitude of the influence, where a larger correlation coefficient indicates a greater influence.

[0056] The type of correlation coefficient can be flexibly selected based on the relationship between the impact factor and the number of recruits. For example, in the scenario where the impact factor and the number of recruits have a linear relationship, the Pearson correlation coefficient can be used. In the scenario where the impact factor and the number of recruits have a monotonic but non-linear relationship, the Spearman's rank correlation coefficient can be used.

[0057] For example, taking order volume as an influencing factor, we can obtain the order volume sequence for different time periods, let's say: (x1, x2, x3, ..., xn). Similarly, we can obtain the recruitment quantity sequence for the corresponding time period, let's say: (y1, y2, y3, ..., yn). Then we can calculate the correlation coefficient between the two sequences. For each influencing factor, we can determine its correlation coefficient with the recruitment quantity, and then rank them according to the magnitude of their influence based on the correlation coefficient.

[0058] In some embodiments, for each influencing factor, analysis of variance can also be used to determine whether there is a significant difference in the number of recruits at different levels of influencing factor. The more significant the difference, the greater the degree of influence.

[0059] In some embodiments, when determining the influence level of each influencing factor, multiple historical recruitment data points can be obtained. Each historical recruitment data point corresponds to a historical recruitment event, and each historical recruitment data point includes the historical input cost of that historical recruitment event, the historical data of the influencing factor, and the historical actual recruitment quantity of that historical recruitment event. For each historical recruitment data point, the input cost in the historical recruitment data can be input into a trained base sub-model. The base sub-model is used to predict the recruitment quantity of the target object, and the difference between the predicted result and the actual result (historical actual recruitment quantity) is determined. The difference sequence formed by the differences corresponding to the multiple historical recruitment data points is used as the dependent variable, and the historical data sequence formed by the historical data of the influencing factor in the multiple historical recruitment data points is used as the independent variable. The dependent variable and the independent variable are fitted, and the goodness of fit between the independent variable and the dependent variable is calculated. The goodness of fit is positively correlated with the influence level. For example, assuming there are n historical recruitment data points, their corresponding difference sequence can be represented as: Y = (y1, y2, ..., yn). Meanwhile, historical data of this influencing factor can be extracted from n historical recruitment data to obtain the historical data sequence of the influencing factor: X = (x1, x2, ..., xn). Then, a regression equation between X and Y can be constructed, X and Y can be fitted, and the goodness of fit R² can be calculated. The goodness of fit R² can be used to determine how much of the predicted difference (i.e. the difference between the predicted result and the actual result) fluctuation can be explained by X. The larger the R², the more significant the influence of this influencing factor on the recruitment quantity.

[0060] For example, suppose there are n historical data points (n is the number of historical recruitment events; for example, n=12 for 12 months of monthly data). Extract n historical values ​​of a certain influencing factor from the historical data: x1, x2, ..., xn (e.g., the monthly ratios of the order trend factor: 1.2, 1.1, 0.9...). Calculate the corresponding n predicted differences: y1, y2, ..., yn, forming a paired dataset: (x1, y1), (x2, y2)...(xn, yn). Based on the relationship between the influencing factor and the predicted differences, select the corresponding regression function. If it is a linear relationship, linear regression can be used, with the equation y=k×x+b (k is the regression coefficient, b is the intercept). If it is a non-linear relationship (e.g., order trend factor, seasonality factor), then multinomial regression (y=k1×x+k2×x²+b) or logarithmic regression (y=k×ln(x)+b) should be used, prioritizing the form that yields a better fit. The parameters (k, b, etc.) of the regression equation are solved by minimizing the sum of squared predicted differences using the least squares method. Then, the goodness-of-fit R² can be calculated. R² is the core evaluation indicator, directly reflecting the factor's explanatory power for the residuals. The formula for calculating the goodness of fit R² is as follows: in: The predicted difference calculated through a regression equation (such as in linear regression). ); : Mean of the predicted difference ( ); molecular : Sum of squared predicted differences from the regression equation (fluctuations in predicted differences not explained by factors); denominator : The sum of squares of the predicted differences (all fluctuations in the predicted differences).

[0061] The R² value ranges from [0,1]. The closer it is to 1, the more fluctuations in the prediction difference that the factor can explain, and the more significant its actual impact on the number of recruits. For example, if the R² of the order trend factor is 0.65 and the R² of the seasonal factor is 0.32, it means that the order trend factor can explain 65% of the error of the basic sub-model, and its impact on the number of recruits is much greater than that of the seasonal factor.

[0062] After obtaining the goodness of fit for each influencing factor, the training order of the dedicated regulation sub-models for each influencing factor can be determined based on the goodness of fit. In some embodiments, when training the dedicated regulation sub-model for each influencing factor, multiple sets of historical recruitment data can be obtained. Each set of historical recruitment data corresponds to a historical recruitment event. Each set of historical recruitment data includes at least the historical investment cost of the historical recruitment event, the historical data of the influencing factor, and the historical actual recruitment quantity of the target object corresponding to the historical recruitment event. Then, the historical investment cost of the historical recruitment event can be input into the trained base sub-model to predict the prediction result. At least based on the prediction result, the historical predicted recruitment quantity without considering the influencing factor can be determined, and the ratio of the historical actual recruitment quantity to the historical predicted recruitment quantity can be determined. The historical data of the influencing factor is used as the input of the dedicated regulation sub-model, and the ratio is used as the label to train the dedicated regulation sub-model.

[0063] For example, historical data of the influencing factor can be obtained as model input to ensure that the input data matches the factor type. For instance, continuous factors maintain their numerical form, while discrete factors are transformed into numerical features through categorical embedding. A "ratio" is used as the training label, and its calculation strictly adheres to the logic of "independent influence of a single factor." The "historical predicted recruitment number without considering this influencing factor" is determined at least based on the prediction results of the base sub-model (in some embodiments, if other influencing factors with higher influence exist, the adjustment results of dedicated adjustment sub-models for other influencing factors can also be superimposed). This ensures that the value only reflects the effect of factors with high cost and importance, excluding interference from the current training influencing factor. The "historical actual recruitment number under the influence of this influencing factor" is the actual number of recruits during the corresponding historical period. The ratio of these two values ​​directly quantifies the proportion by which the current factor causes the recruitment number to deviate from the "baseline predicted value" (a ratio < 1 indicates that the factor promotes recruitment growth, and a ratio > 1 indicates inhibiting growth). During model training, the corresponding regression model can be adapted according to the type of influencing factor (e.g., linear / multinomial regression for continuous factors, and ARIMA-derived model for time-series factors). The model parameters are optimized with the goal of minimizing the error between the predicted ratio and the actual ratio (e.g., MAPE). The resulting dedicated adjustment sub-model can accurately learn the correlation between the historical data of the influencing factor and the corresponding adjustment magnitude, providing reliable support for outputting adjustment coefficients with clear business meaning in subsequent predictions. At the same time, it realizes the independent isolation and quantification of the influence of a single factor, avoiding the influence confusion problem caused by multi-factor mixed training.

[0064] In some embodiments, historical recruitment data may further include historical data of at least one other influencing factor besides the influencing factor. The specific adjustment sub-model of the other influencing factor is trained before the specific adjustment sub-model of the influencing factor. When determining the historical predicted recruitment quantity without considering the influencing factor based on the prediction results, the historical data of the other influencing factor can be input into the trained specific adjustment sub-model of the other influencing factor for each other influencing factor to output the corresponding historical adjustment coefficient. The prediction results are adjusted sequentially using the historical adjustment coefficients corresponding to each other historical influencing factor to obtain the historical predicted recruitment quantity.

[0065] When determining the "historical predicted recruitment quantity without considering this influencing factor," in addition to considering the prediction results of the base sub-model, the prediction results of the dedicated adjustment sub-model, which incorporates other pre-trained influencing factors (for example, in some scenarios, the importance of these other influencing factors may be higher than that of the currently trained influencing factor), can also be incorporated. For instance, for each of the other influencing factors that have been trained in the dedicated adjustment sub-model, their corresponding historical data can be input into their respective dedicated adjustment sub-models, outputting historical adjustment coefficients that quantify their adjustment magnitude. These historical adjustment coefficients are then used to correct the prediction results output by the base sub-model, and the final value obtained is the "historical predicted recruitment quantity excluding the interference of this influencing factor, but incorporating the core cost driver and the moderating effects of other influencing factors." Compared to a single prediction relying solely on the base sub-model, it fully utilizes the results of the already trained adjustment sub-model, making the "predicted value without considering the target factor" more closely reflect the comprehensive impact of historical actual business scenarios, significantly reducing the error of the baseline prediction. Furthermore, it can ensure that the training labels (ratios) corresponding to the current influencing factors can accurately quantify their independent moderating effects, avoiding the confusion caused by the mutual interference of multiple factors, improving the training accuracy of subsequent dedicated adjustment sub-models, and providing more reliable support for the accuracy of the final prediction results.

[0066] The following example illustrates the training process of a recruitment quantity prediction model. Taking delivery capacity recruitment quantity prediction as an example, we assume that in addition to input costs, other influencing factors include order trend factors, seasonal factors, and subsidy cost factors. Therefore, the recruitment quantity prediction model can include one basic sub-model and three specific adjustment sub-models, which we assume are: specific adjustment sub-model 1 for order trend factors, specific adjustment sub-model 2 for seasonal factors, and specific adjustment sub-model 3 for subsidy cost factors. We also assume that the order of influence of these three factors on recruitment quantity, from largest to smallest, is: order trend factor, seasonal factor, and subsidy cost factor.

[0067] like Figure 4 As shown, the entire recruitment quantity prediction model training process is as follows: (1) Training the basic sub-model A large amount of sample data can be obtained. Each set of sample data can include the investment cost corresponding to a historical recruitment event and the actual number of recruits. Then, the investment cost (x1) can be used as the input of the basic sub-model and the actual number of recruits can be used as the label to train the basic sub-model.

[0068] (2) Training the dedicated regulation sub-model 1 Similarly, a large amount of sample data can be obtained. Each set of sample data can include the investment cost, order trend factor, and actual recruitment quantity corresponding to a historical recruitment event. The investment cost can be input into the pre-trained base sub-model to output the predicted recruitment quantity y1 without considering the order trend factor. The ratio y / y1 of the actual recruitment quantity y to the predicted recruitment quantity y1 is determined. Then, the order trend factor (x2) is used as the input of the dedicated adjustment sub-model 1, and the ratio y / y1 is used as the label to train the dedicated adjustment sub-model 1.

[0069] (3) Training a dedicated regulation sub-model 2 A large amount of sample data can be obtained. Each set of sample data can include the input cost, order trend factor, seasonal factor, and actual recruitment quantity corresponding to a historical recruitment event. The input cost can be input into the pre-trained base sub-model to output the predicted recruitment quantity y1 without considering the order trend factor. The order trend factor can be input into the pre-trained dedicated adjustment sub-model 1 to output the adjustment coefficient y2. The adjustment coefficient y2 is used to adjust the predicted recruitment quantity y1 to obtain the final prediction result y2y1. The ratio y / y2y1 of the actual recruitment quantity y and the prediction result y2y1 is determined. Then, the seasonal factor (x3) is used as the input of the dedicated adjustment sub-model 2, and the ratio y / y2y1 is used as the label to train the dedicated adjustment sub-model 2.

[0070] (4) Training a dedicated regulation sub-model 3 A large amount of sample data can be obtained. Each set of sample data can include the input cost, order trend factor, seasonal factor, subsidy cost factor, and actual recruitment quantity corresponding to a historical recruitment event. The input cost can be input into the pre-trained base sub-model to output the predicted recruitment quantity y1 without considering the order trend factor. The order trend factor can be input into the pre-trained dedicated adjustment sub-model 1 to output the adjustment coefficient y2. The seasonal factor can be input into the pre-trained dedicated adjustment sub-model 2 to output the adjustment coefficient y3. The predicted recruitment quantity y1 is adjusted using the adjustment coefficients y2 and y3 to obtain the final prediction result y3y2y1. The ratio y / y3y2y1 of the actual recruitment quantity y to the prediction result y3y2y1 is determined. Then, the subsidy cost factor (x4) is used as the input of the dedicated adjustment sub-model 3, and the ratio y / y3y2y1 is used as the label to train the dedicated adjustment sub-model 3.

[0071] In the reasoning stage, such as Figure 4As shown, the input cost, order trend factor, seasonal factor, and subsidy cost factor corresponding to the current recruitment event can be obtained and input into the basic sub-model, exclusive adjustment sub-model 1, exclusive adjustment sub-model 2, and exclusive adjustment sub-model 3 respectively to obtain the basic recruitment quantity, adjustment coefficient 1, adjustment coefficient 2, and adjustment coefficient 3 respectively. By adjusting the basic recruitment quantity in sequence using adjustment coefficient 1, adjustment coefficient 2, and adjustment coefficient 3, the final predicted recruitment quantity can be obtained.

[0072] In some embodiments, if the impact factor is a sequential impact factor, then the specific moderating sub-model for the impact factor is a time series model. If the impact factor is a continuous numerical impact factor, then the specific moderating sub-model for the impact factor is one of an autolinear regression model, a multinomial regression model, or a logarithmic regression model. If the impact factor is a categorical impact factor, then the specific moderating sub-model for the impact factor is one of a logistic regression model, a LightGBM classification-derived regression model, or a categorical embedding regression model.

[0073] The selection of the dedicated adjustment sub-model follows the core logic of "precise matching of factor characteristics and model capabilities," that is, selecting a suitable model based on the specific data type of the influencing factor. If the influencing factor is a time-dependent serial factor (such as an order volume ratio series, a periodic demand fluctuation series, etc.), then its dedicated adjustment sub-model is a time series model that can capture time series trends and periods to match the serial change pattern of the factor. If the influencing factor is a continuous numerical factor with continuous values ​​and numerical distribution (such as subsidy costs, order trend ratios, etc.), then the dedicated adjustment sub-model is selected from linear regression, multinomial regression, or logarithmic regression models, and can be flexibly matched according to the linear, nonlinear, or diminishing marginal effect correlation patterns between the factor and the adjustment magnitude. If the influencing factor is a discrete categorical factor without clear numerical meaning (such as seasonal factors, regional type factors, etc.), then the dedicated adjustment sub-model is a logistic regression model, a regression model derived from LightGBM classification, or a categorical embedding regression model. Discrete categories are transformed into modelable features through categorical encoding, embedding transformation, etc., and then the regression prediction of the adjustment coefficient is achieved.

[0074] By precisely matching factor types with model capabilities, each dedicated regulation sub-model can fully leverage its modeling advantages for the data type it is suited for, effectively capturing the correlation between different types of factors and regulation magnitude, significantly reducing fitting bias caused by improper model selection, and improving the accuracy of regulation coefficient prediction. This avoids the limitation of using a single model in a "one-size-fits-all" manner in traditional multi-factor prediction.

[0075] In some embodiments, the target object includes delivery capacity, and the at least one influencing factor includes one or more of the following: an order trend factor describing the trend of changes in delivery orders, a seasonal factor describing seasonal attributes, and a subsidy cost for subsidizing delivery capacity.

[0076] In some embodiments, the order trend factor can be characterized by the ratio of the order volume in the current period to the order volume in the previous period, for example, by the ratio of the order volume in the current month to the order volume in the previous month.

[0077] In some embodiments, in order to more accurately reflect the changing trend of order volume, the order trend factor is characterized by an order volume ratio sequence, which is determined based on the following method: obtaining a historical order volume sequence, which includes the historical order volume corresponding to multiple historical time periods; obtaining the current order volume of the current time period; determining the ratio of each historical order volume in the historical order volume sequence to the current order volume; and obtaining the order volume ratio sequence.

[0078] For example, to accurately capture the dynamic trends in order volume to better support recruitment forecasting, the order trend factor is represented by an order volume ratio sequence. Its determination process follows the logic of "using the current period as a benchmark and quantifying historical relative changes." This involves obtaining historical order volume sequences covering multiple continuous or discrete historical periods (such as recent weeks or months) to ensure the sequences cover sufficient time dimensions to reflect trend patterns. Then, the current order volume for the current period to which the prediction node belongs is obtained and used as a benchmark to measure historical order volume changes. Finally, by calculating the ratio of the historical order volume for each historical period in the historical order volume sequence to the current order volume, an order volume ratio sequence consisting of a series of relative ratios is formed, serving as the final representation of the order trend factor. Compared to directly using the absolute value of historical order volume to represent the trend, the ratio form can effectively eliminate the interference of differences in the absolute value of order volume (such as differences in the base of order volume in different business stages and different regions), and focuses more on the "growth or decline of order volume relative to the current situation", accurately reflecting the essence of the trend. The accurate representation of order trend can provide more reliable input for the prediction of subsequent adjustment coefficients, enabling the model to more accurately quantify the adjustment effect of order trend on recruitment quantity, thereby improving the accuracy of overall recruitment quantity prediction and providing a more scientific basis for the platform to optimize recruitment strategies based on order trends.

[0079] In some embodiments, the target includes delivery capacity capable of handling specific types of waybills, wherein the specific types of waybills include long-haul orders (i.e., waybills with a distance exceeding a certain limit), heavy freight waybills, night-time long-haul orders, etc. The input costs are multiple, and the predicted recruitment quantity of the delivery capacity capable of handling the specific types of waybills can be determined under each of the multiple input costs to obtain a mapping relationship between the input costs corresponding to the delivery capacity capable of handling the specific types of waybills and the predicted recruitment quantity. Historical waybill data for the specific type of waybills and the predicted waybill volume for the specific type of waybills in the future can be obtained. Based on the waybill data and the predicted waybill volume, the gap in delivery capacity capable of handling the specific types of waybills can be determined. Based on the gap and the mapping relationship, the actual input costs in the recruitment scheme for recruiting delivery capacity capable of handling the specific types of waybills can be determined.

[0080] For example, for delivery capacity that can handle specific types of orders such as long-distance orders (orders with a distance exceeding a certain distance), heavy freight orders, and nighttime long-distance orders, multiple tiered input costs are first set for them. The predicted recruitment quantity is calculated for each input cost using the aforementioned recruitment quantity prediction scheme. Then, a mapping relationship of "input cost - predicted recruitment quantity" is constructed specifically for this type of segmented delivery capacity (such as the elasticity curve of the correlation between cost growth and recruitment quantity increase). Simultaneously, historical order data for this specific type of waybill (including historical order volume, fulfillment timeliness, capacity handling efficiency, and delivery capacity attrition rate) can be obtained. Based on this data, the average order volume of delivery capacity can be determined. Then, based on the historical order volume, the predicted order volume of this specific type of waybill in the future can be predicted (for example, an order volume prediction model can be pre-trained to predict the predicted order volume of this specific type of waybill in the future). Then, based on the average order volume of delivery capacity, the predicted order volume, the current delivery capacity available to handle this specific type of waybill, and the delivery capacity attrition rate, the gap in delivery capacity for this type of waybill in the future period can be calculated (i.e., the difference between the capacity required to meet the fulfillment needs of this specific type of waybill and the currently available effective capacity). Then, combined with the established mapping relationship, the minimum input cost required to fill the gap can be matched in reverse, and this cost is determined as the actual input cost in the recruitment plan for this type of delivery capacity.

[0081] By subdividing capacity types and binding them to specific waybills, the pain point of mismatch between capacity supply and demand in subdivided scenarios such as long-distance orders and heavy freight is accurately addressed. This avoids the inadequacy of general recruitment solutions to meet specific needs. The logical closed loop of "multi-cost gradient prediction - mapping relationship establishment - gap-oriented cost determination" ensures that recruitment investment costs are precisely linked to gap demand. This avoids capacity gaps caused by insufficient investment and prevents resource waste caused by excessive investment, achieving a better allocation of recruitment costs.

[0082] In some embodiments, for each influencing factor, after determining the adjustment coefficient corresponding to the influencing factor, the basic recruitment quantity can be adjusted separately using the adjustment coefficient corresponding to the influencing factor to obtain the adjusted recruitment quantity. Then, the change in the recruitment quantity of the target object caused by the influencing factor can be determined based on the difference between the adjusted recruitment quantity and the basic recruitment quantity.

[0083] For each influencing factor, the base recruitment quantity output by the base sub-model is individually corrected by using only the adjustment coefficient corresponding to that influencing factor, so as to obtain the adjusted recruitment quantity that only reflects the adjustment effect of that factor. Then, by calculating the difference between the adjusted recruitment quantity and the base recruitment quantity (if the adjustment coefficient < 1, the adjusted recruitment quantity is greater than the base recruitment quantity, and the difference is positive, that is, the recruitment quantity increases; if the adjustment coefficient > 1, the adjusted recruitment quantity is less than the base recruitment quantity, and the difference is negative, that is, the recruitment quantity decreases), the change in the recruitment quantity of the target object under the sole effect of that influencing factor is accurately determined.

[0084] This scheme quantifies the moderating effect of individual influencing factors on recruitment numbers, overcoming the limitations of traditional multi-factor mixed prediction where "multiple factors' influences are intertwined, making it impossible to distinguish the contribution of individual factors." It clearly reveals the specific impact of each factor (e.g., order trend factors can increase recruitment by 150 people, while seasonal factors can decrease recruitment by 80 people), providing direct evidence for refined recruitment decisions on the platform and facilitating targeted optimization strategies (e.g., if the subsidy cost factor shows the largest change, the subsidy level can be adjusted first to improve recruitment effectiveness). Furthermore, it significantly improves the model's interpretability, making the prediction results no longer a "black box output" but allowing for clear tracing of the impact path and magnitude of each factor, lowering the barrier to understanding the model results for users. In some embodiments, recruitment channels for recruiting target groups typically include multiple channels, such as WeChat official account promotion, offline flyer distribution, and APP advertising. For each recruitment channel, multiple input costs can be set. Then, the above scheme is used to predict the number of recruits under each input cost, thus obtaining the mapping relationship between the input cost and the predicted recruitment number for that recruitment channel. For example, an elasticity curve can be obtained for that recruitment channel, which represents the mapping relationship between input cost and predicted recruitment number. After obtaining the mapping relationship between input cost and predicted recruitment number for each recruitment channel, an optimization model can be constructed based on this mapping relationship and the total number of target recruits (i.e., the target number). The optimization model can use the actual input cost corresponding to each recruitment channel as the optimization variable, with the goal of minimizing the total input cost of multiple recruitment channels. Simultaneously, the above mapping relationship and the constraint that the sum of the predicted recruitment numbers for each of the multiple recruitment channels is not less than the target number can be used together. Solving this optimization model yields the actual input cost corresponding to each recruitment channel.

[0085] By deeply integrating recruitment quantity forecasting with cost optimization, a complete "prediction-optimization" closed loop is formed, solving the problem of blind estimation based on experience and lack of data support in traditional recruitment cost allocation, and achieving precise minimization of total investment costs. Through the unique mapping relationship (elastic curve) of each channel, the differences in recruitment efficiency of different channels are fully taken into account (e.g., some channels can achieve high recruitment volume with low investment, while others need to break through the cost saturation point to increase recruitment volume), ensuring that cost allocation is tilted towards efficient channels and improving resource utilization efficiency. With the total target quantity as a rigid constraint, the optimal cost is achieved while ensuring the completion of recruitment tasks, avoiding both the waste of resources caused by excessive investment in a single channel and the recruitment gap caused by insufficient investment.

[0086] Furthermore, this specification also provides a method for training a recruitment quantity prediction model. This recruitment quantity prediction model includes a base sub-model and at least one dedicated adjustment sub-model. Each dedicated adjustment sub-model corresponds to an influencing factor, which is different from the input cost and can affect the recruitment quantity of the target group. Figure 5 As shown, the method may include the following steps: S502. Obtain first sample data, which includes the historical investment cost of the first historical recruitment event and the historical actual recruitment quantity corresponding to the first historical recruitment event. S504. Using the historical investment cost of the first historical recruitment event as input and the historical actual recruitment quantity corresponding to the first historical recruitment event as label, train the basic sub-model. S506. Obtain second sample data, which includes the historical investment cost of the second historical recruitment event, the historical data of at least one of the influencing factors, and the historical actual recruitment quantity corresponding to the second historical recruitment event. S508. Input the historical input cost of the second historical recruitment event into the trained basic sub-model to obtain the prediction result; S510. For each dedicated adjustment sub-model, at least based on the prediction results, determine the historical predicted recruitment quantity corresponding to the second historical recruitment event when the influence factor corresponding to the dedicated adjustment sub-model is not considered, and determine the ratio of the historical actual recruitment quantity corresponding to the second historical recruitment event to the historical predicted recruitment quantity; use the historical data of the influence factor corresponding to the dedicated adjustment sub-model as the input of the dedicated adjustment sub-model, and use the ratio as the label to train the dedicated adjustment sub-model.

[0087] The first historical recruitment event and the second historical recruitment event can be the same recruitment event, meaning the first sample data and the second sample data are the same sample data. Alternatively, the first historical recruitment event and the second historical recruitment event can be different recruitment events, meaning the first sample data and the second sample data are different sample data.

[0088] In some embodiments, the dedicated moderating sub-models of each of the at least one influencing factor are trained sequentially, wherein the dedicated moderating sub-models of the influencing factor that have a greater impact on the recruitment quantity of the target object are trained earlier, and the step of determining the historical predicted recruitment quantity of the second historical recruitment event without considering the influencing factor corresponding to the dedicated moderating model, based at least on the prediction results, includes: For each other dedicated moderating sub-model trained prior to this dedicated moderating sub-model, the historical data of the influencing factors corresponding to the other dedicated moderating sub-model are input into the other dedicated moderating sub-model to output the corresponding historical moderating coefficients; The prediction results are adjusted sequentially using the historical adjustment coefficients output by each of the other dedicated adjustment sub-models to obtain the historical predicted recruitment quantity.

[0089] The specific training process of the recruitment quantity prediction model can be referred to the description in the above embodiments, and will not be repeated here.

[0090] Corresponding to the method embodiments provided in the embodiments of this specification, the embodiments of this specification also provide a computer program product, including a computer program that, when executed by a processor, implements the methods mentioned in any of the above embodiments.

[0091] This description also provides an electronic device, such as... Figure 6 The diagram shown is a structural schematic of an electronic device according to an embodiment of this specification, except... Figure 6 In addition to the processor 62 and memory 64 shown, the device may also include other hardware, such as a forwarding chip responsible for processing messages; from a hardware structure perspective, the device may also be a distributed device, possibly including multiple interface cards to extend message processing at the hardware level. The memory 64 stores computer instructions, and the processor 72 executes the computer instructions to implement the methods mentioned in any of the above embodiments.

[0092] The user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, use and processing of the relevant data must comply with the relevant laws, regulations and standards of the relevant countries and regions, and corresponding operation entry points are provided for users to choose to authorize or refuse.

[0093] Since the parts of the embodiments in this specification that contribute to the prior art, or all or part of the technical solution, can be embodied in the form of a software product, the computer software product is stored in a storage medium and includes several instructions to cause a terminal device to execute all or part of the steps of the methods in the various embodiments of this specification. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0094] The above description is merely a preferred embodiment of the embodiments of this specification and is not intended to limit the embodiments of this specification. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the embodiments of this specification should be included within the scope of protection of the embodiments of this specification.

Claims

1. A method for predicting the number of recruits for a target group, the method comprising: Obtain the input cost of recruiting target individuals, and at least one other influencing factor besides the input cost that can affect the number of target individuals recruited; The input cost is fed into the base sub-model of the pre-trained recruitment quantity prediction model to predict the base recruitment quantity of the target object. The base sub-model is trained with historical input cost as input and historical actual recruitment quantity corresponding to historical input cost as label. For each influencing factor, the influencing factor is input into the dedicated adjustment sub-model of the recruitment quantity prediction model to output the adjustment coefficient corresponding to the influencing factor; wherein, the dedicated adjustment sub-model of each influencing factor is trained with the historical data of the influencing factor as input and the ratio of the historical actual recruitment quantity under the influence of the influencing factor to the historical predicted recruitment quantity without considering the influencing factor as the label. The basic recruitment quantity is adjusted sequentially using the adjustment coefficients corresponding to each of the at least one influencing factor to obtain the predicted recruitment quantity corresponding to the input cost.

2. The method according to claim 1, wherein the base sub-model is trained prior to the dedicated adjustment sub-models for each influencing factor, and the historical predicted recruitment quantity without considering the influencing factor is determined at least based on the prediction results of the trained base sub-model.

3. The method according to claim 2, wherein the dedicated regulation sub-models of each of the at least one influencing factor are trained sequentially in a chronological order, wherein, The training order of the dedicated regulation sub-models for the factors that have a greater impact on the recruitment quantity of the target group is higher. The historical predicted recruitment quantity without considering the factor is determined by combining the prediction results of the trained base sub-model and the prediction results of the dedicated regulation sub-models for other factors that have been trained.

4. The method according to claim 2 or 3, wherein the dedicated regulation sub-model for each influencing factor is trained in the following manner: Acquire multiple sets of historical recruitment data, each set of historical recruitment data corresponding to one historical recruitment event. Each set of historical recruitment data includes at least the historical investment cost of that historical recruitment event, the historical data of that influencing factor, and the historical actual number of the target objects recruited for that historical recruitment event. The historical input cost of this recruitment event is fed into the trained base sub-model to obtain the prediction results; At least based on the prediction results, determine the historical predicted recruitment numbers without considering this influencing factor; Determine the ratio of the historical actual recruitment number to the historical predicted recruitment number; The historical data of the influencing factor is used as the input to the specific regulation sub-model, and the ratio is used as the label to train the specific regulation sub-model.

5. The method according to claim 4, wherein the historical recruitment data further includes historical data of other influencing factors besides the at least one influencing factor, wherein the specific moderating sub-models for the other influencing factors are trained prior to the specific moderating sub-model for the influencing factor, and the step of determining the historical predicted recruitment quantity without considering the influencing factor based at least on the prediction result includes: For each other influencing factor, the historical data of that other influencing factor is input into the trained dedicated adjustment sub-model of that other historical influencing factor to output the corresponding historical adjustment coefficient; The prediction results are adjusted sequentially using the historical adjustment coefficients corresponding to each of the other historical influencing factors to obtain the historical predicted recruitment number.

6. The method according to claim 3, wherein the degree of influence of each influencing factor on the recruitment quantity of the target object is characterized by a goodness-of-fit score, wherein the goodness-of-fit score is determined based on the following method: Obtain multiple historical recruitment data points, each corresponding to a historical recruitment event. Each historical recruitment data point includes the historical investment cost of that recruitment event, the historical data of that influencing factor, and the historical actual recruitment quantity of that recruitment event. For each historical recruitment data point, the historical investment cost in that historical recruitment data point is input into the trained base sub-model to output a prediction result, and the difference between the prediction result and the actual historical recruitment number is determined. The difference sequence formed by the differences corresponding to each of the multiple historical recruitment data is used as the dependent variable, and the historical data sequence formed by the historical data of the influencing factor in the multiple historical recruitment data is used as the independent variable; The dependent and independent variables are fitted, and the goodness of fit between the independent and dependent variables is calculated, wherein, The goodness of fit is positively correlated with the degree of influence.

7. The method according to claim 1, wherein if the impact factor is a sequential impact factor, then the specific moderating sub-model of the impact factor is a time series model; If the impact factor is a continuous numerical impact factor, then the specific moderating sub-model of the impact factor shall be selected from one of the following: autolinear regression model, multinomial regression model, or logarithmic regression model. If the impact factor is a categorical impact factor, then the specific moderating sub-model for the impact factor shall be one of the following: logistic regression model, LightGBM classification-derived regression model, or categorical embedding regression model.

8. The method according to claim 1, wherein the target object includes delivery capacity, and the at least one influencing factor includes one or more of the following: The order trend factor describes the changing trend of delivery orders, the seasonal factor describes seasonal attributes, and the subsidy cost subsidizes delivery capacity.

9. The method according to claim 8, wherein the order trend factor is characterized by an order volume ratio sequence, the order volume ratio sequence being determined based on the following method: Obtain the historical order volume sequence, which includes the historical order volume corresponding to each of multiple historical time periods; Get the current order volume for the current time period; The ratio of each historical order volume to the current order volume in the historical order volume sequence is determined to obtain the order volume ratio sequence.

10. The method according to claim 1, wherein the target object includes delivery capacity capable of handling specific types of waybills, the input cost includes multiple components, and the method further includes: Determine the predicted recruitment quantity of the delivery capacity capable of handling specific types of waybills under multiple input costs, so as to obtain the mapping relationship between input costs and predicted recruitment quantity; Obtain historical waybill data for this specific type of waybill, as well as the predicted waybill volume for this specific type of waybill over a future period. Based on the waybill data and the predicted waybill volume, determine the quantity of delivery capacity gap that can handle the specific type of waybill; The actual input cost in the recruitment scheme for recruiting delivery capacity capable of handling this specific type of waybill is determined based on the number of gaps and the mapping relationship.

Images