Model training method and apparatus

Abstract

The application provides a model training method and device, and relates to the technical field of logistics. The method comprises the following steps: acquiring log information and activity information corresponding to a plurality of initial cargos respectively, each initial cargo being a cargo in a driver-cargo pair that generates an exposure behavior in a freight platform; performing preliminary screening on the plurality of initial cargos according to the log information and the activity information, and determining a plurality of candidate cargos; acquiring feature information corresponding to the plurality of candidate cargos respectively; performing fine screening on the plurality of candidate cargos according to driver behavior features and cargo owner behavior features in the feature information, and determining a plurality of sample cargos; taking the feature information corresponding to the plurality of sample cargos respectively as training samples, taking cargo states corresponding to the plurality of sample cargos respectively as sample labels, training a basic model, and obtaining a click rate prediction model, so that the model learns healthy, real and valuable behavior patterns, and the prediction accuracy and rationality of the click rate prediction model obtained through training are improved.

Description

Technical Field

[0001] This application relates to the field of logistics technology, and in particular to a model training method and apparatus. Background Technology

[0002] With the rapid development of online freight, matching drivers and freight using click-through rate (CTR) prediction models has become a core technology for improving platform matching efficiency and transaction conversion rates. This model predicts the probability of drivers clicking on freight listings, ranks them, and prioritizes exposing the most likely-to-be-clicked freight listings to drivers.

[0003] In related technologies, click-through rate (CTR) prediction models are typically trained based on historical data of cargo exposed on freight platforms. However, historical data often contains data generated by behaviors such as random exposure and fraudulent orders, as well as data from "black hole users." Black hole users are those who, despite generating clicks and feedback on a large scale on the freight platform, do not actually intend to complete a transaction, regardless of whether they are drivers or cargo owners. This data fails to reflect genuine user preferences, and the behavior of "black hole users" deviates from normal transaction logic, thus contaminating the training samples. This causes the model to learn unhealthy, fraudulent, and worthless behavioral patterns, resulting in poor accuracy and reasonableness in the CTR prediction model. Summary of the Invention

[0004] This application provides a model training method and apparatus to address the problem that the model learns unhealthy, false, and worthless behavioral patterns, resulting in poor prediction accuracy and reasonableness of the trained click-through rate prediction model. It achieves the goal of enabling the model to learn healthy, real, and valuable behavioral patterns, thereby improving the prediction accuracy and reasonableness of the trained click-through rate prediction model.

[0005] Firstly, this application provides a model training method, the method comprising:

[0006] Obtain log information and activity information corresponding to multiple initial cargo sources; where each initial cargo source is a cargo source in a driver-cargo pair that generates exposure behavior in the freight platform;

[0007] Based on log information and activity information, multiple initial sources of goods are screened to identify multiple candidate sources of goods.

[0008] Obtain the feature information corresponding to multiple candidate sources of goods;

[0009] Based on the driver and cargo owner behavior characteristics in the feature information, multiple candidate cargo sources are screened to determine multiple sample cargo sources.

[0010] The feature information corresponding to multiple sample cargo sources is used as training samples, and the cargo source status corresponding to multiple sample cargo sources is used as sample labels to train the basic model, thereby obtaining the click-through rate prediction model; the click-through rate prediction model is used to predict the probability of drivers clicking on cargo sources.

[0011] In one possible design, based on log information and activity information, multiple initial sources of goods are screened to determine multiple candidate sources of goods, including:

[0012] Based on the log information corresponding to multiple initial sources, click window sampling is performed on multiple initial sources based on multiple clicked sources to determine multiple clicked sampled sources; clicked sources are the sources among multiple initial sources that generated click behavior;

[0013] Based on the activity information corresponding to multiple click-sampled sources, the activity-level sampling of multiple click-sampled sources is performed to determine multiple candidate sources.

[0014] In one possible design, the log information includes exposure time and exposure location; the exposure location indicates the position of the cargo in the cargo list when it is exposed to the driver; based on the log information corresponding to multiple initial cargo sources, click window sampling is performed on multiple initial cargo sources based on multiple clicked cargo sources to determine multiple clicked sampled cargo sources, including:

[0015] For a single clicked product source, based on the exposure time and location of the clicked product source, multiple related product sources are identified from multiple initial product sources; the exposure time of the related product sources is the same as that of the clicked product source.

[0016] Based on multiple click-based product sources, multiple click-sampled product sources are identified; the multiple click-sampled product sources include multiple click-based product sources and multiple related product sources corresponding to each click-based product source.

[0017] In one possible design, based on the exposure time and location corresponding to the clicked product source, multiple related product sources are determined from multiple initial product sources, including:

[0018] Among the multiple initial product sources, all product sources that have the same exposure time as the clicked product source and whose position in the product source list is above the clicked product source at the time of exposure, and N product sources that have the same exposure time as the clicked product source and whose position in the product source list is below the clicked product source at the time of exposure, are identified as multiple related product sources; where N is a positive integer.

[0019] In one possible design, activity information includes driver activity rating and shipper activity rating; based on the activity information corresponding to multiple click-sampled cargo sources, stratified activity sampling is performed on the multiple click-sampled cargo sources to determine multiple candidate cargo sources, including:

[0020] Based on the driver activity rating and shipper activity rating corresponding to multiple click-sampled cargo sources, the multiple click-sampled cargo sources are stratified to obtain a set of top cargo sources, a set of mid-tier cargo sources, and a set of long-tail cargo sources;

[0021] Multiple candidate sources are obtained by sampling a first proportion of sources from the head source set, a second proportion of sources from the mid-range source set, and a third proportion of sources from the long-tail source set.

[0022] In one possible design, based on driver and cargo owner behavior characteristics in the feature information, multiple candidate cargo sources are finely screened to determine multiple sample cargo sources, including:

[0023] For a candidate cargo source, determine whether the candidate cargo source meets the first preset condition based on the driver behavior characteristics corresponding to the candidate cargo source, and determine whether the candidate cargo source meets the second preset condition based on the shipper behavior characteristics corresponding to the candidate cargo source; wherein, the driver corresponding to the cargo source that meets the first preset condition is a black hole user, and the shipper corresponding to the cargo source that meets the second preset condition is a black hole user.

[0024] From multiple candidate sources, those that meet both the first and second preset conditions are removed, and the remaining candidate sources are determined as multiple sample sources.

[0025] In one possible design, driver behavioral characteristics include driver fulfillment volume, driver outbound call volume, driver outbound call success volume, and driver identity tag, which indicates whether the driver is a team leader or not; shipper behavioral characteristics include the amount of cargo posted by the shipper and the amount of cargo completed by the shipper.

[0026] The first preset condition is: the driver's outbound call fulfillment rate is less than the first threshold, the driver's outbound call success rate is greater than the second threshold, and the driver's identity tag indicates that the driver is a non-vehicle team leader; wherein, the driver's outbound call fulfillment rate is the ratio of the driver's fulfillment rate to the driver's outbound call volume;

[0027] The second preset condition is: the cargo owner's transaction rate is less than the third threshold, and the cargo owner's posted cargo volume is greater than the fourth threshold; where the cargo owner's transaction rate is the ratio of the cargo owner's transaction volume to the cargo owner's posted cargo volume.

[0028] In one possible design, the feature information also includes driver behavior tags and shipper behavior tags, with the driver behavior tags indicating the rating of the driver's transaction behavior and the shipper behavior tags indicating the rating of the shipper's transaction behavior.

[0029] In one possible design, sample labels are used to indicate the status of the sample goods: unclicked, clicked, feedback received, or sold. Before training the base model, the method also includes:

[0030] The sample weight of the sample goods whose corresponding sample label indicates that the source status is "transaction" is set as the first weight; the sample weight of the sample goods whose corresponding sample label indicates that the source status is "feedback" is set as the second weight; and the sample weight of the sample goods whose corresponding sample label indicates that the source status is "not clicked" or "clicked" is set as the third weight; wherein, the first weight is greater than the second weight, and the second weight is greater than the third weight.

[0031] Using the method provided in the first aspect, log information and activity information corresponding to multiple initial cargo sources are obtained. Based on this information, the initial cargo sources are screened to determine multiple candidate cargo sources. This initial screening filters out cargo sources where drivers do not have genuine interest, avoiding invalid traffic from random exposure and fraudulent orders from contaminating the training samples. It also selects cargo sources with different activity levels, balancing the training samples and avoiding the Matthew effect. Next, feature information corresponding to each candidate cargo source is obtained. Based on driver and cargo owner behavior characteristics within this feature information, the candidate cargo sources are further screened to determine multiple sample cargo sources. This further screening filters out cargo sources where drivers do not have genuine transaction intentions, avoiding black hole user-related data from contaminating the training samples. Using the feature information corresponding to each sample cargo source as training samples and the cargo status corresponding to each sample cargo source as sample labels, the basic model is trained to obtain a click-through rate (CTR) prediction model. This allows the model to focus on cargo sources where drivers have genuine interest and transaction intentions during the learning process, thereby learning healthy, authentic, and valuable behavioral patterns and improving the accuracy and rationality of the CTR prediction model.

[0032] Secondly, this application provides a model training apparatus, the apparatus comprising:

[0033] The first acquisition module is used to acquire log information and activity information corresponding to multiple initial cargo sources; where each initial cargo source is a cargo source in a driver-cargo source pair that generates exposure behavior in the freight platform;

[0034] The initial screening module is used to screen multiple initial sources of goods based on log information and activity information to determine multiple candidate sources of goods;

[0035] The second acquisition module is used to acquire the feature information corresponding to multiple candidate sources of goods;

[0036] The fine screening module is used to fine screen multiple candidate cargo sources based on the driver behavior characteristics and cargo owner behavior characteristics in the feature information, and determine multiple sample cargo sources.

[0037] The training module is used to train the basic model by using the feature information corresponding to multiple sample cargo sources as training samples and the cargo source status corresponding to multiple sample cargo sources as sample labels, thereby obtaining the click-through rate prediction model. The click-through rate prediction model is used to predict the probability of drivers clicking on cargo sources.

[0038] In one possible design, the initial screening module is specifically used for:

[0039] Based on the log information corresponding to multiple initial sources, click window sampling is performed on multiple initial sources based on multiple clicked sources to determine multiple clicked sampled sources; clicked sources are the sources among multiple initial sources that generated click behavior;

[0040] Based on the activity information corresponding to multiple click-sampled sources, the activity-level sampling of multiple click-sampled sources is performed to determine multiple candidate sources.

[0041] In one possible design, the log information includes exposure time and exposure location; the exposure location indicates the position of the cargo in the cargo list when it is exposed to the driver; the initial screening module is specifically used for:

[0042] For a single clicked product source, based on the exposure time and location of the clicked product source, multiple related product sources are identified from multiple initial product sources; the exposure time of the related product sources is the same as that of the clicked product source.

[0043] Based on multiple click-based product sources, multiple click-sampled product sources are identified; the multiple click-sampled product sources include multiple click-based product sources and multiple related product sources corresponding to each click-based product source.

[0044] In one possible design, the initial screening module is specifically used for:

[0045] Among the multiple initial product sources, all product sources that have the same exposure time as the clicked product source and whose position in the product source list is above the clicked product source at the time of exposure, and N product sources that have the same exposure time as the clicked product source and whose position in the product source list is below the clicked product source at the time of exposure, are identified as multiple related product sources; where N is a positive integer.

[0046] In one possible design, activity information includes driver activity ratings and shipper activity ratings; the initial screening module is specifically used for:

[0047] Based on the driver activity rating and shipper activity rating corresponding to multiple click-sampled cargo sources, the multiple click-sampled cargo sources are stratified to obtain a set of top cargo sources, a set of mid-tier cargo sources, and a set of long-tail cargo sources;

[0048] Multiple candidate sources are obtained by sampling a first proportion of sources from the head source set, a second proportion of sources from the mid-range source set, and a third proportion of sources from the long-tail source set.

[0049] In one possible design, the fine screening module is specifically used for:

[0050] For a candidate cargo source, determine whether the candidate cargo source meets the first preset condition based on the driver behavior characteristics corresponding to the candidate cargo source, and determine whether the candidate cargo source meets the second preset condition based on the shipper behavior characteristics corresponding to the candidate cargo source; wherein, the driver corresponding to the cargo source that meets the first preset condition is a black hole user, and the shipper corresponding to the cargo source that meets the second preset condition is a black hole user.

[0051] From multiple candidate sources, those that meet both the first and second preset conditions are removed, and the remaining candidate sources are determined as multiple sample sources.

[0052] In one possible design, driver behavioral characteristics include driver fulfillment volume, driver outbound call volume, driver outbound call success volume, and driver identity tag, which indicates whether the driver is a team leader or not; shipper behavioral characteristics include the amount of cargo posted by the shipper and the amount of cargo completed by the shipper.

[0053] The first preset condition is: the driver's outbound call fulfillment rate is less than the first threshold, the driver's outbound call success rate is greater than the second threshold, and the driver's identity tag indicates that the driver is a non-vehicle team leader; wherein, the driver's outbound call fulfillment rate is the ratio of the driver's fulfillment rate to the driver's outbound call volume;

[0054] The second preset condition is: the cargo owner's transaction rate is less than the third threshold, and the cargo owner's posted cargo volume is greater than the fourth threshold; where the cargo owner's transaction rate is the ratio of the cargo owner's transaction volume to the cargo owner's posted cargo volume.

[0055] In one possible design, the feature information also includes driver behavior tags and shipper behavior tags, with the driver behavior tags indicating the rating of the driver's transaction behavior and the shipper behavior tags indicating the rating of the shipper's transaction behavior.

[0056] In one possible design, sample labels are used to indicate the status of the sample goods as unclicked, clicked, feedback received, or sold. The model training device also includes a weight adjustment module, which is used for:

[0057] The sample weight of the sample goods whose corresponding sample label indicates that the source status is "transaction" is set as the first weight; the sample weight of the sample goods whose corresponding sample label indicates that the source status is "feedback" is set as the second weight; and the sample weight of the sample goods whose corresponding sample label indicates that the source status is "not clicked" or "clicked" is set as the third weight; wherein, the first weight is greater than the second weight, and the second weight is greater than the third weight.

[0058] The beneficial effects of the apparatus provided in the second aspect and the various possible designs of the second aspect can be found in the first aspect and the various possible implementations of the first aspect, and will not be repeated here.

[0059] Thirdly, this application provides an electronic device including a first processor, which, when executing a computer-executable program or instructions in a memory, implements a model training method as described in the first aspect and any possible design of the first aspect.

[0060] Fourthly, this application provides an electronic device including at least one memory and at least one second processor. The memory stores a computer-executable program or instructions, and the second processor executes the computer-executable program or instructions to implement the model training method as described in the first aspect and any possible design of the first aspect.

[0061] Fifthly, this application provides a computer-readable storage medium storing a computer-executable program or instructions, which, when executed by a processor, implement the model training method as described in the first aspect and any possible design of the first aspect.

[0062] In a sixth aspect, this application provides a computer program product comprising: execution instructions stored in a readable storage medium, at least one processor of an electronic device being able to read the execution instructions from the readable storage medium, and the at least one processor executing the execution instructions causing the electronic device to implement the model training method as described in the first aspect and any possible design of the first aspect.

[0063] The above description is merely an overview of the technical solutions of the embodiments of this application. In order to better understand the technical means of the embodiments of this application and to implement them in accordance with the contents of the specification, and to make the above and other objects, features and advantages of the embodiments of this application more obvious and understandable, specific implementation methods of this application are described below. Attached Figure Description

[0064] Figure 1 This is a flowchart of a model training method provided in an embodiment of this application.

[0065] Figure 2 This is a flowchart illustrating a method for initial screening of multiple initial sources of goods, provided as an embodiment of this application.

[0066] Figure 3 This is a flowchart of a click window sampling method provided in an embodiment of this application.

[0067] Figure 4This is a flowchart of an activity-based hierarchical sampling method provided in an embodiment of this application.

[0068] Figure 5 This is a flowchart illustrating a method for fine screening of multiple candidate sources, provided as an embodiment of this application.

[0069] Figure 6 This diagram illustrates a method for setting sample weights according to an embodiment of this application.

[0070] Figure 7 This is a schematic diagram of the structure of a model training device provided in an embodiment of this application.

[0071] Figure 8 A schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 1 .

[0072] Figure 9 A schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 2 . Detailed Implementation

[0073] In this application, "at least one" means one or more, and "more than one" means two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can mean: A alone, A and B simultaneously, or B alone, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one of a, b, or c alone can mean: a alone, b alone, c alone, a combination of a and b, a combination of a and c, a combination of b and c, or a, b, and c, where a, b, and c can be single or multiple. Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0074] The terms "connected" and "connected" should be interpreted broadly. For example, in circuit structures, "connected" or "connected" can refer not only to physical connections but also to electrical or signal connections. This could be a direct connection (physical connection) or an indirect connection via at least one intermediate component, as long as the circuit is connected. It could also refer to the internal connection between two components. Similarly, a signal connection can refer to a connection via a circuit or a medium, such as radio waves. Those skilled in the art will understand the specific meaning of these terms in this application based on the specific circumstances.

[0075] For example, this application provides a model training method, apparatus, electronic device, and storage medium. The method involves initially screening the data used for training a click-through rate (CTR) prediction model based on log and activity information to avoid data contamination from random invalid exposures, fraudulent orders, and other behaviors. Then, it further refines the screening based on driver and cargo owner behavioral characteristics to prevent data contamination from "black hole users," resulting in a filtered training sample. The basic model is then trained using this filtered training sample to obtain the CTR prediction model. This allows the model to learn healthy, authentic, and valuable behavioral patterns, improving the accuracy and rationality of the CTR prediction model.

[0076] The model training method provided in this application can be executed by an electronic device or by a model training device in an electronic device.

[0077] Among them, electronic devices can be servers, desktop computers, mobile phones, tablets, laptops, wearable devices, in-vehicle devices, augmented reality (AR) / virtual reality (VR) devices, etc.

[0078] The model training device can be implemented through a combination of software and / or hardware. For example, the model training device can be an application (APP), a webpage, or a public account. Another example is a freight platform. For simplicity, this application's embodiments use the execution of a model training device as an example.

[0079] Below, in conjunction with Figures 1 to 6 The model training method provided in the embodiments of this application will be described.

[0080] Please see Figure 1 , Figure 1 This is a flowchart illustrating a model training method provided in one embodiment of this application. Figure 1 As shown, the method includes:

[0081] S101, The model training device acquires log information and activity information corresponding to multiple initial sources of goods.

[0082] Each initial source of freight is a driver-freight pair that generates exposure on the freight platform. Within a historical timeframe, when a source of freight is exposed to a driver on the freight platform, that source and driver form a driver-freight pair that generates exposure; this source is considered an initial source of freight. The status of an initial source of freight can be: unclicked (exposed but not clicked), clicked (exposed and clicked but without feedback), feedback (exposed, clicked, and feedback-received but not yet transacted), or transacted (exposed, clicked, feedback-received, and transacted). In other words, regardless of whether a source that generates exposure subsequently receives clicks, feedback, or a transact, it can still be considered an initial source of freight.

[0083] In addition, it should be noted that a source of cargo can be exposed to multiple different drivers, and a source of cargo can also be exposed to the same driver multiple times at different times. Each time it is exposed, the source of cargo can be used as a new initial source of cargo.

[0084] For example, multiple initial sources of goods can include: sources of goods , source of goods , source of goods , source of goods , source of goods and sources of goods Among them, the source of goods This indicates that cargo source A is exposed to driver j1 at time t1; cargo source This indicates that cargo source A is exposed to driver j1 at time t2; cargo source This indicates that cargo source A is exposed to driver j2 at time t3; cargo source This indicates that cargo source B was exposed to driver J1 at time t4; cargo source This indicates that cargo source B was exposed to driver j3 at time t5; cargo source This indicates that the cargo source C was exposed to driver j4 at time t1.

[0085] The historical time range can be the past year, the past six months, or the past month, etc.

[0086] The log information records relevant data when a cargo source generates exposure, with one log entry corresponding to each initial cargo source. Log information may include driver identifier, cargo owner identifier, cargo source identifier, exposure time, and exposure location. The freight platform generates corresponding log data during operation, and the model training device can directly read the required log information from the freight platform's log data.

[0087] The driver identifier identifies the driver corresponding to the cargo source when it is exposed. The cargo owner identifier identifies the cargo owner. The cargo source identifier identifies the cargo source. The exposure time refers to the time when the cargo source is exposed to the corresponding driver. The exposure position refers to the position of the cargo source in the cargo source list when it is exposed to the corresponding driver, such as its position in the cargo source list.

[0088] The activity level information indicates the activity level of the shipper and / or driver corresponding to the cargo listing. The activity level of the shipper / driver is determined based on their historical behavior on the freight platform. For example, a driver's activity level can be determined by at least one of the following: the number of days the driver logged into the freight platform in a recent period, the number of times the driver clicked on cargo listings, and the number of transactions completed by the driver. Similarly, a shipper's activity level can be determined by at least one of the following: the number of days the shipper logged into the freight platform in a recent period, the number of cargo listings posted by the shipper, and the number of transactions completed by the shipper.

[0089] The model training device can obtain historical behavior information from the freight platform to determine activity level information. Alternatively, the freight platform can determine the activity level of the freight owner / driver based on the historical behavior information of the driver / freight owner on the freight platform in real time or periodically, and store it in a server, cloud, or memory. The model training device can then directly read the required activity level information from the server, cloud, or memory.

[0090] Based on this, the model training device obtains log information and activity information corresponding to multiple initial sources of goods, so as to perform initial screening of the initial sources of goods based on the log information and activity information.

[0091] S102. The model training device performs preliminary screening on multiple initial sources of goods based on log information and activity information, and determines multiple candidate sources of goods.

[0092] Considering the existence of random exposure and order manipulation on freight platforms, these behaviors can lead to some freight listings being exposed to drivers without their genuine interest. Log information records relevant data generated during these exposures, and the model training device can use this log information to filter out freight listings that drivers may have genuine interest in. Furthermore, considering the Matthew effect, the model training device can stratify freight listings that drivers may have genuine interest in based on their activity levels, and sample freight listings from different strata proportionally to identify multiple candidate freight listings.

[0093] Based on this, the model training device can filter out cargo sources that drivers do not have a genuine intention to pay attention to through initial screening, avoiding the pollution of training samples by invalid traffic generated by random exposure and fraudulent orders. It also selects cargo sources with different activity levels, balancing the training samples and avoiding the Matthew effect, so that the model can learn healthy, real and valuable behavioral patterns, thereby helping to improve the prediction accuracy of the click-through rate prediction model obtained from the training.

[0094] S103, The model training device acquires the feature information corresponding to multiple candidate sources of goods.

[0095] The feature information may include driver behavior characteristics, cargo owner behavior characteristics, and cargo source characteristics.

[0096] Driver behavioral characteristics may include driver fulfillment volume, driver outbound call volume, driver outbound call success rate, driver identity tags, and driver outbound call fulfillment rate.

[0097] Driver fulfillment volume refers to the quantity of cargo that a driver fulfills within a recent period.

[0098] Driver outbound call volume refers to the number of cargo listings that drivers actively call or text the cargo owners after clicking on the cargo listings within a recent period.

[0099] Driver outbound call success rate refers to the number of cargo listings in a recent period where the driver proactively calls or messages the cargo owner and successfully contacts them.

[0100] Driver identification tags are used to indicate whether a driver is a team leader or not. Since team leaders may only click, provide feedback, or complete a transaction before handing the goods over to team members for transportation, the team leader may have a high number of clicks, feedback, and transactions, but a low fulfillment rate. Therefore, driver identification tags are needed to distinguish between them and avoid misjudgment.

[0101] The driver's outbound call fulfillment rate is the ratio of the number of fulfilled calls to the number of outbound calls made by the driver.

[0102] The model training device can learn about the transaction logic of the driver corresponding to the candidate cargo source in a recent period based on the driver's behavioral characteristics, so as to determine whether the driver is a black hole user through the driver's behavioral characteristics.

[0103] Among them, the characteristics of cargo owners' behavior can include the amount of cargo they post, the amount of cargo they transact, and the transaction rate.

[0104] The volume of cargo posted by cargo owners refers to the number of cargoes posted by cargo owners on freight platforms in a recent period.

[0105] The volume of goods traded by a cargo owner refers to the quantity of goods traded by the cargo owner within a recent period.

[0106] The cargo owner's transaction rate is the ratio of the cargo owner's transaction volume to the cargo owner's posted cargo volume.

[0107] The model training device can learn about the transaction logic of the cargo owners corresponding to the candidate cargo sources in a recent period based on the cargo owner's behavioral characteristics, so as to determine whether the cargo owner is a black hole user through the cargo owner's behavioral characteristics.

[0108] The aforementioned "recent period" can refer to within 30 days, 3 months, or 6 months, etc.

[0109] The characteristics of the goods can include the type of goods, weight, volume, price, transportation distance, and the time when the goods were posted.

[0110] In some examples, the feature information may also include spatiotemporal features and cross-features. Spatiotemporal features indicate the temporal and spatial characteristics of the cargo source. Examples of spatiotemporal features include the current time (the time the cargo source is exposed to the driver), the distance between the driver's current location and the origin of the cargo source, and the estimated arrival time. Cross-features indicate the cross-features between the driver and the cargo source, as well as the transaction characteristics between the driver and the cargo owner. Examples of cross-features between the driver and the cargo source include the driver's recent click-through rate, conversion rate, and fulfillment rate for cargo sources of the same type as the cargo source, and the driver's preference for the origin of the cargo source. Examples of transaction characteristics between the driver and the cargo owner include the number of times the driver and the cargo owner have historically cooperated.

[0111] Based on this, the model training device acquires feature information corresponding to multiple candidate cargo sources, so as to finely screen multiple candidate cargo sources according to the driver behavior features and cargo owner behavior features in the feature information, and to construct rich feature vectors according to the feature information corresponding to multiple sample cargo sources, thereby improving feature richness.

[0112] S104. The model training device uses the driver behavior characteristics and cargo owner behavior characteristics in the feature information to screen multiple candidate cargo sources and determine multiple sample cargo sources.

[0113] Considering that the driver behavior characteristics of candidate freight sources can reflect the transaction logic of the corresponding drivers in recent times, and the shipper behavior characteristics of candidate freight sources can reflect the transaction logic of the corresponding shippers in recent times, "black hole users" refer to users who, although generating clicks and feedback on a large scale on the freight platform, at least one of the drivers or shippers has no real intention to complete the transaction. Black hole users consume a large amount of traffic resources of the freight platform but damage the efficiency of driver-freight matching and transaction conversion rate. Therefore, by using driver and shipper behavior characteristics, the model training device can identify black hole users, thereby filtering out freight sources related to black hole users from the candidate freight sources.

[0114] For example, the model training device can set black hole driver identification rules and black hole cargo owner identification rules. If the model training device can determine that the driver corresponding to the candidate cargo source meets the black hole driver identification rules based on the driver's behavior characteristics, then the driver corresponding to the candidate cargo source is judged to be a black hole user. If the model training device determines that the cargo owner corresponding to the candidate cargo source meets the black hole cargo owner identification rules based on the cargo owner's behavior characteristics, then the cargo owner corresponding to the candidate cargo source is judged to be a black hole user.

[0115] The black hole driver identification rule can be set based on the driver's outbound call volume, outbound call success rate, and order fulfillment rate. For example, if a driver has a large outbound call volume and a large outbound call success rate, but a low order fulfillment rate, that is, the driver makes frequent outbound calls but does not fulfill orders, it indicates that the driver does not have a genuine intention to complete transactions. In this case, the black hole driver identification rule could be: the driver's outbound call fulfillment rate is less than a certain threshold and the driver's outbound call success rate is greater than a certain threshold.

[0116] The black hole shipper identification rule can be set based on the number of shipments posted by the shipper and the number of transactions completed by the shipper. For example, if a shipper posts a large number of shipments but has a very low number of transactions, it means that the shipper does not have a real intention to complete transactions despite posting a large number of shipments. In this case, the black hole shipper identification rule could be: the shipper's transaction rate is less than a certain threshold and the number of shipments posted by the shipper is greater than a certain threshold.

[0117] The model training device can determine whether to eliminate or retain candidate sources in a variety of ways.

[0118] As a feasible implementation, if the model training device determines that either the driver or the shipper corresponding to a candidate cargo source is a "black hole user," then the model training device will eliminate that candidate cargo source. Based on this, the model training device can ensure the purity of the data used to train the model, enabling the model to focus on learning the most realistic and healthy behavioral patterns and avoiding interference from data related to "black hole users."

[0119] As another feasible implementation, if the model training device determines that both the driver and the cargo owner corresponding to a candidate cargo source are black hole users, then the model training device will eliminate that candidate cargo source. Based on this, the model training device can increase the diversity of samples used to train the model and improve data utilization.

[0120] Based on this, the model training device can filter out cargo sources from which drivers do not have a real intention to make a transaction through fine screening, avoiding the contamination of training samples by black hole user-related data, so that the model can learn healthy, real and valuable behavioral patterns, thereby helping to improve the predictive rationality of the click-through rate prediction model obtained from the training.

[0121] S105. The model training device uses the feature information corresponding to multiple sample sources as training samples and the source status corresponding to multiple sample sources as sample labels to train the basic model and obtain the click-through rate prediction model.

[0122] Among them, the click-through rate prediction model of the model training device is used to predict the probability of drivers clicking on cargo.

[0123] The product status refers to the click, feedback, and transaction status of the sample product. Product status can be categorized as: No Click, Click, Feedback, or Transaction. For example, a sample product that has been exposed but not clicked will have the status: No Click; a sample product that has been exposed and clicked but not received feedback will have the status: Click; a sample product that has been exposed, clicked, and received feedback but not transacted will have the status: Feedback; and a sample product that has been exposed, clicked, received feedback, and transacted will have the status: Transaction.

[0124] Based on sample cargo sources with different cargo statuses, the model training device enables the model to understand the value differences of different behaviors during the model training process, ensuring that the optimization direction of the model can take into account the transaction conversion needs of the freight platform.

[0125] The basic model is, for example, the Light Gradient Boosting Machine (LightGBM) model.

[0126] Based on this, the model training device trains the basic model based on the feature information corresponding to multiple sample cargo sources. This enables the model to focus on cargo sources that drivers are genuinely interested in and intend to transact with during the learning process, filtering out invalid traffic generated by random exposure and fraudulent orders, as well as low-quality traffic generated by black hole users. In this way, the model learns healthy, real, and valuable behavioral patterns, thereby improving the accuracy and rationality of the click-through rate prediction model obtained through training.

[0127] In this embodiment, the model training device acquires log information and activity information corresponding to multiple initial cargo sources. Based on the log information and activity information, it performs initial screening on the multiple initial cargo sources to determine multiple candidate cargo sources. Through this initial screening, the model training device can filter out cargo sources where drivers do not have a genuine intention to engage, avoiding the pollution of training samples by invalid traffic generated by random exposure and fraudulent orders. Furthermore, it selects cargo sources with different activity levels, balancing the training samples and avoiding the Matthew effect. Next, the model training device acquires feature information corresponding to each of the multiple candidate cargo sources. Based on the driver behavior characteristics and cargo owner behavior characteristics in the feature information, it performs fine screening on the multiple candidate cargo sources to determine multiple sample cargo sources. Through this fine screening, the model training device can filter out cargo sources where drivers do not have a genuine intention to transact, avoiding the pollution of training samples by data related to "black hole" users. The model training device uses the feature information corresponding to multiple sample cargo sources as training samples and the cargo source status corresponding to multiple sample cargo sources as sample labels to train the basic model and obtain the click-through rate prediction model. This allows the model to focus on cargo sources that drivers are genuinely interested in and intend to make a purchase during the learning process, thereby learning healthy, authentic, and valuable behavioral patterns and improving the prediction accuracy and rationality of the trained click-through rate prediction model.

[0128] Furthermore, ranking and recommending cargo based on the prediction results of the click-through rate prediction model can recommend high-quality cargo to high-quality drivers, avoiding black hole users consuming a lot of traffic resources and improving the efficiency of driver-cargo matching and transaction conversion rate.

[0129] Based on the above exemplary description, the model training device can be used as follows: Figure 2 The method shown is used to initially screen multiple initial sources of goods and identify multiple candidate sources of goods.

[0130] Please see Figure 2 , Figure 2 This is a flowchart illustrating a method for initial screening of multiple initial sources of goods, provided as an embodiment of this application. Figure 2 As shown, the method includes:

[0131] S201. The model training device, based on the log information corresponding to multiple initial sources of goods, performs click window sampling on multiple initial sources of goods based on multiple clicked sources of goods, and determines multiple clicked sampled sources of goods.

[0132] Among them, the clicked source is the source that generates click behavior among multiple initial sources.

[0133] Among multiple initial cargo sources, due to random exposure and order manipulation, some cargo sources that drivers do not have a genuine intention to pay attention to are exposed to drivers. However, considering that most of the cargo sources that drivers click on are those that drivers have a genuine intention to pay attention to, the model training device can sample the click windows based on the cargo sources that generate click behavior among multiple initial cargo sources, thereby filtering out a portion of cargo sources that drivers have a genuine intention to pay attention to.

[0134] As a feasible approach, the model training device can identify all clicked sources as click-sampling sources, thereby obtaining high-quality training samples and improving the training efficiency of the model.

[0135] As another feasible implementation, the model training device can use a single clicked item as a base and, according to a predetermined strategy, extract related items from the item list as click sampling items. This expands the range of training samples and helps improve the model's learning ability. Items related to the clicked item include, for example, several items adjacent to the clicked item.

[0136] Based on this, by sampling through the click window, the model training device can eliminate invalid traffic generated by random exposure and fraudulent orders based on log information, ensuring that the source of goods that can reflect the driver's true attention intentions can be screened out, thereby improving the efficiency and accuracy of subsequent screening.

[0137] S202. The model training device performs activity-level sampling on multiple click-sampled goods sources based on the activity information corresponding to each of the multiple click-sampled goods sources, and determines multiple candidate goods sources.

[0138] In freight platforms, considering that click-through rate prediction models need to serve users with different activity levels, the model training device performs activity-level stratified sampling. For drivers and shippers with different activity levels, multiple click-sampled cargo sources are divided into different levels, and cargo sources are sampled from each level.

[0139] For example, activity information includes the number of days drivers and shippers have logged into the freight platform. The model training device can identify drivers whose login days on the freight platform exceed a preset number of days as active drivers and those whose login days are less than a preset number of days as inactive drivers. Similarly, shippers whose login days on the freight platform exceed a preset number of days are identified as active shippers and those whose login days are less than a preset number of days are identified as inactive shippers. Based on this, the model training device can group click-sampled freight sources where both the corresponding shipper and driver are active into one set, into another set where either the corresponding shipper or driver is active, and into yet another set where both the corresponding shipper and driver are inactive. From these three sets, a portion of the freight sources can be selected in the same or different proportions to determine multiple candidate freight sources.

[0140] Based on this, by using activity-based stratified sampling, the model training device can stratify multiple click-sampled cargo sources based on different activity levels, and sample a portion of cargo sources from each of the different activity levels, thereby ensuring the diversity of samples and enabling candidate cargo sources to truly reflect the behavior of different users on the freight platform, thus improving the accuracy and robustness of subsequent training.

[0141] Furthermore, the click-through rate prediction model can make accurate predictions when serving users at different activity levels, which helps to build a healthy platform ecosystem.

[0142] Based on the above exemplary description, the log information includes exposure time and exposure location.

[0143] The exposure location is used to indicate the position of the cargo in the cargo list when it is exposed to the driver.

[0144] Based on this, the model training device can be used as follows: Figure 3 The method shown involves clicking on multiple initial sources of goods to sample them through a window, thereby identifying multiple sources of goods for which click sampling is performed.

[0145] Please see Figure 3 , Figure 3 This is a flowchart illustrating a click window sampling method provided in one embodiment of this application. Figure 3 As shown, the method includes:

[0146] S301, The model training device, for a single clicked product source, determines multiple related product sources from multiple initial product sources based on the exposure time and exposure location corresponding to the clicked product source.

[0147] The exposure time for related products and clicked products is the same.

[0148] The model training device can determine the relevant sources of goods in a variety of ways.

[0149] As a feasible implementation, considering that drivers, after browsing the clicked freight list, are likely to focus on freight adjacent to it, these adjacent freight items are generally of higher quality. Based on this, the model training device can identify multiple relevant freight items from among the initial freight items that have the same exposure time as the clicked freight item and are located adjacent to it in the freight list at the time of exposure. This improves the quality of the samples and enhances the training effect.

[0150] As another feasible implementation, considering that drivers typically browse cargo listings from top to bottom, cargo listings above the clicked listing, even if not clicked, are of relatively high quality and can provide negative samples for model training. Several nearby cargo listings below the clicked listing, which may be of potential interest to drivers, can also be used as training samples, while cargo listings farther away may be of lower quality and can be discarded. Based on this, the model training device can identify multiple related cargo listings from multiple initial listings: all cargo listings with the same exposure time as the clicked listing and positioned above it in the listing list at the time of exposure; and N cargo listings with the same exposure time as the clicked listing and positioned below it in the listing list at the time of exposure. Here, N is a positive integer. This improves the authenticity and diversity of the samples, thus enhancing the training effect.

[0151] Source of goods by click For example, the source of goods The exposure time is t1, and the source of goods Located at position 3 in the source list, the source list is shown in Table 1 below:

[0152] Table 1

[0153]

[0154] As can be seen from Table 1, the source of goods , source of goods , source of goods , source of goods In order to source goods Sources with the same exposure time. The model training device can train the source at position 2. The source of goods at location 4 Multiple related sources of goods have been identified. Alternatively, the model training device can identify the source of goods at position 1. The source of goods at location 2 The source of goods at location 4 Multiple related sources of goods have been identified.

[0155] S302, The model training device determines multiple click sampling sources based on multiple click sources.

[0156] Among them, multiple click sampling sources include multiple click sources and multiple related sources corresponding to each click source.

[0157] Based on this, the S301 operation is performed for each clicked product source. The model training device can determine the relevant product source corresponding to each clicked product source, thereby completing the click window sampling and obtaining multiple clicked sampled product sources.

[0158] Based on the above exemplary description, the activity information includes driver activity rating and cargo owner activity rating.

[0159] Freight platforms will rate the activity of drivers and cargo owners based on information such as the number of days they log in to the platform, the number of times they click on cargo listings, the volume of transactions, and the number of cargo listings they post.

[0160] For example, Level 1 represents inactivity, Level 2 represents moderate activity, and Level 3 represents very active. Alternatively, 0-30 points represent inactivity, 30-70 points represent moderate activity, and 70-100 points represent very active.

[0161] Based on this, the model training device can be used as follows: Figure 4 The method shown performs activity-based stratified sampling on multiple click-sampled product sources to determine multiple candidate product sources.

[0162] Please see Figure 4 , Figure 4 This is a flowchart illustrating an activity-based hierarchical sampling method provided in one embodiment of this application. Figure 4 As shown, the method includes:

[0163] S401 The model training device stratifies multiple click-sampled cargo sources according to the driver activity rating and cargo owner activity rating corresponding to each source, resulting in a set of top cargo sources, a set of mid-tier cargo sources, and a set of long-tail cargo sources.

[0164] In some examples, the model training device can first calculate a comprehensive activity rating corresponding to a click-sampled cargo source based on the driver activity rating and the cargo owner activity rating, and then stratify based on the comprehensive activity rating.

[0165] The overall activity rating can be the median or average of driver activity rating and cargo owner activity rating.

[0166] Specifically, the model training device can categorize click-sampled products with a comprehensive activity rating greater than the first comprehensive threshold into the top product set, click-sampled products with a comprehensive activity rating greater than the second comprehensive threshold and less than or equal to the first comprehensive threshold into the mid-tier product set, and click-sampled products with a comprehensive activity rating less than or equal to the second comprehensive threshold into the long-tail product set.

[0167] In other examples, the model training device can directly stratify based on driver activity ratings and shipper activity ratings.

[0168] Specifically, the model training device can categorize click-sampled cargo sources whose corresponding driver activity rating and cargo owner activity rating are both greater than the first activity threshold into the top cargo source set, click-sampled cargo sources whose corresponding driver activity rating and cargo owner activity rating are both greater than the second activity threshold and less than or equal to the first activity threshold into the mid-tier cargo source set, click-sampled cargo sources whose corresponding driver activity rating and cargo owner activity rating are both less than or equal to the second activity threshold into the long-tail cargo source set, and the remaining click-sampled cargo sources into the mid-tier cargo source set.

[0169] S402, the model training device samples a first proportion of goods from the head goods set, a second proportion of goods from the waist goods set, and a third proportion of goods from the long tail goods set, respectively, to obtain multiple candidate goods.

[0170] The first, second, and third ratios can be the same or different, and the model training device can set the first, second, and third ratios according to business needs. For example, when it is necessary to improve the user experience of active top users, the first ratio can be set to be greater than the second and third ratios.

[0171] The model training device identifies multiple candidate sources, including a first proportion of sources sampled from the head source set, a second proportion of sources sampled from the mid-range source set, and a third proportion of sources sampled from the long-tail source set.

[0172] Based on this, the model training device divides the activity into different activity levels and samples proportionally at each level to complete the activity stratification sampling, thereby obtaining multiple candidate sources and ensuring the accuracy and balance of the sampling, so as to obtain more realistic and healthy training samples.

[0173] Based on the above exemplary description, the model training device can be used as follows: Figure 5 The method shown is used to screen multiple candidate sources of goods and determine multiple sample sources of goods.

[0174] Please see Figure 5 , Figure 5This is a flowchart illustrating a method for finely screening multiple candidate sources, as provided in one embodiment of this application. Figure 5 As shown, the method includes:

[0175] S501, the model training device, for a candidate cargo source, determines whether the candidate cargo source meets the first preset condition based on the driver behavior characteristics corresponding to the candidate cargo source, and determines whether the candidate cargo source meets the second preset condition based on the cargo owner behavior characteristics corresponding to the candidate cargo source.

[0176] Among them, the drivers corresponding to the cargo that meets the first preset condition are black hole users, and the cargo owners corresponding to the cargo that meets the second preset condition are black hole users.

[0177] In some examples, driver behavioral characteristics include driver fulfillment volume, driver outbound call volume, driver outbound call success volume, and driver identity tags, which indicate whether the driver is a team leader or not. Shipper behavioral characteristics include the amount of cargo posted by the shipper and the number of cargo transactions completed by the shipper.

[0178] The first preset condition can be: the driver's outbound call fulfillment rate is less than the first threshold, the driver's outbound call success rate is greater than the second threshold, and the driver's identity tag indicates that the driver is not a fleet leader; wherein, the driver's outbound call fulfillment rate is the ratio of the driver's fulfillment rate to the driver's outbound call volume.

[0179] The second preset condition can be: the cargo owner's transaction rate is less than the third threshold, and the cargo owner's posted cargo volume is greater than the fourth threshold; where the cargo owner's transaction rate is the ratio of the cargo owner's transaction volume to the cargo owner's posted cargo volume.

[0180] Based on this, the model training device can identify black hole drivers through the first preset condition and black hole cargo owners through the second preset condition. By utilizing the behavioral characteristics of drivers and cargo owners, the device can accurately identify black hole users, so as to remove the relevant data of black hole users.

[0181] S502, The model training device removes the goods that meet the first preset condition and the second preset condition from multiple candidate goods sources, and determines the remaining candidate goods sources as multiple sample goods sources.

[0182] The model training device removes sources that meet both the first and second preset conditions, thereby eliminating the lowest quality and least authentic sources among all candidate sources and identifying the remaining candidate sources as multiple sample sources, thus achieving accurate and comprehensive data cleaning.

[0183] Based on the above exemplary description, the feature information also includes driver behavior tags and cargo owner behavior tags, whereby the driver behavior tags are used to indicate the rating of the driver's transaction behavior and the cargo owner behavior tags are used to indicate the rating of the cargo owner's transaction behavior.

[0184] The freight platform will rate the drivers and shippers on the platform based on their transaction behavior, thereby obtaining driver behavior tags and shipper behavior tags.

[0185] For example, if a shipper's behavior is tagged as "Class A Shipper," "Class B Shipper," or "Class C Shipper," the freight platform can label shippers with a transaction rate greater than a first transaction rate threshold as "Class A Shipper," shippers with a transaction rate greater than or equal to a second transaction rate threshold but less than or equal to the first transaction rate threshold as "Class B Shipper," and shippers with a transaction rate less than the second transaction rate threshold as "Class C Shipper." Similarly, if a driver's behavior is tagged as "Class A Driver," "Class B Driver," or "Class C Driver," the freight platform can label drivers with a transaction rate greater than a first transaction rate threshold as "Class A Driver," drivers with a transaction rate greater than or equal to a second transaction rate threshold but less than or equal to the first transaction rate threshold as "Class B Driver," and drivers with a transaction rate less than the second transaction rate threshold as "Class C Driver."

[0186] Based on this, the model training device can obtain driver behavior tags and cargo owner behavior tags from the freight platform as feature information for model training. During the training process, the quality of cargo sources can be distinguished based on driver behavior tags and cargo owner behavior tags, thereby enhancing the model's perception and discrimination capabilities and helping to improve the predictive rationality of the click-through rate prediction model obtained through training.

[0187] Based on the above exemplary description, the sample label is used to indicate whether the status of the sample product is unclicked, clicked, feedback received, or completed.

[0188] Based on this, before training the base model, the model training device can also perform other operations such as... Figure 6 The method shown is used to set sample weights.

[0189] Please see Figure 6 , Figure 6 This diagram illustrates a method for setting sample weights according to an embodiment of this application. Figure 6 As shown, the method includes:

[0190] S601, set the sample weight of the sample source with the corresponding sample label indicating the source status as "transaction" as the first weight, set the sample weight of the sample source with the corresponding sample label indicating the source status as "feedback" as the second weight, and set the sample weight of the sample source with the corresponding sample label indicating the source status as "not clicked" or "clicked" as the third weight.

[0191] Considering that the target of the click-through rate prediction model is the click-through rate of the goods, but the ultimate goal of predicting the click-through rate of the goods is to improve the transaction conversion rate of the freight platform, the model training device sets different sample weights for sample goods with different statuses to ensure that the optimization direction of the model can be adjusted to click-through rate prediction with the final goal of transaction.

[0192] Among them, the first weight is greater than the second weight, and the second weight is greater than the third weight.

[0193] For example, the first weight is 10, the second weight is 5, and the third weight is 1.

[0194] Based on this, the model training device sets different sample weights for sample sources with different source states before training the basic model. Without changing the number of samples, it adjusts the data distribution, enhances the model's ability to learn about the sources of goods that are likely to be sold, so as to prioritize the allocation of traffic resources to the sources of goods that are most likely to be sold, thereby improving resource utilization and transaction conversion rate.

[0195] For example, this application provides a model training apparatus.

[0196] Figure 7 This is a schematic diagram of a model training device provided in one embodiment of this application. Figure 7 As shown, the device includes: a first acquisition module 101, a primary screening module 102, a second acquisition module 103, a fine screening module 104, and a training module 105.

[0197] The first acquisition module 101 is used to acquire log information and activity information corresponding to multiple initial cargo sources; wherein, each initial cargo source is a cargo source in a driver and cargo source pair that generates exposure behavior in the freight platform;

[0198] The initial screening module 102 is used to perform initial screening on multiple initial sources of goods based on log information and activity information, and to determine multiple candidate sources of goods.

[0199] The second acquisition module 103 is used to acquire feature information corresponding to multiple candidate sources of goods;

[0200] The fine screening module 104 is used to fine screen multiple candidate cargo sources based on the driver behavior characteristics and cargo owner behavior characteristics in the feature information, and determine multiple sample cargo sources.

[0201] The training module 105 is used to train the basic model by using the feature information corresponding to multiple sample cargo sources as training samples and the cargo source status corresponding to multiple sample cargo sources as sample labels, so as to obtain the click-through rate prediction model; the click-through rate prediction model is used to predict the probability of drivers clicking on cargo sources.

[0202] It should be noted that the model training device in this application embodiment can be used to execute the technical solutions of the above method embodiments, and its implementation principle and technical effect are similar, so it will not be repeated here.

[0203] In some examples, the initial screening module 102 is specifically used for:

[0204] Based on the log information corresponding to multiple initial sources, click window sampling is performed on multiple initial sources based on multiple clicked sources to determine multiple clicked sampled sources; clicked sources are the sources among multiple initial sources that generated click behavior;

[0205] Based on the activity information corresponding to multiple click-sampled sources, the activity-level sampling of multiple click-sampled sources is performed to determine multiple candidate sources.

[0206] In some examples, log information includes exposure time and exposure location; the exposure location indicates the position of the cargo in the cargo list when it is exposed to the driver; the initial screening module 102 is specifically used for:

[0207] For a single clicked product source, based on the exposure time and location of the clicked product source, multiple related product sources are identified from multiple initial product sources; the exposure time of the related product sources is the same as that of the clicked product source.

[0208] Based on multiple click-based product sources, multiple click-sampled product sources are identified; the multiple click-sampled product sources include multiple click-based product sources and multiple related product sources corresponding to each click-based product source.

[0209] In some examples, the initial screening module 102 is specifically used for:

[0210] Among the multiple initial product sources, all product sources that have the same exposure time as the clicked product source and whose position in the product source list is above the clicked product source at the time of exposure, and N product sources that have the same exposure time as the clicked product source and whose position in the product source list is below the clicked product source at the time of exposure, are identified as multiple related product sources; where N is a positive integer.

[0211] In some examples, activity information includes driver activity ratings and shipper activity ratings; the initial screening module 102 is specifically used for:

[0212] Based on the driver activity rating and shipper activity rating corresponding to multiple click-sampled cargo sources, the multiple click-sampled cargo sources are stratified to obtain a set of top cargo sources, a set of mid-tier cargo sources, and a set of long-tail cargo sources;

[0213] Multiple candidate sources are obtained by sampling a first proportion of sources from the head source set, a second proportion of sources from the mid-range source set, and a third proportion of sources from the long-tail source set.

[0214] In some examples, the fine screening module 104 is specifically used for:

[0215] For a candidate cargo source, determine whether the candidate cargo source meets the first preset condition based on the driver behavior characteristics corresponding to the candidate cargo source, and determine whether the candidate cargo source meets the second preset condition based on the shipper behavior characteristics corresponding to the candidate cargo source; wherein, the driver corresponding to the cargo source that meets the first preset condition is a black hole user, and the shipper corresponding to the cargo source that meets the second preset condition is a black hole user.

[0216] From multiple candidate sources, those that meet both the first and second preset conditions are removed, and the remaining candidate sources are determined as multiple sample sources.

[0217] In some examples, driver behavioral characteristics include driver fulfillment volume, driver outbound call volume, driver outbound call success volume, and driver identity tag, which indicates whether the driver is a team leader or not; shipper behavioral characteristics include the amount of cargo posted by the shipper and the amount of cargo completed by the shipper.

[0218] The first preset condition is: the driver's outbound call fulfillment rate is less than the first threshold, the driver's outbound call success rate is greater than the second threshold, and the driver's identity tag indicates that the driver is a non-vehicle team leader; wherein, the driver's outbound call fulfillment rate is the ratio of the driver's fulfillment rate to the driver's outbound call volume;

[0219] The second preset condition is: the cargo owner's transaction rate is less than the third threshold, and the cargo owner's posted cargo volume is greater than the fourth threshold; where the cargo owner's transaction rate is the ratio of the cargo owner's transaction volume to the cargo owner's posted cargo volume.

[0220] In some examples, the feature information also includes driver behavior tags and shipper behavior tags, with driver behavior tags used to indicate the rating of the driver's transaction behavior and shipper behavior tags used to indicate the rating of the shipper's transaction behavior.

[0221] In some examples, sample labels are used to indicate the status of the sample goods as unclicked, clicked, feedback received, or sold. The model training device also includes a weight adjustment module, which is used for:

[0222] The sample weight of the sample goods whose corresponding sample label indicates that the source status is "transaction" is set as the first weight; the sample weight of the sample goods whose corresponding sample label indicates that the source status is "feedback" is set as the second weight; and the sample weight of the sample goods whose corresponding sample label indicates that the source status is "not clicked" or "clicked" is set as the third weight; wherein, the first weight is greater than the second weight, and the second weight is greater than the third weight.

[0223] By way of example, this application also provides an electronic device.

[0224] Figure 8A schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 1 .like Figure 8 As shown, the electronic device may include a first processor 201, which, when executing a computer-executable program or instruction stored in a memory, implements the embodiments of this application. Figures 1 to 6 The model training method shown.

[0225] The electronic device can be used to perform the various steps and / or processes corresponding to the electronic devices in the above method embodiments.

[0226] Figure 9 A schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 2 .like Figure 9 As shown, the electronic device may include a second processor 301 and a memory 302. The memory 302 stores a computer program. When the second processor 301 executes the computer program, it implements the embodiments of this application. Figures 1 to 6 The model training method shown.

[0227] The electronic device can be used to perform the various steps and / or processes corresponding to the electronic devices in the above method embodiments.

[0228] The electronic device of this application can be used to execute the technical solutions of the method embodiments described above. Its implementation principle and technical effects are similar. The operations implemented by each module can be further referred to the relevant descriptions of the method embodiments, which will not be repeated here. The modules here can also be replaced by components or circuits.

[0229] This application can divide electronic devices into functional modules based on the above method examples. For example, each function can be divided into its own functional module, or two or more functions can be integrated into one processing module. The integrated module can be implemented in hardware or as a software functional module. It should be noted that the module division in the embodiments of this application is illustrative and only represents one logical functional division. In actual implementation, there may be other division methods.

[0230] Another embodiment of this application provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it can implement the embodiments of this application. Figures 1 to 6 The model training method shown.

[0231] This application also provides a program product including executable instructions stored in a computer-readable storage medium. At least one processor of an electronic device can read the executable instructions from the computer-readable storage medium, and the at least one processor executes the executable instructions to cause the electronic device to implement embodiments of this application. Figures 1 to 6 The model training method shown.

[0232] This application also provides a chip that is connected to a memory, or a chip that integrates a memory. When a software program stored in the memory is executed, it implements the embodiments of this application. Figures 1 to 6 The model training method shown.

[0233] In the embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or modules may be electrical, mechanical, or other forms.

[0234] Those skilled in the art will understand that although some embodiments herein include certain features included in other embodiments, combinations of features from different embodiments are intended to be within the scope of this application and form different embodiments. For example, in the claims, any of the claimed embodiments can be used in any combination.

[0235] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A model training method, characterized in that, The method comprises: obtaining log information and activity information corresponding to a plurality of initial cargos respectively; wherein each of the initial cargos is a cargo in a driver-cargo pair that generates an exposure behavior in a freight platform; performing preliminary screening on the plurality of initial cargos according to the log information and the activity information, to determine a plurality of candidate cargos; obtaining feature information corresponding to the plurality of candidate cargos respectively; performing fine screening on the plurality of candidate cargos according to driver behavior features and cargo owner behavior features in the feature information, to determine a plurality of sample cargos; using the feature information corresponding to the plurality of sample cargos respectively as training samples, using cargo states corresponding to the plurality of sample cargos respectively as sample labels, training a basic model to obtain a click rate prediction model; the click rate prediction model is used to predict a click probability of a driver on a cargo; wherein the fine screening on the plurality of candidate cargos according to the driver behavior features and the cargo owner behavior features in the feature information to determine the plurality of sample cargos comprises: for a candidate cargo, determining whether the candidate cargo satisfies a first preset condition according to a driver behavior feature corresponding to the candidate cargo, and determining whether the candidate cargo satisfies a second preset condition according to a cargo owner behavior feature corresponding to the candidate cargo; wherein a driver corresponding to a cargo satisfying the first preset condition is a black hole user, and a cargo owner corresponding to a cargo satisfying the second preset condition is a black hole user; wherein the black hole user refers to a user who generates a click and feedback behavior on a large scale in the freight platform, but at least one of the driver and the cargo owner does not have a real transaction intention; from the plurality of candidate cargos, removing cargos satisfying the first preset condition and the second preset condition, and determining the remaining candidate cargos as the plurality of sample cargos; wherein the driver behavior features comprise a driver performance, a driver outbound call volume, a driver outbound call success volume, and a driver identity label, the driver identity label is used to indicate whether the driver is a team leader or not; the cargo owner behavior features comprise a cargo owner published cargo volume and a cargo owner transaction volume; the first preset condition is that the driver outbound call performance rate is less than a first threshold value, the driver outbound call success volume is greater than a second threshold value, and the driver identity label indicates that the driver is not a team leader; wherein the driver outbound call performance rate is a ratio of the driver performance to the driver outbound call volume; the second preset condition is that the cargo owner transaction rate is less than a third threshold value, and the cargo owner published cargo volume is greater than a fourth threshold value; wherein the cargo owner transaction rate is a ratio of the cargo owner transaction volume to the cargo owner published cargo volume.

2. The method of claim 1, wherein, the preliminary screening on the plurality of initial cargos according to the log information and the activity information to determine the plurality of candidate cargos comprises: based on a plurality of clicked cargos, performing click window sampling on the plurality of initial cargos according to log information corresponding to the plurality of initial cargos respectively, to determine a plurality of clicked sampling cargos; the clicked cargo is a cargo that generates a click behavior in the plurality of initial cargos; According to the activity information corresponding to the plurality of click sampling sources respectively, the plurality of click sampling sources are subjected to activity stratified sampling to determine a plurality of candidate sources.

3. The method of claim 2, wherein, The log information includes exposure time and exposure position; the exposure position is used to indicate the position of the source in the source list when exposed to the driver; and the plurality of initial sources are subjected to click window sampling based on the plurality of click sources according to the log information corresponding to the plurality of initial sources respectively to determine a plurality of click sampling sources, including: For one of the click sources, a plurality of related sources are determined from the plurality of initial sources according to the exposure time and the exposure position corresponding to the click source; the related sources have the same exposure time as the click source; Based on the plurality of click sources, a plurality of click sampling sources are determined; the plurality of click sampling sources include the plurality of click sources and the plurality of related sources corresponding to each of the click sources.

4. The method of claim 3, wherein, The plurality of related sources are determined from the plurality of initial sources according to the exposure time and the exposure position corresponding to the click source, including: All sources in the plurality of initial sources that have the same exposure time as the click source and are located above the click source in the source list when exposed, and N sources that have the same exposure time as the click source and are located below the click source in the source list when exposed are determined as the plurality of related sources; wherein N is a positive integer.

5. The method of claim 2, wherein, The activity information includes driver activity rating and owner activity rating; The plurality of click sampling sources are subjected to activity stratified sampling according to the activity information corresponding to the plurality of click sampling sources respectively to determine a plurality of candidate sources, including: The plurality of click sampling sources are stratified according to the driver activity rating and the owner activity rating corresponding to the plurality of click sampling sources respectively to obtain a head source set, a waist source set and a long tail source set; A first proportion of sources are sampled from the head source set, a second proportion of sources are sampled from the waist source set, and a third proportion of sources are sampled from the long tail source set to obtain a plurality of candidate sources.

6. The method of claim 1, wherein, The feature information further includes driver behavior label and owner behavior label; the driver behavior label is used to indicate the rating of the transaction behavior of the driver; and the owner behavior label is used to indicate the rating of the transaction behavior of the owner.

7. The method according to any one of claims 1 to 6, characterized in that, The sample label is used to indicate that the source state of the sample source is not clicked, clicked, feedback or transaction; and before the basic model is trained, the method further includes: The sample weight of the sample source corresponding to the sample label indicating that the source state is transaction is set as a first weight, the sample weight of the sample source corresponding to the sample label indicating that the source state is feedback is set as a second weight, and the sample weight of the sample source corresponding to the sample label indicating that the source state is not clicked or clicked is set as a third weight; wherein the first weight is greater than the second weight, and the second weight is greater than the third weight.

8. A model training apparatus, comprising: The device includes: The first obtaining module is configured to obtain log information and activity information corresponding to a plurality of initial freight sources respectively; each of the initial freight sources is a freight source in a driver and freight source pair that generates an exposure behavior in a freight transportation platform; The preliminary screening module is configured to preliminarily screen the plurality of initial freight sources according to the log information and the activity information, and determine a plurality of candidate freight sources; The second obtaining module is configured to obtain feature information corresponding to the plurality of candidate freight sources respectively; The fine screening module is configured to fine screen the plurality of candidate freight sources according to driver behavior features and freight owner behavior features in the feature information, and determine a plurality of sample freight sources; The training module is configured to train a base model by taking the feature information corresponding to the plurality of sample freight sources as training samples and taking freight source states corresponding to the plurality of sample freight sources as sample labels, and obtain a click rate prediction model; the click rate prediction model is configured to predict a click probability of a driver on a freight source; The fine screening module is configured to, for a candidate freight source, determine whether the candidate freight source satisfies a first preset condition according to driver behavior features corresponding to the candidate freight source, and determine whether the candidate freight source satisfies a second preset condition according to freight owner behavior features corresponding to the candidate freight source; a driver corresponding to a freight source that satisfies the first preset condition is a black hole user, and a freight owner corresponding to a freight source that satisfies the second preset condition is a black hole user; the black hole user refers to a user who generates a click and feedback behavior on a large scale in a freight transportation platform, but at least one of the driver and the freight owner does not have a real transaction intention; the fine screening module is further configured to remove a freight source that satisfies the first preset condition and satisfies the second preset condition from the plurality of candidate freight sources, and determine remaining candidate freight sources as the plurality of sample freight sources; The driver behavior features include a driver performance, a driver outbound call volume, a driver outbound call success volume, and a driver identity label, and the driver identity label is configured to indicate whether the driver is a team leader or not; the freight owner behavior features include a freight owner published freight source volume and a freight owner transaction volume; The first preset condition is that the driver outbound call performance rate is less than a first threshold value, the driver outbound call success volume is greater than a second threshold value, and the driver identity label indicates that the driver is not a team leader; the driver outbound call performance rate is a ratio of the driver performance to the driver outbound call volume; The second preset condition is that the freight owner transaction rate is less than a third threshold value, and the freight owner published freight source volume is greater than a fourth threshold value; the freight owner transaction rate is a ratio of the freight owner transaction volume to the freight owner published freight source volume.

Images