Object push method, medium, computer device, and computer program product

Abstract

The present disclosure provides an object push method, a medium, a computer device, and a computer program product. The method comprises: acquiring a target feature set corresponding to a current operation object of a target user, wherein the target feature set comprises a first feature of the current operation object, a second feature of at least one historical operation object of which the feature similarity to the current operation object is greater than a preset similarity threshold, and a third feature of at least one other object having a co-occurrence relationship with the current operation object; processing the target feature set by a push model to obtain an intent parameter of the target user to the current operation object, wherein the intent parameter is used for representing the correlation between an operation intent of the target user and the current operation object; and on the basis of the intent parameter, pushing to the target user at least one object to be pushed.

Description

Object push method, medium, computer device and computer program product

[0001] This disclosure claims priority to the Chinese patent application filed with the China Patent Office on January 25, 2024, with application number 202410112451.8 and application name “Object Push Method, Medium, Computer Device and Computer Program Product”, the entire contents of which are incorporated by reference into this disclosure. Technical Field

[0002] The present disclosure relates to the field of computer technology, and in particular to an object pushing method, medium, computer device, and computer program product. Background Art

[0003] In a trigger-induced recommendation (TIR) ​​scenario, the push platform pushes objects based on the user's triggering behavior (such as browsing, searching, purchasing, etc.). The object triggered by the user's triggering behavior is the triggering item. Current push methods typically push objects with similar attributes based on the attributes of the triggering item, but lack understanding of user intent, making it difficult to accurately push objects based on user intent. Summary of the Invention

[0004] In a first aspect, an embodiment of the present disclosure provides an object push method, the method comprising: obtaining a target feature set corresponding to a current operation object of a target user; the target feature set comprises a first feature of the current operation object, a second feature of at least one historical operation object whose feature similarity with the current operation object is greater than a preset similarity threshold, and a third feature of at least one other object that has a co-occurrence relationship with the current operation object; processing the target feature set through a push model to obtain an intention parameter of the target user for the current operation object; the intention parameter is used to characterize the correlation between the operation intention of the target user and the current operation object; and pushing at least one object to be pushed to the target user based on the intention parameter.

[0005] In some embodiments, the larger the intention parameter is, the stronger the correlation between the object to be pushed and the current operation object is; the smaller the intention parameter is, the stronger the correlation between the object to be pushed and the historical operation sequence of the target user is.

[0006] In some embodiments, the target feature set is processed by the push model to obtain the target user's intention parameters for the current operation object, including: obtaining the expectation and variance determined by the push model based on the target feature set, the expectation is used to represent the intention strength of the target user's operation intention for the current operation object, and the variance is used to represent the uncertainty of the target user's operation intention for the current operation object; obtaining the target user's intention parameters for the current operation object based on the expectation and the variance.

[0007] In some embodiments, obtaining the target user's intention parameters for the current operation object based on the expectation and the variance includes: generating a random distribution based on the expectation and the variance; sampling the random distribution to obtain a sampling feature; fusing the first feature, the user feature of the target user, and the sampling feature to obtain a first fusion feature; and obtaining the target user's intention parameters for the current operation object based on the first fusion feature.

[0008] In some embodiments, obtaining the expectation and variance determined by the push model based on the target feature set includes: processing the first feature and the second feature through a first multi-head attention layer to obtain a first output vector; processing the first feature and the third feature through a second multi-head attention layer to obtain a second output vector; fusing the first output vector and the second output vector to obtain a second fused feature; and determining the expectation and the variance based on the second fused feature.

[0009] In some embodiments, the expectation and the variance are obtained through a contrastive learning model; the contrastive learning model is trained based on at least one pair of positive samples and at least one pair of negative samples; wherein each pair of positive samples includes a target feature set corresponding to the same current operation object and an augmented feature set corresponding to the current operation object; the features in the augmented feature set are obtained by performing feature transformation on the features in the target feature set; each pair of negative samples includes a target feature set corresponding to a different current operation object.

[0010] In some embodiments, the features in the augmented feature set are obtained by randomly deleting the second features of several historical operation objects and / or the third features of several other objects in the target feature set; randomly deleting several feature dimensions of the second features and / or third features in the target feature set.

[0011] In some embodiments, pushing at least one object to be pushed to the target user based on the intention parameter includes: pushing the at least one object to be pushed to the target user based on the push score of the at least one object to be pushed; wherein the push score of each object to be pushed in the at least one object to be pushed is obtained based on the following method: determining a first weight corresponding to a first feature and a second weight corresponding to a fourth feature of the object to be pushed based on the intention parameter; the first weight is used to characterize the probability that the operation intention of the target user is related to the current operation object, and the second weight is used to characterize the probability that the operation intention of the target user is unrelated to the current operation object; weighting the first feature based on the first weight, and weighting the fourth feature based on the second weight; obtaining the push score of the object to be pushed based on the weighted first feature and the weighted fourth feature.

[0012] In some embodiments, the weighted processing of the first feature based on the first weight and the weighted processing of the fourth feature based on the second weight include: obtaining the historical operation sequence of the target user; inputting the first feature and the historical operation sequence into the fourth multi-head attention layer for processing; inputting the fourth feature and the historical operation sequence into the fifth multi-head attention layer for processing; weighting the processed first feature based on the first weight, and weighting the processed fourth feature based on the second weight.

[0013] In some embodiments, obtaining the push score of the object to be pushed based on the weighted feature includes: fusing the first feature and the fourth feature of the object to be pushed to obtain a third fused feature; and obtaining the push score of the object to be pushed based on the third fused feature, the weighted first feature, and the weighted fourth feature.

[0014] In some embodiments, the fusing the first feature and the fourth feature to obtain a third fused feature includes: obtaining the Hadamard product of the first feature and the fourth feature; and fusing the first feature, the fourth feature and the Hadamard product to obtain the third fused feature.

[0015] In some embodiments, the obtaining of the push score of the object to be pushed based on the third fusion feature, the weighted first feature and the weighted fourth feature includes: obtaining the push score of the object to be pushed based on the third fusion feature, the fourth fusion feature, the weighted first feature and the weighted fourth feature; wherein, the fourth fusion feature is obtained based on the following manner: obtaining the expectation and variance determined by the push model based on the target feature set, the expectation is used to represent the intention intensity of the target user's operation intention for the current operation object, and the variance is used to represent the uncertainty of the target user's operation intention for the current operation object; generating a random distribution based on the expectation and the variance; sampling the random distribution to obtain a sampling feature; fusing the fourth feature and the sampling feature to obtain the fourth fusion feature.

[0016] In a second aspect, an embodiment of the present disclosure provides a computer-readable storage medium having computer instructions stored thereon, which, when executed by a processor, implements the method described in any embodiment of the present disclosure.

[0017] In a third aspect, an embodiment of the present disclosure provides a computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the method described in any embodiment of the present disclosure when executing the program.

[0018] In a fourth aspect, an embodiment of the present disclosure provides a computer program product, including a computer program, which implements the method described in any embodiment of the present disclosure when executed by a processor.

[0019] The disclosed embodiment adopts the second feature and the third feature in the input data of the push model, wherein the second feature is the feature of the historical operation object whose feature similarity with the current operation object is greater than the preset similarity threshold. Since the user intention usually has the characteristic of being cross-session, the historical operation object that is relatively similar to the current operation object may be related to the user's current immediate intention. Therefore, by obtaining the second feature of the historical operation object whose feature similarity with the current operation object is greater than the preset similarity threshold, the user's immediate intention can be reflected to a certain extent; the third feature is the feature of other objects that have a co-occurrence relationship with the current operation object. By obtaining the third feature, the co-occurrence relationship of the user behavior can be reflected, thereby effectively mining the user's potential intention. In summary, the above method can effectively mine the user's immediate intention and potential intention, thereby accurately pushing objects according to the user's intention. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] FIG1 is a schematic diagram of an application scenario of an embodiment of the present disclosure;

[0021] FIG2 is a flow chart of an object push method according to an embodiment of the present disclosure;

[0022] 3 and 4 are schematic structural diagrams of push models according to embodiments of the present disclosure;

[0023] FIG5 is a block diagram of an object pushing device according to an embodiment of the present disclosure;

[0024] FIG6 is a schematic diagram of a computer device according to an embodiment of the present disclosure. DETAILED DESCRIPTION

[0025] Exemplary embodiments will be described in detail herein, with examples illustrated in the accompanying drawings. When the following description refers to the drawings, identical numerals in different figures represent identical or similar elements unless otherwise indicated. The embodiments described in the following exemplary embodiments are not intended to represent all possible implementations consistent with one or more embodiments of the present disclosure. Rather, they are merely examples of apparatuses and methods consistent with certain aspects of one or more embodiments of the present disclosure, as detailed in the appended claims.

[0026] It should be noted that, in other embodiments, the steps of the corresponding method are not necessarily performed in the order shown and described in this disclosure. In some other embodiments, the method may include more or fewer steps than those described in this disclosure. In addition, a single step described in this disclosure may be broken down into multiple steps for description in other embodiments; and multiple steps described in this disclosure may be combined into a single step for description in other embodiments.

[0027] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this disclosure are all information and data 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 regions, and provide corresponding operation entrances for users to choose to authorize or refuse.

[0028] In a push scenario based on trigger items, recommended items are usually guided by trigger items (Trigger). Trigger items refer to the objects triggered by the user's behavioral operations on the platform. Behavioral operations include but are not limited to clicks, searches, purchases, favorites, forwarding, and other operations. These behavioral operations can be regarded as the user's interest in a certain category, theme, or product, and can therefore serve as input signals for the recommendation system. For example, in the e-commerce scenario shown in Figure 1, the server 104 can push products to the user terminal 102. After the user clicks on a product (such as a skirt) displayed on the user terminal, the product clicked by the user can be used as a trigger item, and the server 104 can further push products to the user based on the trigger item.

[0029] It is understood that the above e-commerce scenario is merely illustrative and is not intended to limit the present disclosure. In addition to e-commerce scenarios, push scenarios can also include news push scenarios, music push scenarios, e-book push scenarios, and so on. Accordingly, the triggering product can be news, music, e-books, and so on. For ease of description, the following uses the e-commerce scenario as an example to illustrate the solutions of the embodiments of the present disclosure.

[0030] In related art, objects with similar attributes are typically pushed based on the properties of the trigger item. For example, if the trigger item is a skirt with certain characteristics, skirts with similar characteristics will be pushed to the user. However, this push method lacks understanding of user intent, resulting in low push accuracy. For example, in some scenarios, the user's intent is strongly correlated with the trigger item. In these scenarios, objects with similar characteristics to the trigger item (usually objects in the same category as the trigger item) should be pushed to the user. However, in other scenarios, the user's intent is sometimes to combine and match items, and the user's intent is not strongly correlated with the trigger item. For example, the trigger item is a skirt, but the user's intent is to find a top that matches the skirt; or the trigger item is a phone case, but the user's intent is to find other phone accessories such as phone screen protectors and phone pendants. In other examples, the user's intent may also be to find new products, hot-selling products, or similar products. In these scenarios, if objects with similar characteristics to the trigger item are still pushed to the user, the pushed objects may not match the user's intent, resulting in inaccurate push results.

[0031] Based on this, an embodiment of the present disclosure provides an object push method. Referring to FIG2 , the method includes:

[0032] Step S202: Obtain a target feature set corresponding to the current operation object of the target user; the target feature set includes a first feature of the current operation object, a second feature of at least one historical operation object whose feature similarity with the current operation object is greater than a preset similarity threshold, and a third feature of at least one other object that has a co-occurrence relationship with the current operation object;

[0033] Step S204: Processing the target feature set through the push model to obtain the target user's intention parameter for the current operation object; the intention parameter is used to characterize the correlation between the target user's operation intention and the current operation object;

[0034] Step S206: Push at least one object to be pushed to the target user based on the intention parameter.

[0035] In step S202, the target user can be a single user or multiple users. For example, in an e-commerce scenario, each user on the e-commerce platform can be a target user. However, in some cases, user data is sparse. Therefore, multiple users can be clustered based on user characteristics, and all users in the same category obtained through clustering can be collectively used as target users.

[0036] The target feature set includes the first feature of the current operation object, the second feature of the historical operation object, and the third feature of other objects. Among them, the current operation object is the trigger product, which can be the object targeted by the user's click, search, favorite, purchase, forwarding and other behaviors. The first feature of the current operation object is the feature of the trigger product itself. The historical operation object refers to the object targeted by the user's click, search, favorite, purchase, forwarding and other behaviors within a historical time period (for example, 1 day ago, the last week, etc.). Other objects refer to objects that have a co-occurrence relationship with the current operation object, wherein the current operation object and other objects have a co-occurrence relationship means that the current operation object and other objects usually appear together. For example, when a merchant sells the current operation object, it will sell other objects and the current operation object as a group of goods together, or when the user operates (such as searching or purchasing) the current operation object, it will operate other objects at the same time. The above-mentioned first feature, second feature and third feature may include but are not limited to the brand, price, color, size, function, purpose, material, quantity and other features of the corresponding object.

[0037] The second feature is a feature of a historical operation object whose feature similarity with the current operation object is greater than a preset similarity threshold, wherein the similarity threshold can be set according to actual needs, for example, set to 80%. User intentions usually have the characteristic of being cross-session, wherein a session refers to the process from a user logging in once to logging out. That is to say, if a user had an operation intention for an object when logging in historically, he may still have an operation intention for the object when the user is currently logged in. Therefore, a historical operation object that is relatively similar to the features of the current operation object may be a continuation of the user's historical intention in the current session, and is related to the user's current immediate intention. By obtaining the second feature, the user's immediate intention can be reflected to a certain extent. In some embodiments, the features of each historical operation object operated by the user can be stored. After obtaining the current operation object, the current operation object can be compared with the features of each pre-stored historical operation object for similarity. If the feature of a historical operation object whose similarity with the first feature of the current operation object is higher than the preset similarity threshold is obtained, the feature is used as the second feature.

[0038] The third feature is the feature of other objects that have a co-occurrence relationship with the current operation object. For example, when a user clicks on a mobile phone case, he or she will usually also click on mobile phone accessories such as mobile phone films and mobile phone pendants. By obtaining the third feature, the co-occurrence relationship of user behavior can be reflected, and this co-occurrence relationship can reflect the user's potential combination intention for objects of different categories. In some embodiments, the features of each object and other objects that have a co-occurrence relationship with the object can be stored in the form of key-value pairs, wherein the key of the key-value pair is each object, and the value of the key-value pair is the feature of other objects. After obtaining the current operation object, the key-value pair corresponding to the current operation object can be retrieved from the pre-established key-value pairs, and the value in the retrieved key-value pair is determined as the third feature. For example, the pre-established key-value pairs include:

[0039] Key-value pair 1: {key = object 1, value = feature 1};

[0040] Key-value pair 2: {key = object 2, value = feature 2};

[0041] Key-value pair 3: {key = object 3, value = feature 3}.

[0042] Assuming that the current operation object is object 2, a key-value pair 2 including object 2 can be retrieved, and the value in the key-value pair 2 (ie, feature 2) is determined as the third feature.

[0043] In step S204, the target feature set can be input into the push model for processing to obtain the target user's intention parameter for the current operation object. The intention parameter is used to characterize the correlation between the target user's operation intention and the current operation object. A larger intention parameter indicates a stronger correlation between the target user's operation intention and the current operation object, and thus a greater influence of the current operation object on the object to be pushed. A smaller intention parameter indicates a weaker correlation between the target user's operation intention and the current operation object, and thus a smaller influence of the current operation object on the object to be pushed, and a stronger correlation between the object to be pushed and the target user's historical operation sequence.

[0044] In the related art, user intention is often modeled as a static feature vector. The user intention obtained by modeling does not have strong generalization, which is contrary to the uncertainty of user intention. To this end, the embodiment of the present disclosure models user intention as a random variable. The expectation of the random variable is used to represent the intention strength of the target user's operation intention for the current operation object. The greater the expectation, the stronger the intention strength of the target user's operation intention for the current operation object. The smaller the expectation, the weaker the intention strength of the target user's operation intention for the current operation object. The variance of the random variable is used to represent the uncertainty of the target user's operation intention for the current operation object. The larger the variance, the greater the uncertainty of the target user's operation intention for the current operation object. The smaller the variance, the smaller the uncertainty of the target user's operation intention for the current operation object. By modeling user intention as a random variable, the generalization of user intention can be improved, and the modeled user intention has uncertainty, which is consistent with the characteristics of user intention.

[0045] 3 and 4 , the push model includes a first feature fusion layer (Interaction Layer), a second feature fusion layer (Fusing Layer), and a user instant intent layer (Instant Intent Layer). The specific structure and working principle of each layer will be explained below.

[0046] The user's immediate intention layer can process the first and second features through the first multi-head attention layer to obtain a first output vector, process the first and third features through the second multi-head attention layer to obtain a second output vector, fuse the first and second output vectors to obtain a second fused feature, and determine the expectation and variance of the random variable based on the second fused feature. By adopting a multi-head attention mechanism, it is possible to focus on the second feature related to the first feature and the third feature related to the first feature, reducing the impact of the second feature and the third feature unrelated to the first feature on the expectation and variance, thereby reducing sensitivity to noise.

[0047] Referring to Figure 4, the user's immediate intention layer may include a contrastive learning model, and the expectation and variance may be obtained using the contrastive learning model. In an embodiment in which a multi-head attention mechanism is used to obtain the expectation and variance, the contrastive learning model may include the above-mentioned first multi-head attention layer and second multi-head attention layer. By adopting the contrastive learning model, the problem of inaccurate intention estimation caused by sparse user behavior can be alleviated. Furthermore, the contrastive learning model may also include a third multi-head attention layer and a mapping layer (projector head). It should be noted that the mapping layer can be used only during the training process of the contrastive learning model and not during the inference phase of the contrastive learning model. After obtaining the second fused feature, the second fused feature can be input into the third multi-head attention layer for processing to further integrate the information of different attention heads and provide a richer feature representation. This helps the contrastive learning model better capture the relevance and importance of the input features, thereby improving the modeling capability of the contrastive learning model. The output features of the third multi-head attention layer can be input into the mapping layer, which is typically composed of a fully connected layer, a convolutional layer, or other types of network layers, and is used to learn to map the output features of the multi-head attention layer to the target representation space. The dimension of the target representation space is usually lower than the dimension of the input features of the mapping layer, so the feature dimension can be reduced and the features can be abstracted. The output features of the mapping layer can be output to a fully connected layer of several layers (for example, two layers) in cascade to obtain the expectation of the random variable. On the other hand, it can be output to the fully connected layer and the activation layer to obtain the variance of the random variable. Specifically, the expectation and variance can be recorded as: μ i =f μ (h i ),∑ i =ELU(f ∑ (h i ))+(1+ε)

[0048] Among them, μ i represents the expectation of the random variable corresponding to the i-th target user, f μ () represents the activation function used to obtain the desired fully connected layer, h i represents the input features of the fully connected layer, Σ i represents the variance of the random variable corresponding to the i-th target user, ELU() represents the ELU activation function, and f ∑ () represents the activation function of the fully connected layer used to obtain the variance, and ε is a positive number to ensure that the variance value is non-negative.

[0049] The contrastive learning model can be trained based on at least one pair of positive samples and at least one pair of negative samples. Each pair of positive samples includes a target feature set corresponding to the same current operation object and an augmented feature set corresponding to the current operation object, where the features in the augmented feature set are obtained by performing feature transformation on the features in the target feature set. Each pair of negative samples includes a target feature set corresponding to a different current operation object.

[0050] The features in the augmented feature set can be obtained by at least one of the following methods:

[0051] Method 1: Randomly delete the second features of several historical operation objects and / or the third features of several other objects in the target feature set. For example, assuming that the target feature set includes the second feature of object 1, the second feature of object 2, the second feature of object 3, the third feature of object 4, and the third feature of object 5, then the second feature of one object (assuming it is object 1) and the third feature of one object (assuming it is object 4) can be randomly deleted to obtain a feature set including the second feature of object 2, the second feature of object 3, and the third feature of object 5. This feature set is the augmented feature set.

[0052] Method 2: Randomly delete several feature dimensions of the second feature and / or third feature in the target feature set. Assuming that the target feature set includes the second feature of object 1, the second feature of object 2, the second feature of object 3, the third feature of object 4, and the third feature of object 5, and each of the above second features and each of the third features includes features of the four dimensions of price, size, color, and function, then one feature dimension (assuming it is size) can be randomly deleted to obtain the second and third features including the three feature dimensions of price, color, and function. The set of the second and third features including the above three feature dimensions is the augmented feature set.

[0053] It is understood that although two identical dotted boxes are shown in the figure (both including the first multi-head attention layer, the second multi-head attention layer, the third multi-head attention layer, and the mapping layer), this is to intuitively illustrate the principle of the contrastive learning model. The network structures in the two dotted boxes can share network parameters. In actual applications, the contrastive learning model can only include the network structure in one of the dotted boxes.

[0054] Through training, the model parameters of the contrastive learning model can be adjusted to make the distance between the random distributions corresponding to positive samples (i.e., similar operation intentions) closer, and the distance between the random distributions corresponding to negative samples (i.e., dissimilar operation intentions) farther apart. The distance between two operation intentions can be measured using the KL divergence, specifically expressed as:

[0055] Among them, N q and Np Represents two different random distributions, D KL [N q ||N p ] represents random distribution N q and random distribution N p KL divergence between q and ∑ q Represents random distribution N q The expectation and variance of μ p and ∑ p Represents random distribution N p The expectation and variance of , Tr represents the trace of the matrix, and -1 represents the matrix inversion operation.

[0056] After obtaining the expectation and variance of the random variable, the target user's intention parameters for the current operation object can be obtained based on the expectation and variance of the random variable. Specifically, a random distribution can be generated based on the expectation and variance, and the random distribution can be sampled to obtain a sampled feature. The first feature of the current operation object, the user feature of the target user, and the sampled feature can be fused to obtain a first fused feature. Based on the first fused feature, the target user's intention parameters for the current operation object can be obtained.

[0057] Among them, the random distribution can be a Gaussian distribution. Of course, other random distributions can also be used according to actual needs. The user characteristics of the target user may include but are not limited to at least one of the following: the hierarchical characteristics of the target user, the activity characteristics, the length characteristics of the objects of the historical behavior sequence of the target user that belong to the same leaf category as the current operation object (hereinafter referred to as the same type of object length characteristics), and the length characteristics of the objects of the historical behavior sequence of the target user that belong to different leaf categories under the same parent category as the current operation object (hereinafter referred to as the same parent category object length characteristics), and the source characteristics of the objects operated by the target user (hereinafter referred to as the object source characteristics). Various user characteristics are illustrated below.

[0058] Hierarchical features

[0059] Users can be divided into different user layers based on their historical purchase volume. For example, if the user's historical purchase volume is less than or equal to the first threshold, the user is divided into the L1 user layer; if the user's historical purchase volume is greater than the first threshold and less than or equal to the second purchase volume threshold (the second purchase volume threshold is greater than the first purchase volume threshold), the user is divided into the L2 user layer; if the user's historical purchase volume is greater than the second purchase volume threshold, the user is divided into the L3 user layer. The hierarchical characteristics of the target user are the user layer to which the target user belongs. It will be understood that the hierarchical method and the number of user layers here are only exemplary and are not intended to limit the present disclosure.

[0060] Activity characteristics

[0061] The user's activity level can be obtained based on the number of historical operations. Historical operations may include clicks, browsing product details, and / or other operations. The greater the number of historical operations, the more active the user was during the corresponding historical time period. Conversely, the fewer historical operations, the less active the user was during the corresponding historical time period.

[0062] Length characteristics of similar objects

[0063] Assuming the leaf category of the current action object is Category A, the number of objects in the target user's historical action sequence that also belong to Category A is the same-object length feature. For example, if Category A is mobile phone cases, and the target user's historical action sequence includes clicks on three different models of mobile phone cases, the same-object length feature is 3.

[0064] Length characteristics of objects of the same parent class

[0065] Assume that the leaf category to which the current operation object belongs is Category A, and Category A's parent category is Category B. In addition to Category A, the leaf categories under Category B also include Category C and Category D. Then, among the objects targeted by the target user's historical behavior sequence, the total number of objects belonging to Category C and Category D is the same-parent-category object length feature. For example, suppose Category A is mobile phone cases, Category B is mobile phone accessories, Category C is mobile phone gadgets, and Category D is mobile phone screen protectors. The target user's historical behavior sequence includes click operations on two different models of mobile phone gadgets and click operations on three different models of mobile phone screen protectors. In this case, the same-parent-category object length feature is 5.

[0066] Object source characteristics

[0067] The object source feature is used to indicate the exposure channel through which the object operated by the target user is exposed to the target object. The above-mentioned exposure channels include but are not limited to search channels and recommendation channels. Among them, the object exposed to the target user through the search channel is the object obtained by the target user through the search. When the target user searches for an object, the target user's operation intention usually has a strong correlation with the object. The object exposed to the target user through the recommendation channel is the object pushed to the target user by the e-commerce platform based on user preferences, user historical operation sequences, trigger products and other information. The correlation between the object exposed to the target user through the recommendation channel and the target user's operation intention is usually lower than the object exposed to the target user through the search channel.

[0068] When fusing the first feature, the user feature of the target user, and the sampling feature, the first feature, the user feature of the target user, and the sampling feature can be concatenated and flattened to obtain a first fused feature, which is then input into a first multilayer perceptron (MLP) for processing. After the output feature of the first multilayer perceptron is nonlinearly processed by a sigmoid function, the intent parameter can be obtained.

[0069] In step S206, the at least one object to be pushed may be pushed to the target user based on its push score. For example, the top k objects to be pushed, ranked from highest to lowest push score, may be pushed to the target user. Alternatively, if the push score of an object to be pushed exceeds a preset score threshold, the object to be pushed is pushed to the target user.

[0070] Among them, the push score of each object to be pushed is obtained based on the following method: First, the first weight corresponding to the first feature and the second weight corresponding to the fourth feature of the object to be pushed are determined based on the intention parameter, the first weight is used to characterize the probability that the target user's operation intention is related to the current operation object, and the second weight is used to characterize the probability that the target user's operation intention is unrelated to the current operation object. Then, the first feature is weighted based on the first weight, and the fourth feature is weighted based on the second weight, and the push score of the object to be pushed is obtained based on the weighted first feature and the weighted fourth feature. The process of weighting the first feature and the fourth feature based on the first weight and the second weight can be achieved through the second feature fusion layer.

[0071] When the target user's operation intention is strongly correlated with the current operation object, for example, when the current operation object is the object that the target user operates after searching, the first weight is larger and the second weight is smaller (that is, the second weight is smaller than the first weight), the current operation object has a greater impact on the push score, and the characteristics of the object to be pushed itself have a smaller impact on the push score. When the target user's operation intention is weakly correlated with the current operation object, the first weight is smaller and the second weight is larger (that is, the second weight is larger than the first weight), the current operation object has a smaller impact on the push score, and the characteristics of the object to be pushed itself have a greater impact on the push score. In other words, the stronger the correlation between the target user's operation intention and the current operation object, the larger the first weight and the smaller the second weight; the weaker the correlation between the target user's operation intention and the current operation object, the smaller the first weight and the larger the second weight. In some embodiments, the first weight and the second weight are both real numbers between 0 and 1, wherein the first weight can be the intention parameter itself, and the second weight can be equal to 1 minus the first weight.

[0072] In some embodiments, the second feature fusion layer may adopt a multi-head attention mechanism. Specifically, the historical operation sequence of the target user can be obtained, the first feature and the historical operation sequence are input into the fourth multi-head attention layer for processing, and the fourth feature and the historical operation sequence are input into the fifth multi-head attention layer for processing. Then, the processed first feature is weighted based on the first weight, and the processed fourth feature is weighted based on the second weight, and the push score of the object to be pushed is obtained based on the weighted first feature and the weighted fourth feature. By inputting the first feature and the historical operation sequence into the fourth multi-head attention layer for processing, it is possible to focus on the historical operation sequence related to the first feature, and by inputting the fourth feature and the historical operation sequence into the fifth multi-head attention layer for processing, it is possible to focus on the historical operation sequence related to the fourth feature, thereby reducing the impact of historical operation sequences unrelated to the first and fourth features on the push results.

[0073] Furthermore, before inputting the historical operation sequence into the fourth multi-head attention layer and the fifth multi-head attention layer, the historical operation sequence can also be input into the sixth multi-head attention layer for processing, so as to better capture the important features in the historical operation sequence and improve the modeling ability of the model.

[0074] As shown in Figure 3, in the first feature fusion layer, the first and fourth features can be fused to obtain a third fused feature. A push score for the object to be pushed is then obtained based on the third fused feature, the weighted first feature, and the weighted fourth feature. The Hadamard product of the first and fourth features can be obtained, and the first, fourth, and Hadamard products can be fused to obtain the third fused feature. Specifically, the first, fourth, and Hadamard products can be input into a second multi-layer perceptron for processing to obtain the third fused feature. The Hadamard product, also known as the element-by-element product, is a new vector obtained by multiplying the elements at the same position of two vectors one by one. The Hadamard product multiplies corresponding elements of the two vectors, preserving the information of each element. This allows for better fusion of the features of the two vectors, allowing the fused vector to better represent the original information. Furthermore, the Hadamard product multiplies corresponding elements of the two vectors. When two vectors are similar, the elements at corresponding positions are also similar. Therefore, Hadamard product fusion can enhance similarity information and deepen the expression of similarity features. By adjusting the size of the elements of the two vectors, you can control the strength of each element in the fused vector. If a vector has a larger element, the Hadamard product with the other vector will increase the value of that element in the fused vector, thereby emphasizing its contribution to the fusion result.

[0075] Furthermore, before fusing the first feature, the fourth feature, and their Hadamard product, the first feature may be input into a third multilayer perceptron for processing, and the fourth feature may be input into a fourth multilayer perceptron for processing. This improves the feature extraction capability and the nonlinear mapping capability of features, reduces feature dimensionality, and improves the generalization capability of features.

[0076] Furthermore, the push score of the object to be pushed can also be obtained based on the third fused feature, the fourth fused feature, the weighted first feature, and the weighted fourth feature. The fourth fused feature is obtained based on the following method: obtaining the expectation and variance determined by the push model based on the target feature set, generating a random distribution based on the expectation and variance, sampling the random distribution to obtain a sampled feature, and fusing the fourth feature with the sampled feature to obtain a fourth fused feature. The fusion of the fourth feature and the sampled feature can be element-wise multiplication or addition of the fourth feature and the fourth feature. The fourth feature and the sampled feature can be fused using a feature fusion submodel included in the push model.

[0077] After obtaining the third fused feature, the fourth fused feature, the weighted first feature, and the weighted fourth feature, these features can be concatenated and flattened, and then the processed features are input into the fifth multi-layer perceptron to obtain the push score of the object to be pushed.

[0078] This disclosed embodiment explores key features that influence users' positive feedback behaviors. By incorporating user stratification, activity, length of similar objects, length of objects of the same parent class, and object origin into the immediate user intent layer, the system characterizes user intent strength from multiple dimensions. Furthermore, the random variable representing dynamic intent strength is derived by combining expectation and variance.

[0079] After obtaining the random variable that characterizes the dynamic intent intensity, the sampling features are obtained based on the random variable sampling, and the explicit interaction between the intent and the item to be scored is obtained through the first feature fusion layer with the trigger item. In the second feature fusion layer, the historical operation sequence of the user by the trigger item and the object to be pushed is filtered respectively. At the same time, the influence weights of the trigger item and the object to be pushed on the final score (i.e., the aforementioned first weight and second weight) are adjusted based on the intent parameters. The final feature vectors are spliced ​​and sent to the MLP for score prediction.

[0080] The present disclosure has the following advantages:

[0081] (1) To address the problem that existing solutions do not clearly characterize users' immediate and potential intentions, this solution obtains the second and third features and refines user intentions through a comparative learning model. This solution adopts a comparative learning strategy to ensure the robustness of the push model in extracting user intentions, enhance the push model's ability to resist user noise clicks, and alleviate the problem of inaccurate intention estimation caused by sparse user behavior.

[0082] (2) In response to the problem of insufficient generalization of user intent modeling in existing solutions, this solution does not model user intent as a static representation, but instead models user intent as a Gaussian distribution. This expands the semantic scope of user intent in the form of distribution and enhances the generalization ability of the network to extract user intent.

[0083] (3) To address the problem that existing solutions cannot accurately characterize the dynamic changes in user intention intensity, this solution explores the key influencing factors that affect user decisions and explicitly models the dynamic changes in user intention intensity.

[0084] 5 , an embodiment of the present disclosure further provides an object pushing device, the device comprising:

[0085] An acquisition module 12 is configured to acquire a target feature set corresponding to a current operation object of a target user; the target feature set includes a first feature of the current operation object, a second feature of at least one historical operation object having a feature similarity with the current operation object greater than a preset similarity threshold, and a third feature of at least one other object having a co-occurrence relationship with the current operation object.

[0086] A processing module 14 is configured to process the target feature set using a push model to obtain an intention parameter of the target user with respect to the current operation object; the intention parameter is used to characterize the correlation between the operation intention of the target user and the current operation object;

[0087] The push module 16 is configured to push at least one object to be pushed to the target user based on the intention parameter.

[0088] In some embodiments, the functions or modules included in the device provided by the embodiments of the present disclosure can be used to execute the method described in the above method embodiments. The specific implementation can refer to the description of the above method embodiments. For the sake of brevity, it will not be repeated here.

[0089] An embodiment of the present disclosure further provides a computer device, which includes at least a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the method described in any of the aforementioned embodiments when executing the program.

[0090] FIG6 shows a more specific hardware structure diagram of a computing device provided by an embodiment of the present disclosure. The device may include: a processor 22, a memory 24, an input / output interface 26, a communication interface 28, and a bus 30. The processor 22, the memory 24, the input / output interface 26, and the communication interface 28 are connected to each other within the device via the bus 30.

[0091] The processor 22 can be implemented using a general-purpose CPU (Central Processing Unit), a microprocessor, an application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided by the embodiments of the present disclosure. The processor 22 may also include a graphics card, which may be an Nvidia Titan X graphics card or a 1080Ti graphics card.

[0092] The memory 24 can be implemented in the form of ROM (Read Only Memory), RAM (Random Access Memory), static storage devices, dynamic storage devices, etc. The memory 24 can store an operating system and other application programs. When the technical solutions provided by the embodiments of the present disclosure are implemented through software or firmware, the relevant program codes are stored in the memory 24 and called and executed by the processor 22.

[0093] The input / output interface 26 is used to connect to input / output modules to enable information input and output. The input / output modules can be configured as components within the device (not shown) or can be externally connected to the device to provide corresponding functions. Input devices may include a keyboard, mouse, touch screen, microphone, various sensors, etc., and output devices may include a display, speaker, vibrator, indicator light, etc.

[0094] The communication interface 28 is used to connect to a communication module (not shown) to enable communication between the device and other devices. The communication module can communicate via a wired method (such as a USB (Universal Serial Bus), a network cable, etc.) or a wireless method (such as a mobile network, Wi-Fi (Wireless Fidelity), Bluetooth, etc.).

[0095] The bus 30 comprises a pathway for transmitting information between the various components of the device, such as the processor 22 , the memory 24 , the input / output interface 26 , and the communication interface 28 .

[0096] It should be noted that although the above device only shows the processor 22, memory 24, input / output interface 26, communication interface 28, and bus 30, in a specific implementation, the device may also include other components necessary for normal operation. In addition, those skilled in the art will understand that the above device may only include the components necessary to implement the embodiments of the present disclosure, and does not necessarily include all the components shown in the figure.

[0097] An embodiment of the present disclosure further provides a computer-readable storage medium having a computer program stored thereon, which implements the method described in any of the aforementioned embodiments when the program is executed by a processor.

[0098] Computer-readable media include permanent and non-permanent, removable and non-removable media that can be used to store information by any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, Phase-Change Random Access Memory (PRAM), Static Random Access Memory (SRAM), Dynamic Random Access Memory (DRAM), other types of Random Access Memory (RAM), Read-Only Memory (ROM), Electrically Erasable Programmable Read-Only Memory (EEPROM), flash memory or other memory technology, Compact Disc Read-Only Memory (CD-ROM), Digital Versatile Disc (DVD) or other optical storage, magnetic cassettes, tape disk storage or other magnetic storage devices, or any other non-transmission medium that can be used to store information that can be accessed by a computing device. According to the definition in this article, computer-readable media does not include transitory media such as modulated data signals and carrier waves.

[0099] The present disclosure also provides a computer program product, including a computer program, which, when executed by a processor, implements the method described in any of the embodiments of the present disclosure. The computer program product can be stored in the computer-readable medium described in the aforementioned embodiments, or can be applied to the computer device described in the aforementioned embodiments.

[0100] Through the description of the above implementation methods, it can be seen that those skilled in the art can clearly understand that the embodiments of the present disclosure can be implemented by means of software plus the necessary general hardware platform. Based on this understanding, the technical solution of the embodiments of the present disclosure, or the part that contributes to the prior art, can be embodied in the form of a software product, which can be stored in a storage medium such as ROM / RAM, a magnetic disk, an optical disk, etc., and includes a number of instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or certain parts of the embodiments of the present disclosure.

[0101] The systems, devices, modules, or units described in the above embodiments may be implemented by computer chips or entities, or by products having certain functions. A typical implementation device is a computer, which may be in the form of a personal computer, laptop computer, cellular phone, image acquisition device phone, smartphone, personal digital assistant, media player, navigation device, email transceiver, game console, tablet computer, wearable device, or any combination of these devices.

[0102] Each embodiment in the present disclosure is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the device embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment. The device embodiment described above is merely illustrative, wherein the modules described as separate components may or may not be physically separated, and when implementing the embodiment of the present disclosure, the functions of each module can be implemented in the same one or more software and / or hardware. It is also possible to select some or all of the modules according to actual needs to achieve the purpose of the embodiment. A person of ordinary skill in the art can understand and implement it without paying any creative work.

[0103] The above is only a specific implementation of the embodiment of the present disclosure. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the embodiment of the present disclosure. These improvements and modifications should also be regarded as the scope of protection of the embodiment of the present disclosure.

Claims

1. A method for pushing an object, the method comprising: Get the target feature set corresponding to the current operation object of the target user; The target feature set includes a first feature of a current operation object, a second feature of at least one historical operation object having a feature similarity with the current operation object greater than a preset similarity threshold, and a third feature of at least one other object having a co-occurrence relationship with the current operation object; Processing the target feature set through a push model to obtain an intention parameter of the target user with respect to the current operation object; the intention parameter is used to characterize the correlation between the operation intention of the target user and the current operation object; At least one object to be pushed is pushed to the target user based on the intention parameter.

2. According to the method according to claim 1, the larger the intention parameter is, the stronger the correlation between the object to be pushed and the current operation object is; the smaller the intention parameter is, the stronger the correlation between the object to be pushed and the historical operation sequence of the target user is.

3. The method according to claim 1, wherein the processing of the target feature set by a push model to obtain the target user's intention parameter for the current operation object comprises: Obtaining an expectation and a variance determined by the push model based on the target feature set, wherein the expectation is used to represent the intensity of the target user's intention to operate the current operation object, and the variance is used to represent the uncertainty of the target user's intention to operate the current operation object; An intention parameter of the target user with respect to the current operation object is obtained based on the expectation and the variance.

4. The method according to claim 3, wherein the step of obtaining the target user's intention parameter for the current operation object based on the expectation and the variance comprises: generating a random distribution based on the expectation and the variance; Sampling the random distribution to obtain sampling features; fusing the first feature, the user feature of the target user, and the sampling feature to obtain a first fused feature; Acquire the target user's intention parameter for the current operation object based on the first fusion feature.

5. The method according to claim 3, wherein obtaining the expectation and variance determined by the push model based on the target feature set comprises: Processing the first feature and the second feature through a first multi-head attention layer to obtain a first output vector; Processing the first feature and the third feature through a second multi-head attention layer to obtain a second output vector; fusing the first output vector and the second output vector to obtain a second fused feature; The expectation and the variance are determined based on the second fused features.

6. The method according to claim 3, wherein the expectation and the variance are obtained by a contrastive learning model; the contrastive learning model is trained based on at least one pair of positive samples and at least one pair of negative samples; in, Each pair of positive samples includes a target feature set corresponding to the same current operation object and an augmented feature set corresponding to the current operation object; the features in the augmented feature set are obtained by performing feature transformation on the features in the target feature set; Each pair of negative samples includes a different set of target features corresponding to the current operation object.

7. The method according to claim 6, wherein the features in the augmented feature set are obtained by: Randomly deleting the second features of several historical operation objects and / or the third features of several other objects in the target feature set; Randomly delete several feature dimensions of the second feature and / or the third feature in the target feature set.

8. The method according to any one of claims 1 to 7, wherein the pushing at least one to-be-pushed object to the target user based on the intention parameter comprises: Pushing the at least one object to be pushed to the target user based on the push score of the at least one object to be pushed; wherein the push score of each object to be pushed in the at least one object to be pushed is obtained based on the following method: Determining, based on the intention parameter, a first weight corresponding to the first feature and a second weight corresponding to the fourth feature of the object to be pushed; the first weight is used to represent the probability that the target user's operation intention is related to the current operation object, and the second weight is used to represent the probability that the target user's operation intention is unrelated to the current operation object; Performing weighted processing on the first feature based on the first weight, and performing weighted processing on the fourth feature based on the second weight; A push score of the object to be pushed is obtained based on the weighted first feature and the weighted fourth feature.

9. The method according to claim 8, wherein weighting the first feature based on the first weight and weighting the fourth feature based on the second weight comprises: Obtaining the historical operation sequence of the target user; Inputting the first feature and the historical operation sequence into a fourth multi-head attention layer for processing; Inputting the fourth feature and the historical operation sequence into a fifth multi-head attention layer for processing; The processed first feature is weighted based on the first weight, and the processed fourth feature is weighted based on the second weight.

10. The method according to claim 8, wherein obtaining the push score of the object to be pushed based on the weighted first feature and the weighted fourth feature comprises: fusing the first feature and the fourth feature of the object to be pushed to obtain a third fused feature; A push score of the object to be pushed is obtained based on the third fusion feature, the weighted first feature, and the weighted fourth feature.

11. The method according to claim 10, wherein fusing the first feature and the fourth feature to obtain a third fused feature comprises: obtaining a Hadamard product of the first feature and the fourth feature; The first feature, the fourth feature and the Hadamard product are fused to obtain the third fused feature.

12. The method according to claim 10, wherein obtaining the push score of the object to be pushed based on the third fused feature, the weighted first feature, and the weighted fourth feature comprises: Obtaining a push score for the object to be pushed based on the third fused feature, the fourth fused feature, the weighted first feature, and the weighted fourth feature; The fourth fusion feature is obtained based on the following method: Obtaining an expectation and a variance determined by the push model based on the target feature set, wherein the expectation is used to represent the intensity of the target user's intention to operate the current operation object, and the variance is used to represent the uncertainty of the target user's intention to operate the current operation object; generating a random distribution based on the expectation and the variance; Sampling the random distribution to obtain sampling features; The fourth feature and the sampling feature are fused to obtain the fourth fused feature.

13. A computer-readable storage medium having computer instructions stored thereon, wherein when the computer instructions are executed by a processor, the method according to any one of claims 1 to 12 is implemented.

14. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the method according to any one of claims 1 to 12 when executing the program.

15. A computer program product comprising a computer program, wherein when the computer program is executed by a processor, the method according to any one of claims 1 to 12 is implemented.

Images