Method and apparatus for determining target object recommended by service platform to user

Abstract

The embodiment of the present specification provides a method and device for determining a target object recommended by a service platform to a user, the method comprising: obtaining a user portrait feature of a first user, a sequence feature of an interaction sequence of each object with which the first user has interacted in a history of the service platform, and a candidate object feature of a candidate object recommended to the first user; the candidate object being any one of a plurality of objects; inputting the user portrait feature, the sequence feature, and the candidate object feature into a pre-trained neural network model, outputting a first predicted index value of at least one first index about the first user's current interaction with the candidate object, and a second predicted index value of at least one second index related to the first user's re-registration of the service platform; and determining a target object recommended by the service platform to the first user from the plurality of objects according to the first predicted index value and the second predicted index value corresponding to each object. The determined target object can be more in line with the user's preferences.

Description

Technical Field

[0001] This specification relates to one or more embodiments in the field of computers, and more particularly to a method and apparatus for determining target objects that a service platform recommends to a user. Background Technology

[0002] The service platform offers multiple interactive objects, each corresponding to a product or service provided to the user. The platform often recommends target objects to users after they log in, highlighting these objects to encourage interaction. To improve user experience, it's desirable to determine the target objects recommended by the platform in a way that aligns with user preferences.

[0003] In existing technologies, neural network models are typically used to predict the probability of a user's interaction with each object, selecting the object with the highest probability of interaction as the target object recommended by the service platform. With increasing emphasis on privacy data protection, these neural network models may involve privacy data during training and utilization, necessitating careful attention to prevent privacy data leakage. It is understandable that existing solutions predict the probability of a user's interaction with each object; that is, if the user interacts with a target object, the target object is considered to match the user's preferences; if no interaction occurs, the target object is considered not to match the user's preferences. Since the reason for the user's interaction with a target object might simply be due to the object's title attracting the user's attention, the identified target object may not actually match the user's preferences.

[0004] Therefore, an improved solution is needed to make the identified target objects more in line with user preferences. Summary of the Invention

[0005] This specification describes one or more embodiments of a method and apparatus for determining target objects that a service platform recommends to a user, enabling the determined target objects to better match user preferences.

[0006] In a first aspect, a method is provided for determining target objects recommended by a service platform to a user, wherein the service platform provides multiple interactive objects, each corresponding to a product or service offered to the user, the method comprising:

[0007] The system acquires the user profile features of the first user, the sequence features of the interaction sequence consisting of all objects that the first user has historically interacted with on the service platform, and the candidate object features of the candidate objects recommended to the first user; the candidate object is any one of the plurality of objects.

[0008] The user profile features, the sequence features, and the candidate object features are input into a pre-trained neural network model for model processing. The neural network model outputs a first predicted index value for at least one first indicator related to the first user's current interaction with the candidate object, and a second predicted index value for at least one second indicator related to the first user's re-login to the service platform.

[0009] Based on the first prediction index value and the second prediction index value corresponding to each of the multiple objects, the target object recommended by the service platform to the first user is determined from the multiple objects.

[0010] In one possible implementation, the interaction includes one or more of the following: browsing, adding to favorites, adding to cart, purchasing, and paying.

[0011] In one possible implementation, the sequence features include interaction object features corresponding to each interacted object, the interaction object features including object identifiers and time intervals between historical interaction behaviors and the current time; the candidate object features include object identifiers of candidate objects.

[0012] Furthermore, the time interval is encoded as a time embedding vector, which is formed by superimposing time information into components for multiple time periods and then concatenating these components.

[0013] Furthermore, the neural network model includes multiple branch networks and a sequence representation module shared by each branch network. The model processing includes: using the sequence representation module to encode the features of each interaction object and the features of the candidate object using an attention mechanism to obtain a sequence representation vector; and using the multiple branch networks to predict the first prediction index value and the second prediction index value based on the sequence representation vector.

[0014] Furthermore, the attention-based encoding includes:

[0015] An attention mechanism is used to represent the features of the interactive objects and the candidate objects that belong to the same object category, and to obtain the sequence representation sub-vector corresponding to the object category.

[0016] The sequence representation vector is obtained by concatenating the sequence representation subvectors corresponding to each object category.

[0017] Furthermore, the attention-based encoding includes:

[0018] An attention mechanism is used to represent the features of the interactive objects and the candidate objects that belong to the same object category, and to obtain the sequence representation sub-vector corresponding to the object category.

[0019] An attention mechanism is used to represent the features of each interacted object and the features of the candidate objects, respectively, to obtain a global sequence representation sub-vector.

[0020] The sequence representation vector is obtained by concatenating the sequence representation subvectors corresponding to each object category and the global sequence representation subvector.

[0021] In one possible implementation, the first indicator includes:

[0022] Click-through rate.

[0023] In one possible implementation, the second indicator includes:

[0024] The probability of a user logging back into the service platform after clicking on a candidate object, the time interval between a user logging back in and clicking on a candidate object, or the object clicked by the user upon logging back in.

[0025] In one possible implementation, the at least one second indicator includes:

[0026] The time interval between a user logging back in and clicking on a candidate object; the object clicked by the user upon logging back in.

[0027] The neural network model predicts the time interval between a user's re-login and the click on a candidate object based on the predicted object clicked when the user re-logs in.

[0028] In one possible implementation, determining the target object that the service platform recommends to the first user from the plurality of objects based on the first prediction index value and the second prediction index value corresponding to each object among the plurality of objects includes:

[0029] A first score is determined based on the first predictive index value of any object;

[0030] The second score is determined based on the object's second predictive index value;

[0031] Based on pre-set weights, the first score and the second score are weighted and summed to obtain the comprehensive score of the object;

[0032] Based on the comprehensive score of each object among multiple objects, the target object recommended by the service platform to the first user is determined from the multiple objects.

[0033] Furthermore, the second indicator includes:

[0034] The probability of a user logging back into the service platform after clicking on a candidate item, and the time interval between a user logging back in and clicking on a candidate item;

[0035] The determination of the second score includes:

[0036] The second score is obtained based on the positive correlation between the probability value and the time interval value in the second predictive index.

[0037] Secondly, a method for training a neural network model is provided, the method including:

[0038] The user profile features of the sample user are obtained, the sequence features of the interaction sequence of the objects that the sample user has interacted with in the past on the service platform are obtained, the first object features of the first object recommended to the sample user are obtained, the first tag value of at least one first indicator of the sample user’s current interaction with the first object is obtained, and the second tag value of at least one second indicator related to the sample user’s re-login to the service platform is obtained.

[0039] The user profile features, the sequence features, and the first object features are input into a neural network model for model processing. The neural network model outputs a first predicted index value for the at least one first indicator and a second predicted index value for the at least one second indicator.

[0040] A first prediction loss is determined based on the first label value and the first prediction index value of the sample users; a second prediction loss is determined based on the second label value and the second prediction index value of the sample users; the first prediction loss and the second prediction loss constitute the total loss.

[0041] The model parameters of the neural network model are adjusted with the goal of minimizing the total loss.

[0042] In one possible implementation, the at least one second indicator includes:

[0043] The time interval between the user logging back in and clicking the first object; the object clicked by the user when logging back in;

[0044] The neural network model predicts the time interval between a user's re-login and the first object clicked, based on the prediction results of the object clicked when the user re-logs in.

[0045] Thirdly, an apparatus is provided for determining target objects recommended by a service platform to a user, the service platform providing multiple interactive objects, each object corresponding to a product or service offered to the user, the apparatus comprising:

[0046] The acquisition unit is used to acquire the user profile features of the first user, the sequence features of the interaction sequence consisting of all objects that the first user has historically interacted with on the service platform, and the candidate object features of the candidate objects recommended to the first user; the candidate object is any one of the plurality of objects.

[0047] The prediction unit is used to input the user profile features, sequence features, and candidate object features acquired by the acquisition unit into a pre-trained neural network model for model processing, and output a first prediction index value of at least one first indicator related to the first user's current interaction with the candidate object, and a second prediction index value of at least one second indicator related to the first user's re-login to the service platform through the neural network model.

[0048] The determining unit is used to determine the target object recommended by the service platform to the first user from the plurality of objects based on the first prediction index value and the second prediction index value obtained by the prediction unit corresponding to each of the plurality of objects.

[0049] Fourthly, a training device for a neural network model is provided, the device comprising:

[0050] The acquisition unit is used to acquire user profile features of the sample user, sequence features of the interaction sequence consisting of objects that the sample user has historically interacted with on the service platform, first object features of the first object recommended to the sample user, first tag value of at least one first indicator of the sample user’s current interaction with the first object, and second tag value of at least one second indicator related to the sample user’s re-login to the service platform.

[0051] The prediction unit is used to input the user profile features, the sequence features, and the first object features acquired by the acquisition unit into a neural network model for model processing, and output the first predicted index value of the at least one first indicator and the second predicted index value of the at least one second indicator through the neural network model.

[0052] The determining unit is configured to determine a first prediction loss based on the first label value of the sample user obtained by the acquiring unit and the first prediction index value obtained by the prediction unit; and to determine a second prediction loss based on the second label value of the sample user obtained by the acquiring unit and the second prediction index value obtained by the prediction unit; the first prediction loss and the second prediction loss constitute the total loss.

[0053] A training unit is used to adjust the model parameters of the neural network model with the training objective of minimizing the total loss obtained by the determining unit.

[0054] Fifthly, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed in a computer, causes the computer to perform the method of the first or second aspect.

[0055] In a sixth aspect, a computing device is provided, including a memory and a processor, wherein executable code is stored in the memory, and when the processor executes the executable code, it implements the method of the first aspect or the second aspect.

[0056] The method and apparatus for determining target objects recommended by a service platform to a user, as provided in the embodiments of this specification, firstly acquire user profile features of a first user, sequence features of an interaction sequence consisting of objects that the first user has historically interacted with on the service platform, and candidate object features of candidate objects to be recommended to the first user; the candidate object is any one of the plurality of objects. Then, the user profile features, the sequence features, and the candidate object features are input into a pre-trained neural network model for model processing. The neural network model outputs a first predicted index value for at least one first indicator related to the first user's current interaction with the candidate object, and a second predicted index value for at least one second indicator related to the first user's re-login to the service platform. Finally, based on the first predicted index value and the second predicted index value corresponding to each of the plurality of objects, the target object recommended by the service platform to the first user is determined from the plurality of objects. As can be seen from the above, the embodiments of this specification, by predicting multiple index values ​​through the neural network model, wherein the first predicted index value is a prediction based on short-term user feedback, and the second predicted index value is a prediction based on long-term user feedback, can comprehensively determine the target object by combining short-term and long-term user feedback, making the determined target object more consistent with user preferences.

[0057] The training method and apparatus for the neural network model provided in the embodiments of this specification first acquire user profile features of a sample user, sequence features of an interaction sequence consisting of objects that the sample user has historically interacted with on the service platform, first object features of a first object recommended to the sample user, first label values ​​of at least one first indicator of the sample user's current interaction with the first object, and second label values ​​of at least one second indicator related to the sample user's re-login to the service platform. Then, the user profile features, the sequence features, and the first object features are input into the neural network model for model processing. The neural network model outputs first predicted indicator values ​​of the at least one first indicator and second predicted indicator values ​​of the at least one second indicator. Next, a first prediction loss is determined based on the sample user's first label value and first predicted indicator value. A second prediction loss is determined based on the sample user's second label value and second predicted indicator value. The first prediction loss and the second prediction loss constitute the total loss. Finally, the model parameters of the neural network model are adjusted with minimizing the total loss as the training objective. As can be seen from the above, in the embodiments of this specification, when training the neural network model, multiple indicator values ​​are predicted through the neural network model. The first predicted indicator value is a prediction of short-term user feedback, and the second predicted indicator value is a prediction of long-term user feedback. The total loss is determined based on the prediction loss of the above two predictions. The model parameters of the neural network model are adjusted with minimizing the total loss as the training objective. This allows the trained neural network model to comprehensively determine the target object by combining short-term and long-term user feedback, making the determined target object more in line with user preferences. Attached Figure Description

[0058] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0059] Figure 1 This is a schematic diagram illustrating an implementation scenario of one embodiment disclosed in this specification;

[0060] Figure 2 A flowchart illustrating a method for determining target objects recommended to a user by a service platform according to one embodiment is shown.

[0061] Figure 3 A schematic diagram of the structure of a neural network model according to one embodiment is shown;

[0062] Figure 4 A schematic diagram of the structure of a sequence characterization module according to one embodiment is shown;

[0063] Figure 5 A flowchart illustrating a training method for a neural network model according to one embodiment is shown.

[0064] Figure 6 A schematic block diagram illustrating an apparatus for determining target objects recommended to a user by a service platform according to one embodiment;

[0065] Figure 7 A schematic block diagram of a training apparatus for a neural network model according to one embodiment is shown. Detailed Implementation

[0066] The solution provided in this specification will now be described with reference to the accompanying drawings.

[0067] Figure 1 This is a schematic diagram illustrating an implementation scenario of one embodiment disclosed in this specification. This implementation scenario involves determining target objects that a service platform recommends to a user. The service platform provides multiple interactive objects, each corresponding to a product or service offered to the user. (Refer to...) Figure 1 After a user logs into the service platform, the user interface will display icons representing several interactive objects, for example, Figure 1 The document displays icons corresponding to objects 1 through 8, allowing users to interact with the corresponding objects by clicking the icons. These interactions can take many forms. For example, when an object corresponds to a product offered to the user, the user can click the object's icon to purchase the product. Such products can include, but are not limited to, tangible goods like clothing or intangible goods like computer programs. When an object corresponds to a service offered to the user, the user can click the object's icon to access or use the service. Such services can include, but are not limited to, games, shopping services, or financial services.

[0068] In this embodiment of the specification, the service platform recommends target objects to the user after login and prompts the user to interact with the target object by highlighting it. The highlighting method can be something like a floating badge on the corresponding icon. For example, Figure 1 Object 4 in the diagram represents the target object recommended to the user. A solution is proposed to determine the target object recommended by the service platform to the user, aiming to make the determined target object more aligned with user preferences.

[0069] It is understood that the embodiments in this specification need to implement personalized recommendations. That is, the target objects determined for different users may be different, and the target objects determined for the same user at different times may also be different.

[0070] Figure 2The flowchart illustrates a method for determining target objects recommended to a user by a service platform according to one embodiment. The service platform provides multiple interactive objects, each corresponding to a product or service offered to the user. The method can be based on... Figure 1 The implementation scenario is shown. For example... Figure 2 As shown, the method for determining the target object recommended by the service platform to the user in this embodiment includes the following steps: Step 21, obtaining the user profile features of the first user, the sequence features of the interaction sequence consisting of all objects that the first user has historically interacted with on the service platform, and the candidate object features of the candidate objects to be recommended to the first user; the candidate object is any one of the plurality of objects; Step 22, inputting the user profile features, the sequence features, and the candidate object features into a pre-trained neural network model for model processing, and outputting a first predicted index value of at least one first indicator related to the first user's current interaction with the candidate object, and a second predicted index value of at least one second indicator related to the first user's re-login to the service platform; Step 23, determining the target object recommended by the service platform to the first user from the plurality of objects based on the first predicted index value and the second predicted index value corresponding to each object among the plurality of objects. The specific execution method of each of the above steps is described below.

[0071] First, in step 21, the user profile features of the first user, the sequence features of the interaction sequence consisting of all objects that the first user has historically interacted with on the service platform, and the candidate object features of the candidate objects recommended to the first user are obtained; the candidate object is any one of the multiple objects. It can be understood that each object in the interaction sequence has a corresponding historical interaction time, and a certain order is formed based on the aforementioned historical interaction times.

[0072] In the embodiments described in this specification, the target objects recommended by the service platform to the user are determined, which is to determine the future items in the interaction sequence, often referred to as sequence recommendation.

[0073] In one example, the interaction includes one or more of the following: browsing, adding to favorites, adding to cart, purchasing, and payment. It is understood that all of these interactions can be achieved through user clicks.

[0074] In one example, the sequence features include interaction object features corresponding to each interacted object, and the interaction object features include object identifiers and the time interval between the historical interaction behavior and the current time. The candidate object features include object identifiers of candidate objects. It can be understood that the corresponding object embedding vector can be found using a lookup table based on the object identifier.

[0075] Furthermore, the time interval is encoded as a time embedding vector, which is formed by superimposing time information into components for multiple time periods and then concatenating these components.

[0076] Understandably, user preferences are influenced by their historical interactions with various objects, and this influence changes over time. Since modeling these preferences is highly time-sensitive, temporal encoding should be generalized to any unseen time period. For example, temporal information can be encoded first as a function of a frequency set {ω1, ..., ωk}, representing the time intervals in a continuous high-dimensional space to preserve temporal patterns. It is understood that the frequency corresponds to the time period. The temporal embedding vector is expressed by the following formula:

[0077]

[0078] in, For the time information component with respect to frequency ω1, For time information relative to frequency ω k The ingredients.

[0079] The time information is related to the frequency ω component. The corresponding components can also be obtained using sine and cosine functions, where the frequency ω can be any frequency in the frequency set {ω1, ..., ωk}. It can be represented as follows:

[0080]

[0081] In the embodiments of this specification, the encoding method for encoding the time interval is also called periodic encoding.

[0082] Then, in step 22, the user profile features, the sequence features, and the candidate object features are input into a pre-trained neural network model for model processing. The neural network model outputs a first predicted indicator value for at least one first indicator related to the first user's current interaction with the candidate object, and a second predicted indicator value for at least one second indicator related to the first user's re-login to the service platform. It can be understood that the first predicted indicator value is a prediction of short-term user feedback, and the second predicted indicator value is a prediction of long-term user feedback. Predicting the second predicted indicator value is beneficial for the user's long-term stable login.

[0083] In one example, the neural network model includes multiple branch networks and a sequence representation module shared by each branch network. The model processing includes: using the sequence representation module to encode the features of each interaction object and the features of the candidate object using an attention mechanism to obtain a sequence representation vector; and using the multiple branch networks to predict the first prediction index value and the second prediction index value based on the sequence representation vector.

[0084] In this example, each branch network corresponds to a first indicator or a second indicator. Through the multiple branch networks, predictions can be made for multiple indicators, which can better balance short-term and long-term user feedback to determine the target object.

[0085] Furthermore, the attention-based encoding includes:

[0086] An attention mechanism is used to represent the features of the interactive objects and the candidate objects that belong to the same object category, and to obtain the sequence representation sub-vector corresponding to the object category.

[0087] The sequence representation vector is obtained by concatenating the sequence representation subvectors corresponding to each object category.

[0088] In this example, the interacted objects and the candidate objects are divided into multiple subsequences according to the object category. Multiple objects in the same subsequence belong to the same object category. An attention mechanism is used to represent each object in each subsequence to obtain the sequence representation subvector corresponding to the object category. The subvectors are then concatenated to obtain the sequence representation vector. This method of processing according to the object category can better capture the sequence representation under different object categories.

[0089] For example, the interacted objects and the candidate objects are arranged sequentially according to the time of interaction as object 1, object 2, object 3, object 4, object 5, and object 6. It can be understood that object 6 is a candidate object, corresponding to the current moment. Objects 1, 3, and 4 belong to category 1, while objects 2, 5, and 6 belong to category 2. An attention mechanism is used to represent objects 1, 3, and 4, resulting in a sequence representation vector 1 corresponding to category 1. Similarly, an attention mechanism is used to represent objects 2, 5, and 6, resulting in a sequence representation vector 2 corresponding to category 2. The sequence representation vector 1 and sequence representation vector 2 are then concatenated to obtain the sequence representation vector.

[0090] In the embodiments described in this specification, different object categories may correspond to different service periods or product periods, such as months, weeks, days, etc.

[0091] Furthermore, the attention-based encoding includes:

[0092] An attention mechanism is used to represent the features of the interactive objects and the candidate objects that belong to the same object category, and to obtain the sequence representation sub-vector corresponding to the object category.

[0093] An attention mechanism is used to represent the features of each interacted object and the features of the candidate objects, respectively, to obtain a global sequence representation sub-vector.

[0094] The sequence representation vector is obtained by concatenating the sequence representation subvectors corresponding to each object category and the global sequence representation subvector.

[0095] In this example, a method of processing objects separately according to their categories is adopted, which can better capture the sequence representation under different object categories. Furthermore, an attention mechanism is used to represent the interactive object features and the candidate object features of each interactive object and the candidate object respectively to obtain a global sequence representation sub-vector. Then, the sequence representation sub-vectors corresponding to each object category and the global sequence representation sub-vector are concatenated to obtain the sequence representation vector, which can take into account the global sequence representation under the entire sequence.

[0096] For example, the interacted objects and the candidate objects are arranged sequentially according to the time of interaction as object 1, object 2, object 3, object 4, object 5, and object 6. It can be understood that object 6 is a candidate object, corresponding to the current moment. Objects 1, 3, and 4 belong to category 1, while objects 2, 5, and 6 belong to category 2. An attention mechanism is applied to represent objects 1, 3, and 4 to obtain sequence representation sub-vector 1 corresponding to category 1; the same attention mechanism is applied to objects 2, 5, and 6 to obtain sequence representation sub-vector 2 corresponding to category 2; and the same attention mechanism is applied to objects 1, 2, 3, 4, 5, and 6 to obtain a global sequence representation sub-vector 3. Sequence representation sub-vectors 1, 2, and 3 are concatenated to obtain the sequence representation vector.

[0097] In the embodiments of this specification, the attention mechanism can be used to represent the entire sequence or its subsequences consisting of the interacted objects and the candidate objects, and a similar processing method can be adopted.

[0098] For example, the process of performing representation processing on the interaction object features respectively corresponding to each interacted object and the candidate object features of the candidate object by using an attention mechanism may include: constructing a triplet (u, v i , t i ), wherein u represents a user, v i represents an object, t i represents an interaction time. It can be understood that the interaction time corresponding to the candidate object feature can be considered as the current time. Mark historical interaction records before current time t as S u,v ={(vi, ti)|ti<t}. Under the attention mechanism, considering the relationship between user behavior S u、v and the object v, the time series matrix at current time t is obtained as follows:

[0099]

[0100] wherein, is an object embedding vector corresponding to object v0, t-t0 is the time interval between the historical interaction behavior of user u and object v0 and the current time, is a time embedding vector corresponding to the time interval t-t0; is an object embedding vector corresponding to object v N-1 , t-t N-1 is the time interval between the historical interaction behavior of user u and object v N-1 and the current time, time embedding vector corresponding to time interval t-t N-1 ; e v is an object embedding vector corresponding to the candidate object v, since the candidate object corresponds to the current time, the time interval is 0, is a time embedding vector corresponding to time interval 0.

[0101] Then, the representation form of the sequence representation vector summarizing the influence of user behaviors is generated by using dot-product attention:

[0102]

[0103] wherein, Q u,v (t), K u,v (t) and V u,v (t) are linear mappings of the time series matrix Z u,v (t), respectively.

[0104] In one example, the first indicator includes:

[0105] Click-through rate.

[0106] In this example, click-through rate corresponds to short-term user feedback.

[0107] In one example, the second metric includes:

[0108] The probability of a user logging back into the service platform after clicking on a candidate object, the time interval between a user logging back in and clicking on a candidate object, or the object clicked by the user upon logging back in.

[0109] In this example, the probability of a user logging back into the service platform after clicking on a candidate object, the time interval between the user logging back in and clicking on a candidate object, or the object the user clicks upon logging back in, all correspond to long-term user feedback.

[0110] In one example, the at least one second indicator includes:

[0111] The time interval between a user logging back in and clicking on a candidate object; the object clicked by the user upon logging back in.

[0112] The neural network model predicts the time interval between a user's re-login and the click on a candidate object based on the predicted object clicked when the user re-logs in.

[0113] In this example, the prediction of the aforementioned time interval depends on the prediction result of the object clicked when the user logs back in; that is, the two predictions are sequential. In this embodiment of the specification, if the second indicator includes both the time interval between the user's login and clicking on a candidate object, and the object clicked when the user logs back in, the object clicked when the user logs back in can be used only as an intermediate result and not as the basis for subsequently determining the target object.

[0114] Finally, in step 23, based on the first and second predicted indicator values ​​corresponding to each of the multiple objects, the target object recommended by the service platform to the first user is determined from the multiple objects. It is understood that the first and second predicted indicator values ​​can reflect user preferences, thereby determining the target object that better matches the user's preferences.

[0115] In one example, determining the target object that the service platform recommends to the first user from the plurality of objects based on the first prediction index value and the second prediction index value corresponding to each object among the plurality of objects includes:

[0116] A first score is determined based on the first predictive index value of any object;

[0117] The second score is determined based on the object's second predictive index value;

[0118] Based on pre-set weights, the first score and the second score are weighted and summed to obtain the comprehensive score of the object;

[0119] Based on the comprehensive score of each object among multiple objects, the target object recommended by the service platform to the first user is determined from the multiple objects.

[0120] Furthermore, the second indicator includes:

[0121] The probability of a user logging back into the service platform after clicking on a candidate item, and the time interval between a user logging back in and clicking on a candidate item;

[0122] The determination of the second score includes:

[0123] The second score is obtained based on the positive correlation between the probability value and the time interval value in the second predictive index.

[0124] Understandably, the greater the probability that a user will log back into the service platform after clicking on a candidate, and the shorter the time interval between the user logging back in and clicking on the candidate, the more the candidate matches the user's preferences, and the easier it is for the candidate to be identified as the target.

[0125] The method for determining the target object recommended by a service platform to a user, as provided in the embodiments of this specification, firstly obtains the user profile features of a first user, the sequence features of the interaction sequence formed by the objects that the first user has historically interacted with on the service platform, and the candidate object features of candidate objects to be recommended to the first user; the candidate object is any one of the plurality of objects; then, the user profile features, the sequence features, and the candidate object features are input into a pre-trained neural network model for model processing, and the neural network model outputs a first predicted index value for at least one first indicator related to the first user's current interaction with the candidate object, and a second predicted index value for at least one second indicator related to the first user's re-login to the service platform; finally, based on the first predicted index value and the second predicted index value corresponding to each of the plurality of objects, the target object recommended by the service platform to the first user is determined from the plurality of objects. As can be seen from the above, the embodiments of this specification, by predicting multiple index values ​​through the neural network model, wherein the first predicted index value is a prediction for short-term user feedback, and the second predicted index value is a prediction for long-term user feedback, can comprehensively determine the target object by combining short-term and long-term user feedback, making the determined target object more in line with user preferences.

[0126] Figure 3 A schematic diagram of a neural network model according to one embodiment is shown. (Refer to...) Figure 3The neural network model comprises multiple branch networks and a sequence representation module shared by each branch network. Each branch network is used to predict the value of an indicator, achieving multi-objective prediction. There are four prediction objectives: the click-through rate (CTR) of the currently recommended candidate, the probability (CRR) of a user re-logging into the service platform after clicking on a candidate, the time interval between the user's re-login and clicking on a candidate (RTI), and the object clicked by the user upon re-login (RV). The neural network model makes further predictions based on the prediction results of RV to obtain the prediction results of RTI.

[0127] Figure 4 A schematic diagram of the structure of a sequence characterization module according to one embodiment is shown. (Refer to...) Figure 4 The sequence representation module encodes the features of each interactive object and the features of the candidate objects using an attention-based mechanism to obtain a sequence representation vector. The interactive object features include object identifiers and the time interval between historical interaction actions and the current time; the candidate object features include object identifiers of the candidate objects. It is understood that the time interval corresponding to the candidate object features is 0. In the attention-based encoding, a method of processing according to object category is adopted. That is, for the subsequences composed of objects belonging to the same object category, a self-attention mechanism is used for representation processing to obtain the sequence representation sub-vector corresponding to that object category. Furthermore, a self-attention mechanism is used for representation processing of the entire interaction sequence composed of the interacted objects and the candidate objects to obtain a global sequence representation sub-vector. Finally, the sequence representation sub-vectors corresponding to each object category and the global sequence representation sub-vector are concatenated to obtain the sequence representation vector.

[0128] Figure 5 This diagram illustrates a training method for a neural network model according to one embodiment. The trained neural network model can be applied to... Figure 2 In the method shown, the service platform provides multiple interactive objects, each corresponding to a product or service offered to the user. This method can be based on... Figure 1 The implementation scenario is shown. For example... Figure 5As shown, the training method of the neural network model in this embodiment includes the following steps: Step 51, obtaining the user profile features of the sample user, the sequence features of the interaction sequence composed of the objects that the sample user has historically interacted with on the service platform, the first object features of the first object recommended to the sample user, the first label value of at least one first indicator of the sample user's current interaction with the first object, and the second label value of at least one second indicator related to the sample user's re-login to the service platform; Step 52, inputting the user profile features, the sequence features, and the first object features into the neural network model for model processing, and outputting the first predicted indicator value of the at least one first indicator and the second predicted indicator value of the at least one second indicator through the neural network model; Step 53, determining the first prediction loss based on the first label value and the first predicted indicator value of the sample user; determining the second prediction loss based on the second label value and the second predicted indicator value of the sample user; the first prediction loss and the second prediction loss constitute the total loss; Step 54, adjusting the model parameters of the neural network model with minimizing the total loss as the training objective. The specific execution methods of the above steps are described below.

[0129] First, in step 51, the user profile features of the sample user, the sequence features of the interaction sequence consisting of all objects that the sample user has historically interacted with on the service platform, the first object features of the first object recommended to the sample user, the first tag value of at least one first indicator of the sample user's current interaction with the first object, and the second tag value of at least one second indicator related to the sample user's re-login to the service platform are obtained. It is understood that the aforementioned first object is an object that has been historically recommended to the sample user, and the sample user has already provided feedback on this object, such as whether they clicked on the object or whether they re-logged in.

[0130] In the embodiments described in this specification, each object in the interaction sequence has a temporal order with the first object mentioned above.

[0131] Then, in step 52, the user profile features, the sequence features, and the first object features are input into a neural network model for model processing. The neural network model outputs a first predicted index value for the at least one first indicator and a second predicted index value for the at least one second indicator. It is understood that, since the accuracy of the neural network model's predictions needs to be improved through training, the first predicted index value is usually different from the first label value, and the second predicted index value is different from the second label value.

[0132] In one example, the at least one second indicator includes:

[0133] The time interval between the user logging back in and clicking the first object; the object clicked by the user when logging back in;

[0134] The neural network model predicts the time interval between a user's re-login and the first object clicked, based on the prediction results of the object clicked when the user re-logs in.

[0135] Next, in step 53, a first prediction loss is determined based on the first label value and the first prediction index value of the sample user; a second prediction loss is determined based on the second label value and the second prediction index value of the sample user; the first prediction loss and the second prediction loss constitute the total loss. It is understandable that the neural network model needs to achieve multi-objective prediction, therefore the prediction losses of various indicators must be combined to determine the total loss.

[0136] For example, there are four objectives for modeling and optimization: CTR, CRR, RTI, and RV. The total loss function used to determine the total loss is:

[0137] Total Loss=LossCTR+LossCRR+LossRTI+LossRV.

[0138] Finally, in step 54, the model parameters of the neural network model are adjusted with minimizing the total loss as the training objective. It is understood that the neural network model includes multiple branch networks, and adjusting the model parameters of the neural network model includes adjusting the model parameters of each branch network.

[0139] The training method for the neural network model provided in the embodiments of this specification first obtains the user profile features of the sample user, the sequence features of the interaction sequence consisting of all objects that the sample user has historically interacted with on the service platform, the first object features of the first object recommended to the sample user, the first label value of at least one first indicator of the sample user's current interaction with the first object, and the second label value of at least one second indicator related to the sample user's re-login to the service platform; then, the user profile features, the sequence features, and the first object features are input into the neural network model for model processing, and the neural network model outputs the first predicted indicator value of the at least one first indicator and the second predicted indicator value of the at least one second indicator; next, a first prediction loss is determined based on the first label value and the first predicted indicator value of the sample user; a second prediction loss is determined based on the second label value and the second predicted indicator value of the sample user; the first prediction loss and the second prediction loss constitute the total loss; finally, the model parameters of the neural network model are adjusted with minimizing the total loss as the training objective. As can be seen from the above, in the embodiments of this specification, when training the neural network model, multiple indicator values ​​are predicted through the neural network model. The first predicted indicator value is a prediction of short-term user feedback, and the second predicted indicator value is a prediction of long-term user feedback. The total loss is determined based on the prediction loss of the above two predictions. The model parameters of the neural network model are adjusted with minimizing the total loss as the training objective. This allows the trained neural network model to comprehensively determine the target object by combining short-term and long-term user feedback, making the determined target object more in line with user preferences.

[0140] According to another embodiment, an apparatus is also provided for determining target objects recommended by a service platform to a user, the service platform providing a plurality of interactive objects, each object corresponding to a product or service provided to the user, the apparatus being used to perform the method provided in the embodiments of this specification for determining target objects recommended by a service platform to a user. Figure 6 A schematic block diagram of an apparatus for determining target objects recommended to a user by a service platform, according to one embodiment, is shown. Figure 6 As shown, the device 600 includes:

[0141] The acquisition unit 61 is used to acquire the user profile features of the first user, the sequence features of the interaction sequence consisting of the objects that the first user has historically interacted with on the service platform, and the candidate object features of the candidate objects recommended to the first user; the candidate object is any one of the plurality of objects.

[0142] Prediction unit 62 is used to input the user profile features, sequence features, and candidate object features acquired by acquisition unit 61 into a pre-trained neural network model for model processing, and output a first prediction index value of at least one first indicator related to the first user's current interaction with the candidate object, and a second prediction index value of at least one second indicator related to the first user's re-login to the service platform through the neural network model.

[0143] The determining unit 63 is used to determine the target object recommended by the service platform to the first user from the multiple objects based on the first prediction index value and the second prediction index value obtained by the prediction unit 62 corresponding to each of the multiple objects.

[0144] Optionally, as an example, the interaction includes one or more of the following: browsing, adding to favorites, adding to cart, purchasing, and paying.

[0145] Optionally, as an embodiment, the sequence features include the interaction object features corresponding to each of the interacted objects, the interaction object features including object identifiers and the time interval between the historical interaction behavior and the current time; the candidate object features include the object identifiers of the candidate objects.

[0146] Furthermore, the time interval is encoded as a time embedding vector, which is formed by superimposing time information into components for multiple time periods and then concatenating these components.

[0147] Furthermore, the neural network model includes multiple branch networks and a sequence representation module shared by each branch network. The prediction unit 62 is specifically used to encode the features of each interaction object and the features of the candidate object using an attention mechanism through the sequence representation module to obtain a sequence representation vector; and to predict the first prediction index value and the second prediction index value based on the sequence representation vector through the multiple branch networks.

[0148] Furthermore, the prediction unit 62 includes:

[0149] The attention subunit is used to perform representation processing on the features of the interactive objects and the candidate objects that belong to the same object category using an attention mechanism, so as to obtain the sequence representation subvector corresponding to the object category.

[0150] The splicing subunit is used to splice the sequence representation subvectors corresponding to each object category obtained by the attention subunit to obtain the sequence representation vector.

[0151] Furthermore, the prediction unit 62 includes:

[0152] The first attention subunit is used to perform representation processing on the features of the interactive objects and the features of the candidate objects that belong to the same object category using an attention mechanism, so as to obtain the sequence representation subvector corresponding to the object category.

[0153] The second attention subunit is used to perform representation processing on the interaction object features and the candidate object features corresponding to each interacted object using an attention mechanism, so as to obtain a global sequence representation subvector.

[0154] The splicing subunit is used to splice the sequence representation subvectors corresponding to each object category obtained by the first attention subunit and the global sequence representation subvector obtained by the second attention subunit to obtain the sequence representation vector.

[0155] Optionally, as an embodiment, the first indicator includes:

[0156] Click-through rate.

[0157] Optionally, as an embodiment, the second indicator includes:

[0158] The probability of a user logging back into the service platform after clicking on a candidate object, the time interval between a user logging back in and clicking on a candidate object, or the object clicked by the user upon logging back in.

[0159] Optionally, as an embodiment, the at least one second indicator includes:

[0160] The time interval between a user logging back in and clicking on a candidate object; the object clicked by the user upon logging back in.

[0161] The neural network model predicts the time interval between a user's re-login and the click on a candidate object based on the predicted object clicked when the user re-logs in.

[0162] Optionally, as an embodiment, the determining unit 63 includes:

[0163] The first determining subunit is used to determine a first score based on a first predictive index value for any object;

[0164] The second determining subunit is used to determine the second score based on the second predictive index value of the object;

[0165] The comprehensive processing subunit is used to calculate the comprehensive score of the object by weighted summing of the first score obtained by the first determining subunit and the second score obtained by the second determining subunit, according to the preset weights.

[0166] The selection subunit is used to determine the target object that the service platform recommends to the first user from among the multiple objects obtained by the comprehensive processing subunit, based on the comprehensive score corresponding to each object.

[0167] Furthermore, the second indicator includes:

[0168] The probability of a user logging back into the service platform after clicking on a candidate item, and the time interval between a user logging back in and clicking on a candidate item;

[0169] The second determining subunit is specifically used to obtain the second score based on the positive correlation between the probability value and the time interval value in the second prediction index value.

[0170] The apparatus for determining the target object recommended by a service platform to a user, as provided in the embodiments of this specification, firstly, acquires user profile features of a first user, sequence features of an interaction sequence consisting of objects that the first user has historically interacted with on the service platform, and candidate object features of candidate objects to be recommended to the first user; the candidate object is any one of the plurality of objects. Then, prediction unit 62 inputs the user profile features, the sequence features, and the candidate object features into a pre-trained neural network model for model processing, and outputs a first predicted index value for at least one first indicator related to the first user's current interaction with the candidate object, and a second predicted index value for at least one second indicator related to the first user's re-login to the service platform. Finally, determination unit 63 determines the target object recommended by the service platform to the first user from the plurality of objects based on the first predicted index value and the second predicted index value corresponding to each object among the plurality of objects. As can be seen from the above, in the embodiments of this specification, the neural network model is used to predict multiple indicator values. The first predicted indicator value is a prediction of short-term user feedback, and the second predicted indicator value is a prediction of long-term user feedback. This allows for the comprehensive analysis of both short-term and long-term user feedback to determine the target object, making the determined target object more in line with user preferences.

[0171] According to another embodiment, a training apparatus for a neural network model is also provided, which is used to perform the training method for the neural network model provided in the embodiments of this specification. Figure 7 A schematic block diagram of a training apparatus for a neural network model according to one embodiment is shown. Figure 7 As shown, the device 700 includes:

[0172] The acquisition unit 71 is used to acquire the user profile features of the sample user, the sequence features of the interaction sequence composed of the objects that the sample user has historically interacted with on the service platform, the first object features of the first object recommended to the sample user, the first tag value of at least one first indicator of the sample user's current interaction with the first object, and the second tag value of at least one second indicator related to the sample user's re-login to the service platform.

[0173] The prediction unit 72 is used to input the user profile features, the sequence features, and the first object features acquired by the acquisition unit 71 into a neural network model for model processing, and output the first predicted index value of the at least one first index and the second predicted index value of the at least one second index through the neural network model.

[0174] The determining unit 73 is configured to determine a first prediction loss based on the first label value of the sample user obtained by the obtaining unit 71 and the first prediction index value obtained by the prediction unit 72; and to determine a second prediction loss based on the second label value of the sample user obtained by the obtaining unit 71 and the second prediction index value obtained by the prediction unit 72; the first prediction loss and the second prediction loss constitute the total loss.

[0175] The training unit 74 is used to adjust the model parameters of the neural network model with the goal of minimizing the total loss obtained by the determining unit 73.

[0176] Optionally, as an embodiment, the at least one second indicator includes:

[0177] The time interval between the user logging back in and clicking the first object; the object clicked by the user when logging back in;

[0178] The neural network model predicts the time interval between a user's re-login and the first object clicked, based on the prediction results of the object clicked when the user re-logs in.

[0179] The training apparatus for the neural network model provided in the embodiments of this specification first acquires user profile features of the sample user, sequence features of the interaction sequence consisting of objects that the sample user has historically interacted with on the service platform, first object features of the first object recommended to the sample user, first label values ​​of at least one first indicator of the sample user's current interaction with the first object, and second label values ​​of at least one second indicator related to the sample user's re-login to the service platform. Then, the prediction unit 72 inputs the user profile features, the sequence features, and the first object features into the neural network model for model processing, and outputs first predicted indicator values ​​of the at least one first indicator and second predicted indicator values ​​of the at least one second indicator through the neural network model. Next, the determination unit 73 determines a first prediction loss based on the first label value and the first predicted indicator value of the sample user; and determines a second prediction loss based on the second label value and the second predicted indicator value of the sample user. The first prediction loss and the second prediction loss constitute the total loss. Finally, the training unit 74 adjusts the model parameters of the neural network model with minimizing the total loss as the training objective. As can be seen from the above, in the embodiments of this specification, when training the neural network model, multiple indicator values ​​are predicted through the neural network model. The first predicted indicator value is a prediction of short-term user feedback, and the second predicted indicator value is a prediction of long-term user feedback. The total loss is determined based on the prediction loss of the above two predictions. The model parameters of the neural network model are adjusted with minimizing the total loss as the training objective. This allows the trained neural network model to comprehensively determine the target object by combining short-term and long-term user feedback, making the determined target object more in line with user preferences.

[0180] According to another embodiment, a computer-readable storage medium is also provided, on which a computer program is stored, which, when executed in a computer, causes the computer to perform a combination Figure 2 or Figure 5 The method described.

[0181] According to another embodiment, a computing device is also provided, including a memory and a processor, wherein the memory stores executable code, and when the processor executes the executable code, it implements a combination... Figure 2 or Figure 5 The method described.

[0182] Those skilled in the art will recognize that, in one or more of the examples above, the functions described in this invention can be implemented using hardware, software, firmware, or any combination thereof. When implemented in software, these functions can be stored in a computer-readable medium or transmitted as one or more instructions or code on a computer-readable medium.

[0183] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made on the basis of the technical solution of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for determining target objects recommended by a service platform to a user, the service platform providing multiple interactive objects, each object corresponding to a product or service offered to the user, the method comprising: The system acquires the user profile features of the first user, the sequence features of the interaction sequence consisting of all objects that the first user has historically interacted with on the service platform, and the candidate object features of the candidate objects recommended to the first user; the candidate object is any one of the multiple objects; the sequence features include the interaction object features corresponding to each of the interacted objects, and the interaction object features include the time interval between the historical interaction behavior and the current time. The user profile features, the sequence features, and the candidate object features are input into a pre-trained neural network model for model processing. The neural network model outputs a first predicted index value for at least one first indicator related to the first user's current interaction with the candidate object, and a second predicted index value for at least one second indicator related to the first user's re-login to the service platform. The at least one second indicator includes the time interval between the user's re-login and clicking on the candidate object, and the object clicked by the user upon re-login. The prediction of the time interval between the user's re-login and clicking on the candidate object depends on the prediction result of the object clicked by the user upon re-login. Based on the first prediction index value and the second prediction index value corresponding to each of the multiple objects, the target object recommended by the service platform to the first user is determined from the multiple objects.

2. The method as described in claim 1, wherein, The interaction includes one or more of the following: browsing, adding to favorites, adding to cart, purchasing, and payment.

3. The method as described in claim 1, wherein, The interactive object features include an object identifier; the alternative object features include the object identifier of the alternative object.

4. The method of claim 3, wherein, The time interval is encoded as a time embedding vector, which is formed by splitting time information into components for multiple time periods and then splicing the components together.

5. The method of claim 3, wherein, The neural network model includes multiple branch networks and a sequence representation module shared by each branch network. The model processing includes: using the sequence representation module to encode the features of each interaction object and the features of the candidate object using an attention mechanism to obtain a sequence representation vector; and using the multiple branch networks to predict the first prediction index value and the second prediction index value based on the sequence representation vector.

6. The method of claim 5, wherein, The attention-based encoding includes: An attention mechanism is used to represent the features of the interactive objects and the candidate objects that belong to the same object category, and to obtain the sequence representation sub-vector corresponding to the object category. The sequence representation vector is obtained by concatenating the sequence representation subvectors corresponding to each object category.

7. The method of claim 5, wherein, The attention-based encoding includes: An attention mechanism is used to represent the features of the interactive objects and the candidate objects that belong to the same object category, and to obtain the sequence representation sub-vector corresponding to the object category. An attention mechanism is used to represent the features of each interacted object and the features of the candidate objects, respectively, to obtain a global sequence representation sub-vector. The sequence representation vector is obtained by concatenating the sequence representation subvectors corresponding to each object category and the global sequence representation subvector.

8. The method of claim 1, wherein, The first indicator includes: Click-through rate.

9. The method of claim 1, wherein, The second indicator also includes: The probability that a user will log back into the service platform after clicking on an alternative.

10. The method of claim 1, wherein, The neural network model predicts the time interval between a user's re-login and the click on a candidate object based on the predicted object clicked when the user re-logs in.

11. The method of claim 1, wherein, The step of determining the target object recommended by the service platform to the first user from the plurality of objects based on the first prediction index value and the second prediction index value corresponding to each object among the plurality of objects includes: A first score is determined based on the first predictive index value of any object; The second score is determined based on the object's second predictive index value; Based on pre-set weights, the first score and the second score are weighted and summed to obtain the comprehensive score of the object; Based on the comprehensive score of each object among multiple objects, the target object recommended by the service platform to the first user is determined from the multiple objects.

12. The method of claim 11, wherein, The second indicator also includes: The probability that a user will log back into the service platform after clicking on a candidate; The determination of the second score includes: The second score is obtained based on the positive correlation between the probability value and the time interval value in the second predictive index.

13. A method for training a neural network model, the method comprising: The system acquires user profile features of the sample user, sequence features of the interaction sequence formed by the sample user's historical interactions with various objects on the service platform, first object features of the first object recommended to the sample user, first tag values ​​of at least one first indicator of the sample user's current interaction with the first object, and second tag values ​​of at least one second indicator related to the sample user's re-login to the service platform; the sequence features include interaction object features corresponding to each of the interacted objects, and the interaction object features include the time interval between the historical interaction behavior and the current time. The user profile features, the sequence features, and the first object features are input into a neural network model for model processing. The neural network model outputs a first predicted index value for at least one first indicator and a second predicted index value for at least one second indicator. The at least one second indicator includes the time interval between the user's re-login and the click on the first object, and the object clicked by the user when re-login. The prediction of the time interval between the user's re-login and the click on the first object depends on the prediction result of the object clicked by the user when re-login. A first prediction loss is determined based on the first label value and the first prediction index value of the sample users; a second prediction loss is determined based on the second label value and the second prediction index value of the sample users; the first prediction loss and the second prediction loss constitute the total loss. The model parameters of the neural network model are adjusted with the goal of minimizing the total loss.

14. The method of claim 13, wherein, The neural network model predicts the time interval between a user's re-login and the first object clicked, based on the prediction results of the object clicked when the user re-logs in.

15. An apparatus for determining target objects recommended by a service platform to a user, the service platform providing a plurality of interactive objects, each object corresponding to a product or service offered to the user, the apparatus comprising: The acquisition unit is used to acquire user profile features of a first user, sequence features of an interaction sequence consisting of objects that the first user has historically interacted with on the service platform, and candidate object features of candidate objects recommended to the first user; the candidate object is any one of the plurality of objects; the sequence features include interaction object features corresponding to each of the interacted objects, and the interaction object features include the time interval between the historical interaction behavior and the current time. The prediction unit is used to input the user profile features, sequence features, and candidate object features acquired by the acquisition unit into a pre-trained neural network model for model processing. The neural network model outputs a first predicted index value for at least one first indicator related to the first user's current interaction with the candidate object, and a second predicted index value for at least one second indicator related to the first user's re-login to the service platform. The at least one second indicator includes the time interval between the user's re-login and clicking on the candidate object, and the object clicked by the user upon re-login. The prediction of the time interval between the user's re-login and clicking on the candidate object depends on the prediction result of the object clicked by the user upon re-login. The determining unit is used to determine the target object recommended by the service platform to the first user from the plurality of objects based on the first prediction index value and the second prediction index value obtained by the prediction unit corresponding to each of the plurality of objects.

16. A training apparatus for a neural network model, the apparatus comprising: The acquisition unit is used to acquire user profile features of the sample user, sequence features of the interaction sequence consisting of all objects that the sample user has historically interacted with on the service platform, first object features of the first object recommended to the sample user, first tag values ​​of at least one first indicator of the sample user's current interaction with the first object, and second tag values ​​of at least one second indicator related to the sample user's re-login to the service platform; the sequence features include interaction object features corresponding to each of the interacted objects, and the interaction object features include the time interval between the historical interaction behavior and the current time. The prediction unit is used to input the user profile features, sequence features, and first object features acquired by the acquisition unit into a neural network model for model processing, and output a first predicted index value of the at least one first indicator and a second predicted index value of the at least one second indicator through the neural network model; the at least one second indicator includes the time interval between the user's re-login and the click on the first object, and the object clicked by the user when re-login; the prediction of the time interval between the user's re-login and the click on the first object depends on the prediction result of the object clicked by the user when re-login; The determining unit is configured to determine a first prediction loss based on the first label value of the sample user obtained by the acquiring unit and the first prediction index value obtained by the prediction unit; and to determine a second prediction loss based on the second label value of the sample user obtained by the acquiring unit and the second prediction index value obtained by the prediction unit; the first prediction loss and the second prediction loss constitute the total loss. A training unit is used to adjust the model parameters of the neural network model with the training objective of minimizing the total loss obtained by the determining unit.

17. A computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to perform the method of any one of claims 1-14.

18. A computing device comprising a memory and a processor, wherein the memory stores executable code, and the processor, when executing the executable code, implements the method of any one of claims 1-14.

Images