Goods source recommendation method and device, electronic equipment and storage medium

Abstract

The invention provides a goods source recommendation method and device, electronic equipment and a storage medium. The method comprises the steps that first information is acquired, the first information comprises user information, goods source information, scene identification and matching information, the user information comprises user attribute information, user statistical information and user behavior sequence information of a first user, and the goods source information comprises goods source attribute information and goods source statistical information of a first goods source; according to the first information, through a multi-task multi-scene model, the click rate and the conversion rate of each goods source in the first goods source are determined, and the multi-task multi-scene model is obtained by training a basic model through a machine learning method. And according to the click rate and the conversion rate of each goods source in the first goods sources, a goods source list is obtained, and the goods source list is used for recommending goods sources to the first user, so that the goods source recommendation accuracy is improved.

Description

Technical Field

[0001] The present application relates to the field of logistics technology, and in particular to a method, device, electronic device and storage medium for recommending a source of goods. Background Art

[0002] With the development of the internet, digital freight has emerged. In digital freight, efficient matching between drivers and cargo sources is crucial to improving transportation efficiency and reducing costs.

[0003] In related technologies, transportation platforms can match drivers with suitable sources of goods based on the drivers' historical behavioral characteristics, such as the number of clicks, exposures, and searches for different types of sources of goods, and recommend these sources of goods to the drivers.

[0004] However, for different business scenarios, such as same-day cargo or subscription routes, drivers' cargo needs are diverse. Traditional cargo recommendation methods cannot take into account drivers' cargo needs in different business scenarios, resulting in poor accuracy of cargo recommendations. Summary of the Invention

[0005] The present application provides a cargo source recommendation method, device, electronic device and storage medium to solve the problem that traditional cargo source recommendation methods cannot take into account the drivers' cargo search needs in different business scenarios, resulting in poor accuracy of cargo source recommendation, thereby achieving the purpose of improving the accuracy of cargo source recommendation.

[0006] In a first aspect, the present application provides a method for recommending a source of goods, the method comprising:

[0007] Obtaining first information, where the first information includes user information, supply source information, a scenario identifier, and matching information. The user information includes user attribute information, user statistical information, and user behavior sequence information of the first user. The supply source information includes supply attribute information and supply statistical information of the first supply source. The scenario identifier is used to identify a business scenario corresponding to the first user. The matching information is used to indicate a degree of matching between the first user and the first supply source.

[0008] Determining, based on the first information, a click-through rate and a conversion rate for each of the first sources of supply using a multi-task, multi-scenario model, wherein the multi-task, multi-scenario model is obtained by training a basic model using a machine learning method, and the multi-task, multi-scenario model is used to predict the click-through rate and conversion rate of the sources of supply;

[0009] A supply list is obtained based on the click rate and conversion rate of each supply source in the first supply source, and the supply list is used to recommend supply sources to the first user.

[0010] Through the supply source recommendation method provided in the first aspect, first information is obtained, and the first information includes user information, supply source information, scenario identification and matching information. The user information can be used to learn the first user's long-term demand for finding goods, as well as the supply search preferences for specific types of supply sources in historical periods, specific business scenarios, and the market demand, user groups, transaction trends, user demand preferences, etc. of the first supply source can be learned through the supply source information. The matching information can take into account the degree of matching between the first supply source and the first user. The scenario identification can be used to learn the business scenario in which the first user is currently located, so that when recommending supply sources, targeted recommendations can be made to the first user based on the business scenario. Based on the first information, the click-through rate and conversion rate of each supply source in the first supply source are determined through a multi-task and multi-scenario model, so that the user's demand for finding goods can be analyzed based on the first information, so that the user-level information, the supply source-level information, and the degree of matching between the user and the supply source are taken into account when calculating the click-through rate and conversion rate, thereby taking into account the diversity of the user's demand for finding goods, the diversity of the supply source, and the differences in the business scenarios, making the predicted click-through rate and conversion rate more accurate. A supply list is obtained based on the click-through rate and conversion rate of each supply in the first supply source, so that the supply sources with high transaction or click probability can be placed at the front of the supply list through the click-through rate and conversion rate of each supply in the first supply source. The supply sources at the front of the supply list will be displayed to the first user first when they are at the front end, so that suitable supply sources can be recommended to the first user quickly and accurately through the supply list, thereby improving the accuracy of supply recommendation.

[0011] Furthermore, the first user can quickly obtain relevant information about the supply that is more suitable for him / herself, thereby improving the transaction rate and transaction speed of the supply.

[0012] In one possible design, based on the click-through rate and conversion rate of each source in the first source, a source list is obtained, including:

[0013] According to the click rate and conversion rate of each source in the first source, the ranking score of each source is determined by formula 1;

[0014] According to the sorting points, the first source of goods is sorted in descending order to obtain a source list;

[0015] Among them, formula 1 is: R = a × C1 b +c×(C1×C2) d ;

[0016] Among them, R is the ranking score of the source of goods, a is the first hyperparameter, b is the second hyperparameter, c is the third hyperparameter, d is the fourth hyperparameter, C1 is the click-through rate of the source of goods, and C2 is the conversion rate of the source of goods.

[0017] In one possible design, the method further includes:

[0018] Determine a first indicator and a second indicator of a precise source of goods in the first source of goods, and determine a third indicator and a fourth indicator of a generalized source of goods in the first source of goods, wherein the precise source of goods is used to indicate a source of goods that fully meets the user's search criteria, and the generalized source of goods is used to indicate a source of goods that meets the generalized user's search criteria, the first indicator is used to indicate the click probability of the precise source of goods, the second indicator is used to indicate the transaction probability of the precise source of goods, the third indicator is used to indicate the click probability of the generalized source of goods, and the fourth indicator is used to indicate the transaction probability of the generalized source of goods;

[0019] Determine an expected minimization loss function based on the first indicator, the second indicator, the third indicator, and the fourth indicator;

[0020] According to the expected minimization loss function, the first hyperparameter, the second hyperparameter, the third hyperparameter, and the fourth hyperparameter are optimized, and Formula 1 is updated.

[0021] In one possible design, the multi-task multi-scenario model includes a vector processing unit, an information extraction unit, a deep matching unit, a sequence conversion unit, a feature connection unit, a click rate tower unit, and a conversion rate tower unit. The vector processing unit is connected to the information extraction unit, the deep matching unit, and the sequence conversion unit, respectively. The information extraction unit, the deep matching unit, and the sequence conversion unit are respectively connected to the feature connection unit, and the feature connection unit is respectively connected to the click rate tower unit and the conversion rate tower unit.

[0022] a vector processing unit for processing user attribute information and user statistical information to obtain a first feature vector, processing supply information to obtain a second feature vector, processing scene identification to obtain a third feature vector, processing matching information to obtain a fourth feature vector, processing user behavior sequence information to obtain a fifth feature vector, and inputting the first, second, and third feature vectors into the information extraction unit, inputting the fourth feature vector into the deep matching unit, and inputting the fifth feature vector into the sequence conversion unit;

[0023] an information extraction unit, configured to extract information from the first eigenvector, the second eigenvector, and the third eigenvector to obtain a sixth eigenvector, and input the sixth eigenvector into a feature connection unit;

[0024] a depth matching unit, configured to perform depth matching on the fourth eigenvector to obtain a seventh eigenvector, and input the seventh eigenvector into the feature connection unit;

[0025] a sequence conversion unit, configured to perform sequence conversion on the fifth eigenvector to obtain an eighth eigenvector, and input the eighth eigenvector into the feature connection unit;

[0026] a feature connection unit, configured to connect and merge the sixth feature vector, the seventh feature vector, and the eighth feature vector to obtain a ninth feature vector, and input the ninth feature vector into the click rate tower unit and the conversion rate tower unit respectively;

[0027] The click rate tower unit is used to process the ninth eigenvector to obtain the click rate;

[0028] The conversion rate tower unit is used to process the ninth eigenvector to obtain the conversion rate.

[0029] In one possible design, the information extraction unit includes a scene information extraction unit and a task information extraction unit, and the scene information extraction unit and the task information extraction unit are connected in sequence;

[0030] a scene information extraction unit, configured to perform scene information extraction processing on the first eigenvector, the second eigenvector, and the third eigenvector to obtain a tenth eigenvector, and input the tenth eigenvector to the task information extraction unit;

[0031] The task information extraction unit is used to perform task information extraction processing on the tenth eigenvector to obtain a sixth eigenvector.

[0032] In one possible design, the scene information extraction unit includes a first scene-specific expert unit, a second scene-specific expert unit and a scene-sharing expert unit, and the task information extraction unit includes a click-through rate task-specific expert unit, a conversion rate task-specific expert unit and a task-sharing expert unit.

[0033] In one possible design, the method further includes:

[0034] Determine the sample supply from the historical supply corresponding to the historical user. The historical supply is the supply that has been exposed to the historical user.

[0035] Determine sample information based on the sample supply. The sample information includes: historical user information, sample supply information, sample matching information, sample scenario information, and a first tag, a second tag, and a historical scenario identifier corresponding to the sample supply. The historical user information is used to indicate the user information of the historical user, the sample supply information is used to indicate the supply information of the sample supply, the sample matching information is used to indicate the degree of matching between the historical user and the sample supply, the first tag is used to indicate whether the historical user clicked on the sample supply, the second tag is used to indicate whether the historical user traded the sample supply, and the historical scenario identifier is used to identify the business scenario corresponding to the sample supply;

[0036] The basic model is trained according to the sample information to obtain a multi-task multi-scenario model. The loss function of the multi-task multi-scenario model is: f(x i θ ctr )×f(xi θ cvr )];

[0037] Among them, θ ctr is the first parameter of the multi-task multi-scenario model, θ cvr is the second parameter of the multi-task multi-scenario model, N is a positive integer, y i is the first label, x i is the sample information, z i is the second label, and l(·) is the cross entropy loss function.

[0038] In a possible design, the sample supply source is determined from the historical supply sources corresponding to historical users, including:

[0039] Determine the average click-through rate and average transaction rate for each supply location in each business scenario;

[0040] Determine the propensity score corresponding to the second source of supply based on the average click-through rate, average transaction rate, and the first position corresponding to the second source of supply. The first position indicates the source position of the second source of supply when it was exposed to historical users. The second source of supply is a source of supply that has been exposed to historical users but not clicked.

[0041] According to the propensity score, the third source of supply is determined from the second source through the reservoir algorithm;

[0042] The third and fourth sources of supply are determined as sample sources of supply, and the fourth source of supply is the remaining source of supply in the historical sources of supply except the second source of supply.

[0043] In a second aspect, the present application provides a supply recommendation device, comprising: a module for executing the method of the first aspect and any possible involvement of the first aspect.

[0044] In a third aspect, the present application provides an electronic device, comprising: a processor;

[0045] The processor is used to execute the computer-executable program or instructions in the memory, so that the electronic device executes the method in the first aspect and any possible design of the first aspect.

[0046] In a fourth aspect, the present application provides an electronic device, comprising: a memory and a processor;

[0047] The memory is used to store program instructions; the processor is used to call the program instructions in the memory so that the electronic device executes the method in the first aspect and any possible design of the first aspect.

[0048] In a fifth aspect, the present application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor in an electronic device, implements the method in the first aspect and any possible design of the first aspect. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 A flowchart of a method for recommending a source of goods provided in one embodiment of the present application.

[0050] Figure 2 A schematic diagram of a supply list provided in accordance with an embodiment of the present application.

[0051] Figure 3 A flowchart of a method for obtaining a supply source list provided in one embodiment of the present application.

[0052] Figure 4 A flow chart of a method for updating parameters provided in one embodiment of the present application.

[0053] Figure 5 A schematic diagram of the structure of a multi-task and multi-scenario model provided in one embodiment of the present application.

[0054] Figure 6 A schematic diagram of the structure of an information extraction unit provided in one embodiment of the present application.

[0055] Figure 7 A schematic diagram of the structure of an information extraction unit provided in one embodiment of the present application.

[0056] Figure 8 A schematic diagram of the structure of a MaskNet Block network provided in one embodiment of the present application.

[0057] Figure 9 A flowchart of a method for obtaining a multi-task multi-scenario model provided in one embodiment of the present application.

[0058] Figure 10 A flow chart of a method for determining the source of a sample provided in one embodiment of the present application.

[0059] Figure 11 A schematic structural diagram of a supply source recommendation device provided in one embodiment of the present application.

[0060] Figure 12 A schematic structural diagram of an electronic device provided in one embodiment of the present application.

[0061] Figure 13 A schematic structural diagram of an electronic device provided in one embodiment of the present application. DETAILED DESCRIPTION

[0062] In this application, "at least one" refers to one or more, and "plurality" refers to two or more. "And / or" describes the association relationship of associated objects, indicating that three relationships may exist. For example, A and / or B can represent: the existence of A alone, the existence of A and B at the same time, and the existence of B alone, where A and B can be singular or plural. The character " / " generally indicates that the objects associated before and after are in an "or" relationship. "At least one of the following" or similar expressions refers to any combination of these items, including any combination of single or plural items. For example, at least one of a alone, b alone, or c alone can represent: a alone, b alone, c alone, a and b in combination, a and c in combination, b and c in combination, or a, b, and c in combination, where a, b, and c can be single or multiple. In addition, the terms "first" and "second" are used for descriptive purposes only and are not to be understood as indicating or implying relative importance.

[0063] The directions or positional relationships indicated by terms such as "center", "longitudinal", "lateral", "up", "down", "left", "right", "front", and "back" are based on the directions or positional relationships shown in the accompanying drawings and are only for the convenience of describing the present application and simplifying the description. They do not indicate or imply that the device or element referred to must have a specific direction, be constructed and operated in a specific direction. Therefore, they should not be understood as limiting the present application.

[0064] The terms "connected" and "connect" should be interpreted broadly. For example, "connected" or "connected" in a circuit structure can refer not only to a physical connection, but also to an electrical connection or a signal connection. For example, it can be a direct connection, i.e., a physical connection, or an indirect connection through at least one intermediate component, as long as the circuit is interconnected. It can also refer to internal connectivity between two components. Signal connection can refer not only to signal connection through circuits but also to signal connection through media, such as radio waves. Those skilled in the art will understand the specific meanings of the above terms in this application on a case-by-case basis.

[0065] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the accompanying drawings.

[0066] The present application provides a source of goods recommendation method, device, electronic device, computer-readable storage medium and computer program product. Based on static features such as user attribute information, user statistical information, and user behavior sequence information, combined with source of goods characteristics such as source of goods attribute information and source of goods statistical information, taking into account features such as scene identification and matching information, through a multi-task and multi-scenario model, the click-through rate and conversion rate of the source of goods are predicted, so that multiple different prediction tasks are taken into account in the prediction process, multiple different business scenarios are adapted, and the user's long-term preferences for finding goods can be taken into account. According to the click-through rate and conversion rate of the source of goods, the source of goods are sorted to obtain a source of goods list. Based on the source of goods list, suitable sources of goods can be recommended to users efficiently and accurately, thereby improving the accuracy of source of goods recommendations, and further, helping to improve the transaction speed and transaction rate of the source of goods.

[0067] Among them, the supply source recommendation method provided in this application can be executed by an electronic device, or by a supply source recommendation device in an electronic device.

[0068] The electronic device may be a server, a desktop computer, a mobile phone, a tablet computer, a laptop computer, a wearable device, an in-vehicle device, an augmented reality (AR) / virtual reality (VR) device, or the like.

[0069] The supply recommendation device can be implemented through a combination of software and / or hardware. For example, the supply recommendation device can be an application (APP), a webpage, or a public account. For example, the supply recommendation device can be a transportation platform. To simplify the description, the embodiment of the present application is described using the transportation platform as an example.

[0070] Next, combine Figure 1 , the supply recommendation method provided in the embodiment of the present application is described.

[0071] See also Figure 1 , Figure 1 This is a flow chart of a method for recommending a source of goods provided in one embodiment of the present application. Figure 1 As shown, the method includes:

[0072] S101. Obtain first information, where the first information includes user information, supply information, scenario identifier, and matching information.

[0073] The user information includes user attribute information, user statistical information and user behavior sequence information of the first user.

[0074] The first user is, for example, a driver.

[0075] When a user registers on a transportation platform, he or she will input user attribute information into the transportation platform, and the transportation platform will store the user attribute information.

[0076] Among them, user attribute information may include vehicle length and vehicle model.

[0077] For example, the user attribute information is: 4.5 meters, van.

[0078] Based on this, the transportation platform can obtain the user attribute information of the first user.

[0079] During the process of users using the transportation platform, the transportation platform can record relevant data generated by the users' behavior.

[0080] Among them, user behavior may include, for example, browsing sources, clicking on sources, completing sources, using search functions, and providing feedback on sources, etc. Among them, providing feedback on sources means that the user contacts the owner of the source through telephone or other means.

[0081] The transportation platform obtains user statistical information by recording relevant data generated by user behavior. User statistical information is used to indicate statistics related to the user. User statistical information can reflect the user's preferences for finding goods within a historical period, specific business scenarios, and specific types of sources of goods.

[0082] User statistics can include the following types.

[0083] 1. User category, which represents the overall statistics of users within the first preset time window. For example, user statistical information may include: the total number of product browsing, total number of product clicks, number of searches, total number of product transactions, and total number of product feedback, etc. within the first preset time window.

[0084] The total number of product views is the number of products viewed by users within the first preset time window. The total number of product clicks is the number of times users clicked on a product within the first preset time window. The number of searches is the number of times users used the search function within the first preset time window. The total number of product transactions is the number of products purchased by users within the first preset time window. The total number of product feedback is the number of times users provided product feedback within the first preset time window.

[0085] The first preset time window is, for example, 3 days, 7 days or 30 days.

[0086] 2. User-business scenario class, which represents the statistics of users in different types of business scenarios.

[0087] Among them, business scenarios include "daily supply", "subscription route", "home page recommendation" and "grab good products in seconds" and other scenarios.

[0088] The transportation platform will record relevant data generated by users' behaviors in these business scenarios.

[0089] For example, user statistical information may include: the user's supply transaction volume, supply transaction ratio, supply browsing volume, supply click count, and supply feedback count in the "Daily Supply" / "Subscription Route" / "Home Page Recommendation" / "Grab Good Products in Seconds" scenarios within the first preset time window.

[0090] Among them, the proportion of supply transactions is the proportion of supply transactions / views to total transactions / views in the scenarios of "Day's Supply" / "Subscription Route" / "Home Page Recommendation" / "Grab Good Products in Seconds".

[0091] In addition, user statistical information may also include: the proportion of user's supply clicks / feedback times in the "Today's Supply" / "Subscription Route" / "Home Page Recommendation" / "Grab Good Products in Seconds" scenarios within the first preset time window to the total supply clicks.

[0092] 3. User-source category, which represents the statistics of users for different types of sources within the first preset time window.

[0093] Among them, different types of cargo sources can be determined based on cargo source attribute information, which includes the weight, volume, category, length of the required vehicle, model of the required vehicle, departure information, and destination information of the cargo source.

[0094] Among them, the categories of goods can include agricultural products, industrial products and daily chemical products.

[0095] The departure information is used to indicate the location of the departure of the goods. For example, the departure information can be the location coordinates of the departure.

[0096] The destination information is used to indicate the destination location of the source of goods. For example, the destination information may be the location coordinates of the destination.

[0097] The transportation platform will record relevant data generated by users' behaviors regarding the above different types of sources of goods.

[0098] For example, user statistical information may include: the user's transaction volume, source browsing volume, source clicks, and source feedback volume for sources of goods of different categories / vehicle lengths / weights / volumes / lengths of required vehicles / models of required vehicles / departure information / destination information within the first preset time window.

[0099] 4. User-shipper category, which represents the statistics of users for different shippers within the first preset time window.

[0100] The transportation platform will record the relevant data generated by the user's behavior on the sources of goods uploaded by different shippers within the first preset time window.

[0101] For example, user statistical information may include: the user's supply transaction volume, supply fulfillment volume, supply default volume, and supply fulfillment rate with different supply owners within the first preset time window.

[0102] 5. User-route class, which represents the statistics of different transportation routes of users within the first preset time window.

[0103] The transportation platform will record relevant data generated by users' behavior on cargo sources on different transportation routes within the first preset time window.

[0104] Among them, the transportation routes are, for example, Xi'an→Shanghai, Beijing→Guangzhou, etc.

[0105] For example, user statistical information may include: the transaction volume, browsing volume, clicks and feedback of users on goods on different transportation routes within the first preset time window.

[0106] In addition, the types of user statistical information may also include user-route-vehicle length category, such as the number of clicks and transaction volume of the user's cargo supply for the vehicle length on a certain route in the first preset time window; user-fixed price category, such as the volume of cargo supply transaction completed by the user through a fixed price within the first preset time window; user-telephone negotiation category, such as the volume of cargo supply transaction completed by the user through telephone negotiation within the first preset time window.

[0107] Based on this, the transportation platform can obtain user statistical information of the first user.

[0108] During the process of the user using the transportation platform, the transportation platform can also record the user's behavior within the first preset time window to obtain user behavior sequence information.

[0109] For example, the behavior sequence formed by the transportation platform recording the user's latest N behaviors can be:

[0110] "Click on the destination: Shanghai, click on the departure place: Beijing, click on the cargo category: agricultural products, click on the vehicle length: 4.5 meters... Click on the source of goods, feedback on the source of goods, and complete the source of goods..."

[0111] Based on this, the transportation platform obtains the user behavior sequence information of the first user.

[0112] In summary, the transportation platform obtains user information. Through this information, the transportation platform can learn about the first user's long-term cargo needs on the transportation platform, as well as their preferences for specific cargo types within historical time periods, specific business scenarios, and specific types of cargo. This allows the transportation platform to consider the diversity of the first user's cargo needs when making cargo recommendations, helping to improve the accuracy of cargo recommendations.

[0113] In addition to user information, the transportation platform can also obtain supply information related to the first source of supply.

[0114] The first source of goods may include multiple sources of goods determined according to the user search conditions of the first user.

[0115] The transportation platform can determine the first source of goods based on the user search conditions input by the first user, ensuring that when recommending sources of goods, sources of goods that meet the user search conditions are recommended to the first user.

[0116] The supply source information includes supply attribute information and supply source statistical information of the first supply source.

[0117] For each source of goods, when the source of goods is put on the platform, the owner of the goods will input the source attributes of the source of goods to the transportation platform, so that the transportation platform can determine the source attribute information of the source of goods.

[0118] Among them, the cargo attribute information includes, for example, the loading and unloading method, the shipper's quotation, the category, the length of the required vehicle, the model of the required vehicle, the departure information, the destination information, the weight, the volume, the quality grade, whether it is a less-than-truckload cargo source, the creation time, the update time, the packaging type, etc.

[0119] Examples of loading and unloading methods include packaged and bulk. Product categories include primary categories, agricultural products, steel, and coal. Required vehicle lengths include 4.5 meters, 5 meters, and 10 meters. Required vehicle models include trucks, vans, and minivans. Packaging types include bulk and full case. Creation duration refers to the time it takes for a consignor to list a product. Update duration refers to the time it takes for a consignor to update product information.

[0120] Based on this, the transportation platform obtains the source attribute information of the first source of goods.

[0121] When the first source of supply includes multiple sources of supply, the source of supply attribute information of the first source of supply includes source of supply attribute information corresponding to each source of supply in the first source of supply.

[0122] For each source of goods, the transportation platform will record the relevant data generated after each source of goods is put on the shelves, so that the transportation platform can determine the source of goods statistical information.

[0123] Among them, there are multiple types of supply statistics information.

[0124] 1. Supply category, which represents the overall statistics of the supply within the second preset time window. For example, supply statistics may include: the number of exposures, clicks, and feedback received for the supply within the second preset time window.

[0125] The second preset time window is, for example, 3 days, 7 days, 30 days, etc.

[0126] 2. Source-owner category, which represents the behavior statistics of the owner corresponding to the source within the second preset time window.

[0127] For example, the supply source statistical information may include: the number of user clicks, the number of user browsing, the rate of unsuccessful orders, the shipment volume, and the registration duration of the cargo owner within the second preset time window.

[0128] Among them, the "flying order rate" refers to the ratio of the number of orders that the shipper fails to execute within the second preset time window to the total transaction volume.

[0129] 3. Cargo source-route category, which represents the cargo source statistics of different routes within the second preset time window.

[0130] For example, supply statistics may include: the total quantity of supplies in different routes, the supply-demand ratio, the quantity of supplies in different categories, the number of supply exposures in different business scenarios, the number of clicks on supplies in different business scenarios, etc.

[0131] Based on this, the transportation platform obtains the supply statistics information of the first source of goods.

[0132] When the first source of supply includes multiple sources of supply, the source of supply statistical information of the first source of supply includes source of supply statistical information corresponding to each source of supply in the first source of supply.

[0133] In summary, the transportation platform obtains supply information. Through this supply information, the transportation platform can learn about the market demand, user groups, transaction trends, user demand preferences, etc. of the first supply source. This allows the transportation platform to consider the characteristics of the first supply source when making supply recommendations, thereby recommending more suitable supplies to the first user, which helps improve the accuracy of supply recommendations.

[0134] In addition to user information and cargo source information, the transportation platform also obtains scene identification and matching information.

[0135] The scenario identifier is used to identify the business scenario corresponding to the first user.

[0136] The matching information is used to indicate the matching degree between the first user and the first source of goods.

[0137] Taking into account that when the transportation platform determines the first source of goods based on the user search conditions of the first user, in order to expand the scope of sources of goods, the first source of goods may have generalized sources of goods, that is, sources of goods that do not fully meet the user search conditions. Based on this, the transportation platform can score the first source of goods according to the degree of matching between the first user and the first source of goods, thereby obtaining matching information.

[0138] For example, the first source of goods includes source 1, source 2 and source 3.

[0139] The user's search criteria are: "vehicle length 10 meters", "agricultural products", and "van truck".

[0140] The length of the vehicle required by cargo source 1 is 5 meters, the category of cargo source 1 is steel, and the model of the vehicle required by cargo source 1 is a van.

[0141] The length of the vehicle required by source 2 is 10 meters, the category of source 1 is agricultural products, and the type of vehicle required by source 1 is a truck.

[0142] The length of the vehicle required by cargo source 3 is 10 meters, the category of cargo source 1 is agricultural products, and the type of vehicle required by cargo source 1 is a van.

[0143] Then the matching information of the first source of goods can be shown in Table 1 below:

[0144] Table 1

[0145] Source 1 Source 2 Source 3 10 meters 0 1 1 agricultural products 0 1 1 van 1 0 1 Matching score 1 point 2 points 3 points

[0146] In Table 1, "1" indicates a match, and "0" indicates a mismatch. The "1," "2," and "3" scores in Table 1 are used to measure the degree of match between the first user and the first source. A higher score indicates a higher degree of match.

[0147] As shown in Table 1, source 1 has the lowest matching degree, source 3 has the highest matching degree, and source 2 has a medium matching degree.

[0148] In summary, the transportation platform obtains scenario identification and matching information. Using the scenario identification, the transportation platform can determine the first user's current business scenario, enabling it to make targeted recommendations based on the business scenario when recommending sources of supply. The transportation platform can also consider the degree of match between the first source of supply and the first user based on the matching information when recommending sources of supply, thereby recommending more suitable sources of supply to the first user and improving the accuracy of source of supply recommendations.

[0149] S102: Determine the click-through rate and conversion rate of each source of goods in the first source of goods through a multi-task multi-scenario model based on the first information.

[0150] Among them, the multi-task and multi-scenario model is obtained by training the basic model through machine learning methods.

[0151] Among them, the multi-task and multi-scenario model is used to predict the click-through rate (CTR) and conversion rate (CVR) of the supply.

[0152] Among them, the click-through rate represents the probability that the first user clicks on the source of goods, and the conversion rate represents the probability that the first user converts the source of goods.

[0153] In addition, the multi-task and multi-scenario model can also predict the click-through conversion rate (CTCVR) of the supply source. The click-through conversion rate is the product of the click rate and the conversion rate.

[0154] The click-through conversion rate is used to indicate the probability that the first user clicks and successfully completes the delivery.

[0155] The transportation platform can use a multi-task and multi-scenario model to analyze users' demand for finding goods based on the first information, so that the transportation platform can take into account user-level information, supply-level information, and the degree of matching between users and supplies when obtaining click-through rate and conversion rate, thereby taking into account the diversity of users' demand for finding goods and the diversity of supplies, making the predicted click-through rate and conversion rate more accurate, thereby improving the accuracy of supply recommendation when making supply recommendation.

[0156] S103. Obtain a supply source list according to the click rate and conversion rate of each supply source in the first supply source.

[0157] The supply list is used to recommend supply sources to the first user.

[0158] The cargo source list is obtained by arranging each cargo source in the first cargo source in a certain order by the transportation platform.

[0159] Among them, the click-through rate and conversion rate can reflect the probability that the first user clicks on the source of goods, makes a transaction, or clicks and makes a transaction. The transportation platform can place the sources of goods with high probability at the front of the source list based on the click-through rate and conversion rate of each source of goods in the first source of goods. The sources of goods at the front of the source list will be displayed to the first user first when they are at the front.

[0160] Among them, the transportation platform can obtain the source list through a variety of methods.

[0161] As a feasible implementation method, the transportation platform can obtain the click-through rate of each source of goods according to the click-through rate and conversion rate of each source of goods in the first source of goods, and sort the first sources of goods in descending order according to the size of the click-through rate to obtain a source of goods list.

[0162] For example, if Source 1 has a CTR of 0.8 and a conversion rate of 0.7, the CTR is 0.56. Source 2 has a CTR of 0.5 and a conversion rate of 0.3, the CTR is 0.15. Source 3 has a CTR of 0.7 and a conversion rate of 0.7, the CTR is 0.49.

[0163] The transportation platform can sort the supply source 1, supply source 2 and supply source 3 in descending order according to the click conversion rate, and obtain the supply source list as follows: supply source 1, supply source 3, supply source 2. Figure 2 shown.

[0164] As another feasible implementation method, the transportation platform can set preset weights for click-through rate and conversion rate, and obtain the following formula 2. The proportion of click-through rate and conversion rate during sorting is adjusted according to formula 2, thereby obtaining the ranking score of each source of goods in the first source of goods, and sorting the first source of goods in descending order according to the ranking score to obtain a source of goods list.

[0165] R=k×C1+j×C2 Formula 2;

[0166] Among them, R is the ranking score of the source of goods, k is the first preset weight, j is the second preset weight, C1 is the click rate, and C2 is the conversion rate.

[0167] Based on this, the transportation platform obtains a list of sources of goods, so that the transportation platform can quickly and accurately recommend suitable sources of goods to the first user through the list of sources of goods. The first user can quickly receive sources of goods that are more suitable for him / herself, thereby improving the transaction rate and transaction speed of the sources of goods.

[0168] In an embodiment of the present application, the transportation platform obtains first information, which includes user information, source information, scenario identification and matching information. The transportation platform can obtain the first user's long-term demand for finding goods on the transportation platform through the user information, as well as the preference for finding goods for a specific type of source in a historical period, a specific business scenario, and a specific source of goods through the source information. The market demand, user group, transaction trend, user demand preference, etc. of the first source of goods are obtained through the source information. The matching degree between the first source of goods and the first user is taken into account through the matching information, which helps to recommend a more suitable source of goods to the first user, so as to improve the accuracy of the source of goods recommendation. The transportation platform obtains the current business scenario of the first user through the scenario identification, so that when making source of goods recommendations, targeted recommendations can be made to the first user based on the business scenario. Based on the first information, the transportation platform determines the click-through rate and conversion rate of each source of goods in the first source of goods through a multi-task and multi-scenario model. The transportation platform can analyze the user's demand for finding goods based on the first information, so that when calculating the click-through rate and conversion rate, the user-level information, the supply-level information, and the degree of match between the user and the supply are taken into account, thereby taking into account the diversity of the user's demand for finding goods, the diversity of the supply, and the different business scenarios, making the predicted click-through rate and conversion rate more accurate. The transportation platform obtains a supply list based on the click-through rate and conversion rate of each supply in the first supply. The transportation platform can place the supply with a high probability at the front of the supply list based on the click-through rate and conversion rate of each supply in the first supply. The supply at the front of the supply list will be displayed to the first user first when it is at the front, so that the transportation platform can quickly and accurately recommend suitable supply to the first user through the supply list, making it easier for the first user to find a supply with a high transaction willingness, thereby improving the accuracy of the supply recommendation.

[0169] Furthermore, the first user can quickly receive goods that are more compatible with him / herself and have a higher transaction willingness, thereby improving the transaction rate and transaction speed of the goods.

[0170] Based on the above exemplary description, in S103, the transport platform can be Figure 3 Use the method shown to get the supply list.

[0171] See also Figure 3 , Figure 3 This is a flow chart of a method for obtaining a supply list provided in one embodiment of the present application. Figure 3 As shown, the method includes:

[0172] S201. Determine a ranking score for each source of goods according to Formula 1 based on the click-through rate and conversion rate of each source of goods in the first source of goods.

[0173] R=a×C1 b +c×(C1×C2) d Formula 1;

[0174] Among them, R is the ranking score of the source of goods, a is the first hyperparameter, b is the second hyperparameter, c is the third hyperparameter, d is the fourth hyperparameter, C1 is the click-through rate of the source of goods, and C2 is the conversion rate of the source of goods.

[0175] Initially, the initial values ​​of the first hyperparameter, the second hyperparameter, the third hyperparameter, and the fourth hyperparameter may be set based on experience.

[0176] The transportation platform obtains the ranking score of each source of goods according to Formula 1, so that the determined ranking score takes into account the combined impact of click-through rate, conversion rate, and click-through rate, avoiding determining the ranking score based on a single parameter, thereby making the determined ranking score more accurate and helping to obtain a list of sources of goods that better meets the needs of the first user.

[0177] S202. Arrange the first source of goods in descending order according to the ranking points to obtain a source of goods list.

[0178] The transportation platform sorts the top sources of goods in descending order based on their ranking scores, placing higher-scoring sources at the front of the list. These sources are then displayed to the first user first. This allows the platform to accurately recommend suitable sources of goods to the first user, improving both the transaction rate and the transaction speed.

[0179] Based on the above exemplary description, Figure 3 Based on the embodiment, the transport platform can also Figure 4 The method shown updates the hyperparameters in Equation 1.

[0180] See also Figure 4 , Figure 4 This is a flow chart of a method for updating parameters provided in one embodiment of the present application. Figure 4 As shown, the method includes:

[0181] S301. Determine the first index and the second index of the precise source of goods in the first source of goods, and determine the third index and the fourth index of the generalized source of goods in the first source of goods.

[0182] Among them, precise supply sources are used to indicate supply sources that fully meet the user's search conditions.

[0183] The generalized supply source is used to indicate the supply source that meets the generalized user search conditions.

[0184] When the transportation platform determines the first source of goods based on the user search conditions of the first user, it will determine a part of the sources of goods that completely meet the user search conditions and a part of the sources of goods that meet the generalized user search conditions.

[0185] For example, the user search conditions are: "vehicle length 10 meters", "agricultural products", and "van".

[0186] The transportation platform filters out the first part of the cargo that meets the requirements of "vehicle length 10 meters", "agricultural products", and "van" according to the user's search conditions. The first part of the cargo is the accurate cargo.

[0187] The transportation platform generalizes the user search conditions, that is, based on the above user search conditions, the search scope is appropriately expanded. For example, the generalized user search conditions can be: "vehicle length 10 meters", "van truck".

[0188] The transportation platform filters out all the second part of cargo sources that meet the requirements of "vehicle length 10 meters" and "van" based on the generalized user search conditions. The second part of cargo sources will also include the first part of cargo sources. Therefore, the cargo sources in the second part of cargo sources except the first part of cargo sources are generalized cargo sources.

[0189] The transportation platform jointly identifies the first part of the cargo source and the second part of the cargo source as the first source of cargo.

[0190] For the generalized and precise sources of goods in the first source of goods, the transportation platform evaluates the first indicator, the second indicator, the third indicator and the fourth indicator respectively.

[0191] Among them, the first indicator is used to indicate the click probability of the precise source of goods.

[0192] Among them, the second indicator is used to indicate the transaction probability of accurate supply.

[0193] Among them, the third indicator is used to indicate the click probability of the generalized supply source.

[0194] Among them, the fourth indicator is used to indicate the transaction probability of generalized supply.

[0195] The first indicator, the second indicator, the third indicator and the fourth indicator may be the area under the curve (Area Under Curve, AUC).

[0196] Wherein, AUC is the area under the receiver operating characteristic curve (ROC).

[0197] For the precise sources of goods and the generalized sources of goods in the first source of goods, the transportation platform determines the click AUC of the precise source of goods as the first indicator, the transaction AUC of the precise source of goods as the second indicator, the click AUC of the generalized source of goods as the third indicator, and the transaction AUC of the generalized source of goods as the fourth indicator.

[0198] Among them, AUC represents the probability that the predicted value of the positive sample is greater than the predicted value of the negative sample when randomly obtaining a positive and negative sample pair.

[0199] Among them, for the task of predicting the click probability of precise sources of goods, the positive samples are precise sources of goods that the user has clicked, and the negative samples are precise sources of goods that the user has not clicked.

[0200] For the task of predicting the transaction probability of precise supply sources, the positive samples are precise supply sources that the user has already transacted with, and the negative samples are precise supply sources that the user has not transacted with.

[0201] For the task of predicting the click probability of a generalized source of goods, the positive sample is the generalized source of goods that the user has clicked, and the negative sample is the generalized source of goods that the user has not clicked.

[0202] For the task of predicting the transaction probability of generalized goods, the positive samples are generalized goods that the user has already traded, and the negative samples are generalized goods that the user has not traded.

[0203] Based on this, the transportation platform determines the first indicator, the second indicator, the third indicator and the fourth indicator.

[0204] S302: Determine an expected minimization loss function based on the first indicator, the second indicator, the third indicator, and the fourth indicator.

[0205] The expected minimization loss function can be expressed as follows:

[0206] Loss1=-(F D +J D )×A×[-(F C +J C )]Formula 3;

[0207] Among them, Loss1 is the expected minimization loss, F D is the third indicator, J D is the first indicator, A is the fifth hyperparameter, F C is the fourth indicator, J C is the second indicator.

[0208] Among them, the fifth hyperparameter is pre-set based on experience.

[0209] S303. According to the expected minimization loss function, optimize the first hyperparameter, the second hyperparameter, the third hyperparameter, and the fourth hyperparameter, and update Formula 1.

[0210] The transportation platform can optimize the first hyperparameter, the second hyperparameter, the third hyperparameter, and the fourth hyperparameter based on historical cargo transaction volume, exposure, click volume, and other data.

[0211] The transportation platform optimizes the first hyperparameter, the second hyperparameter, the third hyperparameter, and the fourth hyperparameter based on the expected minimization loss function. When recommending sources of goods to the first user, the optimized first hyperparameter, the second hyperparameter, the third hyperparameter, and the fourth hyperparameter are substituted into the above formula one, thereby updating formula one and improving the accuracy of the ranking score of each source of goods determined according to formula one, thereby improving the accuracy of source of goods recommendations, and recommending sources of goods with higher transaction probabilities and click probabilities, that is, sources of goods that the first user is more interested in, to the first user, thereby improving the transaction rate and transaction speed of the sources of goods.

[0212] Based on the above exemplary description, below, Figures 5 to 7 The embodiment introduces the multi-task and multi-scenario model provided by the embodiment of this application.

[0213] See also Figure 5 , Figure 5 This is a structural diagram of a multi-task multi-scenario model provided in one embodiment of the present application. Figure 5 As shown, the multi-task multi-scenario model includes: a vector processing unit 10, an information extraction unit 20, a depth matching unit 30, a sequence conversion unit 40, a feature connection unit 50, a click rate tower unit 60 and a conversion rate tower unit 70.

[0214] Among them, the vector processing unit 10 is respectively connected to the information extraction unit 20, the depth matching unit 30 and the sequence conversion unit 40, the information extraction unit 20, the depth matching unit 30 and the sequence conversion unit 40 are respectively connected to the feature connection unit 50, and the feature connection unit 50 is respectively connected to the click rate tower unit 60 and the conversion rate tower unit 70.

[0215] Among them, the vector processing unit 10 is used to process user attribute information and user statistical information to obtain a first feature vector, process the supply information to obtain a second feature vector, process the scene identification to obtain a third feature vector, process the matching information to obtain a fourth feature vector, process the user behavior sequence information to obtain a fifth feature vector, and input the first feature vector, the second feature vector and the third feature vector into the information extraction unit 20, input the fourth feature vector into the deep matching unit 30, and input the fifth feature vector into the sequence conversion unit 40.

[0216] The transportation platform uses the vector processing unit 10 to perform low-dimensional vector conversion processing on user attribute information and user statistical information, cargo source information, scene identification, user behavior sequence information, and matching information, thereby mapping the information from data in high-dimensional space to low-dimensional space to reduce the dimension of the data and simplify data processing, thereby improving the efficiency of data processing, and removing noise and irrelevant information, thereby simplifying the complexity of the model and improving the interpretability of the model.

[0217] The information extraction unit 20 is used to extract information from the first eigenvector, the second eigenvector, and the third eigenvector to obtain a sixth eigenvector, and input the sixth eigenvector to the feature connection unit 50 .

[0218] The transportation platform processes the first feature vector, the second feature vector and the third feature vector through the information extraction unit 20, thereby extracting the features under the specific business scenario based on user attribute information, user statistical information, source information and scenario identification, and establishing click-through rate and conversion rate tasks, thereby capturing the first user's long-term and short-term preferences for finding goods.

[0219] The depth matching unit 30 is configured to perform depth matching on the fourth eigenvector to obtain a seventh eigenvector, and input the seventh eigenvector to the feature connection unit 50 .

[0220] The transportation platform performs deep matching on the fourth feature vector through the deep matching unit 30, so as to calculate the matching degree between the first user and the first source of goods based on the matching information.

[0221] The sequence conversion unit 40 is configured to perform sequence conversion on the fifth eigenvector to obtain an eighth eigenvector, and input the eighth eigenvector into the feature connection unit 50 .

[0222] The transportation platform performs sequence conversion on the fifth feature vector through the sequence conversion unit 40, and can capture the first user's long-term and short-term interest preferences in finding goods based on the user behavior sequence information, thereby being able to more accurately predict the click rate and conversion rate.

[0223] The feature connection unit 50 is used to connect and merge the sixth feature vector, the seventh feature vector and the eighth feature vector to obtain a ninth feature vector, and input the ninth feature vector into the click rate tower unit 60 and the conversion rate tower unit 70 respectively.

[0224] The transportation platform connects and merges the sixth eigenvector, the seventh eigenvector and the eighth eigenvector through the feature connection unit 50, and comprehensively considers user attribute information and user statistical information, source information, scene identification, user behavior sequence information, and matching information to obtain the ninth eigenvector, so as to accurately calculate the click-through rate and conversion rate of each source of goods.

[0225] The click rate tower unit 60 is used to process the ninth eigenvector to obtain the click rate.

[0226] The conversion rate tower unit 70 is used to process the ninth eigenvector to obtain the conversion rate.

[0227] The transportation platform processes the ninth eigenvector based on the click rate tower unit 60 and the conversion rate tower unit 70 respectively, thereby obtaining the click rate and conversion rate of each source of goods in the first source of goods.

[0228] exist Figure 5 Based on Figure 6 , Figure 6 This is a structural diagram of an information extraction unit provided in one embodiment of the present application. Figure 6 As shown, the information extraction unit 20 includes a scene information extraction unit 21 and a task information extraction unit 22, and the scene information extraction unit 21 and the task information extraction unit 22 are connected in sequence.

[0229] The scene information extraction unit 21 is configured to perform scene information extraction processing on the first eigenvector, the second eigenvector, and the third eigenvector to obtain a tenth eigenvector, and input the tenth eigenvector to the task information extraction unit 22 .

[0230] The task information extraction unit 22 is configured to perform task information extraction processing on the tenth eigenvector to obtain a sixth eigenvector.

[0231] Based on this, the transportation platform can first extract the scene features of the first feature vector, the second feature vector and the third feature vector according to the scene information extraction unit 21 to obtain the tenth feature vector, and then extract the task features related to the click-through rate and conversion rate through the task information extraction unit 22 based on the tenth feature vector, so as to take into account the needs of multiple scenes and multiple tasks.

[0232] exist Figure 6 Based on Figure 7 , Figure 7 This is a structural diagram of an information extraction unit provided in one embodiment of the present application. Figure 7 As shown, the scene information extraction unit 21 includes a first scene-specific expert unit 211, a second scene-specific expert unit 212 and a scene-sharing expert unit 213, and the task information extraction unit 22 includes a click-through rate task-specific expert unit 221, a conversion rate task-specific expert unit 222 and a task-sharing expert unit 223.

[0233] The first scenario-specific expert unit 211 and the second scenario-specific expert unit 212 are respectively used to extract features of a specific business scenario based on the first feature vector, the second feature vector and the third feature vector.

[0234] The first scenario may be, for example, business scenarios such as "daily supply" and "subscription route".

[0235] The second scenario may be, for example, business scenarios such as "daily supply" and "subscription route".

[0236] Among them, the scene sharing expert unit 213 is used to extract the cross-features in the first scene and the second scene based on the first feature vector, the second feature vector and the third feature vector, so that the transportation platform can more comprehensively consider the mutual influence between various business scenarios when making click-through rate and conversion rate predictions.

[0237] The click-through rate task-specific expert unit 221 is used to extract features related to the click-through rate based on the tenth feature vector, and the conversion rate task-specific expert unit 222 is used to extract features related to the conversion rate based on the tenth feature vector.

[0238] Among them, the task sharing expert unit 223 is used to extract cross-features related to click-through rate and conversion rate based on the tenth eigenvector, so that the transportation platform can more comprehensively consider the mutual influence between different tasks when making click-through rate and conversion rate predictions.

[0239] In some examples, the first scenario-specific expert unit 211, the second scenario-specific expert unit 212, the scenario-sharing expert unit 213, the click-through rate task-specific expert unit 221, the conversion rate task-specific expert unit 222, and the task-sharing expert unit 223 may all be implemented as follows: Figure 8 The MaskNet Block network shown.

[0240] Based on the above exemplary description, before using the multi-task multi-scenario model to determine the click rate and conversion rate, it is necessary to Figure 9 The method shown is used to train the basic model to obtain a multi-task and multi-scenario model.

[0241] See also Figure 9 , Figure 9 A flowchart of a method for obtaining a multi-task multi-scenario model provided in one embodiment of the present application.

[0242] like Figure 9 As shown, the method includes:

[0243] S401. Determine a sample supply from historical supply sources corresponding to historical users.

[0244] Among them, historical sources of goods are sources of goods that have been exposed to historical users.

[0245] The transportation platform's server stores a large amount of historical user data, which can be used for model training.

[0246] To ensure the accuracy of model training, the transportation platform can determine sample cargo sources from the historical cargo sources corresponding to multiple historical users. The transportation platform can determine all historical cargo sources as sample cargo sources, or it can select some historical cargo sources from the historical cargo sources corresponding to each historical user as sample cargo sources.

[0247] S402. Determine sample information based on the sample source.

[0248] The sample information includes: historical user information, sample supply information, sample matching information, sample scenario information, and the first label, second label, and historical scenario identifier corresponding to the sample supply.

[0249] The historical user information is used to indicate the user information of historical users.

[0250] The sample supply information is used to indicate the supply information of the sample supply.

[0251] Among them, the sample matching information is used to indicate the matching degree between historical users and sample sources.

[0252] The first tag is used to indicate whether the historical user clicked on the sample source. For example, the first tag can be "1" or "0", where "1" indicates that the historical user clicked on the sample source, and "0" indicates that the historical user did not click on the sample source.

[0253] The second tag is used to indicate whether the historical user has completed a transaction with the sample source. For example, the second tag can be "1" or "0", where "1" indicates that the historical user has completed a transaction with the sample source, and "0" indicates that the historical user has not completed a transaction with the sample source.

[0254] Among them, the historical scenario identifier is used to identify the business scenario corresponding to the sample supply source.

[0255] S403: Train the basic model according to the sample information to obtain a multi-task multi-scenario model.

[0256] Among them, the loss function of the multi-task multi-scenario model can be expressed as follows:

[0257]

[0258] Among them, θ ctr is the first parameter of the multi-task multi-scene model, θ cvr is the second parameter of the multi-task multi-scenario model, N is a positive integer, y i is the first label, x i is the sample information, z i is the second label, and l(·) is the cross entropy loss function.

[0259] In some examples, the first parameter is a network parameter of a click-through rate portion of a network in a multi-task multi-scenario model, and the second parameter is a network parameter of a conversion rate portion of a network in a multi-task multi-scenario model.

[0260] During training, the base model is trained based on the aforementioned loss function. When the loss function converges, training is complete, resulting in a multi-task, multi-scenario model. This loss function balances both the click-through rate and conversion rate tasks within the model, helping to improve the accuracy of determining conversion rates and click-through rates using the multi-task, multi-scenario model.

[0261] Based on the above exemplary description, in S401, the transport platform can be Figure 10 Determine the source of the sample using the method shown.

[0262] See also Figure 10 , Figure 10 This is a flow chart of a method for determining the source of a sample provided in one embodiment of the present application. Figure 10 As shown, the method includes:

[0263] S501: Determine the average click rate and average transaction rate of each supply source location in each business scenario.

[0264] In different business scenarios, there are multiple supply locations, and users may have different preferences for different supply locations. For example, users may be more likely to click on supply locations that are closer to the front. In order to ensure sample uniformity and training accuracy, the transportation platform can determine the average click-through rate and average transaction rate of each supply location in each business scenario.

[0265] For each business scenario, the transportation platform can determine the average click-through rate and average transaction rate of each source location based on the exposure, click volume and transaction volume of each source location in the business scenario within the historical period.

[0266] S502: Determine the propensity score corresponding to the second source of goods based on the average click rate, the average transaction rate, and the first position corresponding to the second source of goods.

[0267] The first position is used to indicate the source location of the second source when it was exposed to historical users.

[0268] Among them, the second source of goods is the source of goods in the historical sources that have been exposed to historical users but not clicked.

[0269] The transportation platform identifies historical sources of goods that have been exposed to users but not clicked on by them, i.e., secondary sources. These secondary sources are screened to ensure sample uniformity and avoid bias in the training sample. Furthermore, all historical sources of goods that have been exposed to users and clicked on by users can be used as samples, without considering the impact of user attention on different source locations.

[0270] For each secondary source of goods, the transportation platform determines the source position of the secondary source when it was previously exposed to users, i.e., the primary position. Based on the average click-through rate and average transaction rate of the primary position, the platform determines the corresponding propensity score for the secondary source of goods according to the following formula 5.

[0271] S=p×C R +q×D R Formula 5;

[0272] Among them, S is the propensity score, p is the sixth hyperparameter, and C R is the average click rate of the first position, q is the seventh hyperparameter, D R is the average transaction rate of the first position.

[0273] The sixth and seventh hyperparameters are predetermined based on experience. For example, the sixth hyperparameter may be 1, and the seventh hyperparameter may be 35.

[0274] When there are multiple second sources of cargo, the transportation platform calculates the propensity score of each second source of cargo to obtain multiple propensity scores.

[0275] Based on this, the transportation platform obtains a propensity score for each second source of cargo.

[0276] S503. Determine a third source of supply from the second source of supply using a reservoir algorithm based on the propensity score.

[0277] Specifically, the transportation platform selects K sources of goods from the second source of goods as initial samples, where K is a positive integer, and puts them into the reservoir. According to the propensity score corresponding to the initial sample, the sampling weight of each initial sample is calculated. From the remaining second sources of goods, one source of goods is selected, and the propensity score corresponding to the source of goods is calculated to calculate the sampling weight of the source of goods. If the sampling weights of the initial samples in the reservoir are all less than or equal to the sampling weight of the source of goods, the sample corresponding to the minimum sampling weight in the reservoir is deleted, and the source of goods is added to the reservoir as a new sample. Repeat the above operation until all second sources of goods have been processed. At this point, all samples in the reservoir are the third source of goods.

[0278] S504. Determine the third source of goods and the fourth source of goods as sample sources of goods.

[0279] Among them, the fourth source of supply is the remaining sources of supply in the historical sources of supply except the second source of supply, that is, the sources of supply in the historical sources of supply that have been exposed to users and clicked by users.

[0280] Based on this, the transportation platform identifies the third and fourth sources of cargo as sample sources of cargo. The transportation platform screens the sources of cargo based on the reservoir algorithm, which can correct the bias of the sources of cargo in different source locations, avoid the training sample deviation caused by users' different attention to different source locations, and screen out sample sources of cargo that are more conducive to training the model, thereby reducing the training error, improving the accuracy of the training model, and making the obtained multi-task and multi-scenario model more accurate when making predictions.

[0281] Figure 11 This is a schematic diagram of the structure of a supply recommendation device provided in one embodiment of the present application. Figure 11 As shown, the device includes: an acquisition module 101, a determination module 102 and a recommendation module 103.

[0282] The acquisition module 101 is used to obtain first information, where the first information includes user information, supply source information, scenario identification and matching information. The user information includes user attribute information, user statistical information and user behavior sequence information of the first user. The supply source information includes supply attribute information and supply statistical information of the first supply source. The scenario identification is used to identify the business scenario corresponding to the first user. The matching information is used to indicate the degree of matching between the first user and the first supply source.

[0283] Determination module 102 is used to determine the click-through rate and conversion rate of each source of supply in the first source of supply based on the first information through a multi-task multi-scenario model, wherein the multi-task multi-scenario model is obtained by training a basic model through a machine learning method, and the multi-task multi-scenario model is used to predict the click-through rate and conversion rate of the source of supply.

[0284] The recommendation module 103 is used to obtain a supply list based on the click rate and conversion rate of each supply in the first supply source, and the supply list is used to recommend the supply source to the first user.

[0285] It should be noted that the supply recommendation device of the embodiment of the present application can be used to implement the technical solution of the above-mentioned method embodiment. Its implementation principle and technical effects are similar and will not be repeated here.

[0286] In some examples, the recommendation module 103 is specifically configured to:

[0287] According to the click rate and conversion rate of each source in the first source, the ranking score of each source is determined by formula 1;

[0288] According to the sorting points, the first source of goods is sorted in descending order to obtain a source list;

[0289] Among them, formula 1 is: R = a × C1 b +c×(C1×C2) d ;

[0290] Among them, R is the ranking score of the source of goods, a is the first hyperparameter, b is the second hyperparameter, c is the third hyperparameter, d is the fourth hyperparameter, C1 is the click-through rate of the source of goods, and C2 is the conversion rate of the source of goods.

[0291] In some examples, the recommendation module 103 is further configured to:

[0292] Determine a first indicator and a second indicator of a precise source of goods in the first source of goods, and determine a third indicator and a fourth indicator of a generalized source of goods in the first source of goods, wherein the precise source of goods is used to indicate a source of goods that fully meets the user's search criteria, and the generalized source of goods is used to indicate a source of goods that meets the generalized user's search criteria, the first indicator is used to indicate the click probability of the precise source of goods, the second indicator is used to indicate the transaction probability of the precise source of goods, the third indicator is used to indicate the click probability of the generalized source of goods, and the fourth indicator is used to indicate the transaction probability of the generalized source of goods;

[0293] Determine an expected minimization loss function based on the first indicator, the second indicator, the third indicator, and the fourth indicator;

[0294] According to the expected minimization loss function, the first hyperparameter, the second hyperparameter, the third hyperparameter, and the fourth hyperparameter are optimized, and Formula 1 is updated.

[0295] In some examples, the multi-task multi-scenario model includes a vector processing unit, an information extraction unit, a deep matching unit, a sequence conversion unit, a feature connection unit, a click rate tower unit, and a conversion rate tower unit, wherein the vector processing unit is respectively connected to the information extraction unit, the deep matching unit, and the sequence conversion unit, the information extraction unit, the deep matching unit, and the sequence conversion unit are respectively connected to the feature connection unit, and the feature connection unit is respectively connected to the click rate tower unit and the conversion rate tower unit;

[0296] a vector processing unit for processing user attribute information and user statistical information to obtain a first feature vector, processing supply information to obtain a second feature vector, processing scene identification to obtain a third feature vector, processing matching information to obtain a fourth feature vector, processing user behavior sequence information to obtain a fifth feature vector, and inputting the first, second, and third feature vectors into the information extraction unit, inputting the fourth feature vector into the deep matching unit, and inputting the fifth feature vector into the sequence conversion unit;

[0297] an information extraction unit, configured to extract information from the first eigenvector, the second eigenvector, and the third eigenvector to obtain a sixth eigenvector, and input the sixth eigenvector into a feature connection unit;

[0298] a depth matching unit, configured to perform depth matching on the fourth eigenvector to obtain a seventh eigenvector, and input the seventh eigenvector into the feature connection unit;

[0299] a sequence conversion unit, configured to perform sequence conversion on the fifth eigenvector to obtain an eighth eigenvector, and input the eighth eigenvector into the feature connection unit;

[0300] a feature connection unit, configured to connect and merge the sixth feature vector, the seventh feature vector, and the eighth feature vector to obtain a ninth feature vector, and input the ninth feature vector into the click rate tower unit and the conversion rate tower unit respectively;

[0301] The click rate tower unit is used to process the ninth eigenvector to obtain the click rate;

[0302] The conversion rate tower unit is used to process the ninth eigenvector to obtain the conversion rate.

[0303] In some examples, the information extraction unit includes a scene information extraction unit and a task information extraction unit, and the scene information extraction unit and the task information extraction unit are sequentially connected;

[0304] a scene information extraction unit, configured to perform scene information extraction processing on the first eigenvector, the second eigenvector, and the third eigenvector to obtain a tenth eigenvector, and input the tenth eigenvector to the task information extraction unit;

[0305] The task information extraction unit is used to perform task information extraction processing on the tenth eigenvector to obtain a sixth eigenvector.

[0306] In some examples, the scenario information extraction unit includes a first scenario-specific expert unit, a second scenario-specific expert unit, and a scenario-shared expert unit, and the task information extraction unit includes a click-through rate task-specific expert unit, a conversion rate task-specific expert unit, and a task-shared expert unit.

[0307] In some examples, the apparatus also includes a training module.

[0308] The training module is used to determine sample supply sources from historical supply sources corresponding to historical users, where historical supply sources are supply sources that have been exposed to historical users;

[0309] Determine sample information based on the sample supply. The sample information includes: historical user information, sample supply information, sample matching information, sample scenario information, and a first tag, a second tag, and a historical scenario identifier corresponding to the sample supply. The historical user information is used to indicate the user information of the historical user, the sample supply information is used to indicate the supply information of the sample supply, the sample matching information is used to indicate the degree of matching between the historical user and the sample supply, the first tag is used to indicate whether the historical user clicked on the sample supply, the second tag is used to indicate whether the historical user traded the sample supply, and the historical scenario identifier is used to identify the business scenario corresponding to the sample supply;

[0310] The basic model is trained according to the sample information to obtain a multi-task multi-scenario model. The loss function of the multi-task multi-scenario model is:

[0311] Among them, θ ctr is the first parameter of the multi-task multi-scene model, θ cvr is the second parameter of the multi-task multi-scenario model, N is a positive integer, y i is the first label, x i is the sample information, z i is the second label, and l(·) is the cross entropy loss function.

[0312] In some examples, the training module is specifically configured to:

[0313] Determine the average click-through rate and average transaction rate for each supply location in each business scenario;

[0314] Determine the propensity score corresponding to the second source of supply based on the average click-through rate, average transaction rate, and the first position corresponding to the second source of supply. The first position indicates the source position of the second source of supply when it was exposed to historical users. The second source of supply is a source of supply that has been exposed to historical users but not clicked.

[0315] Based on the propensity score, the third source of supply is determined from the second source through the reservoir algorithm.

[0316] The third and fourth sources of supply are determined as sample sources of supply, and the fourth source of supply is the remaining source of supply in the historical sources of supply except the second source of supply.

[0317] Figure 12 This is a schematic diagram of the structure of an electronic device provided in one embodiment of the present application. Figure 12 As shown, the electronic device may include: a processor 201 and a memory 202, wherein the memory 202 stores a computer program, and when the processor 201 executes the computer program, the embodiment of the present application is implemented. Figures 1 to 10 The recommended method of sourcing is shown.

[0318] Figure 13This is a schematic diagram of the structure of an electronic device provided in one embodiment of the present application. Figure 13 As shown, the electronic device may include: a processor 301, and when the processor 301 executes a computer executable program or an instruction in a memory to execute a computer program, the embodiment of the present application is implemented. Figures 1 to 10 The recommended method of sourcing is shown.

[0319] Another embodiment of the present application further provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the embodiment of the present application can be implemented. Figures 1 to 10 The recommended method of sourcing is shown.

[0320] Another embodiment of the present application further provides a computer program product, comprising: an execution instruction, the execution instruction being stored in a readable storage medium, at least one processor of the electronic device being able to read the execution instruction from the readable storage medium, and at least one processor executing the execution instruction so that the electronic device implements the embodiment of the present application Figures 1 to 10 The recommended method of sourcing is shown.

[0321] In the several embodiments provided in this application, it should be understood that the disclosed systems, devices, and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of modules or units is merely a logical functional division. In actual implementation, other division methods may be used. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not performed.

[0322] Those skilled in the art will appreciate 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 to form different embodiments. For example, in the claims, any one of the claimed embodiments may be used in any combination.

[0323] As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for recommending a source of goods, characterized in that: The method comprises: Acquire first information, where the first information includes user information, supply source information, scenario identifier, and matching information, the user information includes user attribute information, user statistical information, and user behavior sequence information of the first user, the supply source information includes supply source attribute information and supply source statistical information of the first supply source, the scenario identifier is used to identify a business scenario corresponding to the first user, and the matching information is used to indicate a matching degree between the first user and the first supply source; According to the first information, determining the click rate and conversion rate of each source of supply in the first source of supply by using a multi-task multi-scenario model, wherein the multi-task multi-scenario model is obtained by training a basic model by a machine learning method, and the multi-task multi-scenario model is used to predict the click rate and conversion rate of the source of supply; A supply list is obtained according to the click rate and conversion rate of each supply source in the first supply source, and the supply list is used to recommend supply sources to the first user.

2. The method according to claim 1, characterized in that The step of obtaining a supply source list according to the click rate and conversion rate of each supply source in the first supply source includes: According to the click rate and conversion rate of each source of goods in the first source of goods, the ranking score of each source of goods is determined by formula 1; Arrange the first sources of goods in descending order according to the ranking points to obtain the source list; Wherein, the formula 1 is: R = a × C1 b +c×(C1×C2) d ; Among them, R is the ranking score of the source of goods, a is the first hyperparameter, b is the second hyperparameter, c is the third hyperparameter, d is the fourth hyperparameter, C1 is the click rate of the source of goods, and C2 is the conversion rate of the source of goods.

3. The method according to claim 2, characterized in that The method further comprises: Determine a first index and a second index of a precise source of goods in the first source of goods, and determine a third index and a fourth index of a generalized source of goods in the first source of goods, wherein the precise source of goods is used to indicate a source of goods that fully meets the user's search conditions, and the generalized source of goods is used to indicate a source of goods that meets the generalized user's search conditions, the first index is used to indicate a click probability of the precise source of goods, the second index is used to indicate a transaction probability of the precise source of goods, the third index is used to indicate a click probability of the generalized source of goods, and the fourth index is used to indicate a transaction probability of the generalized source of goods; Determine an expected minimization loss function according to the first indicator, the second indicator, the third indicator and the fourth indicator; According to the expected minimization loss function, parameter optimization is performed on the first hyperparameter, the second hyperparameter, the third hyperparameter and the fourth hyperparameter, and the formula one is updated.

4. The method according to any one of claims 1 to 3, characterized in that: The multi-task multi-scenario model includes a vector processing unit, an information extraction unit, a deep matching unit, a sequence conversion unit, a feature connection unit, a click rate tower unit and a conversion rate tower unit, wherein the vector processing unit is respectively connected to the information extraction unit, the deep matching unit and the sequence conversion unit, the information extraction unit, the deep matching unit and the sequence conversion unit are respectively connected to the feature connection unit, and the feature connection unit is respectively connected to the click rate tower unit and the conversion rate tower unit; The vector processing unit is used to process the user attribute information and the user statistical information to obtain a first feature vector, process the supply information to obtain a second feature vector, process the scene identifier to obtain a third feature vector, process the matching information to obtain a fourth feature vector, process the user behavior sequence information to obtain a fifth feature vector, and input the first feature vector, the second feature vector and the third feature vector to the information extraction unit, input the fourth feature vector to the deep matching unit, and input the fifth feature vector to the sequence conversion unit; The information extraction unit is used to extract information from the first feature vector, the second feature vector and the third feature vector to obtain a sixth feature vector, and input the sixth feature vector to the feature connection unit; The depth matching unit is used to perform depth matching on the fourth feature vector to obtain a seventh feature vector, and input the seventh feature vector to the feature connection unit; The sequence conversion unit is used to perform sequence conversion on the fifth feature vector to obtain an eighth feature vector, and input the eighth feature vector to the feature connection unit; The feature connection unit is used to connect and merge the sixth feature vector, the seventh feature vector and the eighth feature vector to obtain a ninth feature vector, and input the ninth feature vector into the click rate tower unit and the conversion rate tower unit respectively; The click rate tower unit is used to process the ninth feature vector to obtain the click rate; The conversion rate tower unit is used to process the ninth eigenvector to obtain the conversion rate.

5. The method according to claim 4, characterized in that The information extraction unit includes a scene information extraction unit and a task information extraction unit, and the scene information extraction unit is connected to the task information extraction unit in sequence; The scene information extraction unit is used to perform scene information extraction processing on the first feature vector, the second feature vector and the third feature vector to obtain a tenth feature vector, and input the tenth feature vector to the task information extraction unit; The task information extraction unit is used to perform task information extraction processing on the tenth feature vector to obtain a sixth feature vector.

6. The method according to claim 5, characterized in that The scene information extraction unit includes a first scene-specific expert unit, a second scene-specific expert unit and a scene-sharing expert unit, and the task information extraction unit includes a click-through rate task-specific expert unit, a conversion rate task-specific expert unit and a task-sharing expert unit.

7. The method according to any one of claims 1 to 3, characterized in that: The method further comprises: Determine a sample supply source from historical supply sources corresponding to historical users, where the historical supply sources are supply sources that have been exposed to the historical users; Determine sample information according to the sample source of goods, the sample information including: historical user information, sample source of goods information, sample matching information, sample scenario information, and a first label, a second label, and a historical scenario identifier corresponding to the sample source of goods, the historical user information is used to indicate the user information of the historical user, the sample source of goods information is used to indicate the source information of the sample source of goods, the sample matching information is used to indicate the degree of matching between the historical user and the sample source of goods, the first label is used to indicate whether the historical user clicks on the sample source of goods, the second label is used to indicate whether the historical user trades the sample source of goods, and the historical scenario identifier is used to identify the business scenario corresponding to the sample source of goods; The basic model is trained according to the sample information to obtain the multi-task multi-scenario model, and the loss function of the multi-task multi-scenario model is: Among them, θ ctr is the first parameter of the multi-task multi-scenario model, θ cvr is the second parameter of the multi-task multi-scenario model, N is a positive integer, y i is the first label, x i is the sample information, z i is the second label and l(·) is the cross entropy loss function.

8. The method according to claim 7, characterized in that Determining the sample supply source from the historical supply sources corresponding to the historical users includes: Determine the average click-through rate and average transaction rate for each supply location in each business scenario; Determine the propensity score corresponding to the second source of goods according to the average click-through rate, the average transaction rate, and the first position corresponding to the second source of goods, wherein the first position is used to indicate the source position of the second source of goods when it was exposed to the historical user, and the second source of goods is a source of goods in the historical sources of goods that has been exposed to the historical user but not clicked; According to the propensity score, a third source of supply is determined from the second source of supply by using a reservoir algorithm; The third source of supply and the fourth source of supply are determined as the sample sources of supply, and the fourth source of supply is the remaining sources of supply in the historical sources of supply except the second source of supply.

9. An electronic device, characterized in that: The method comprises a memory and a processor, wherein the memory stores a computer program that can be run on the processor, and when the processor executes the program, the method according to any one of claims 1 to 8 is implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 8 is implemented.

Images