Recommendation model training method, object recommendation method, equipment and storage medium

Abstract

The invention discloses a recommendation model training method, an object recommendation method, equipment and a storage medium, and belongs to the technical field of computers. The method comprises the following steps: acquiring sample data, acquiring daily conversion labels of the sample data from a delayed 0th day to a delayed Nth day, processing the sample data through N + 1 towers respectively, and determining daily predicted conversion rates from the delayed 0th day to the delayed Nth day, and determining a training loss value on the basis of the daily predicted conversion rate from the delayed 0th day to the delayed Nth day and the daily conversion label from the delayed 0th day to the delayed Nth day, and performing parameter optimization on the recommendation model on the basis of the training loss value until a convergence condition is met, thereby obtaining a trained target recommendation model. According to the method, the sample data within N days are supplemented, so that the deviation caused by delayed conversion is reduced, and the recommendation accuracy of the recommendation model is improved.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and in particular to a training method for a recommendation model, a method for recommending objects, a device, and a storage medium. Background Art

[0002] With the continuous development of computer technology and Internet technology, more and more applications have emerged. Most of these applications cannot do without recommendation systems, so it is necessary to improve the recommendation accuracy of recommendation systems.

[0003] In current recommendation systems, recommendation models are used for recommendations. To ensure the accuracy of recommendations, the recommendation models are trained every day based on the latest samples. However, in many scenarios, due to delayed conversions, positive samples may be marked as negative samples, resulting in poor recommendation accuracy of the trained recommendation models. Summary of the Invention

[0004] This application provides a training method for a recommendation model, a method for recommending objects, a device, and a storage medium, which can improve the accuracy of the recommendation model. The technical solution is as follows:

[0005] In one aspect, a method for training a recommendation model is provided, wherein the recommendation model is used to process a conversion rate prediction task, and the recommendation model includes N+1 towers. The method includes:

[0006] Obtaining sample data and obtaining a conversion label for each day of the sample data from the 0th day of the delay to the Nth day of the delay; the sample data includes attribute information of a target user and object information of a target object interacted with by the target user; when the sample data is converted on the i-th day from the 0th day of the delay to the Nth day of the delay, assigning a first label value to the conversion label for each day of the sample data from the i-th day of the delay to the Nth day of the delay, and assigning a second label value to the conversion label for each day of the sample data from the 0th day of the delay to the i-1th day of the delay, where i is less than or equal to N, and N is an integer greater than or equal to 1;

[0007] Processing the sample data through the N+1 towers respectively to determine the predicted conversion rate for each day from the 0th day of delay to the Nth day of delay;

[0008] Determine a training loss value based on the predicted conversion rate for each day from the 0th day of delay to the Nth day of delay and the conversion label for each day from the 0th day of delay to the Nth day of delay;

[0009] Parameters of the recommendation model are optimized based on the training loss value until a convergence condition is reached, thereby obtaining a trained target recommendation model.

[0010] In an optional manner, the processing of the sample data by the N+1 towers to determine the predicted conversion rate for each day from the 0th day of delay to the Nth day of delay respectively includes:

[0011] The sample data are processed by the N+1 towers respectively to obtain the probability of conversion between the 0th day of delay and the Nth day of delay, and the probability of conversion on the jth day of delay, where j is greater than or equal to 0 and less than or equal to N;

[0012] The predicted conversion rate on the Nth day of delay is multiplied by the probability of conversion within the 0th day to the Nth day of delay and conversion on the jth day of delay to obtain the predicted conversion rate on the jth day of delay, wherein the predicted conversion rate on the Nth day of delay is equal to the probability of conversion within the 0th day to the Nth day of delay and conversion on the Nth day of delay.

[0013] In an optional manner, the target recommendation model is a recommendation model deployed on day T+N+1, and the sample data includes sample data within day T+N;

[0014] The sample data are processed by the N+1 towers to determine the predicted conversion rate for each day from the 0th day of delay to the Nth day of delay; a training loss value is determined based on the predicted conversion rate for each day from the 0th day of delay to the Nth day of delay and the conversion label for each day from the 0th day of delay to the Nth day of delay; and parameters of the recommendation model are optimized based on the training loss value until a convergence condition is reached to obtain a trained target recommendation model, including:

[0015] The sample data up to the Tth day are processed by the N+1 towers to determine the predicted conversion rate for each day from the 0th day of delay to the Nth day of delay; a first training loss value is determined based on the predicted conversion rate for each day from the 0th day of delay to the Nth day of delay and the conversion label for each day from the 0th day of delay to the Nth day of delay; parameters of the recommendation model are optimized based on the first training loss value until convergence conditions are reached, thereby obtaining a trained complete model for the Tth day, wherein the structure of the complete model is the same as that of the recommendation model;

[0016] The sample data from the T+1 day to the T+N day are processed respectively by the N+1 towers in the complete model on the T-th day to determine the predicted conversion rate for each day from the 0th day of delay to the Nth day of delay; based on the predicted conversion rate for each day from the 0th day of delay to the Nth day of delay and the conversion label for each day from the 0th day of delay to the Nth day of delay, a second training loss value is determined; based on the second training loss value, the parameters of the complete model are optimized until the convergence condition is reached, thereby obtaining a trained target recommendation model.

[0017] In an optional manner, the recommendation model corresponds to M task objectives, where M is an integer greater than or equal to 2, the N+1 towers correspond to one task objective among the M task objectives, the recommendation model further includes (M-1)*(N+1) towers corresponding to M-1 task objectives among the M task objectives, and the daily conversion labels of the sample data from the 0th day of delay to the Nth day of delay include the daily conversion labels corresponding to the M task objectives from the 0th day of delay to the Nth day of delay;

[0018] The method further comprises:

[0019] Processing the sample data respectively through the (M-1)*(N+1) towers to determine the predicted conversion rate for each day from the 0th day of delay to the Nth day of delay under the M-1 task objectives;

[0020] The determining of the training loss value based on the predicted conversion rate for each day from the 0th day of delay to the Nth day of delay and the conversion label for each day from the 0th day of delay to the Nth day of delay includes:

[0021] A training loss value is determined based on the predicted conversion rate for each day from the 0th day of delay to the Nth day of delay and the conversion label for each day from the 0th day of delay to the Nth day of delay under the M task objectives.

[0022] In an optional manner, the target recommendation model is a music recommendation model;

[0023] The M mission objectives include collection, completion and membership; or,

[0024] The M task objectives include clicks, favorites, completion of broadcasts and membership.

[0025] In another aspect, a method for recommending an object is provided, the method comprising:

[0026] Obtain the user's attribute information and the object information of each object in the recommendation candidate set;

[0027] For each of the objects, determining a predicted conversion rate of the object on the Nth day of delay based on the attribute information, the object information of the object, and the target recommendation model as described in the first aspect or any optional embodiment of the first aspect, and determining a recommendation score for the object based on the determined predicted conversion rate;

[0028] Based on the recommendation score of each object in the recommendation candidate set, an object recommended for the user is output.

[0029] In an optional manner, the target recommendation model corresponds to M task targets, where M is an integer greater than or equal to 2;

[0030] The step of determining, for each object, a predicted conversion rate of the object on the Nth day of delay based on the user information, the object information of the object, and a target recommendation model, and determining a recommendation score of the object based on the determined predicted conversion rate includes:

[0031] For each of the objects, based on the attribute information, the object information of the object, and the target recommendation model, determining, for each of the M task targets, a predicted conversion rate of the object on the Nth day of delay;

[0032] The predicted conversion rates determined for at least two of the M task objectives are fused to obtain a recommendation score for the object.

[0033] On the other hand, a training device for a recommendation model is provided, the device comprising one or more modules, and the one or more modules are used to implement the method described in the first aspect or any optional embodiment of the first aspect.

[0034] On the other hand, a device for recommending an object is provided, the device comprising one or more modules, and the one or more modules are used to implement the method described in the second aspect or any optional manner of the second aspect.

[0035] On the other hand, the present application provides a computer device, comprising a processor and a memory, wherein the memory stores at least one instruction, and the instruction is loaded and executed by the processor to implement the method described in the first aspect or any optional manner of the first aspect, or to implement the method described in the second aspect or any optional manner of the second aspect.

[0036] On the other hand, the present application provides a computer-readable storage medium, which stores at least one instruction, and the instruction is loaded and executed by a processor to implement the method described in the first aspect or any optional manner of the first aspect, or to implement the method described in the second aspect or any optional manner of the second aspect.

[0037] On the other hand, the present application provides a computer program product, which stores at least one instruction, and the instruction is loaded and executed by a processor to implement the method described in the first aspect or any optional manner of the first aspect, or to implement the method described in the second aspect or any optional manner of the second aspect.

[0038] The beneficial effects of the technical solution provided by the embodiments of the present application are:

[0039] When training the recommendation model, we backfill sample data from the past N days to obtain sample data with delayed conversions, thus reducing the sample bias caused by delayed conversions. This way, using more accurate sample data will improve the accuracy of the recommendation model trained. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0041] Figure 1 This is a flow chart of the training method of the recommendation model provided in the embodiment of the present application;

[0042] Figure 2 This is a schematic diagram of the architecture of the recommendation model provided in the embodiment of the present application;

[0043] Figure 3 This is a flow chart of a method for recommending objects provided in an embodiment of the present application;

[0044] Figure 4 Schematic diagram of the structure of the training device of the recommendation model provided in the embodiment of the present application;

[0045] Figure 5 This is a schematic diagram of the structure of the device of the recommended object provided in the embodiment of the present application;

[0046] Figure 6 It is a structural diagram of the computer device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0047] In order to make the objectives, technical solutions and advantages of this application clearer, the implementation methods of this application will be further described in detail below with reference to the accompanying drawings.

[0048] The following is an introduction to the noun concepts involved in the embodiments of this application.

[0049] 1. Bayesian principle means that given a prior probability P(H), when evidence E appears, the posterior probability P(H|E) is calculated, where P(H|E) = P(E|H)*P(E) / P(H), and P(E) is the probability of evidence E appearing.

[0050] 2. MMOE (Multi-gate Mixture-of-Experts) structure: This is an MTL (Multi-Task Learning) model that can characterize the correlation between multiple tasks, learn multiple tasks based on shared representations, and avoid a surge in model parameters.

[0051] Recommendation systems are being applied in an increasing number of scenarios thanks to recommendation models, also known as ranking models. A good recommendation model can score a user's candidate items and accurately rank them. Users engage in a variety of behaviors in different scenarios, such as clicking, playing, adding to favorites, and signing up for memberships. Therefore, recommendation models must not only accurately characterize a single user behavior but also effectively score multiple behaviors.

[0052] Currently, the multi-task objective recommendation model is MMOE. MMOE primarily uses shared representations from multiple experts to model multiple task objectives without significantly increasing the number of parameters, thereby improving the accuracy of the recommendation model. However, delayed conversions often occur in real-world scenarios. For example, in e-commerce scenarios, users may add items they like to their shopping carts and then wait a few days or purchase them when they see a discount. Another example is music, where users collect songs they like, then play them repeatedly and eventually become members. Therefore, it is crucial to consider delayed conversions in recommendation models.

[0053] Based on this, in an embodiment of the present application, a training method for a recommendation model is provided. In this method, when obtaining sample data, delayed conversion is considered, and the user's behavior under the task goal is effectively captured. By training the recommendation model with sample data considering delayed conversion, the deviation caused by delayed conversion can be alleviated, thereby improving the recommendation accuracy of the recommendation model.

[0054] Next, the training method process of the recommendation model in the embodiment of the present application is described. Figure 1 Provides the training method process of the recommended model, see Figure 1 The execution subject of the method may be a computer device, such as a server, etc. In the embodiment of the present application, the execution subject is described as a server.

[0055] Step S101, obtain sample data, and obtain the conversion label of the sample data for each day from the 0th day of delay to the Nth day of delay; the sample data includes attribute information of the target user and object information of the target object interacted by the target user. When the sample data is converted on the i-th day from the 0th day of delay to the Nth day of delay, the conversion label of the sample data for each day from the i-th day of delay to the Nth day of delay is assigned to a first label value, and the conversion label of the sample data for each day from the 0th day of delay to the i-1th day of delay is assigned to a second label value, where i is less than or equal to N, and N is an integer greater than or equal to 1.

[0056] In this embodiment, the recommendation model is used to process the conversion rate prediction task. The structure of the recommendation model is an MMOE structure. The recommendation model corresponds to one or more task goals. In the case of corresponding to one task goal, the recommendation model includes N+1 towers, which are also called task towers. Generally speaking, in the case of corresponding to one task goal, the task goal is the task goal to achieve the final conversion, such as opening a membership. In the case of corresponding to multiple task goals (such as M task goals, M is an integer greater than or equal to 2), the recommendation model includes M*(N+1) towers. These multiple task goals are associated and are all related to the final desired conversion. N is an integer greater than or equal to 1. For each task goal, N+1 towers correspond to the 0th to the Nth day of delay. The value of N is related to the length of the window period. For example, N is the number of days included in the window period. The number of days included in the window period is 3. The N+1 towers are towers 1 to 4, tower 1 corresponds to the 0th day of delay, tower 2 corresponds to the 1st day of delay, tower 3 corresponds to the 2nd day of delay, and tower 4 corresponds to the 3rd day of delay. In different scenarios, the value of N is set based on empirical values. For example, the number of days generally required for conversion can be counted and this number of days can be determined as N, such as N is equal to 3.

[0057] The structure of the recommended model can be found in Figure 2 ,exist Figure 2 In this paper, we take multiple task objectives as an example to illustrate that the recommendation model includes embedding layer, expert network layer, gated network layer ( Figure 2(not shown) and the task layer, the embedding layer is connected to the expert network layer, the expert network layer is connected to the gating network layer, the gating network layer is connected to the task layer, and is connected to the embedding layer. Among them, the embedding layer is used to vectorize the user's attribute information and object information, and the object information is the object information of the object interacting with the user, and the vectorized user's attribute information and object information are spliced ​​together; the expert network layer includes multiple expert networks, and each expert network learns some unique features; the gating network is used to weight the output of each expert network using a gating mechanism. The gating network layer can be connected to all expert networks, or part of the expert networks. The gating network layer includes a neural network for each task target. The neural network input of each task target is a shared feature, and the output is the weight distribution of the expert network. The weight distribution is used to perform a weighted summation on the output of the expert network as the input of the tower; the task layer includes multiple towers for each task target among multiple task targets, and each tower is connected to the gating network layer. For each task target, different towers in the N+1 towers correspond to different delay days, and the N+1 towers correspond to the 0th day of delay to the Nth day of delay respectively. In Figure 2 In the example, N is equal to 3, and three task objectives are shown, including task objectives 1 to 3. In the embodiment of the present application, the specific structures of the embedding layer, expert network, gated network layer and tower are not limited. For example, the embedding layer is an underlying network architecture for extracting common features of input data, the expert network can be a fully connected layer structure for capturing information of the input vector from different dimensions, and the tower includes a fully connected layer and a nonlinear activation function.

[0058] In different scenarios, task objectives and objects are different. For example, in a music recommendation scenario, the object is a song, and multiple task objectives include collection, completion, and membership, or multiple task objectives include clicks, collection, completion, and membership, where completion refers to complete playback, or multiple task objectives include collection, playback, and membership, and playback refers to playback progress exceeding 80%. For another example, in an e-commerce scenario, the object is an item, and multiple task objectives include collection, adding to a shopping cart, and placing an order, or multiple task objectives include browsing, adding to a shopping cart, and placing an order. For another example, in a video recommendation scenario, the object is a video, which is a short video or a film and television video, and multiple task objectives include playback and membership.

[0059] When training the recommendation model, the server obtains sample data and the daily conversion labels of the sample data from the 0th day to the Nth day of the delay for each task target. The sample data includes the attribute information of the target user and the object information of the target object with which the target user interacts, wherein the attribute information includes but is not limited to the name, age, gender or region of the target user, and the object information includes but is not limited to the object identifier or the object category, etc. For example, if the target object is a song, the object information may include at least one of the song name, the song singer, the song genre or the song era. The attribute information can be the original unprocessed information of the target user or the vectorized information, and the object information can be the original unprocessed information of the object or the vectorized information.

[0060] When the sample data is converted on the i-th day between the 0th day of delay and the Nth day of delay, the conversion label of the sample data for each day between the 1st day of delay and the Nth day of delay is assigned to the first label value, and the conversion label of the sample data for each day between the 0th day of delay and the i-1th day of delay is assigned to the second label value. The first label value and the second label value are different, such as the first label value is 1 and the second label value is 0, or the first label value is 0 and the second label value is 1. For example, if N is 3, for a certain sample data, if it is collected on the 1st day of delay, for the collection task goal, the conversion label of the sample data in the next 3 days is 1, and the conversion label on the 0th day of delay is 0. In this way, the above-mentioned conversion label setting method is equivalent to backfilling user behavior.

[0061] It should be noted that when acquiring sample data, each piece of sample data has at least one conversion tag value under at least one task objective as the first tag value. If there are M task objectives, each of which corresponds to a conversion tag, then, if N is 3, one piece of sample data corresponds to M*4 conversion tags, including the four conversion tags for each task objective.

[0062] Step S102 : Process the sample data through N+1 towers respectively to determine the predicted conversion rate for each day from the 0th day of delay to the Nth day of delay.

[0063] The predicted conversion rate on the j-th day of delay refers to the predicted conversion rate of conversion on the j-th day of delay, where j is greater than or equal to 0 and less than or equal to N.

[0064] In this embodiment, an initial recommendation model is obtained. The initial recommendation model is an untrained model or a trained but incomplete model. Each sample data item is input into the initial recommendation model to determine the predicted conversion rate for each day from the 0th day of delay to the Nth day of delay.

[0065] In an optional approach, when determining the predicted conversion rate, the Bayesian principle can be introduced so that the task target on the delayed Nth day participates in the training of the task targets on the delayed 0th to N-1th days. The predicted conversion rate is determined as follows:

[0066] For each task target, the sample data is input into the initial recommendation model, and the initial recommendation model outputs the probability of conversion between the 0th day of delay and the Nth day of delay, and the conversion on the jth day of delay.

[0067] Then, the probability of conversion between the 0th and Nth days of delay and conversion on the Nth day of delay is determined as the predicted conversion rate on the Nth day of delay. Using the Bayesian principle, the predicted conversion rate on the Nth day of delay is multiplied by the first probability to obtain the predicted conversion rate on the jth day of delay. The first probability is the probability of conversion between the 0th and Nth days of delay and conversion on the jth day of delay. For example, if the value of N is 3, the predicted conversion rate on the jth day of delay (P(j)) is equal to P(3)*P(j|3), where P(3) is the predicted conversion rate on the 3rd day of delay and P(j|3) is the probability of conversion within 3 days of delay and conversion on the jth day of delay. The reason why this formula can be used for calculation is: From the Bayesian principle, we know that P(j|3)=P(3|j)*P(j) / P(3), P(3|j) represents the probability of conversion on the jth day of delay and within 3 days of delay. If the conversion has been completed on the jth day of delay, then the conversion will definitely occur within 3 days of delay, so P(3|j) is equal to 1, then P(j)=P(3)*P(j|3).

[0068] Step S103 : determining a training loss value based on the predicted conversion rate for each day from the 0th day of delay to the Nth day of delay and the conversion label for each day from the 0th day of delay to the Nth day of delay.

[0069] In this embodiment, for each task target, for each day from the 0th to the Nth day of delay, the predicted conversion rate is subtracted from the label value of the conversion label to obtain the training loss value.

[0070] Step S104: Optimize the parameters of the recommendation model based on the training loss value until the convergence condition is reached, thereby obtaining a trained target recommendation model.

[0071] The convergence condition is that the specified number of training times is reached, or the training loss value obtained by using the recommendation model obtained from the latest training to calculate the sample data is less than the specified value.

[0072] In this embodiment, the server sums the absolute values ​​of the training loss values ​​from the 0th to the Nth day of delay to obtain the total loss. The total loss is used to update the parameters of the embedding layer, the expert network layer, and the gating network layer. The training loss value of the 0th day of delay is used to update the parameters of the tower of the 0th day of delay. The training loss value of the 1st day of delay is used to update the parameters of the tower of the 1st day of delay. The training loss value of the Nth day of delay is used to update the parameters of the tower of the Nth day of delay. Then, the next one or more sample data are input and training continues until convergence conditions are reached, and the trained target recommendation model is obtained.

[0073] In one optional approach, the target recommendation model is the one deployed on day T+N+1. To obtain the most up-to-date recommendation model, the sample data includes sample data from days T+N. Before or at the end of day T+N, the target recommendation model is trained using sample data from days T+N (including day T+N). The following example uses a target recommendation model corresponding to multiple task objectives.

[0074] Optionally, the labels of sample data before (including) day T are completely determined. This is because the user behavior from day T+1 to day T+N has been obtained, but not all labels of sample data from day T+1 to day T+N can be determined. For example, when N is 3, day T+1 is the 0th day of delay, day T+2 is the 1st day of delay, day T+3 is the 2nd day of delay, and day T+4 is the 3rd day of delay. However, there is currently no data for day T+4, so the label cannot be completely determined. Here, "cannot be completely determined" can be understood as: for a certain task goal, if a certain sample data corresponds to a conversion on day T+1, then the label of the sample data in the subsequent N days is the first label value, and the label is completely determined. However, for days T+1 to T+3, the sample data does not correspond to a conversion, so on day T+4, the label of the sample data is uncertain. Therefore, we can first use the sample data before (including) day T to train the complete model for day T (the structure of the complete model is exactly the same as that of the recommended model, see Figure 2 ), and then use the complete model on day T and the sample data from day T+1 to day T+N to train the target recommendation model for day T+N+1. The target recommendation model for day T+N+1 can be used as the recommendation model for online recommendation on day T+N+1.

[0075] Optionally, when training the complete model on the Tth day, for each sample data up to the Tth day, for each task target, use multiple towers of each task target in the last obtained complete model (the last obtained complete model can be the initial recommendation model, which corresponds to the first training, or it can be the complete model obtained in the last update) to determine the predicted conversion rate corresponding to each day of the delay from the 0th day to the Nth day of the delay. Use the difference between each determined predicted conversion rate and the conversion label corresponding to the sample data to calculate the training loss value corresponding to each task target, which is called the first training loss value. Determine whether each first training loss value is less than a specified value. If it is less than the first specified value, stop training. If it is not less than the first specified value, use the first training loss value to update the last obtained complete model until the convergence condition is reached.

[0076] Here, when using the first training loss value to update the parameters of the tower, you can only use your own first training loss value to update the parameters of the tower to which the first training loss value belongs, or you can add the absolute values ​​of the first training loss values ​​of the towers with the same task target and update the parameters of each tower with the same task target. When updating the parameters of the expert network layer, the parameters of the gated network layer, and the parameters of the embedding layer, you can add the absolute values ​​of the first training loss values ​​corresponding to all towers to obtain the total loss, and use the total loss to update the parameters of the expert network layer, the parameters of the gated network layer, and the embedding layer.

[0077] Optionally, when training the target recommendation model for day T+N+1, sample data from day T+1 to day T+N is obtained. Since day T+1 is the 0th day of delay, the Nth day of delay is day T+N+1, and the sample data for day T+N+1 is not obtained. For each task target, when a conversion has occurred from day T+1 to day T+N, the conversion label of day T+N+1 is assigned to the first label value. For task targets that have not been converted from day T+1 to day T+N, the conversion label of day T+N+1 is alienated. The alienation processing includes but is not limited to setting the conversion label to a preset value, such as 999.

[0078] When training the target recommendation model for day T+N+1, for each sample data from day T+1 to day T+N, for each task target, use multiple towers of each task target in the complete model on day T to determine the predicted conversion rate corresponding to each day of the sample data from the 0th day of delay to the Nth day of delay. Use the difference between the determined predicted conversion rate and the conversion label corresponding to the sample data to calculate the training loss value corresponding to each tower, which is called the second training loss value. Determine whether each second training loss value is less than the second specified value. If it is less than the second specified value, stop training. If it is not less than the second specified value, use the second training loss value to update the complete model of day T, and use the next one or more sample data to train the updated complete model until the convergence condition is reached to obtain the target recommendation model for day T+N+1.

[0079] Here, when using the second training loss value to update the parameters of the tower, you can only use your own second training loss value to update the parameters of the tower to which the second training loss value belongs, or you can add the absolute values ​​of the second training loss values ​​of the towers with the same task target and update the parameters of each tower with the same task target. When updating the parameters of the expert network layer, the parameters of the gated network layer, and the parameters of the embedding layer, you can add the absolute values ​​of the second training loss values ​​corresponding to all towers to obtain the total loss, and use the total loss to update the parameters of the expert network layer, the parameters of the gated network layer, and the embedding layer.

[0080] Here, in the process of using the complete model to obtain the target recommendation model, since there may be no labels on some delayed days, the parameters of these towers may not be updated, but the parameters of the expert network and embedding layer can still be updated.

[0081] It should be noted that in order to improve the accuracy of recommendations, the complete model can be updated every w days, and the target recommendation model can be obtained based on the updated complete model, where w is relatively small, such as w equal to 1.

[0082] If the complete model is not trained for the first time on day T+N, the complete model from day T-1 can be updated using sample data up to day T to obtain the complete model for day T. This means that each complete model update is based on the most recent complete model. The recommendation model is updated daily based on the most recent complete model. Because the recommendation model is incrementally trained based on the complete model, it can be considered an incremental model.

[0083] In another optional manner, the complete model on day T mentioned above can be directly determined as the recommended model on day T+N+1.

[0084] The above training process based on sample data is only an example and is not limited in the embodiments of the present application.

[0085] In an embodiment of the present application, a method for recommending objects is also provided. As mentioned above, in different applications, the objects are different, such as songs in a music application, videos in a video application, and items in a shopping application. Figure 3 Provides a method flow for recommending objects, see Figure 3 Steps S201 to S203.

[0086] Step S201: Acquire the user's attribute information and the object information of each object in the recommendation candidate set.

[0087] In this embodiment, after the user opens the application, the server detects that the user is logged in and obtains the user's attribute information and the object information of each object in the recommendation candidate set. The concepts of attribute information and object information are described above and will not be repeated here.

[0088] In a music application, a recommendation candidate set can be a collection of objects selected based on the user's historical playback history. For example, a coarse-grained ranking model is used to input the user's historical playback history and information about the recommended objects. This model then selects objects from the available list to form a recommendation candidate set. For another example, a music application can identify songs that are highly similar to the user's historically played songs and form a recommendation candidate set.

[0089] In a shopping application, the recommendation candidate set can be a collection of objects selected based on the user's historical browsing behavior.

[0090] Step S202 : For each object, based on the attribute information, the object information of the object and the target recommendation model, determine the predicted conversion rate of the object on the Nth day of delay, and determine the recommendation score of the object based on the determined predicted conversion rate.

[0091] In this embodiment, for each object in the recommendation candidate set, the user's attribute information and the object's object information are input into the target recommendation model. If the target recommendation model targets a single task target, the target recommendation model outputs the predicted conversion rate for the object on the Nth day of delay. If the target recommendation model targets multiple task targets, the predicted conversion rate for the object on the Nth day of delay is output for each task target. Based on the determined predicted conversion rate, the recommendation score for the object is then determined.

[0092] In an optional manner, when the target recommendation model corresponds to a task target, the determined predicted conversion rate is determined as the recommendation score of the object.

[0093] In an optional manner, when the target recommendation model corresponds to multiple task targets, for each object, the predicted conversion rates determined for at least two task targets are fused to obtain a recommendation score for the object.

[0094] Optionally, for each object, the predicted conversion rates determined for at least two task objectives are multiplied together to obtain a product, which is then used as the recommendation score for the object. For example, if the multiple task objectives include favorites, complete broadcasts, and membership registration, the predicted conversion rates determined for favorites, broadcasts, and membership registration are multiplied together to obtain a product, which is then used as the recommendation score for the object.

[0095] Optionally, for each object, the predicted conversion rates determined for at least two task objectives are added together to obtain an addition result, which is determined as the recommendation score of the object.

[0096] Step S203: outputting the object recommended for the user based on the recommendation score of each object in the recommendation candidate set.

[0097] In this embodiment, objects in the recommendation candidate set are sorted according to the recommendation scores from high to low, and the top K objects are recommended to the user, where K is an integer greater than or equal to 1. This method is the top-K method.

[0098] Alternatively, the objects in the recommendation candidate set are sorted from high to low according to the recommendation scores, and objects with recommendation scores greater than a recommendation threshold are recommended to the user. The recommendation threshold can be determined based on an empirical value.

[0099] Here, when the server recommends to the user, it may send the recommended content to the user's terminal, and the terminal displays the content after receiving the content.

[0100] use Figure 3 The process shown, based on modeling the task objectives, effectively models the problem of delayed conversion, improves the fitting ability of the recommendation model, thereby improving the accuracy of recommendations for users and further improving the conversion rate.

[0101] In the embodiment of the present application, a sample-replenishing delay modeling scheme is used, and combined with the Bayesian principle, the deviation caused by delayed conversion can be alleviated in the modeling process. Moreover, the use of multi-task objectives can also alleviate the problem of inaccurate estimation caused by sample sparsity (such as the small amount of sample data for membership, we can obtain sample data for task objectives such as completion and collection to alleviate the problem of inaccurate estimation caused by sample sparsity). Ultimately, the online platform can accurately estimate user interests while improving the conversion rate of the overall platform.

[0102] Moreover, the multi-task target delayed conversion model has strong scalability. The recommendation model structure can be designed according to the different numbers of task targets and the different dependencies between task targets. Ultimately, the recommendation model can alleviate the problem of inaccurate estimation in modeling caused by delayed conversion.

[0103] All of the above optional technical solutions can be combined in any way to form optional embodiments of the present application, and will not be described in detail here.

[0104] Based on the same technical concept, the embodiment of the present application also provides a training device for a recommendation model, wherein the recommendation model is used to process a conversion rate prediction task, and the recommendation model includes N+1 towers; Figure 4 As shown, the apparatus includes a sample acquisition module 410 and a training module 420, wherein:

[0105] The sample acquisition module 410 is used to acquire sample data and a conversion label for each day of the sample data from the 0th day of the delay to the Nth day of the delay; the sample data includes attribute information of a target user and object information of a target object interacted with by the target user; when the sample data is converted on the i-th day from the 0th day of the delay to the Nth day of the delay, the conversion label for each day of the sample data from the i-th day of the delay to the Nth day of the delay is assigned a first label value, and the conversion label for each day of the sample data from the 0th day of the delay to the i-1th day of the delay is assigned a second label value, where i is less than or equal to N, and N is an integer greater than or equal to 1;

[0106] The training module 420 is used to:

[0107] Processing the sample data through the N+1 towers respectively to determine the predicted conversion rate for each day from the 0th day of delay to the Nth day of delay;

[0108] Determine a training loss value based on the predicted conversion rate for each day from the 0th day of delay to the Nth day of delay and the conversion label for each day from the 0th day of delay to the Nth day of delay;

[0109] Parameters of the recommendation model are optimized based on the training loss value until a convergence condition is reached, thereby obtaining a trained target recommendation model.

[0110] In an optional manner, the training module 420 is used to:

[0111] The sample data are processed by the N+1 towers respectively to obtain the probability of conversion between the 0th day of delay and the Nth day of delay, and the probability of conversion on the jth day of delay, where j is greater than or equal to 0 and less than or equal to N;

[0112] The predicted conversion rate on the Nth day of delay is multiplied by the probability of conversion within the 0th day to the Nth day of delay and conversion on the jth day of delay to obtain the predicted conversion rate on the jth day of delay, wherein the predicted conversion rate on the Nth day of delay is equal to the probability of conversion within the 0th day to the Nth day of delay and conversion on the Nth day of delay.

[0113] In an optional manner, the target recommendation model is a recommendation model deployed on day T+N+1, and the sample data includes sample data within day T+N;

[0114] The training module 420 is used to:

[0115] The sample data up to the Tth day are processed by the N+1 towers to determine the predicted conversion rate for each day from the 0th day of delay to the Nth day of delay; a first training loss value is determined based on the predicted conversion rate for each day from the 0th day of delay to the Nth day of delay and the conversion label for each day from the 0th day of delay to the Nth day of delay; parameters of the recommendation model are optimized based on the first training loss value until convergence conditions are reached, thereby obtaining a trained complete model for the Tth day, wherein the structure of the complete model is the same as that of the recommendation model;

[0116] The sample data from the T+1 day to the T+N day are processed respectively by the N+1 towers in the complete model on the T-th day to determine the predicted conversion rate for each day from the 0th day of delay to the Nth day of delay; based on the predicted conversion rate for each day from the 0th day of delay to the Nth day of delay and the conversion label for each day from the 0th day of delay to the Nth day of delay, a second training loss value is determined; based on the second training loss value, the parameters of the complete model are optimized until the convergence condition is reached, thereby obtaining a trained target recommendation model.

[0117] In an optional manner, the recommendation model corresponds to M task objectives, where M is an integer greater than or equal to 2, the N+1 towers correspond to one task objective among the M task objectives, the recommendation model further includes (M-1)*(N+1) towers corresponding to M-1 task objectives among the M task objectives, and the daily conversion labels of the sample data from the 0th day of delay to the Nth day of delay include the daily conversion labels corresponding to the M task objectives from the 0th day of delay to the Nth day of delay;

[0118] The training module 420 is further configured to:

[0119] Processing the sample data respectively through the (M-1)*(N+1) towers to determine the predicted conversion rate for each day from the 0th day of delay to the Nth day of delay under the M-1 task objectives;

[0120] The training module 420 is used to:

[0121] A training loss value is determined based on the predicted conversion rate for each day from the 0th day of delay to the Nth day of delay and the conversion label for each day from the 0th day of delay to the Nth day of delay under the M task objectives.

[0122] In an optional manner, the target recommendation model is a music recommendation model;

[0123] The M mission objectives include collection, completion and membership; or,

[0124] The M task objectives include clicks, favorites, completion of broadcasts and membership.

[0125] It should be noted that the recommendation model training device provided in the above embodiment only uses the division of the above functional modules as an example to illustrate the training of the recommendation model. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the recommendation model training device provided in the above embodiment and the recommendation model training method embodiment are based on the same concept. The specific implementation process is detailed in the method embodiment and will not be repeated here.

[0126] Based on the same technical concept, the embodiment of the present application also provides a device for recommending objects, such as Figure 5 As shown, the apparatus includes an acquisition module 510 and a determination module 520, wherein:

[0127] The acquisition module 510 is used to obtain the user's attribute information and the object information of each object in the recommendation candidate set;

[0128] The determination module 520 is used to:

[0129] For each of the objects, determining a predicted conversion rate of the object on the Nth day of delay based on the attribute information, the object information of the object, and the target recommendation model described above, and determining a recommendation score for the object based on the determined predicted conversion rate;

[0130] Based on the recommendation score of each object in the recommendation candidate set, an object recommended for the user is output.

[0131] In an optional manner, the target recommendation model corresponds to M task targets, where M is an integer greater than or equal to 2;

[0132] The determining module 520 is used to:

[0133] For each of the objects, based on the attribute information, the object information of the object, and the target recommendation model, determining, for each of the M task targets, a predicted conversion rate of the object on the Nth day of delay;

[0134] The predicted conversion rates determined for at least two of the M task objectives are fused to obtain a recommendation score for the object.

[0135] It should be noted that the device for recommending objects provided in the above embodiments only uses the division of the above functional modules as an example when recommending objects. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the device for recommending objects provided in the above embodiments and the method embodiment for recommending objects are based on the same concept. The specific implementation process is detailed in the method embodiment and will not be repeated here.

[0136] Figure 6 1 is a schematic diagram of the structure of a computer device provided in an embodiment of the present application. The computer device 1000 may vary significantly due to different configurations or performances, and may include one or more processors (central processing units, CPUs) 1001 and one or more memories 1002, wherein the memories 1002 store at least one instruction, and the at least one instruction is loaded and executed by the processor 1001 to implement the methods provided in the above-mentioned various method embodiments. Of course, the computer device may also have components such as a wired or wireless network interface, a keyboard, and an input / output interface for input and output. The computer device may also include other components for implementing device functions, which will not be described in detail here.

[0137] In an exemplary embodiment, a computer-readable storage medium is further provided, such as a memory including instructions, which can be executed by a processor in a terminal to complete the training method of the recommendation model in the above embodiment, or the method of recommending an object. The computer-readable storage medium can be non-transitory. For example, the computer-readable storage medium can be a ROM (Read-Only Memory), a RAM (Random Access Memory), a CD-ROM (Compact Disc Read-Only Memory), a magnetic tape, a floppy disk, and an optical data storage device.

[0138] It should be noted that the information (including but not limited to user device information, user personal information, etc.), data (including but not limited to data used for analysis, stored data, displayed data, etc.), and signals (including but not limited to signals transmitted between user terminals and other devices, etc.) involved in this application are all authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with the relevant laws, regulations and standards of the relevant countries and regions. For example, the user information and object information involved in this application are obtained with full authorization.

[0139] Those skilled in the art will understand that all or part of the steps to implement the above embodiments may be accomplished by hardware, or by a program to instruct the relevant hardware, and the program may be stored in a computer-readable storage medium, which may be a read-only memory, a disk, or an optical disk, etc.

[0140] The above description is merely an optional embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application shall be included in the scope of protection of the present application.

Claims

1. A training method for a recommendation model, characterized in that: The recommendation model is used to process the conversion rate prediction task, and the recommendation model includes N+1 towers; the method includes: Obtaining sample data and obtaining a conversion label for each day of the sample data from the 0th day of the delay to the Nth day of the delay; the sample data includes attribute information of a target user and object information of a target object interacted with by the target user; when the sample data is converted on the i-th day from the 0th day of the delay to the Nth day of the delay, assigning a first label value to the conversion label for each day of the sample data from the i-th day of the delay to the Nth day of the delay, and assigning a second label value to the conversion label for each day of the sample data from the 0th day of the delay to the i-1th day of the delay, where i is less than or equal to N, and N is an integer greater than or equal to 1; Processing the sample data through the N+1 towers respectively to determine the predicted conversion rate for each day from the 0th day of delay to the Nth day of delay; Determine a training loss value based on the predicted conversion rate for each day from the 0th day of delay to the Nth day of delay and the conversion label for each day from the 0th day of delay to the Nth day of delay; Parameters of the recommendation model are optimized based on the training loss value until a convergence condition is reached, thereby obtaining a trained target recommendation model.

2. The method according to claim 1, characterized in that The processing of the sample data by the N+1 towers to determine the predicted conversion rate for each day from the 0th day of delay to the Nth day of delay includes: The sample data are processed by the N+1 towers respectively to obtain the probability of conversion between the 0th day of delay and the Nth day of delay, and the probability of conversion on the jth day of delay, where j is greater than or equal to 0 and less than or equal to N; The predicted conversion rate on the Nth day of delay is multiplied by the probability of conversion within the 0th day to the Nth day of delay and conversion on the jth day of delay to obtain the predicted conversion rate on the jth day of delay, wherein the predicted conversion rate on the Nth day of delay is equal to the probability of conversion within the 0th day to the Nth day of delay and conversion on the Nth day of delay.

3. The method according to claim 1 or 2, characterized in that The target recommendation model is the recommendation model deployed on day T+N+1, and the sample data includes sample data within day T+N; The sample data are processed by the N+1 towers respectively to determine the predicted conversion rate for each day from the 0th day of delay to the Nth day of delay; a training loss value is determined based on the predicted conversion rate for each day from the 0th day of delay to the Nth day of delay and the conversion label for each day from the 0th day of delay to the Nth day of delay; Optimizing parameters of the recommendation model based on the training loss value until a convergence condition is reached to obtain a trained target recommendation model, including: The sample data up to the Tth day are processed by the N+1 towers to determine the predicted conversion rate for each day from the 0th day of delay to the Nth day of delay; a first training loss value is determined based on the predicted conversion rate for each day from the 0th day of delay to the Nth day of delay and the conversion label for each day from the 0th day of delay to the Nth day of delay; parameters of the recommendation model are optimized based on the first training loss value until convergence conditions are reached, thereby obtaining a trained complete model for the Tth day, wherein the structure of the complete model is the same as that of the recommendation model; The sample data from the T+1 day to the T+N day are processed respectively by the N+1 towers in the complete model on the T-th day to determine the predicted conversion rate for each day from the 0th day of delay to the Nth day of delay; based on the predicted conversion rate for each day from the 0th day of delay to the Nth day of delay and the conversion label for each day from the 0th day of delay to the Nth day of delay, a second training loss value is determined; based on the second training loss value, the parameters of the complete model are optimized until the convergence condition is reached, thereby obtaining a trained target recommendation model.

4. The method according to claim 1 or 2, characterized in that The recommendation model corresponds to M task objectives, where M is an integer greater than or equal to 2, and the N+1 towers correspond to one task objective among the M task objectives. The recommendation model also includes (M-1)*(N+1) towers corresponding to M-1 task objectives among the M task objectives. The daily conversion labels of the sample data from the 0th day of delay to the Nth day of delay include the daily conversion labels corresponding to the M task objectives from the 0th day of delay to the Nth day of delay. The method further comprises: Processing the sample data respectively through the (M-1)*(N+1) towers to determine the predicted conversion rate for each day from the 0th day of delay to the Nth day of delay under the M-1 task objectives; The determining of the training loss value based on the predicted conversion rate for each day from the 0th day of delay to the Nth day of delay and the conversion label for each day from the 0th day of delay to the Nth day of delay includes: A training loss value is determined based on the predicted conversion rate for each day from the 0th day of delay to the Nth day of delay and the conversion label for each day from the 0th day of delay to the Nth day of delay under the M task objectives.

5. The method according to claim 4, characterized in that The target recommendation model is a music recommendation model; The M mission objectives include collection, completion and membership; or, The M task objectives include clicks, favorites, completion of broadcasts and membership.

6. A method for recommending an object, characterized in that: The method comprises: Obtain the user's attribute information and the object information of each object in the recommendation candidate set; For each of the objects, based on the attribute information, the object information of the object, and the target recommendation model according to any one of claims 1 to 5, determining a predicted conversion rate of the object on the Nth day of delay, and determining a recommendation score of the object based on the determined predicted conversion rate; Based on the recommendation score of each object in the recommendation candidate set, an object recommended for the user is output.

7. The method according to claim 6, characterized in that The target recommendation model corresponds to M task targets, where M is an integer greater than or equal to 2; The step of determining, for each object, a predicted conversion rate of the object on the Nth day of delay based on the user information, the object information of the object, and the target recommendation model according to any one of claims 1 to 5, and determining a recommendation score of the object based on the determined predicted conversion rate, comprising: For each of the objects, based on the attribute information, the object information of the object, and the target recommendation model, determining, for each of the M task targets, a predicted conversion rate of the object on the Nth day of delay; The predicted conversion rates determined for at least two of the M task objectives are fused to obtain a recommendation score for the object.

8. A computer device, characterized in that: The computer device includes a processor and a memory, wherein the memory stores at least one instruction, and the instruction is loaded and executed by the processor to implement the operation performed by the method according to any one of claims 1 to 7.

9. A computer-readable storage medium, characterized in that The storage medium stores at least one instruction, which is loaded and executed by the processor to implement the operation performed by the method according to any one of claims 1 to 7.

10. A computer program product, characterized in that The computer program product stores at least one instruction, which is loaded and executed by a processor to implement the operations performed by the method according to any one of claims 1 to 7.

Images