Information push method, device and computer readable storage medium

Abstract

The present invention discloses an information push method, device and computer-readable storage medium, and relates to the field of data processing. The information push method comprises: using the positive features and the inverted negative features of the user to construct a user feature vector of the user, wherein the positive features are positively correlated with the future resource consumption of the user, and the negative features are negatively correlated with the future resource consumption of the user; inputting the user feature vector into a neural network, wherein the values ​​of the parameters of the neural network are all non-negative numbers; obtaining the prediction result of the future resource consumption of the user output by the neural network; and pushing information to the user according to the resource consumption of the user. The embodiments of the present invention can make the prediction result of the resource consumption more accurate, and can deliver the pushed information to the user more accurately.

Description

Technical Field

[0001] The present invention relates to the field of data processing, and in particular to an information pushing method, device and computer-readable storage medium. Background Art

[0002] In the field of information push on websites in the past, information is usually pushed based on user-side information such as user historical behavior, user inherent attributes, and inherent information of the recommended object, so that the pushed information is information that the user is interested in, thereby improving the efficiency of the push. Summary of the invention

[0003] After analysis, the inventors found that in addition to the user's historical behavior and inherent attributes, the user's resource consumption can also be predicted to measure a customer or user's resource consumption level for a website or application in the future, so as to push more targeted information.

[0004] In related technologies, there are two main methods for calculating user resource consumption. One is to directly construct a mathematical formula and calculate it by substituting the corresponding parameter values ​​of the user into the mathematical formula. The other is a machine learning method, which first selects features and labels to construct training data, and then trains a decision tree model or a neural network model to discover the relationship between labels and features.

[0005] However, in actual use, it was found that the prediction accuracy of resource consumption by related technologies was low, resulting in low accuracy of information push, which affected the effectiveness of information push.

[0006] A technical problem to be solved by the embodiments of the present invention is: how to improve the accuracy of information push.

[0007] According to a first aspect of some embodiments of the present invention, there is provided an information push method, comprising: constructing a user feature vector of the user using a positive feature and an inverted negative feature of the user, wherein the positive feature is positively correlated with the user's future resource consumption, and the negative feature is negatively correlated with the user's future resource consumption; inputting the user feature vector into a neural network, wherein the values ​​of the parameters of the neural network are all non-negative numbers; obtaining a prediction result of the user's future resource consumption output by the neural network; and pushing information to the user based on the user's resource consumption.

[0008] In some embodiments, pushing information to a user based on the user's resource consumption includes: determining the association between the user and a preset set of items, or items in a preset set of items, based on the resource consumption; and pushing information to the user based on the association.

[0009] In some embodiments, the resource consumption is the resource consumption of the user for the first set of items, and based on the user's resource consumption, information is pushed to the user, including: determining the association between each user in the alternative user set and the first set of items based on the resource consumption of each user in the alternative user set; and pushing information corresponding to the first set of items to users whose association is higher than a preset association, or a preset number of users with the highest association.

[0010] In some embodiments, the resource consumption is the resource consumption of the user on the second set of items, and information is pushed to the user based on the resource consumption of the user, including: determining the association between the user and each item in the third set of items based on the resource consumption of the user on the second set of items, attributes of the second set of items, and attributes of each item in the third set of items; constructing an item feature vector for each item in the third set of items, wherein the features in the item feature vector include the association; sorting the items in the third set of items based on the item feature vector; and pushing information about the items in the third set of items to the user based on the sorting result.

[0011] In some embodiments, the information push method also includes: constructing a training feature vector for each user in the training user set based on the positive features and the inverted negative features of each user in the training user set in the first time period, wherein the marking value of the training feature vector is the resource consumption of the corresponding user in the second time period, and the first time period is located before the second time period; inputting the training feature vector into the neural network to obtain the predicted value of the training feature vector by the neural network; and adjusting the parameters of the neural network based on the loss function of the neural network, the predicted value and the marked value of the training feature vector.

[0012] In some embodiments, the loss function of the neural network is determined based on the label value and the predicted value of the input sample of the neural network, and includes a first loss part and a second loss part, wherein the first loss part is positively correlated with the absolute value of the difference between the label value and the predicted value, and the second loss part is positively correlated with the difference between the label value and the predicted value.

[0013] In some embodiments, the loss function of the neural network is:

[0014] L=|y true -y pre |+λ*(y true -y pre )

[0015] Among them, L represents the value of the loss function; y true Indicates the tag value; y pre represents the predicted value; λ represents the preset parameter, and λ∈(0,1).

[0016] In some embodiments, the output value of the activation function used by the neural network is a non-negative number.

[0017] According to a second aspect of some embodiments of the present invention, there is provided an information push device, comprising: a vector construction module, configured to construct a user feature vector of a user using a positive feature and an inverted negative feature of the user, wherein the positive feature is positively correlated with the user's future resource consumption, and the negative feature is negatively correlated with the user's future resource consumption; an input module, configured to input the user feature vector into a neural network, wherein the values ​​of the parameters of the neural network are all non-negative numbers; a prediction result acquisition module, configured to obtain a prediction result of the future resource consumption of the user output by the neural network; and a push module, configured to push information to the user according to the user's resource consumption.

[0018] In some embodiments, the push module is further configured to determine the association between the user and a preset set of items, or items in a preset set of items, based on the resource consumption; and push information to the user based on the association.

[0019] In some embodiments, the information push device also includes: a training module, configured to construct a training feature vector for each user in the training user set based on the positive features and the inverted negative features of each user in the training user set in the first time period, wherein the marking value of the training feature vector is the resource consumption of the corresponding user in the second time period, and the first time period is located before the second time period; input the training feature vector into the neural network to obtain the predicted value of the training feature vector by the neural network; and adjust the parameters of the neural network based on the loss function of the neural network, the predicted value and the marked value of the training feature vector.

[0020] In some embodiments, the loss function of the neural network is determined based on the label value and the predicted value of the input sample of the neural network, and includes a first loss part and a second loss part, wherein the first loss part is positively correlated with the absolute value of the difference between the label value and the predicted value, and the second loss part is positively correlated with the difference between the label value and the predicted value.

[0021] According to a third aspect of some embodiments of the present invention, there is provided an information push device, comprising: a memory; and a processor coupled to the memory, wherein the processor is configured to execute any one of the aforementioned information push methods based on instructions stored in the memory.

[0022] According to a fourth aspect of some embodiments of the present invention, there is provided a computer-readable storage medium having a computer program stored thereon, wherein the program, when executed by a processor, implements any one of the aforementioned information push methods.

[0023] Some embodiments of the above invention have the following advantages or beneficial effects. In the user feature vector constructed in the embodiment of the present invention, each feature is positively correlated with the future resource consumption; and the parameter values ​​of the neural network model are also non-negative. Therefore, the prediction process of resource consumption meets the requirement of monotonicity, making the prediction result of resource consumption more accurate. In addition, when pushing information, the predicted future resource consumption is referenced, so that the pushed information can be delivered to the user more accurately.

[0024] Further features and advantages of the present invention will become apparent from the following detailed description of exemplary embodiments of the present invention with reference to the attached drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0026] Figure 1 A schematic flow chart of an information push method according to some embodiments of the present invention is shown.

[0027] Figure 2 A schematic flow chart of a neural network training method according to some embodiments of the present invention is shown.

[0028] Figure 3 A schematic flow chart of a method for information push according to other embodiments of the present invention is shown.

[0029] Figure 4 A schematic flow chart of a method for information push according to yet other embodiments of the present invention is shown.

[0030] Figure 5 A schematic flow chart of a method for information push according to some further embodiments of the present invention is shown.

[0031] Figure 6 A schematic structural diagram of an information push device according to some embodiments of the present invention is shown.

[0032] Figure 7 A schematic structural diagram of an information push device according to some other embodiments of the present invention is shown.

[0033] Figure 8 A schematic structural diagram of an information push device according to some other embodiments of the present invention is shown.

[0034] Fig. 9A schematic structural diagram of an information push device according to some further embodiments of the present invention is shown. DETAILED DESCRIPTION

[0035] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments. The following description of at least one exemplary embodiment is actually only illustrative and is by no means intended to limit the present invention and its application or use. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0036] The relative arrangement of components and steps, the numerical expressions and numerical values ​​set forth in these embodiments do not limit the scope of the present invention unless specifically stated otherwise.

[0037] At the same time, it should be understood that for the convenience of description, the sizes of the various parts shown in the drawings are not drawn according to the actual proportional relationship.

[0038] Technologies, methods, and apparatus known to ordinary technicians in the relevant field may not be discussed in detail, but where appropriate, such technologies, methods, and apparatus should be considered part of the authorization specification.

[0039] In all examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limiting. Therefore, other examples of the exemplary embodiments may have different values.

[0040] It should be noted that like reference numerals and letters refer to similar items in the following figures, and therefore, once an item is defined in one figure, it need not be further discussed in subsequent figures.

[0041] After further analysis, the inventors found that for the method of predicting resource consumption based on mathematical formulas, it is necessary to pre-calculate the values ​​of multiple variables, such as the change in user's annual resource consumption, the change in annual survival rate, risk-free interest rate and other factors. Therefore, there are many variables that need to be estimated in this way. Only when all variables are estimated more accurately, the calculation of resource consumption is more accurate. However, the estimation of these variables itself is also very difficult, so this method is not accurate enough.

[0042] Prediction based on decision tree regression model is also a typical algorithm for predicting resource consumption. It divides the feature space into several units, and each division unit (i.e., leaf node of the tree) has a specific output. Each intermediate node in the decision tree corresponds to a division result of a certain feature, and different feature values ​​may enter different subtrees. This means that different input values ​​near the critical point of the feature correspond to different subtrees. That is, the change of the feature value is very small, but the corresponding output values ​​may be very different, so this model is unstable.

[0043] Neural network model is another commonly used prediction method. However, in the process of predicting resource consumption, the traditional neural network has the problem that the model cannot guarantee monotonicity. For example, when other conditions remain unchanged, when the number of user logins is large and the resource consumption of historical single resource consumption behavior is large, the predicted resource consumption should be greater. However, the traditional neural network is non-monotonic, so the above correspondence cannot be established, and the prediction result of resource consumption is also inaccurate.

[0044] Since the accuracy of resource consumption prediction affects the accuracy of push information, it is necessary to improve the prediction method of resource consumption and thus improve the push efficiency. Figure 1 An embodiment of the information pushing method of the present invention is described.

[0045] Figure 1 FIG. 2 is a flow chart showing a method for pushing information according to some embodiments of the present invention. Figure 1 As shown, the information push method of this embodiment includes steps S102 to S108.

[0046] In step S102, a user feature vector of the user is constructed using the positive features of the user and the inverted negative features.

[0047] The embodiment of the present invention pre-screens, classifies and processes the features when constructing the user feature vector. The collected user features include positive features and negative features, wherein the positive features are positively correlated with the user's future resource consumption, and the negative features are negatively correlated with the user's future resource consumption.

[0048] The collected user features include, for example, user information features (demographic attributes, account, whether a social network account is bound, whether real-name authentication is performed, etc.), resource information statistical features (order information, shopping cart data, payment methods, etc.), and activity features (browsing records, click records, login frequency, terminal type used, etc.). Then, the positive and negative features can be determined manually or through data analysis.

[0049] In some embodiments, determining positive features and negative features by data analysis includes: obtaining the value of the feature to be analyzed for each user in the user data set in the first time period, and the resource consumption of each user in the second time period (the second time period is after the first time period), and then calculating the correlation coefficient (such as the Pearson correlation coefficient) between the feature to be analyzed and the resource consumption through the obtained data. If the correlation coefficient is a positive number, the feature to be analyzed is a positive feature, otherwise it is a negative feature.

[0050] In the constructed user feature vector, by inverting the negative features, each dimension in the vector is positively correlated with the future resource consumption. For example, suppose the initial vector constructed using the user's features is (x0, x1, ..., x n ,y0,...,y m ), where x0, x1, ..., x n is the positive feature, y0,...,y m is a negative feature. Let the feature transformation function Trans be defined as Trans(x0, x1, ..., x n , y0,…,y m )=(x0,x1,...,x n , -y0, ..., -y m ). Thus, each dimension in the transformed vector is positively correlated with the future resource consumption, so the vector after feature transformation is used as the user feature vector for input into the neural network.

[0051] In step S104, the user feature vector is input into a neural network, wherein the values ​​of the parameters of the neural network are all non-negative numbers.

[0052] In some embodiments, if the neural network includes an activation function, the output value of the activation function is a non-negative number to further ensure the monotonicity of the neural network. For example, an activation function whose minimum output value is 0, such as a linear rectifier function (ReLU, whose output value is the maximum value between the input value and 0), a leaky linear rectifier function (Leaky ReLU), an exponential linear unit (ELU), or SoftPlus (a smoothed version of ReLU), is used.

[0053] In step S106, the prediction result of the resource consumption of the user output by the neural network is obtained. The resource in the resource consumption represents, for example, time, bandwidth, GMV (Gross Merchandise Volume, total amount of goods traded), etc.

[0054] In step S108, information is pushed to the user according to the user's resource consumption. The pushed information includes, for example, description information of the item or item collection, links to the item or item collection, etc. This information is sent from the server side to the user's terminal through the network and displayed to the user.

[0055] In the user feature vector constructed in the above embodiment, each feature is positively correlated with the future resource consumption; and the parameter values ​​of the neural network model are all non-negative. Therefore, the prediction process of resource consumption meets the requirement of monotonicity, making the prediction result of resource consumption more accurate. In addition, when pushing information, the predicted future resource consumption is referenced, so that the pushed information can be delivered to the user more accurately.

[0056] The accuracy of the pushed information can be measured in a variety of ways, such as by whether the user clicks on it, whether the user browses the information for a period of time exceeding a threshold, whether the user purchases or subscribes to the product corresponding to the browsed pushed information within a preset time period, etc.

[0057] Reference below Figure 2 An embodiment of the neural network training method of the present invention is described.

[0058] Figure 2 FIG. 2 is a flow chart of a neural network training method according to some embodiments of the present invention. Figure 2 As shown, the training method of this embodiment includes steps S202 to S206.

[0059] In step S202, a training feature vector for each user in the training user set is constructed based on the positive features and the inverted negative features of each user in the training user set in the first time period, wherein the marking value of the training feature vector is the resource consumption of the corresponding user in the second time period, and the first time period is located before the second time period.

[0060] In step S204, the training feature vector is input into the neural network to obtain the prediction value of the neural network for the training feature vector.

[0061] In step S206, the parameters of the neural network are adjusted based on the loss function of the neural network, the predicted value and the labeled value of the training feature vector.

[0062] When adjusting parameters, make sure that the value of each parameter is non-negative. For example, if the value of a parameter to be adjusted is negative after calculation, set it to 0.

[0063] Through the above embodiments, the neural network can accurately predict the future resource consumption of the user and maintain monotonicity.

[0064] In the process of model training, it is necessary to specify the optimization target, that is, the loss function, so that the loss function value of the trained model is as small as possible. The loss function in the related art usually only describes the error between the predicted value and the marked value, that is, the absolute value of the difference between the two. After analyzing the training data, it is found that whether it is the ordinary neural network used in the related art or the neural network whose parameter values ​​are all non-negative numbers, if the above-mentioned default loss function is used, the predicted result is much smaller than the actual value as a whole. For ordinary classification problems, the size of the predicted result value is not important. The present invention aims to predict numerical values. Therefore, in some embodiments, in addition to the need for the overall distribution of the predicted results to be accurate, the predicted value and the true value, that is, the marked value, must also be close in size.

[0065] In some embodiments, the loss function of the neural network is determined based on the label value and the predicted value of the input sample of the neural network, and includes a first loss part and a second loss part, wherein the first loss part is positively correlated with the absolute value of the difference between the label value and the predicted value, and the second loss part is positively correlated with the difference between the label value and the predicted value.

[0066] For example, let y true Indicates the marked value, y pre Represents the predicted value. The training process is regarded as an optimization process, and the optimization goal is to require min θ |y true -y pre |, and satisfy the condition y true -y pre <0. According to the Lagrange multiplier method, the optimization is transformed into an unconstrained optimization problem: CMAE = min θ (|y true -y pre |+λ*(y true -y pre )), where λ represents a preset parameter. Thus, formula (1) can be used as the loss function of the neural network.

[0067] L=|y true -y pre |+λ*(y true -y pre ) (1)

[0068] In formula (1), L represents the value of the loss function; y true Indicates the tag value; y pre represents the predicted value; λ represents the preset parameter, and λ∈(0,1). The first loss part of the loss function is |y true -y pre |, used to make the predicted value and the marked value as close as possible; the second loss part is λ*(y true -ypre ) is used to make the actual value as small as possible as possible to avoid the predicted resource consumption being too small, which will affect the accuracy of subsequent push notifications.

[0069] As needed, those skilled in the art can modify formula (1), such as adding a constant, using other methods to determine the difference between the marked value and the predicted value, etc., which will not be elaborated here.

[0070] The role of λ∈(0,1) is to make the value of L greater than 0, so as to avoid the consequence that the marked value is much smaller than the predicted value. Theoretically, the larger the λ is, the greater the effect of the second loss part is than the effect of the first loss part, so the larger the predicted mean value is, the greater the error is; conversely, the smaller the λ is, the smaller the error is. The present invention has experimentally compared the prediction effects of loss functions using different parameters, and Table 1 exemplarily shows the comparison results, in which the comparison between the prediction error and the average value of the predicted value is mainly examined.

[0071] In some embodiments, the effect of the model is evaluated using the error wMAPE (Weighted Mean Absolute Percentage Error):

[0072]

[0073] In formula (2), wmape represents the error value; m represents the number of fitting points; i represents the identifier of the fitting point; y i represents the label value of fitting point i, p i Represents the predicted value of fitting point i.

[0074] Table 1

[0075] Parameter settings wmape The average of the predicted values λ=0 0.7379 769 λ=0.1 0.7380 860 λ=0.2 0.7483 918 λ=0.3 0.7627 1064 λ=0.4 0.7957 1210 λ=0.5 0.8288 1319

[0076] As can be seen from Table 1, the larger the value of λ is, the greater the impact of the second loss part is. The average value of the predicted value gradually increases, but the error also gradually increases. Therefore, the appropriate λ value can be determined through preliminary experiments and actual business needs.

[0077] Reference below Figure 3 An embodiment of pushing information according to resource consumption is described.

[0078] Figure 3 FIG. 2 is a flow chart showing a method for information push according to other embodiments of the present invention. Figure 3 As shown, the information push method of this embodiment includes steps S302 to S304. In this embodiment, the resource consumption prediction process refers to the above embodiment, which will not be repeated here.

[0079] In step S302, the association between the user and a preset item set or items in the preset item set is determined according to the resource consumption.

[0080] In some embodiments, the resource consumption is non-directional, for example, indicating the user's consumption of resources on the entire platform of a certain platform.

[0081] In this case, the resource consumption of the user is positively correlated with the preset item set or the items in the preset item set.

[0082] In some embodiments, resource consumption is directional, for example, indicating the user's resource consumption for a certain set of items, where the set of items is determined by attributes such as category, brand, etc. For example, if the user's resource consumption for set A is large, then, under other conditions unchanged, the user has a higher association with set A and the items therein, and set B and the items therein, which have similar attributes to set A, while the user has a lower association with set C and the items therein, which have opposite attributes to set A.

[0083] In step S304, information is pushed to the user according to the relevance.

[0084] Through the above embodiments, the correlation between the user and the item set or the items can be determined through the predicted resource consumption, so that information with a higher correlation degree is pushed to the user, thereby improving the accuracy of the push.

[0085] Combine the following Figure 4 and Figure 5 Describe two more specific push scenarios.

[0086] Figure 4 FIG. 2 is a flow chart showing a method for information push according to some other embodiments of the present invention. Figure 4 As shown, the information push method of this embodiment includes steps S402 to S404. In this embodiment, the resource consumption prediction process refers to the above embodiment, which will not be repeated here.

[0087] In step S402, the association degree between each user in the candidate user set and the first item set is determined according to the resource consumption of each user in the candidate user set on the first item set.

[0088] For example, the predicted resource consumption is directly used as the correlation degree, or the calculation result between the predicted resource consumption and other attributes of the user is used as the correlation degree.

[0089] In step S404, information corresponding to the first set of items is pushed to users whose association degree is higher than a preset association degree, or to a preset number of users with the highest association degree.

[0090] For example, users in the candidate user set are sorted according to the relevance, and information corresponding to the first item set is pushed to the first N (N is a positive integer) users. The information corresponding to the first item set is, for example, advertising information, product description information, category or brand entry link, etc.

[0091] Therefore, when a certain set of items needs to be accurately delivered, the above embodiment can be used to determine a user group that is more compatible with the set of items, thereby improving the accuracy of information delivery.

[0092] Figure 5 FIG. 2 is a flow chart showing a method for information push according to some further embodiments of the present invention. Figure 5 As shown, the information push method of this embodiment includes steps S502 to S508. In this embodiment, the resource consumption prediction process refers to the above embodiment, which will not be repeated here.

[0093] In step S502, the association between the user and each item in the third item set is determined according to the resource consumption of the user on the second item set, the attributes of the second item set, and the attributes of each item in the third item set.

[0094] For example, the second item set is a cosmetics category, and the third item set is a set of items recommended to users on the homepage of an e-commerce application, including cosmetics products, women's clothing products, and men's clothing products. If, through preliminary data analysis, cosmetics and women's clothing categories are classified as female attributes, and men's clothing products are classified as male attributes, then when it is predicted that the user's resource consumption for the second item set is high, the user has a higher correlation with cosmetics and women's clothing products in the third item set than with men's clothing products.

[0095] In step S504, an item feature vector of each item in the third item set is constructed, wherein the features in the item feature vector include a correlation degree.

[0096] In step S506, the items in the third item set are sorted according to the item feature vectors.

[0097] In some embodiments, the item feature vectors are input into a ranking prediction model to obtain ranking values ​​of the corresponding items, and the items are ranked according to the ranking values.

[0098] In step S508, information about items in the third item set, such as descriptions and links of the items, is pushed to the user based on the sorting result.

[0099] Through the method of the above embodiment, the items in the set of items to be recommended to the user (ie, the third set of items) that are more likely to interest the user can be screened, thereby improving the accuracy of information push.

[0100] Reference below Figure 6 and Figure 7 An embodiment of the information pushing device of the present invention is described.

[0101] Figure 6 FIG. 2 shows a schematic diagram of the structure of an information push device according to some embodiments of the present invention. Figure 6 As shown, the information push device 60 of this embodiment includes: a vector construction module 610, configured to use the user's positive features and the inverted negative features to construct a user feature vector of the user, wherein the positive feature is positively correlated with the user's future resource consumption, and the negative feature is negatively correlated with the user's future resource consumption; an input module 620, configured to input the user feature vector into a neural network, wherein the values ​​of the parameters of the neural network are all non-negative numbers; a prediction result acquisition module 630, configured to obtain the prediction result of the future user's resource consumption output by the neural network; and a push module 640, configured to push information to the user according to the user's resource consumption.

[0102] In some embodiments, the push module 640 is further configured to determine the association between the user and a preset set of items, or items in a preset set of items, based on the resource consumption; and push information to the user based on the association.

[0103] In some embodiments, the resource consumption is the resource consumption of the user for the first set of items, and the push module 640 is further configured to determine the association between each user in the alternative user set and the first set of items based on the resource consumption of each user in the alternative user set; and push the information corresponding to the first set of items to users whose association is higher than a preset association, or a preset number of users with the highest association.

[0104] In some embodiments, the resource consumption is the resource consumption of the user on the second set of items, and the push module 640 is further configured to determine the association between the user and each item in the third set of items based on the resource consumption of the user on the second set of items, the attributes of the second set of items, and the attributes of each item in the third set of items; construct an item feature vector for each item in the third set of items, wherein the features in the item feature vector include the association; sort the items in the third set of items according to the item feature vector; and push information about the items in the third set of items to the user based on the sorting result.

[0105] Figure 7 FIG. 2 shows a schematic diagram of the structure of an information push device according to some other embodiments of the present invention. Figure 7As shown, the information push device 70 of this embodiment includes, in addition to the vector construction module 610, the input module 620, the prediction result acquisition module 630 and the push module 640, a training module 750, which is configured to construct a training feature vector for each user in the training user set according to the positive features and the inverted negative features of each user in the training user set in the first time period, wherein the marking value of the training feature vector is the resource consumption of the corresponding user in the second time period, and the first time period is located before the second time period; the training feature vector is input into the neural network to obtain the prediction value of the neural network for the training feature vector; and the parameters of the neural network are adjusted based on the loss function of the neural network, the prediction value and the marking value of the training feature vector.

[0106] In some embodiments, the loss function of the neural network is determined based on the label value and the predicted value of the input sample of the neural network, and includes a first loss part and a second loss part, wherein the first loss part is positively correlated with the absolute value of the difference between the label value and the predicted value, and the second loss part is positively correlated with the difference between the label value and the predicted value.

[0107] In some embodiments, the loss function of the neural network is:

[0108] L=|y true -y pre |+λ*(y true -y pre )

[0109] Among them, L represents the value of the loss function; y true Indicates the tag value; y pre represents the predicted value; λ represents the preset parameter, and λ∈(0,1).

[0110] In some embodiments, the output value of the activation function used by the neural network is a non-negative number.

[0111] Figure 8 FIG. 2 shows a schematic diagram of the structure of an information push device according to some other embodiments of the present invention. Figure 8 As shown, the information push device 80 of this embodiment includes: a memory 810 and a processor 820 coupled to the memory 810, and the processor 820 is configured to execute the information push method in any of the aforementioned embodiments based on the instructions stored in the memory 810.

[0112] The memory 810 may include, for example, a system memory, a fixed non-volatile storage medium, etc. The system memory may store, for example, an operating system, an application program, a boot loader, and other programs.

[0113] Fig. 9FIG. 2 shows a schematic diagram of the structure of an information push device according to some other embodiments of the present invention. Fig. 9 As shown, the information push device 90 of this embodiment includes: a memory 910 and a processor 920, and may also include an input / output interface 930, a network interface 940, a storage interface 950, etc. These interfaces 930, 940, 950 and the memory 910 and the processor 920 may be connected, for example, via a bus 960. Among them, the input / output interface 930 provides a connection interface for input / output devices such as a display, a mouse, a keyboard, and a touch screen. The network interface 940 provides a connection interface for various networked devices. The storage interface 950 provides a connection interface for external storage devices such as SD cards and USB flash drives.

[0114] An embodiment of the present invention further provides a computer-readable storage medium having a computer program stored thereon, wherein the program implements any of the aforementioned information push methods when executed by a processor.

[0115] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable non-transient storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0116] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0117] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0118] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0119] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. An information push method, comprising: Using a positive feature and an inverted negative feature of a user to construct a user feature vector of the user, wherein the positive feature is positively correlated with the future resource consumption of the user, and the negative feature is negatively correlated with the future resource consumption of the user; The user feature vector is input into a neural network, wherein the values ​​of the parameters of the neural network are all non-negative numbers, wherein the training method of the neural network comprises: constructing a training feature vector for each user in the training user set according to the positive features and the negated negative features of each user in the training user set in the first time period, wherein the label value of the training feature vector is the resource consumption of the corresponding user in the second time period, and the first time period is located before the second time period; inputting the training feature vector into the neural network to obtain the predicted value of the training feature vector by the neural network; adjusting the parameters of the neural network based on the loss function of the neural network, the predicted value and the label value of the training feature vector; Obtaining a prediction result of the future resource consumption of the user output by the neural network; Information is pushed to the user according to the resource consumption of the user.

2. The information push method according to claim 1, wherein: The pushing of information to the user according to the resource consumption of the user comprises: Determining, according to the resource consumption, a correlation between the user and a preset set of items, or items in the preset set of items; Information is pushed to the user according to the correlation degree.

3. The information push method according to claim 2, wherein: The resource consumption is the resource consumption of the user on the first set of items, and the information push to the user according to the resource consumption of the user includes: Determining, according to the resource consumption of each user in the candidate user set, the degree of association between each user in the candidate user set and the first item set; The information corresponding to the first set of items is pushed to users whose association degree is higher than a preset association degree, or to a preset number of users with the highest association degree.

4. The information push method according to claim 2, wherein: The resource consumption is the resource consumption of the user on the second set of items, and the pushing of information to the user according to the resource consumption of the user includes: Determining the association between the user and each item in the third item set according to the resource consumption of the user on the second item set, the attributes of the second item set, and the attributes of each item in the third item set; constructing an item feature vector for each item in the third item set, wherein the features in the item feature vector include the degree of association; sorting the items in the third item set according to the item feature vector; According to the sorting result, information of items in the third item set is pushed to the user.

5. The information push method according to claim 1, wherein: The loss function of the neural network is determined based on the label value and the predicted value of the input sample of the neural network, and includes a first loss part and a second loss part, wherein the first loss part is positively correlated with the absolute value of the difference between the label value and the predicted value, and the second loss part is positively correlated with the difference between the label value and the predicted value.

6. The information push method according to claim 5, wherein: The loss function of the neural network is: Wherein, L represents the value of the loss function; represents the tag value; represents the predicted value; represents the preset parameters, and .

7. The information push method according to claim 1, wherein: The output value of the activation function used by the neural network is a non-negative number.

8. An information push device, comprising: The training module is configured to construct a training feature vector for each user in the training user set according to the positive features and the negated negative features of each user in the training user set in the first time period, wherein the label value of the training feature vector is the resource consumption of the corresponding user in the second time period, and the first time period is located before the second time period; input the training feature vector into the neural network to obtain the predicted value of the training feature vector by the neural network; and adjust the parameters of the neural network based on the loss function of the neural network, the predicted value and the label value of the training feature vector; A vector construction module is configured to construct a user feature vector of the user using a positive feature and a negated negative feature of the user, wherein the positive feature is positively correlated with the future resource consumption of the user, and the negative feature is negatively correlated with the future resource consumption of the user; An input module is configured to input the user feature vector into a neural network, wherein the values ​​of the parameters of the neural network are all non-negative numbers; A prediction result obtaining module, configured to obtain a prediction result of the resource consumption of the user in the future output by the neural network; The push module is configured to push information to the user according to the resource consumption of the user.

9. The information push device according to claim 8, wherein: The push module is further configured to determine the association between the user and a preset set of items, or items in the preset set of items, based on the resource consumption; and push information to the user based on the association.

10. The information push device according to claim 8, wherein: The loss function of the neural network is determined based on the label value and the predicted value of the input sample of the neural network, and includes a first loss part and a second loss part, wherein the first loss part is positively correlated with the absolute value of the difference between the label value and the predicted value, and the second loss part is positively correlated with the difference between the label value and the predicted value.

11. An information push device, comprising: Memory; as well as A processor coupled to the memory, the processor being configured to execute the information push method according to any one of claims 1 to 7 based on instructions stored in the memory.

12. A computer-readable storage medium having a computer program stored thereon, wherein when the program is executed by a processor, the information push method according to any one of claims 1 to 7 is implemented.

Images