Cargo source ranking factor determination method, model training method, and electronic device

Abstract

The application provides a cargo source ranking factor determination method, a model training method and an electronic device, and relates to the technical field of cargo transportation. The model training method comprises the following steps: determining sample bottom line price features based on obtained sample cargo source historical transaction orders; obtaining sample cargo source price competitiveness features and sample cargo source real-time markup features, and training an initial cargo source ranking factor determination model based on the sample bottom line price features, the sample cargo source price competitiveness features and the sample cargo source real-time markup features to obtain a trained cargo source ranking factor determination model; wherein the cargo source ranking factor determination model is used to determine a cargo source ranking factor of a cargo source, and the cargo source ranking factor comprises a click rate and / or a transaction rate of the cargo source. The scheme of the application can improve the cargo source ranking matching effect of the model and ensure the monotonicity of the model in the cargo source price.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of cargo transportation, and in particular to a cargo source ranking factor determination method, a model training method, and an electronic device. BACKGROUND

[0002] In a logistics transportation service platform, cargo source ranking is an important content. For example, the logistics transportation service platform can recommend a cargo source list to a driver through cargo source ranking, and the driver can call the cargo source list through a search engine or a page click operation, so as to facilitate the driver to undertake an order. The ranking result of the cargo source is usually determined by a cargo source ranking factor, such as a click rate of the cargo source. Therefore, the determination of the cargo source ranking factor is crucial to the cargo source ranking.

[0003] A neural network model based on deep learning plays an important role in cargo source ranking. The determination of the cargo source ranking factor can be performed by using a deep neural network model, and then the cargo source ranking is determined based on the cargo source ranking factor. In theory, the higher the price of the cargo source, the greater the probability of driver clicks and transactions, and accordingly, the higher the matching degree of such cargo source. However, in related technologies, since the freight of the cargo source is mainly used as a price feature for training during model training, there are fewer price-related features, which leads to a situation that when the trained model is used for cargo source ranking, the ranking score of the cargo source decreases after the price of the cargo source increases. When it is required to maintain the monotonicity of the price, the matching effect of the model is obviously decreased. SUMMARY

[0004] The present application provides a cargo source ranking factor determination method, a model training method, and an electronic device to improve the matching effect of the model in cargo source ranking.

[0005] In a first aspect, the present application provides a training method of a cargo source ranking factor determination model, comprising:

[0006] obtaining a sample cargo source historical transaction order, and determining a sample bottom line price feature based on the sample cargo source historical transaction order, the sample bottom line price feature being used to represent a sample bottom line price of a plurality of target dimensions and a relationship between a sample cargo source price and the sample bottom line price;

[0007] obtaining a sample cargo source price competitiveness feature and a sample cargo source real-time markup feature, and training an initial cargo source ranking factor determination model based on the sample bottom line price feature, the sample cargo source price competitiveness feature, and the sample cargo source real-time markup feature to obtain a trained cargo source ranking factor determination model; the cargo source ranking factor determination model is used to determine a cargo source ranking factor of the cargo source, and the cargo source ranking factor includes a click rate and / or a transaction rate of the cargo source.

[0008] Optionally, the plurality of target dimensions include at least two of a driver dimension, a driver-by-route dimension, and a route-by-registered driver dimension; and the determining of the sample baseline price feature based on the sample historical transaction order of the cargo source includes:

[0009] determining a minimum value of a sample cargo source standardized price of the driver dimension in the sample historical transaction order of the cargo source, to obtain a first sample baseline price;

[0010] determining a minimum value of a sample cargo source standardized price of the driver-by-route dimension in the sample historical transaction order of the cargo source, to obtain a second sample baseline price;

[0011] determining a minimum value of a sample cargo source standardized price of the route-by-registered driver dimension in the sample historical transaction order of the cargo source, to obtain a third sample baseline price;

[0012] determining at least two of the first sample baseline price, the second sample baseline price, and the third sample baseline price as the sample baseline price of the plurality of target dimensions;

[0013] determining a sample target difference value between the sample baseline price and the sample cargo source price; the sample target difference value includes a sample absolute difference value and / or a sample relative difference value;

[0014] determining the sample baseline price and the sample target difference value as the sample baseline price feature.

[0015] Optionally, the obtaining of the sample cargo source price competitiveness feature and the sample cargo source real-time price increase feature includes:

[0016] obtaining a price quantile of a first sample cargo source in each sample sub-market divided in a first sample preset time period, to obtain the sample cargo source price competitiveness feature;

[0017] obtaining a sample price increase feature of a second sample cargo source in a second sample preset time period, to obtain the sample cargo source real-time price increase feature; wherein the sample price increase feature includes at least one of a sample price increase amount, a first sample time interval from a price increase time to a release time of the second sample cargo source, and a second sample time interval from a last price increase to a current time.

[0018] Optionally, the initial cargo source ranking factor determination model includes an initial price feature extraction layer, an initial feature encoding layer, an initial scene extraction layer, an initial task extraction layer, and an initial task tower layer; and the training of the initial cargo source ranking factor determination model based on the sample baseline price feature, the sample cargo source price competitiveness feature, and the sample cargo source real-time price increase feature includes:

[0019] The sample bottom line price feature, the sample goods source price competitiveness feature, and the sample goods source real-time markup feature are taken as sample price sharing features, and sample scene related features are obtained;

[0020] The sample price sharing features and the sample scene related features are input into the initial feature encoding layer, the sample price sharing features and the sample scene related features are encoded by the initial feature encoding layer, sample price encoding features and sample scene encoding features output by the initial feature encoding layer are obtained;

[0021] The sample price encoding features are input into the initial price feature extraction layer, the sample price encoding features are extracted by the initial price feature extraction layer, and sample price features output by the initial price feature extraction layer are obtained;

[0022] The sample price encoding features and the sample scene encoding features are input into the initial scene extraction layer, the sample price encoding features and the sample scene encoding features are extracted by the initial scene extraction layer, and sample scene features output by the initial scene extraction layer are obtained;

[0023] The sample scene features are input into the initial task extraction layer, the sample scene features are extracted by the initial task extraction layer, and sample task features output by the initial task extraction layer are obtained;

[0024] The sample task features and the sample price features are input into the initial task tower layer, task recognition is performed by the initial task tower layer, and sample goods source ranking factors output by the initial task tower layer are obtained; the sample goods source ranking factors include sample click rates and / or sample transaction rates;

[0025] The model parameters of the initial goods source ranking factor determination model are adjusted based on the sample goods source ranking factors and sample label data until the initial goods source ranking factor determination model converges, and the trained goods source ranking factor determination model is obtained.

[0026] Optionally, the sample price sharing features and the sample scene related features are input into the initial feature encoding layer, the sample price sharing features and the sample scene related features are encoded by the initial feature encoding layer, and sample price encoding features and sample scene encoding features output by the initial feature encoding layer are obtained, including:

[0027] Each type of sample price sharing feature is taken as a sample to-be-bucketed feature, non-default value price features in the sample to-be-bucketed feature are equally bucketed, and a first sample bucket is obtained;

[0028] The default value price feature in the sample to be binned is taken as a new bin and added to the boundary of the first sample bin, together with the first sample bin to form the second sample bin of the sample to be binned feature;

[0029] Based on a preset gain coefficient, the bucket boundary value of the second sample bucket is increased to obtain the target sample bucket feature corresponding to the sample bucket feature to be bucketed.

[0030] The target sample binning features and the sample scene-related features corresponding to the shared features of all types of sample prices are input into the initial feature encoding layer. The initial feature encoding layer encodes all the target sample binning features and the sample scene-related features to obtain the sample price encoding features and sample scene encoding features output by the initial feature encoding layer.

[0031] Optionally, after obtaining the trained source ranking factor determination model, the training method of the source ranking factor determination model further includes:

[0032] Collect real driver-cargo pairs samples, and increase the prices in the real driver-cargo pairs samples based on a preset price increase ratio to obtain virtual driver-cargo pairs samples;

[0033] The real driver-cargo pair sample and the virtual driver-cargo pair sample are determined as test samples;

[0034] For each test sample group in the test sample, the virtual driver-cargo pair samples in the test sample group are input into the cargo source ranking factor determination model to obtain the virtual cargo source ranking factor output by the cargo source ranking factor determination model; the virtual cargo source ranking factor includes virtual click-through rate and / or virtual transaction rate;

[0035] The virtual ranking score of the virtual driver pair samples in the test sample group is determined based on the virtual cargo source ranking factor.

[0036] Determine the Spearman correlation coefficient between the virtual ranking score of the test sample group and the source price of the real cargo pair sample in the test sample group; the Spearman correlation coefficient is used to characterize the price monotonicity of the source ranking factor determination model.

[0037] Secondly, this application provides a method for determining the source ranking factor, including:

[0038] Retrieve historical transaction orders for goods within the first preset time period;

[0039] determine a bottom line price feature based on the historical transaction orders of the goods source, the bottom line price feature being used to represent a bottom line price of a plurality of target dimensions and a relationship between the goods source price and the bottom line price;

[0040] obtain a goods source price competitiveness feature in a second preset time period and a goods source real-time markup feature in a third preset time period;

[0041] input the bottom line price feature, the goods source price competitiveness feature, and the goods source real-time markup feature as price sharing features into a goods source ranking factor determination model to obtain a goods source ranking factor output by the goods source ranking factor determination model;

[0042] The goods source ranking factor determination model is trained based on the training method of the goods source ranking factor determination model of any one of the first aspect.

[0043] Optionally, the goods source ranking factor determination model includes a price feature extraction layer, a feature encoding layer, a scene extraction layer, a task extraction layer, and a task tower layer. The input of the bottom line price feature, the goods source price competitiveness feature, and the goods source real-time markup feature as price sharing features into the goods source ranking factor determination model to obtain the goods source ranking factor output by the goods source ranking factor determination model includes:

[0044] obtain a scene-related feature;

[0045] input the bottom line price feature, the goods source price competitiveness feature, and the goods source real-time markup feature as price sharing features, together with the scene-related feature, into the feature encoding layer, perform feature encoding on the price sharing features and the scene-related feature through the feature encoding layer to obtain price encoding features and scene encoding features output by the feature encoding layer;

[0046] input the price encoding features into the price feature extraction layer, perform price feature extraction on the price encoding features through the price feature extraction layer to obtain price features output by the price feature extraction layer;

[0047] input the price encoding features and the scene encoding features into the scene extraction layer, perform scene feature extraction on the price encoding features and the scene encoding features through the scene extraction layer to obtain scene features output by the scene extraction layer;

[0048] input the scene features into the task extraction layer, perform task feature extraction on the scene features through the task extraction layer to obtain task features output by the task extraction layer;

[0049] The task feature and the price feature are input into the task tower layer, task recognition is performed through the task tower layer, and a goods source sorting factor output by the task tower layer is determined as a goods source sorting factor output by the goods source sorting factor determination model.

[0050] In a third aspect, the present application provides an electronic device, comprising a memory and a processor, wherein the memory stores a computer program capable of running on the processor, and the processor implements the steps of the training method of the goods source sorting factor determination model according to any one of the first aspect or implements the steps of the goods source sorting factor determination method according to any one of the second aspect.

[0051] In a fourth aspect, the present application provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the steps of the training method of the goods source sorting factor determination model according to any one of the first aspect or implement the steps of the goods source sorting factor determination method according to any one of the second aspect.

[0052] The goods source sorting factor determination method, the training method of the goods source sorting factor determination model and the electronic device provided by the present application can be used to train the goods source sorting factor determination model in the training process of the goods source sorting factor determination model, first determine the sample baseline price feature based on the sample goods source historical transaction order, then obtain the sample goods source price competitiveness feature and the sample goods source real-time price increase feature, and train the initial goods source sorting factor determination model based on the sample baseline price feature, the sample goods source price competitiveness feature and the sample goods source real-time price increase feature to obtain the trained goods source sorting factor determination model. The sample baseline price feature is used to represent the sample baseline price of multiple target dimensions and the relationship between the sample goods source price and the sample baseline price, and comes from the sample goods source historical transaction order. The sample baseline price feature can reflect the characteristics of the goods source price and its relationship with user behavior from multiple different dimensions. The sample goods source price competitiveness feature can reflect the value of the sample goods source, and the sample goods source real-time price increase feature can represent the personalized price increase behavior characteristics of the sample goods source. Therefore, when the model is trained based on the sample baseline price feature, the sample goods source price competitiveness feature and the sample goods source real-time price increase feature, the characteristics of the price factor in multiple different dimensions and the correlation between the price feature and user behavior are comprehensively considered, the price-related features are greatly enriched, the model's ability to capture the price is improved, and the model's goods source sorting matching effect is improved, thereby ensuring the monotonicity of the model in the goods source price as much as possible. BRIEF DESCRIPTION OF DRAWINGS

[0053] Figure 1 A flowchart of the training method of the goods source sorting factor determination model provided by the embodiments of the present application is shown in the figure;

[0054] Figure 2A structure and working principle schematic diagram of an initial cargo source ranking factor determination model provided by an embodiment of the present application are provided.

[0055] Figure 3 A flowchart of a cargo source ranking factor determination method provided by an embodiment of the present application is provided.

[0056] Figure 4 A structure and working principle schematic diagram of an initial cargo source ranking factor determination model provided by an embodiment of the present application are provided.

[0057] Figure 5 A structure schematic diagram of a training device of a cargo source ranking factor determination model provided by an embodiment of the present application is provided.

[0058] Figure 6 A structure schematic diagram of a cargo source ranking factor determination device provided by an embodiment of the present application is provided. DETAILED DESCRIPTION

[0059] In the present application, "at least one" means one or more, and "multiple" means two or more. "And / or" describes the association between the associated objects, indicating that there can be three relationships, for example, A and / or B can represent the following situations: A exists alone, A and B exist together, and B exists alone. A and B can be singular or plural. The character " / " generally represents an "or" relationship between the associated objects. "At least one of the following" or similar expressions means any combination of these items, including any combination of single or multiple items. For example, at least one of a, b, or c alone can represent: a alone, b alone, c alone, combination of a and b, combination of a and c, combination of b and c, or combination of a, b, and c. Wherein, a, b, and c can be single or multiple. In addition, the terms "first", "second" are only used to distinguish different description objects, and cannot be understood as indicating or implying relative importance.

[0060] With the rapid development of Internet technology, the transportation service of the logistics industry gradually transits from the traditional offline operation mode to the online unified logistics transportation service platform operation. More and more drivers and consignors rely on comprehensive logistics transportation service platforms to realize the docking of cargo sources and transportation demand. The logistics transportation service platform integrates massive cargo source information to provide extensive transportation capacity selection for consignors and rich freight opportunities for drivers, and plays an important role in the field of logistics transportation.

[0061] In a logistics transportation service platform, cargo source sorting is an important content. For example, the logistics transportation service platform can recommend a cargo source list to a driver through cargo source sorting, and the driver can call the cargo source list through a search engine or a page click operation, so as to facilitate the driver to undertake an order. The sorting result of the cargo source is usually determined by a cargo source sorting factor, such as a click rate of the cargo source. Therefore, the determination of the cargo source sorting factor is crucial to the cargo source sorting.

[0062] A neural network model based on deep learning plays an important role in cargo source sorting. The determination of the cargo source sorting factor can be performed by using a deep neural network model, and then the cargo source sorting is determined based on the cargo source sorting factor. In theory, the higher the price of the cargo source, the greater the probability of driver clicks and transactions, and accordingly, the higher the matching degree of such cargo source. However, in related technologies, since the freight of the cargo source is mainly used as a price feature for training during model training, there are fewer price-related features, which leads to a situation that when the trained model is used for cargo source sorting, the matching degree of the cargo source decreases after the price of the cargo source increases. When it is required to maintain the monotonicity of the price, the matching effect of the model decreases significantly.

[0063] On the other hand, in a multi-task scenario, different tasks may have different dependencies on the price factor. For example, the click rate of the cargo source may be more affected by the price, while the collection rate of the cargo source may be more related to the quality of the cargo source. It is difficult to coordinate the price factor among multiple tasks by using a single cargo source freight as a price feature to train the model.

[0064] Therefore, an improved training method of a cargo source sorting factor determination model and a method of determining a cargo source sorting factor by using a cargo source sorting factor determination model trained based on the training method are provided. The sample bottom-line price features are extracted from the historical transaction orders of the sample cargo source, the sample bottom-line price features, the sample cargo source price competitiveness features and the sample cargo source real-time price increase features are used as the sample features related to the price for model training, and an initial cargo source sorting factor determination model is trained to obtain a trained cargo source sorting factor determination model. The trained cargo source sorting factor determination model can learn the price-related features from multiple dimensions, enhance the ability of the model to capture the price information, and thus improve the monotonicity and matching effect of the model on the price of the cargo source. In a multi-task scenario, the ability of the model to coordinate the price factor among multiple tasks can also be improved.

[0065] The training method of the cargo source sorting factor determination model provided in the embodiments of the present application can be applied to an electronic device, and can also be applied to a training device of a cargo source sorting factor determination model arranged in the electronic device. The device can be realized by software, hardware or a combination of both. The electronic device can include at least one of a server, a mobile phone, a computer, a vehicle terminal, a tablet computer, a wearable device and a smart home device. The server can include a standalone server, a virtual server and a cluster server.

[0066] The following takes an application to an electronic device as an example, combined with Figures 1-2 The training method of the goods source ranking factor determination model provided in the embodiments of the present application is described in detail.

[0067] Figure 1 The flowchart of the training method of the goods source ranking factor determination model provided in the embodiments of the present application is shown, referring to Figure 1 As shown in the figure, the training method of the goods source ranking factor determination model can include the following steps 110~120.

[0068] Step 110: Obtain sample goods source historical transaction orders, and determine sample bottom line price features based on the sample goods source historical transaction orders.

[0069] Among them, the sample bottom line price features are used to represent the sample bottom line prices of multiple target dimensions and the relationship between the sample goods source prices and the sample bottom line prices. The multiple target dimensions can include at least two of the driver dimension, the driver and route intersection dimension, and the route and registered driver intersection dimension, but are not limited thereto.

[0070] Specifically, the electronic device can obtain the goods source historical transaction orders in a preset historical time period from the logistics transportation service platform to obtain the sample goods source historical transaction orders. The sample goods source historical transaction orders can reflect the behavior information of the sample driver and the transaction information between the sample driver and the sample goods source in the preset historical time period.

[0071] After obtaining the sample goods source historical transaction orders, the sample price information in the sample goods source historical transaction orders can be identified, and then the sample price information is subjected to statistical analysis of multiple target dimensions according to the sample price attribute information corresponding to the sample price information, to obtain the sample bottom line price features. Among them, the sample price attribute information can include at least one of the sample driver information, the sample route information, the sample driver information, and the sample goods source information, but is not limited thereto.

[0072] The number of sample goods source historical transaction orders is large, and the price unit of the sample goods source contained therein can also be large, such as some sample goods sources priced by transportation trips, some sample goods sources priced by weight, etc., and the same pricing unit can also have different pricing methods. At the same time, the price distribution of less than full load and full load is naturally not in the same space, and the price of less than full load is generally lower. Based on this, in one embodiment, before the sample price information is subjected to statistical analysis of multiple target dimensions, the sample price information can be standardized to obtain corresponding sample goods source standardized prices, and then the sample goods source standardized prices are subjected to statistical analysis of multiple target dimensions.

[0073] For example, the sample cargo source price can be standardized in a normalized manner of "calculating the vehicle kilometer price of the whole vehicle calculation sample cargo source, and calculating the ton kilometer price of the less than full load calculation sample cargo source".

[0074] For example, determining the sample baseline price features based on the sample cargo source historical transaction order can include steps 111-116 as follows.

[0075] Step 111: Determine the minimum value of the sample cargo source standardized price in the driver dimension of the sample cargo source historical transaction order to obtain the first sample baseline price.

[0076] For each sample driver in the sample cargo source historical transaction order, the electronic device can obtain the sample order of the sample driver from the sample cargo source historical transaction order, then obtain the sample cargo source price in each sample order, and standardize the sample cargo source price to obtain the sample cargo source standardized price per kilometer of the sample cargo source in each sample order, and then determine the minimum value of the sample cargo source standardized price to obtain the first sample baseline price of the sample driver. In this way, the first sample baseline price corresponding to each sample driver in the sample cargo source historical transaction order can be obtained, that is, the first sample baseline price in the driver dimension of the sample cargo source historical transaction order is obtained.

[0077] Step 112: Determine the minimum value of the sample cargo source standardized price in the driver and route intersection dimension of the sample cargo source historical transaction order to obtain the second sample baseline price.

[0078] For each sample driver in the sample cargo source historical transaction order, the electronic device can obtain the sample order of the sample driver from the sample cargo source historical transaction order, then obtain the first sample route information and the corresponding sample cargo source price from each sample order, and standardize the sample cargo source price to obtain the corresponding sample cargo source standardized price, then for each first sample route in the first sample route information, determine the minimum value of the sample cargo source standardized price corresponding to the first sample route to obtain the second sample baseline price corresponding to the first sample route. In this way, the second sample baseline price of each first sample route of each sample driver can be obtained, that is, the second sample baseline price in the driver and route intersection dimension of the sample cargo source historical transaction order is obtained.

[0079] Step 113: Determine the minimum value of the sample cargo source standardized price in the route and registered truck driver intersection dimension of the sample cargo source historical transaction order to obtain the third sample baseline price.

[0080] The electronic device can obtain all route information from the sample source historical transaction order, and obtain second sample route information. For each second sample route in the second sample route information, the electronic device can obtain all registered trucker information corresponding to the second sample route and the corresponding sample source price from the sample source historical transaction order, and standardize the sample source price to obtain the sample registered trucker corresponding to the second sample route and the sample source standardized price corresponding to each sample registered trucker. Then, for each sample registered trucker corresponding to the second sample route, the minimum value of the sample source standardized price corresponding to the sample registered trucker is determined to obtain the third sample bottom line price of the sample registered trucker corresponding to the second sample route. In this way, the third sample bottom line price of each sample registered trucker of each sample route can be obtained, that is, the third sample bottom line price of the route and registered trucker cross-dimension in the sample source historical transaction order is obtained.

[0081] Step 114: At least two of the first sample bottom line price, the second sample bottom line price and the third sample bottom line price are determined as the sample bottom line price of the plurality of target dimensions.

[0082] After obtaining the first sample bottom line price, the second sample bottom line price and the third sample bottom line price, at least two of the sample bottom line prices can be determined as the sample bottom line price of the plurality of target dimensions.

[0083] Step 115: Determine the sample target difference value between the sample bottom line price and the sample source price.

[0084] Among them, the sample target difference value includes at least one of the sample absolute difference value and the sample relative difference value.

[0085] After obtaining the sample bottom line price, for each sample bottom line price in the sample bottom line price, the sample absolute difference value and / or the sample relative difference value between the sample bottom line price and the corresponding sample source price are determined by feature crossing the sample bottom line price and the corresponding sample source price.

[0086] For example, the first sample bottom line price in the sample bottom line price includes the driver dimension, that is, the first sample bottom line price corresponding to each sample driver. For each sample driver, the absolute difference value between the first sample bottom line price of the sample driver and the sample source price in each sample order corresponding to the sample driver can be determined to obtain the sample absolute difference value of each sample order corresponding to the sample driver. Further, the ratio of the sample absolute difference value to the corresponding sample source price can be determined to obtain the sample relative difference value corresponding to the sample driver.

[0087] In this way, by explicitly crossing the features of the sample bottom line price and the sample source price, the sample absolute difference and / or the sample relative difference between the sample bottom line price and the corresponding sample source price are determined, which can explicitly depict the relationship between the sample bottom line price and the sample source price.

[0088] Step 116: determining the sample bottom line price and the sample target difference value as the sample bottom line price feature.

[0089] The electronic device can determine all the determined sample bottom line prices and sample target difference values as the sample bottom line price feature.

[0090] Step 120: obtaining the sample source price competitiveness feature and the sample source real-time markup feature, and training the initial source ranking factor determination model based on the sample bottom line price feature, the sample source price competitiveness feature and the sample source real-time markup feature to obtain the trained source ranking factor determination model.

[0091] The source ranking factor determination model is used to determine the source ranking factor of the source, which can be used to rank the source. The source ranking factor can include the click rate and / or the transaction rate of the source.

[0092] For example, taking the source ranking factor as including the click rate and the transaction rate of the source, after the click rate and the transaction rate of the source are determined, the click rate and the transaction rate of the source can be power-weighted and summed to obtain the ranking score of the source, and then the sources can be ranked in descending order of the ranking score.

[0093] The sample source price competitiveness feature can represent the competitiveness or value of the sample source. The higher the competitiveness of the sample source, the greater the probability and speed of transaction. The sample source real-time markup feature is used to represent the individualized markup behavior of the sample source.

[0094] Specifically, in an embodiment, obtaining the sample source price competitiveness feature and the sample source real-time markup feature can include steps 1211-1212 as follows.

[0095] Step 1211: obtaining the price quantile of the first sample source in each sample sub-market in a first sample preset time period, to obtain the sample source price competitiveness feature.

[0096] The first sample preset time period can be set according to actual needs, such as 48 hours, which is not specially limited in the embodiment. The sample sub-market can be divided based on the sample attribute information. The sample sources with the same sample attribute are divided in the same sample sub-market, and the sample attribute information can include at least one of the non-price information such as the sample source route and the sample vehicle length, but is not limited thereto.

[0097] Specifically, when a sample source of a ticket is published or the price of the sample source of the ticket is updated on the logistics transportation service platform, the sample source of the ticket can be divided into a corresponding sample sub-market according to sample attribute information of the sample source of the ticket. For example, the sample sources in the sample sub-market can be sorted according to the sample source prices, and the competitiveness of the sample sources can be represented by the ordered price quantiles of the sample sources in the corresponding sample sub-market.

[0098] The electronic device can obtain the price quantiles of the first sample sources in the respective divided sample sub-markets in a first sample preset time period, to obtain the sample source price competitiveness features.

[0099] Step 1212: Obtain the sample markup features of the second sample sources in a second sample preset time period, to obtain the sample source real-time markup features.

[0100] The sample markup features include at least one of the sample markup amount, the first sample time interval from the markup time to the publishing time of the second sample source, and the second sample time interval from the last markup to the current time, but are not limited thereto.

[0101] The second sample preset time period can be set according to actual needs, such as 2 days or 5 days, and the present application embodiment does not specially limit this.

[0102] Specifically, for the sample sources published on the logistics transportation service platform, if no transaction is made in a period of time, the consignor can perform a markup operation. The electronic device can count the sample markup features of the second sample sources in the second sample preset time period, to obtain the sample source real-time markup features. The second sample sources are the sample sources that have performed the markup operation.

[0103] After obtaining the sample bottom line price features, the sample source price competitiveness features, and the sample source real-time markup features, the electronic device can supervise and train the initial source ranking factor determination model based on the sample bottom line price features, the sample source price competitiveness features, and the sample source real-time markup features, to obtain a trained source ranking factor determination model. Then, the source ranking factor determination model can be used to determine the source ranking factor of the real-time sources published on the logistics transportation service platform, and the sources can be ranked based on the determined source ranking factor.

[0104] For example, the initial source ranking factor determination model can be a model of a deep neural network architecture. The electronic device can input the sample bottom line price feature, the sample source price competitiveness feature, and the sample source real-time markup feature into the initial source ranking factor determination model, and perform supervised training on the initial source ranking factor determination model. The sample label data during the training includes a label source ranking factor, and the label source ranking factor includes a label click rate and / or a label transaction rate.

[0105] For example, the initial source ranking factor determination model can be a neural network model of a multi-scene multi-task architecture. For example, a currently existing neural network model of a multi-scene multi-task architecture, such as a Hierarchical information extraction Network (HiNet) model, can be used. Alternatively, a new model can be formed by adding an initial price feature extraction layer for extracting price features to a currently existing neural network model of a multi-scene multi-task architecture. The initial price feature extraction layer is used to extract price features from the sample price features encoded by the neural network model of the multi-scene multi-task architecture, and the extracted price features are input into the last initial task tower layer of the neural network model of the multi-scene multi-task architecture together with other features for task recognition, so as to improve the shallow expression of the price features.

[0106] In the application scenario of the multi-scene multi-task, the electronic device can further obtain sample scene related features, input the sample scene related features together with the sample bottom line price feature, the sample source price competitiveness feature, and the sample source real-time markup feature into the initial source ranking factor determination model, and perform supervised training on the initial source ranking factor determination model to obtain the trained source ranking factor determination model. The sample scene related features can include sample scene shared features, sample scene unique features, sample scene identification features, and sample scene attribute features. The sample scene shared features are used to represent features shared by different scenes. The sample scene unique features are used to represent features unique to a scene. The sample scene identification features are used to identify different scenes. The sample scene attribute features are used to represent attribute information of a scene, such as the name of the scene.

[0107] For example, the sample scenario shared features can include at least one of a sample distance of the sample driver from a sample loading location of the sample consignor, a sample number of consignors clicked by the sample driver in a first preset sample time period, and a ratio of a sample consignor weight to a vehicle carrying capacity of the sample driver, but are not limited to this; the sample scenario specific features can include at least one of a sample number of consignor clicks by the sample driver on a sample route in each scenario in a past second preset sample time period, a sample duration for which the sample driver subscribes to the sample route, and a sample search number of the sample driver in a third preset sample time period, but are not limited to this; the sample scenario identification features can include a corresponding number of the scenario, etc.; and the sample scenario attribute features can include a name of the scenario, etc.

[0108] Figure 2 The structure and working principle of the initial consignor ranking factor determination model provided by the embodiments of the present application are shown. In one embodiment, referring to FIG. 13, the initial consignor ranking factor determination model can include an initial price feature extraction layer 201, an initial feature encoding layer 202, an initial scenario extraction layer 203, an initial task extraction layer 204, and an initial task tower layer 205. Figure 2

[0109] Correspondingly, the step 120 of training the initial consignor ranking factor determination model based on the sample bottom line price features, the sample consignor price competitiveness features, and the sample consignor real-time price increase features can include the following steps 1221-1227.

[0110] Step 1221: Take the sample bottom line price features, the sample consignor price competitiveness features, and the sample consignor real-time price increase features as sample price shared features, and obtain sample scenario related features.

[0111] The sample scenario related features can include sample scenario shared features, sample scenario specific features, sample scenario identification features, and sample scenario attribute features.

[0112] Step 1222: Input the sample price shared features and the sample scenario related features into the initial feature encoding layer 202, and encode the sample price shared features and the sample scenario related features through the initial feature encoding layer 202 to obtain sample price encoding features and sample scenario encoding features output by the initial feature encoding layer 202.

[0113] The initial feature encoding layer 202 is configured to encode the sample price shared features and the sample scenario encoding features.

[0114] ​In an embodiment, the sample bottom line price feature, the sample source price competitiveness feature, and the sample source real-time price increase feature can be directly input into the initial feature encoding layer 202 as sample price sharing features, and the initial feature encoding layer 202 encodes the sample price sharing features to obtain sample price encoding features corresponding to the sample price sharing features output by the initial feature encoding layer 202.

[0115] Considering that the sample price sharing features can be continuous features, the sample price sharing features can be input into the initial feature encoding layer 202 for feature encoding after being binned, that is, the sample price sharing features are discretized and then encoded. The binning method of continuous features affects the sensitivity of the model to prices, which can cause the probability of feature binning change to be smaller when the source price increases than when the source price decreases. At the same time, it is found in actual testing that about 30% of the price feature bins of the samples are divided into the default value bin, and the boundary granularity of the remaining bins is relatively coarse, which also leads to a low sensitivity of the model to prices.

[0116] Based on this, in another embodiment, the binning method can be improved. The default value samples are removed first, and then the remaining samples are re-binned according to the price-related features to refine the granularity of the bins. Then, the price features of the default value samples are appended to the bin boundaries. In addition, after obtaining the new feature bin boundaries, the bin boundaries can be increased by a preset percentage, such as 1%, which can solve the problem of unchanged binning after price increase.

[0117] Specifically, the sample price sharing features and the sample scenario-related features are input into the initial feature encoding layer 202, and the initial feature encoding layer 202 encodes the sample price sharing features and the sample scenario-related features to obtain sample price encoding features and sample scenario encoding features output by the initial feature encoding layer 202, which can include: taking each type of sample price sharing feature as a sample to-be-binned feature, equally binning the non-default value price features in the sample to-be-binned feature to obtain a first sample bin; taking the default value price feature in the sample to-be-binned feature as a new bin, appending it to the boundary of the first sample bin, and forming a second sample bin of the sample to-be-binned feature together with the first sample bin; based on a preset gain coefficient, gain the bin boundary value of the second sample bin to obtain target sample bin features corresponding to the sample to-be-binned feature; the sample scenario-related features and the target sample bin features corresponding to all types of sample price sharing features are input into the initial feature encoding layer 202, and the initial feature encoding layer 202 encodes the target sample bin features and the sample scenario-related features to obtain sample price encoding features and sample scenario encoding features output by the initial feature encoding layer 202.

[0118] The default value price feature is a price feature corresponding to a sample source with no price, such as a 0 price feature. For example, for the sample source real-time price increase feature, some sample sources may have no price increase behavior, and the price-related feature of the price increase behavior at this time is the default value price feature. The non-default value price feature is the remaining price feature excluding the default value price feature. The preset gain coefficient can be set according to actual needs or through experiments, such as 1%, indicating that the gain is 1%.

[0119] For example, taking the sample source real-time price increase feature as an example, some sample sources have no price increase behavior, and the price-related feature (i.e., the default value price feature) corresponding to these sample sources can be first removed from the sample source real-time price increase feature. The remaining features are equally divided into buckets according to the price value, such as dividing into 5 buckets, and the 5 buckets are determined as the first sample bucket. Then, the features removed before are taken as a new bucket and added before the first bucket of the first sample bucket to form the second sample bucket of the sample price sharing feature together with the first sample bucket, that is, 6 buckets are obtained. Then, the bucket boundary values of each of the 6 buckets can be increased by 1% to obtain 6 new buckets, and the 6 new buckets are determined as the target sample bucket feature. Then, the target sample bucket feature and the sample scene-related feature are input into the initial feature encoding layer 202, and the feature encoding is performed through the initial feature encoding layer 202 to obtain the sample price encoding feature and the sample scene encoding feature output by the initial feature encoding layer 202.

[0120] In this way, by first performing the bucketing on the non-default value price feature and then adding the default value price feature as a new bucket to the boundary of the already bucketed bucket, the granularity of the bucketing can be refined, thereby improving the sensitivity of the model to the price. Meanwhile, by increasing the bucket boundary values of each of the divided buckets, the sensitivity of the bucketing to the price change can be improved, and the sensitivity of the model to the price is further improved.

[0121] For example, for the sample source price competitiveness feature, considering that the distribution variance thereof can be large, the sample source price competitiveness feature can be respectively subjected to first equal-interval bucketing and second equal-interval bucketing, such as 5-bucketing and 10-bucketing, and the bucketing results are taken as the target sample bucket feature of the sample source price competitiveness feature. In this way, the bucketing effect of the sample source price competitiveness feature can be improved, and the sensitivity and effect of the model are improved.

[0122] Step 1223: inputting the sample price encoding feature into the initial price feature extraction layer 201, performing price feature extraction on the sample price encoding feature through the initial price feature extraction layer 201, and obtaining the sample price feature output by the initial price feature extraction layer 201.

[0123] The initial price feature extraction layer 201 is used to extract price features. For example, it can be implemented using a multilayer perceptron (MLP) neural network structure.

[0124] Step 1224: Input the sample price coding features and sample scene coding features into the initial scene extraction layer 203. The initial scene extraction layer 203 extracts scene features from the sample price coding features and sample scene coding features to obtain the sample scene features output by the initial scene extraction layer 203.

[0125] The initial scene extraction layer 203 can be used to extract scene-specific information and valuable information shared between scenes from the sample price coding features and sample scene coding features, thus obtaining sample scene features. Extracting scene features from the sample price coding features and sample scene coding features helps improve the information representation capabilities of the task layer.

[0126] For example, such as Figure 2 As shown, the initial scene extraction layer 203 may include an initial scene common expert network and an initial scene specific expert network. The sample scene shared coding features in the sample price coding features and sample scene coding features can be input into the initial scene common expert network and the initial scene specific expert network of the initial scene extraction layer 203, respectively. The sample scene specific features in the sample scene coding features can be input into the initial scene specific expert network. The common information between different scenes is extracted through the initial scene common expert network, and the scene-specific information is learned through the initial scene specific expert network. The initial scene extraction layer 203 outputs the extracted common information between different scenes and the scene-specific information as sample scene features.

[0127] Considering the potential correlation between different scenarios, information from other scenarios may also contribute to the information representation of the current scenario. Furthermore, considering the varying representational capabilities of different scenarios, in one embodiment, the initial scenario extraction layer 203 may further include an initial scenario attention network. The electronic device may also input sample price encoding features and all sample scenario encoding features into the initial scenario attention network for scenario perception. Then, the sample scenario attention features output by the initial scenario attention network, along with the features output by the initial scenario common expert network and the initial scenario specific expert network, are used as the sample scenario features output by the initial scenario extraction layer 203. This measures the importance of the contribution of other scenario information to the information representation of the current scenario, thereby enhancing the information representation capability of the current scenario.

[0128] Step 1225: input the sample scene feature into the initial task extraction layer 204, and perform task feature extraction on the sample scene feature through the initial task extraction layer 204 to obtain the sample task feature output by the initial task extraction layer 204.

[0129] The sample scene feature output by the initial scene extraction layer 203 can be input into the initial task extraction layer 204 after vector splicing, and the sample scene feature is extracted through the initial task extraction layer 204. The initial task extraction layer 204 can be used to extract shared information of all tasks and unique information of each task in the sample scene feature to obtain the sample task feature.

[0130] For example, as shown in Figure 2 The initial task extraction layer 204 can include an initial task shared expert network and an initial task unique expert network, and the sample scene feature can be input into the initial task shared expert network and the initial task unique expert network respectively. The initial task shared expert network learns the shared information in all tasks in the current scene, and the initial task unique expert network extracts the unique information of each task in the current scene. The initial task extraction layer 204 outputs all the shared information in the tasks and the unique information of each task as the sample task feature.

[0131] It can be understood that the initial task unique expert network can be designed according to the number of tasks, and each task corresponds to an initial task unique expert network. For example, as shown in Figure 2 It can include two initial task unique expert networks, i.e., a first initial task unique expert network and a second initial task unique expert network, which are respectively assigned to two tasks, such as click rate and transaction rate.

[0132] Step 1226: input the sample task feature and the sample price feature into the initial task tower layer 205, and perform task identification through the initial task tower layer 205 to obtain the sample source sorting factor output by the initial task tower layer 205.

[0133] The sample source sorting factor includes sample click rate and / or sample transaction rate.

[0134] The sample task feature output by the initial task extraction layer 204 and the sample price feature output by the initial price feature extraction layer 201 can be input into the initial task tower layer 205 after vector splicing according to different tasks, and the initial task tower layer 205 is used to identify the task to obtain the sample source sorting factor output by the initial task tower layer 205.

[0135] In this way, by splicing the additional extracted sample price feature into the input vector of the initial task tower layer 205, the sample task feature output by the initial task extraction layer 204 is used for task identification, which can improve the shallow expression mode of the price feature, and further improve the sensitivity and monotonicity of the initial task tower layer 205 and the model to the price factor.

[0136] Step 1227: Adjust the model parameters of the initial source sorting factor determination model based on the sample source sorting factor and the sample label data until the initial source sorting factor determination model converges, and obtain the trained source sorting factor determination model.

[0137] Specifically, the electronic device can determine the loss value of the loss function of the initial source sorting factor determination model based on the sample source sorting factor and the sample label data corresponding to the sample price sharing feature, and then adjust the model parameters of the initial source sorting factor determination model based on the loss value until the loss value converges. The initial source sorting factor determination model converges, and the trained source sorting factor determination model can be obtained.

[0138] The sample label data corresponding to the sample price sharing feature includes a label source sorting factor, and the label source sorting factor includes a label click rate and / or a label transaction rate.

[0139] After obtaining the sample source sorting factor, the sample source sorting factor and the label source sorting factor can be compared, and the loss value of the loss function of the initial source sorting factor determination model is determined according to the comparison result. The loss function can select the existing loss function, which is not limited in the present application.

[0140] Then, the trained source sorting factor determination model can be used to determine the source sorting factor of the real-time source published on the logistics transportation service platform, and then sort the source based on the determined source sorting factor.

[0141] The training method of the cargo source ranking factor determination model provided in the embodiments of the present application first determines sample baseline price features based on sample cargo source historical transaction orders, then obtains sample cargo source price competitiveness features and sample cargo source real-time price increase features, and trains an initial cargo source ranking factor determination model based on the sample baseline price features, the sample cargo source price competitiveness features and the sample cargo source real-time price increase features to obtain a trained cargo source ranking factor determination model. The sample baseline price features are used to represent the sample baseline prices of multiple target dimensions and the relationship between the sample cargo source prices and the sample baseline prices, and are derived from the sample cargo source historical transaction orders. The sample baseline price features can reflect the characteristics of the cargo source prices and their relationship with user behaviors from multiple different dimensions. The sample cargo source price competitiveness features can reflect the value of the sample cargo source. The sample cargo source real-time price increase features can represent the individualized price increase behavior characteristics of the sample cargo source. In this way, when the model is trained based on the sample baseline price features, the sample cargo source price competitiveness features and the sample cargo source real-time price increase features, the characteristics of the price factor in multiple different dimensions and the correlation between the price features and user behaviors are comprehensively considered, the price-related features are greatly enriched, the price capturing capability of the model is improved, and the cargo source ranking matching effect of the model is improved, thereby ensuring the monotonicity of the model in the cargo source price as much as possible.

[0142] The price factor is an important factor in the cargo source ranking matching scenario, and other features also greatly affect the matching effect of the cargo source ranking. For example, whether it is a new cargo, the matching degree of the cargo source and the driver screening condition, etc. It is difficult to measure the optimization effect of the model using real samples. Moreover, the AUC (Area Under Curve) index and GAUC (Gross Area Under Curve) index commonly used in the ranking model cannot meet the performance of optimizing the price monotonicity of the model.

[0143] Therefore, based on the training method of the cargo source ranking factor determination model in the above embodiments, in an embodiment of the present application, the interference problem caused by other features and factors can be solved by using an adversarial test method. Real samples on the line can be sampled, and only the relevant features of these samples in terms of price are improved to the same degree, that is, virtual price increases, while other features remain unchanged. Each real sample can derive a preset number of virtual samples, such as 30 virtual samples, and the price increase range can be uniformly distributed in a preset interval, such as 1% to 30%, thereby forming test samples. The virtual ranking score of the virtual sample is obtained by offline inference on the trained cargo source ranking factor determination model. The consistency of the cargo source price and the virtual ranking score in the ranking is calculated using the Spearman correlation coefficient in each sample pair, so as to quantitatively analyze whether the monotonic relationship of the price factor is enhanced.

[0144] Specifically, after obtaining the trained supplier ranking factor determination model, the training method of the supplier ranking factor determination model can further include steps 130-170 as follows.

[0145] Step 130: Collect real supplier and commodity pairs, and raise the prices of the real supplier and commodity pairs based on a preset price raising ratio to obtain virtual supplier and commodity pairs.

[0146] The preset price raising ratio can be set according to actual needs, for example, it can be set in the interval of 1% to 30%.

[0147] For each real supplier and commodity pair, the price of the real supplier and commodity pair can be raised for a preset number of times to obtain a corresponding preset number of virtual supplier and commodity pairs. The preset price raising ratio is different each time, for example, it can be uniformly distributed in the preset interval, for example, uniformly distributed in the interval of 1% to 30%. The preset number can be designed as needed, for example, 30 or 50, etc.

[0148] Step 140: Determine the real supplier and commodity pairs and the virtual supplier and commodity pairs as test samples.

[0149] It can be understood that in the test samples, each real supplier and commodity pair and the multiple virtual supplier and commodity pairs derived therefrom form a test sample group.

[0150] Step 150: For each test sample group in the test samples, input the virtual supplier and commodity pairs in the test sample group into the supplier ranking factor determination model to obtain the virtual supplier ranking factor output by the supplier ranking factor determination model.

[0151] The virtual supplier ranking factor includes a virtual click rate and / or a virtual transaction rate.

[0152] Step 160: Determine the virtual ranking score of the virtual supplier and commodity pairs in the test sample group based on the virtual supplier ranking factor.

[0153] Taking the virtual supplier ranking factor including the virtual click rate and the virtual transaction rate as an example, the virtual click rate and the virtual transaction rate can be weighted and summed or power weighted and summed to obtain the virtual ranking score. In this way, for each test sample group, the virtual ranking score of the virtual supplier and commodity pairs in the test sample group can be determined.

[0154] Step 170: Determine the Spearman correlation coefficient between the virtual ranking score of the test sample group and the supplier price of the real supplier and commodity pair in the test sample group.

[0155] The Spearman correlation coefficient is used to represent the price monotonicity of the supplier ranking factor determination model.

[0156] For each test sample set, after obtaining the virtual ranking score of the virtual supplier-cargo pair sample in the test sample set, the Spearman coefficient between the virtual ranking score and the cargo source price of the real supplier-cargo pair sample in the test sample set can be determined, the Spearman coefficient is used to analyze the consistency of the cargo source price and the virtual ranking score in ranking, and whether the monotonic relationship of the price factor is enhanced is quantitatively analyzed.

[0157] Through experiments, for the cargo source ranking factor determination model trained by the training method of the cargo source ranking factor determination model provided in the embodiments of the present application, for the cargo source with a price competitiveness in the range of 10 to 90, the Spearman coefficient is tested by using the anti-test set with a price increase of not more than 10%, the average value of the Spearman coefficient is increased by 2.85pp, and the median is increased by 2.15pp; the Spearman coefficient is tested by using the anti-test set with a price increase of not more than 20%, the average value of the Spearman coefficient is increased by 0.74pp, and the median is increased by 0.86pp. From the matching effect of the model, the GAUC (Grouped Area Under Curve) index of the click rate is increased by 0.22pp, and the GAUC index of the transaction rate is increased by 0.24pp. The cargo source ranking matching effect of the cargo source ranking factor determination model is obviously improved, and the monotonic relationship of the price factor is enhanced. Wherein, "pp" represents the absolute value of the percentage direct subtraction, for example, 3%-2%=1%=1pp. The GAUC is an index for evaluating the performance of the model, and the AUC (Area Under Curve) is calculated in the group after the samples are grouped, and the probability that the positive sample prediction score is higher than the negative sample is evaluated. The AUC is also an index for evaluating the performance of the model, which refers to the area under the ROC (Receiver Operating Characteristic) curve.

[0158] Based on the training method of the cargo source ranking factor determination model of the above-mentioned embodiments, the embodiments of the present application also provide a cargo source ranking factor determination method. The cargo source ranking factor determination method can be applied to an electronic device, and can also be applied to a cargo source ranking factor determination device provided in the electronic device. The device can be realized by software, hardware or a combination of both. Wherein, the electronic device can include at least one of a server, a mobile phone, a computer, a vehicle terminal, a tablet computer, a wearable device and a smart home device, but is not limited thereto. The server can include a standalone server, a virtual server and a cluster server.

[0159] In the following, taking application to an electronic device as an example, combining Figures 3-4 The cargo source ranking factor determination method provided by the embodiments of the present application is described in detail.

[0160] Figure 3A flowchart of a method for determining a cargo source ranking factor is shown. The method can include the following steps 310-340. Figure 3 As shown, the method for determining a cargo source ranking factor can include the following steps 310-340.

[0161] Step 310: Obtain historical transaction orders of a cargo source in a first preset time period.

[0162] The first preset time period can be set according to actual needs, such as 24 hours, 48 hours, etc. The application does not make special limitations.

[0163] Step 320: Determine a bottom line price feature based on the historical transaction orders of the cargo source.

[0164] The bottom line price feature is used to represent the bottom line prices of multiple target dimensions and the relationship between the cargo source price and the bottom line price. The multiple target dimensions can include at least two of the driver dimension, the driver and route intersection dimension, and the route and registered truck driver intersection dimension, but are not limited thereto.

[0165] The electronic device can identify the price information in the historical transaction orders of the cargo source, and then perform statistical analysis of the price information in multiple target dimensions according to the corresponding price attribute information of the price information, to obtain the bottom line price feature. The price attribute information can include at least one of driver information, route information, truck driver information, and cargo source information, but is not limited thereto.

[0166] For historical transaction orders of a cargo source, some are priced by transportation trips, some are priced by weight, etc., and the same pricing unit can also have different pricing methods. At the same time, the price distribution of less-than-truckload and full truckload is naturally not in the same space, and the price of less-than-truckload is generally lower. Based on this, in one embodiment, before performing statistical analysis of the price information in multiple target dimensions, the price information can be standardized to obtain corresponding cargo standardized prices, and then statistical analysis of multiple target dimensions is performed according to the cargo standardized prices.

[0167] Specifically, the step 320 of determining the bottom-line price feature based on the historical transaction order of the cargo source can include: determining the minimum value of the standardized price of the cargo source in the driver dimension in the historical transaction order of the cargo source, to obtain a first bottom-line price; determining the minimum value of the standardized price of the cargo source in the intersection of the driver and the route dimension in the historical transaction order of the cargo source, to obtain a second bottom-line price; determining the minimum value of the standardized price of the cargo source in the intersection of the route and the registered driver dimension in the historical transaction order of the cargo source, to obtain a third bottom-line price; determining at least two of the first bottom-line price, the second bottom-line price and the third bottom-line price as the bottom-line price of the plurality of target dimensions; determining the target difference value between the bottom-line price and the cargo source price; and determining the bottom-line price and the target difference value as the bottom-line price feature.

[0168] The specific implementation principle of determining the bottom-line price feature based on the historical transaction order of the cargo source is the same as the principle of determining the sample bottom-line price feature based on the sample historical transaction order of the cargo source in the steps 111-116, and the specific implementation principle can refer to the steps 111-116, which will not be described here.

[0169] Step 330: Obtain the cargo source price competitiveness feature in the second preset time period and the cargo source real-time price increase feature in the third preset time period.

[0170] The second preset time period and the third preset time period can be set according to actual needs, which can be the same or different. The cargo source price competitiveness feature can represent the competitiveness or value of the cargo source, and the higher the competitiveness, the greater the probability and speed of transaction. The cargo source real-time price increase feature is used to represent the individualized price increase behavior of the cargo source.

[0171] For example, the step 330 of obtaining the cargo source price competitiveness feature in the second preset time period and the cargo source real-time price increase feature in the third preset time period can include: obtaining the price quantile of the first cargo source in each sub-market divided in the second preset time period, to obtain the cargo source price competitiveness feature; and obtaining the price increase feature of the second cargo source in the third preset time period, to obtain the cargo source real-time price increase feature. The specific implementation principle is the same as the principle of obtaining the sample cargo source price competitiveness feature and the sample cargo source real-time price increase feature in the steps 1211-1212, and the specific implementation principle can refer to the steps 1211-1212, which will not be described here.

[0172] Step 340: input the bottom-line price feature, the cargo source price competitiveness feature and the cargo source real-time price increase feature into the cargo source ranking factor determination model as the price sharing feature, to obtain the cargo source ranking factor determined by the cargo source ranking factor determination model.

[0173] The goods source ranking factor determination model is trained based on the training method of the goods source ranking factor determination model in any of the above embodiments. The goods source ranking factor is used to rank the goods source and can include the click rate and / or transaction rate of the goods source.

[0174] For example, when the goods source ranking factor includes the click rate and transaction rate of the goods source, after the click rate and transaction rate of the goods source are determined, the click rate and transaction rate of the goods source can be power-weighted and summed to obtain a ranking score of the goods source. Then, the goods sources can be ranked in descending order of the ranking score, and the ranking result can be used for goods list recommendation and the like.

[0175] For example, the goods source ranking factor determination model can be a single-scene deep neural network architecture model. The electronic device can input the bottom line price feature, the goods source price competitiveness feature, and the goods source real-time markup feature into the goods source ranking factor determination model, determine the goods source ranking factor of the goods source through the goods source ranking factor determination model, and obtain the goods source ranking factor output by the goods source ranking factor determination model.

[0176] For example, the goods source ranking factor determination model can be a multi-scene multi-task architecture neural network model. For example, a currently existing multi-scene multi-task architecture neural network model such as a HiNet model can be used. Alternatively, a new model formed by adding a price feature extraction layer for extracting price features to a currently existing multi-scene multi-task architecture neural network model can be used. Figure 4 The structure and working principle of the goods source ranking factor determination model provided by the embodiment of the application are shown in FIG. 4. Referring to FIG. 4, Figure 4 As shown in FIG. 4, the goods source ranking factor determination model can include a price feature extraction layer 401, a feature encoding layer 402, a scene extraction layer 403, a task extraction layer 404, and a task tower layer 405.

[0177] Referring to FIG. 4, Figure 4 As shown in FIG. 4, the goods source ranking factor determination model, the bottom line price feature, the goods source price competitiveness feature, and the goods source real-time markup feature are input into the goods source ranking factor determination model as price sharing features, and the goods source ranking factor output by the goods source ranking factor determination model is obtained. This can be implemented through steps 341-346 as follows.

[0178] Step 341: Obtain scene-related features.

[0179] The scene-related features can include scene-shared features, scene-specific features, scene-identification features, and scene-attribute features. The scene-shared features are used to represent features shared by different scenes. The scene-specific features are used to represent features specific to a scene. The scene-identification features are used to identify different scenes. The scene-attribute features are used to represent attribute information of a scene, such as a name of the scene.

[0180] For example, the scene-shared features can include at least one of a distance of the driver from a loading location of the freight source, a number of freight sources clicked by the driver in the fourth preset time period, and a ratio of a weight of the freight source to a carrying capacity of the vehicle of the driver, but are not limited thereto. The scene-specific features can include at least one of a number of clicks on the freight source by the driver on a route in each scene in a past fifth preset time period, a duration of subscription to the route by the driver, and a number of searches by the driver in a sixth preset time period, but are not limited thereto. The scene-identification features can include a corresponding number of the scene, and the like. The scene-attribute features can include a name of the scene, and the like.

[0181] In step 342, the bottom-line price feature, the freight source price competitiveness feature, and the freight source real-time markup feature are taken as price-shared features, and are input into the feature encoding layer 402 together with the scene-related features. The feature encoding layer 402 encodes the price-shared features and the scene-related features to obtain price-encoded features and scene-encoded features output by the feature encoding layer 402.

[0182] In an embodiment, the bottom-line price feature, the freight source price competitiveness feature, and the freight source real-time markup feature can be taken as price-shared features, and are directly input into the feature encoding layer 402 together with the scene-related features. The feature encoding layer 402 encodes the price-shared features and the scene-related features to obtain price-encoded features and scene-encoded features output by the feature encoding layer 402.

[0183] In another embodiment, the price-shared features can be input into the feature encoding layer 402 for feature encoding after being binned, that is, the price-shared features are discretized and then encoded. The binning manner can refer to the binning manner described in step 1222, which will not be described here. Through this binning manner, the granularity of binning can be refined, the sensitivity of binning to price changes can be improved, and thus the sensitivity of the model to price can be improved.

[0184] For example, for the freight source price competitiveness feature, first and second equal-interval binning can be performed, such as 5-binning and 10-binning, respectively, and the binning results are taken as target binning features of the freight source price competitiveness feature. In this way, the binning effect of the freight source price competitiveness feature can be improved, and the sensitivity and effect of the model can be improved.

[0185] Step 343: input the price encoding features into the price feature extraction layer 401, and perform price feature extraction on the price encoding features through the price feature extraction layer 401 to obtain price features output by the price feature extraction layer 401.

[0186] The price feature extraction layer 401 is configured to extract price features, which can be, for example, an MLP neural network structure.

[0187] Step 344: input the price encoding features and the scene encoding features into the scene extraction layer 403, and perform scene feature extraction on the price encoding features and the scene encoding features through the scene extraction layer 403 to obtain scene features output by the scene extraction layer 403.

[0188] The scene extraction layer 403 can be configured to extract information features specific to a scene and valuable information features shared between scenes from the price encoding features and the scene encoding features, to obtain the scene features. By performing scene feature extraction on the price encoding features and the scene encoding features, the ability of the information features of the task layer can be improved.

[0189] For example, as shown in FIG. 4B, the scene extraction layer 403 can include a scene-shared expert network and a scene-specific expert network. The scene-shared encoding features in the price encoding features and the scene encoding features can be input into the scene-shared expert network and the scene-specific expert network of the scene extraction layer 403, respectively. Meanwhile, the scene-specific features in the scene encoding features are input into the scene-specific expert network. The scene-shared expert network extracts shared information between different scenes, and the scene-specific expert network learns scene-specific information. The scene extraction layer 403 outputs the extracted shared information between different scenes and the scene-specific information as the scene features. Figure 4 For example, as shown in FIG. 4B, the scene extraction layer 403 can include a scene-shared expert network and a scene-specific expert network. The scene-shared encoding features in the price encoding features and the scene encoding features can be input into the scene-shared expert network and the scene-specific expert network of the scene extraction layer 403, respectively. Meanwhile, the scene-specific features in the scene encoding features are input into the scene-specific expert network. The scene-shared expert network extracts shared information between different scenes, and the scene-specific expert network learns scene-specific information. The scene extraction layer 403 outputs the extracted shared information between different scenes and the scene-specific information as the scene features.

[0190] Figure 4 For example, as shown in FIG. 4B, the scene extraction layer 403 can include a scene attention network. The price encoding features and all the scene encoding features can be input into the scene attention network for scene perception to obtain scene attention features. The scene attention features are output together with the features output by the scene-shared expert network and the scene-specific expert network as the scene features output by the scene extraction layer 403. In this way, the importance of the contribution of other scene information to the information representation of the current scene information is measured, and the information representation ability of the current scene is enhanced.

[0191] Step 345: input the scene features into the task extraction layer 404, and perform task feature extraction on the scene features through the task extraction layer 404 to obtain task features output by the task extraction layer 404.

[0192] ​The scene features output by scene extraction layer 403 can be input into task extraction layer 404 after vector concatenation. Task extraction layer 404 then extracts task features from the scene features. Specifically, task extraction layer 404 can be used to extract shared information of all tasks and task-specific information from the scene features to obtain task features.

[0193] For example, such as Figure 4 As shown, the task extraction layer 404 may include a task-sharing expert network and a task-specific expert network. Scene features can be input into the task-sharing expert network and the task-specific expert network, respectively. The task-sharing expert network learns the shared information of all tasks in the current scene, and the task-specific expert network extracts the specific information of each task in the current scene. The task extraction layer 404 outputs the extracted shared information of all tasks and the specific information of each task as task features. It can be understood that the task-specific expert network can be designed according to the number of tasks, with each task corresponding to one task-specific expert network.

[0194] Step 346: Input the task features and price features into the task tower layer 405, perform task identification through the task tower layer 405, and determine the source ranking factor output by the task tower layer 405 as the source ranking factor output by the source ranking factor determination model.

[0195] In this way, by concatenating the additional extracted price features into the input vector of the task tower layer 405 and performing task recognition together with the task features output by the task extraction layer 404, the shallow expression of price features can be improved, thereby improving the sensitivity and monotonicity of the task tower layer 405 and the model to price factors.

[0196] according to Figure 4 The model for determining the sourcing ranking factor, as shown, involves price-sharing features that, after input discretization and feature encoding, are fed into a scenario-sharing expert network and a scenario-specific expert network for information extraction. These are then processed by a task extraction layer 404 and a task pyramid layer 405 to learn the targets for clicks and / or transactions. However, the price feature, a crucial decision-making factor in the business scenario, faces information decay or gradient vanishing issues in the multi-layer neural network's information extraction process without a price feature extraction layer 401. As the network depth increases, the numerical information and gradient signals of the price feature may gradually weaken or even be overwhelmed by other features, leading to the partial loss of critical price decision information during transmission. This ultimately affects the model's pricing accuracy and business decision-making effectiveness. Figure 4The illustrated goods source ranking factor determination model is based on the existing multi-scene multi-task architecture neural network model, and a unique information extraction network, i.e., a price feature extraction layer 401, is additionally designed for a plurality of price-related features, and the inputs of the scene-shared expert network and the scene-specific expert network are retained. The information extracted by the price feature extraction layer 401 is directly spliced into the vector output by the task extraction layer 404 and enters the task tower layer 405 for task recognition. In this way, the shallow expression of the price feature can be improved, and the sensitivity and monotonicity of the task tower layer 405 and the goods source ranking factor determination model to the price factor can be improved.

[0197] The goods source ranking factor determination method provided by the embodiment of the present application comprehensively considers the characteristics of the price factor in multiple different dimensions and the correlation between the price feature and user behavior, etc., greatly enriches the price-related features, and thus improves the goods source ranking matching effect and ensures the monotonicity of the goods source price.

[0198] It can be understood that the training method of the goods source ranking factor determination model and the goods source ranking factor determination method provided by the embodiment of the present application can be applied in a single scene or in multiple scenes.

[0199] The embodiment of the present application further provides a training device of a goods source ranking factor determination model, Figure 5 The structure schematic diagram of the training device of the goods source ranking factor determination model is shown, and the training device of the goods source ranking factor determination model is shown. Figure 5 The training device of the goods source ranking factor determination model can include:

[0200] The first acquisition module 510 is configured to acquire a sample goods source historical transaction order.

[0201] The first determination module 520 is configured to determine a sample baseline price feature based on the sample goods source historical transaction order, wherein the sample baseline price feature is used to represent a plurality of target dimension sample baseline prices and a relationship between the sample goods source price and the sample baseline price.

[0202] The second acquisition module 530 is configured to acquire a sample goods source price competitiveness feature and a sample goods source real-time markup feature.

[0203] The training module 540 is configured to train the initial goods source ranking factor determination model based on the sample baseline price feature, the sample goods source price competitiveness feature, and the sample goods source real-time markup feature to obtain a trained goods source ranking factor determination model; the goods source ranking factor determination model is used to determine a goods source ranking factor of the goods source, and the goods source ranking factor includes a click rate and / or a transaction rate of the goods source.

[0204] In an embodiment, the plurality of target dimensions include at least two of a driver dimension, a driver-by-route dimension, and a route-by-registered-truck-driver dimension; accordingly, the first determining module 520 is specifically configured to: determine a minimum value of a sample freight source standardized price in the driver dimension in the sample freight source historical transaction order, to obtain a first sample bottom line price; determine a minimum value of a sample freight source standardized price in the driver-by-route dimension in the sample freight source historical transaction order, to obtain a second sample bottom line price; determine a minimum value of a sample freight source standardized price in the route-by-registered-truck-driver dimension in the sample freight source historical transaction order, to obtain a third sample bottom line price; determine at least two of the first sample bottom line price, the second sample bottom line price, and the third sample bottom line price as a sample bottom line price of the plurality of target dimensions; determine a sample target difference value between the sample bottom line price and the sample freight source price; the sample target difference value includes a sample absolute difference value and / or a sample relative difference value; and determine the sample bottom line price and the sample target difference value as a sample bottom line price feature.

[0205] In an embodiment, the second obtaining module 530 is specifically configured to: obtain a price quantile of the first sample freight source in each sample sub-market divided in a first sample preset time period, to obtain a sample freight source price competitiveness feature; and obtain a sample markup feature of the second sample freight source in a second sample preset time period, to obtain a sample freight source real-time markup feature; wherein the sample markup feature includes at least one of a sample markup amount, a first sample time interval from a markup time to the second sample freight source publishing time, and a second sample time interval from the last markup to the current time.

[0206] In an embodiment, the initial goods source ranking factor determination model includes an initial price feature extraction layer, an initial feature encoding layer, an initial scene extraction layer, an initial task extraction layer, and an initial task tower layer; correspondingly, the training module 540 is specifically configured to: take the sample bottom line price feature, the sample goods source price competitiveness feature, and the sample goods source real-time price increase feature as sample price sharing features, and obtain sample scene related features; input the sample price sharing features and the sample scene related features into the initial feature encoding layer, perform feature encoding on the sample price sharing features and the sample scene related features through the initial feature encoding layer, and obtain sample price encoding features and sample scene encoding features output by the initial feature encoding layer; input the sample price encoding features into the initial price feature extraction layer, perform price feature extraction on the sample price encoding features through the initial price feature extraction layer, and obtain sample price features output by the initial price feature extraction layer; input the sample price encoding features and the sample scene encoding features into the initial scene extraction layer, perform scene feature extraction on the sample price encoding features and the sample scene encoding features through the initial scene extraction layer, and obtain sample scene features output by the initial scene extraction layer; input the sample scene features into the initial task extraction layer, perform task feature extraction on the sample scene features through the initial task extraction layer, and obtain sample task features output by the initial task extraction layer; input the sample task features and the sample price features into the initial task tower layer, perform task recognition through the initial task tower layer, and obtain sample goods source ranking factors output by the initial task tower layer; the sample goods source ranking factors include sample click rates and / or sample transaction rates; adjust model parameters of the initial goods source ranking factor determination model based on the sample goods source ranking factors and sample label data until the initial goods source ranking factor determination model converges, and obtain the trained goods source ranking factor determination model.

[0207] In an embodiment, when the training module 540 inputs the sample price sharing features and the sample scenario related features into the initial feature encoding layer, and performs feature encoding on the sample price sharing features and the sample scenario related features through the initial feature encoding layer to obtain the sample price encoding features and the sample scenario encoding features output by the initial feature encoding layer, the training module 540 is specifically configured to: take each type of sample price sharing feature as a sample to-be-bucketed feature, perform equal-frequency bucketing on the non-default value price features in the sample to-be-bucketed feature to obtain a first sample bucket; take the default value price features in the sample to-be-bucketed feature as a new bucket, and append the new bucket to the boundary of the first sample bucket to form a second sample bucket of the sample to-be-bucketed feature together with the first sample bucket; perform gain on the bucket boundary values of the second sample bucket based on a preset gain coefficient to obtain target sample bucketed features corresponding to the sample to-be-bucketed features; and input the target sample bucketed features corresponding to all types of sample price sharing features and the sample scenario related features into the initial feature encoding layer, perform feature encoding on all target sample bucketed features and the sample scenario related features through the initial feature encoding layer, and obtain the sample price encoding features and the sample scenario encoding features output by the initial feature encoding layer.

[0208] In an embodiment, the training device of the cargo source ranking factor determination model further includes a test module configured to: collect real driver-cargo pair samples, and increase the prices in the real driver-cargo pair samples based on a preset price increase ratio to obtain virtual driver-cargo pair samples; determine the real driver-cargo pair samples and the virtual driver-cargo pair samples as test samples; for each test sample group in the test samples, input the virtual driver-cargo pair samples in the test sample group into the cargo source ranking factor determination model to obtain virtual cargo source ranking factors output by the cargo source ranking factor determination model, the virtual cargo source ranking factors including a virtual click rate and / or a virtual transaction rate; determine virtual ranking scores of the virtual driver-cargo pair samples in the test sample group based on the virtual cargo source ranking factors; and determine a Spearman correlation coefficient between the virtual ranking scores of the test sample group and the cargo source prices of the real driver-cargo pair samples in the test sample group, the Spearman correlation coefficient being used to represent the price monotonicity of the cargo source ranking factor determination model.

[0209] The training device of the cargo source ranking factor determination model provided in the embodiments of the present application has the same implementation principles and beneficial effects as the training method of the cargo source ranking factor determination model provided in the above embodiments, and thus will not be described here.

[0210] The embodiments of the present application further provide a cargo source ranking factor determination device, Figure 6 FIG. 1 shows a structural schematic diagram of the cargo source ranking factor determination device, and FIG. 2 shows a structural schematic diagram of the cargo source ranking factor determination device according to another embodiment of the present application. Figure 6 As shown in FIG. 1 and FIG. 2, the cargo source ranking factor determination device can include:

[0211] The third acquisition module 610 is configured to acquire historical transaction orders of the cargo source in a first preset time period.

[0212] The second determining module 620 is configured to determine a bottom line price feature based on historical transaction orders of the goods source, the bottom line price feature being used to represent bottom line prices of a plurality of target dimensions and a relationship between the goods source price and the bottom line price.

[0213] The fourth obtaining module 630 is configured to obtain a goods source price competitiveness feature in a second preset time period and a goods source real-time markup feature in a third preset time period.

[0214] The third determining module 640 is configured to input the bottom line price feature, the goods source price competitiveness feature and the goods source real-time markup feature as price sharing features into a goods source ranking factor determination model to obtain a goods source ranking factor output by the goods source ranking factor determination model, wherein the goods source ranking factor determination model is trained based on the training method of the goods source ranking factor determination model in any of the above embodiments, and the goods source ranking factor includes a click rate and / or a transaction rate of the goods source.

[0215] In an embodiment, the goods source ranking factor determination model includes a price feature extraction layer, a feature encoding layer, a scene extraction layer, a task extraction layer and a task tower layer. Correspondingly, the third determining module 640 is specifically configured to: obtain a scene-related feature; input the bottom line price feature, the goods source price competitiveness feature and the goods source real-time markup feature as price sharing features into the feature encoding layer together with the scene-related feature, perform feature encoding on the price sharing features and the scene-related feature through the feature encoding layer to obtain price encoding features and scene encoding features output by the feature encoding layer; input the price encoding features into the price feature extraction layer, perform price feature extraction on the price encoding features through the price feature extraction layer to obtain price features output by the price feature extraction layer; input the price encoding features and the scene encoding features into the scene extraction layer, perform scene feature extraction on the price encoding features and the scene encoding features through the scene extraction layer to obtain scene features output by the scene extraction layer; input the scene features into the task extraction layer, perform task feature extraction on the scene features through the task extraction layer to obtain task features output by the task extraction layer; and input the task features and the price features into the task tower layer, perform task recognition through the task tower layer, and determine a goods source ranking factor output by the task tower layer as the goods source ranking factor output by the goods source ranking factor determination model.

[0216] The goods source ranking factor determination apparatus provided by the embodiments of the present application has the same implementation principles and beneficial effects as the goods source ranking factor determination method provided by the above embodiments, and thus will not be described here again.

[0217] The apparatus embodiments described above are merely illustrative, wherein the units described as separate components can or can not be physically separate, and the components displayed as units can or can not be physical units, i.e., can be located in one place, or can be distributed on multiple network units. Part or all of the modules can be selected to achieve the purpose of the embodiment scheme according to actual needs. Those skilled in the art can understand and implement without creative labor.

[0218] The embodiment of the present application also provides an electronic device, comprising a memory and a processor, the memory stores a computer program capable of running on the processor, and the processor implements the steps of the training method of the cargo source ranking factor determination model of any one of the above method embodiments or the steps of the cargo source ranking factor determination method of any one of the above method embodiments when executing the program, which will not be repeated here.

[0219] Based on the training method of the cargo source ranking factor determination model or the cargo source ranking factor determination method described in any one of the above embodiments, the embodiment of the present application also provides a computer readable storage medium, for example, a non-transitory computer readable storage medium can be a read only memory (Read Only Memory, ROM), a random access memory (Random Access Memory, RAM), a CD-ROM, a magnetic tape, a floppy disk and an optical data storage device, etc. The storage medium stores computer instructions for executing the training method of the cargo source ranking factor determination model described in any one of the above embodiments or executing the cargo source ranking factor determination method described in any one of the above embodiments, which will not be repeated here.

[0220] Those skilled in the art can understand that all or part of the steps of the above embodiments can be completed by hardware, or by program to instruct related hardware to complete, and the program can be stored in a computer readable storage medium, and the above mentioned storage medium can be a read only memory, a magnetic disk or an optical disk, etc.

[0221] Other embodiments of the application will be apparent to those skilled in the art from consideration of the specification and practice of the application disclosed herein. The specification and examples given are considered exemplary only, and the true scope and spirit of the application is indicated by the claims.

Claims

1. A training method for a source ranking factor determination model, characterized in that, The method comprises the following steps: obtaining sample source historical transaction orders, and determining sample bottom line price features based on the sample source historical transaction orders, wherein the sample bottom line price features are used to represent sample bottom line prices of multiple target dimensions and the relationship between sample source prices and the sample bottom line prices; the multiple target dimensions include at least two of the following dimensions: driver dimension, driver and route intersection dimension, and route and registered truck driver intersection dimension; obtaining sample source price competitiveness features and sample source real-time price increase features, and training an initial source ranking factor determination model based on the sample bottom line price features, the sample source price competitiveness features and the sample source real-time price increase features to obtain a trained source ranking factor determination model; the source ranking factor determination model is used to determine a source ranking factor of a source, and the source ranking factor includes a click rate and / or a transaction rate of the source; wherein the initial source ranking factor determination model comprises an initial price feature extraction layer, an initial feature encoding layer, an initial scene extraction layer, an initial task extraction layer and an initial task tower layer; the training of the initial source ranking factor determination model based on the sample bottom line price features, the sample source price competitiveness features and the sample source real-time price increase features comprises: inputting the sample bottom line price features, the sample source price competitiveness features and the sample source real-time price increase features as sample price sharing features, and obtaining sample scene related features; inputting the sample price sharing features and the sample scene related features into the initial feature encoding layer, and performing feature encoding on the sample price sharing features and the sample scene related features through the initial feature encoding layer to obtain sample price encoding features and sample scene encoding features output by the initial feature encoding layer; inputting the sample price encoding features into the initial price feature extraction layer, and performing price feature extraction on the sample price encoding features through the initial price feature extraction layer to obtain sample price features output by the initial price feature extraction layer; inputting the sample price encoding features and the sample scene encoding features into the initial scene extraction layer, and performing scene feature extraction on the sample price encoding features and the sample scene encoding features through the initial scene extraction layer to obtain sample scene features output by the initial scene extraction layer; inputting the sample scene features into the initial task extraction layer, and performing task feature extraction on the sample scene features through the initial task extraction layer to obtain sample task features output by the initial task extraction layer; inputting the sample task features and the sample price features into the initial task tower layer, and performing task recognition through the initial task tower layer to obtain sample source ranking factors output by the initial task tower layer; the sample source ranking factors include sample click rates and / or sample transaction rates; adjusting model parameters of the initial source ranking factor determination model based on the sample source ranking factors and sample label data until the initial source ranking factor determination model converges, and obtaining the trained source ranking factor determination model. The sample price sharing feature and the sample scene related feature are input into the initial feature encoding layer, the sample price sharing feature and the sample scene related feature are encoded by the initial feature encoding layer, and sample price encoding features and sample scene encoding features output by the initial feature encoding layer are obtained. Each type of sample price sharing feature is taken as a sample to-be-bucketed feature, non-default value price features in the sample to-be-bucketed feature are equally bucketed, and a first sample bucket is obtained. The default value price feature in the sample to-be-bucketed feature is taken as a new bucket, and is appended to the boundary of the first sample bucket to form a second sample bucket of the sample to-be-bucketed feature together with the first sample bucket. Based on a preset gain coefficient, the bucket boundary values of the second sample bucket are gained, and target sample bucket features corresponding to the sample to-be-bucketed feature are obtained. The target sample bucket features corresponding to all types of sample price sharing features and the sample scene related feature are input into the initial feature encoding layer, and all the target sample bucket features and the sample scene related feature are encoded by the initial feature encoding layer, and sample price encoding features and sample scene encoding features output by the initial feature encoding layer are obtained. 2.The method of Claim 1, wherein, The sample baseline price feature is determined based on the sample source historical transaction order, and includes: A minimum value of a sample source standardized price in the driver dimension of the sample source historical transaction order is determined, and a first sample baseline price is obtained. A minimum value of a sample source standardized price in the intersection of the driver and the route dimension of the sample source historical transaction order is determined, and a second sample baseline price is obtained. A minimum value of a sample source standardized price in the intersection of the route and the registered driver dimension of the sample source historical transaction order is determined, and a third sample baseline price is obtained. At least two of the first sample baseline price, the second sample baseline price and the third sample baseline price are determined as the sample baseline price of the plurality of target dimensions. A sample target difference value between the sample baseline price and the sample source price is determined; the sample target difference value includes a sample absolute difference value and / or a sample relative difference value. The sample baseline price and the sample target difference value are determined as the sample baseline price feature. 3.The method of Claim 1, wherein, The sample source price competitiveness feature and the sample source real-time price increase feature are obtained, and include: Price quantiles of a first sample source in each sample sub-market in a first sample preset time period are obtained, and the sample source price competitiveness feature is obtained. Sample price increase features of a second sample source in a second sample preset time period are obtained, and the sample source real-time price increase feature is obtained; the sample price increase features include at least one of a sample price increase amount, a first sample time interval from a price increase time to a release time of the second sample source, and a second sample time interval from the last price increase to a current time. 4.The method of Claim 1 to 3, wherein, After obtaining the trained source sorting factor determination model, the training method of the source sorting factor determination model further includes: collecting real seller-goods pair samples, and raising prices in the real seller-goods pair samples based on a preset price raising ratio to obtain virtual seller-goods pair samples; determining the real seller-goods pair samples and the virtual seller-goods pair samples as test samples; for each test sample group in the test samples, inputting the virtual seller-goods pair samples in the test sample group into the goods source ranking factor determination model to obtain a virtual goods source ranking factor output by the goods source ranking factor determination model; the virtual goods source ranking factor includes a virtual click rate and / or a virtual transaction rate; determining a virtual ranking score of the virtual seller-goods pair samples in the test sample group based on the virtual goods source ranking factor; determining a Spearman correlation coefficient between the virtual ranking score and a goods source price of the real seller-goods pair samples in the test sample group; the Spearman correlation coefficient is used to represent the price monotonicity of the goods source ranking factor determination model.

5. A method of determining a source ranking factor, the method comprising: comprising: obtaining goods history transaction orders in a first preset time period; determining a bottom line price feature based on the goods history transaction orders, the bottom line price feature being used to represent a bottom line price of a plurality of target dimensions and a relationship between a goods source price and the bottom line price; obtaining a goods price competitiveness feature in a second preset time period and a goods real-time price increase feature in a third preset time period; inputting the bottom line price feature, the goods price competitiveness feature and the goods real-time price increase feature as price sharing features into a goods source ranking factor determination model to obtain a goods source ranking factor output by the goods source ranking factor determination model; wherein the goods source ranking factor determination model is trained based on a training method of the goods source ranking factor determination model according to any one of claims 1 to 4; the goods source ranking factor includes a click rate and / or a transaction rate of a goods source.

6. The method of claim 5, wherein, The goods source ranking factor determination model includes a price feature extraction layer, a feature encoding layer, a scene extraction layer, a task extraction layer and a task tower layer; the inputting the bottom line price feature, the goods price competitiveness feature and the goods real-time price increase feature as price sharing features into a goods source ranking factor determination model to obtain a goods source ranking factor output by the goods source ranking factor determination model includes: obtaining scene-related features; inputting the bottom line price feature, the goods price competitiveness feature and the goods real-time price increase feature as price sharing features into the feature encoding layer together with the scene-related features, and performing feature encoding on the price sharing features and the scene-related features through the feature encoding layer to obtain price encoding features and scene encoding features output by the feature encoding layer; inputting the price encoding features into the price feature extraction layer to perform price feature extraction on the price encoding features through the price feature extraction layer to obtain price features output by the price feature extraction layer; inputting the price encoding features and the scene encoding features into the scene extraction layer to perform scene feature extraction on the price encoding features and the scene encoding features through the scene extraction layer to obtain scene features output by the scene extraction layer; The scene features are input into the task extraction layer, and task feature extraction is performed on the scene features by the task extraction layer to obtain task features output by the task extraction layer; The task features and the price features are input into the task tower layer, task recognition is performed by the task tower layer, and a cargo source ranking factor output by the task tower layer is determined as a cargo source ranking factor output by the cargo source ranking factor determination model.

7. An electronic device comprising a memory and a processor, said memory storing a computer program operable on said processor, characterized in that, The processor executes the computer program to implement the steps of the cargo source ranking factor determination model training method according to any one of claims 1 to 4, or implement the steps of the cargo source ranking factor determination method according to claim 5 or 6.

8. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the steps of the cargo source ranking factor determination model training method according to any one of claims 1 to 4, or implement the steps of the cargo source ranking factor determination method according to claim 5 or 6.

Images