Personalized Recommendation Method and System for New Project Promotion

Abstract

The present invention provides a personalized recommendation method and system for promoting new projects, which relates to the technical field of personalized recommendation. First, the present invention obtains candidate user log data and new project data to be promoted published by publishers, and then obtains the probability that candidate users share the new project based on a pre-constructed multi-task model, the candidate user log data, and the new project data to be promoted, and selects candidate users with the top-N sharing probabilities as recommended users. The present invention comprehensively considers the interests of users and project publishers, recommends new projects to users who may take sharing actions, can not only recommend new projects of interest to users, but also improve the promotion effect of new projects, enhance the popularity and activity of project publishers, and promote the healthy development of social media websites and online communities.

Description

Technical Field

[0001] The present invention relates to the technical field of personalized recommendation, and specifically relates to a personalized recommendation method and system for promoting new projects. Background Art

[0002] With the development of Internet technology, various social media websites and online communities are emerging continuously. More and more people like to create content and share information on these platforms, and data is being generated at an unprecedented speed. This has led to the so-called information overload phenomenon, and it has become increasingly difficult for users to find truly useful information for themselves in the vast amount of information. To overcome the problem of information overload, personalized recommendation systems have become an effective solution. Currently, recommendation systems have been widely applied in various fields, such as movies, news, music, and so on.

[0003] As one of the most successful methods for implementing recommendations, collaborative filtering (CF) has been studied by a large number of researchers. CF mainly includes two categories: user-based CF and item-based CF. The basic assumption of user-based CF is that if two users like similar items or have similar behaviors, then they will also have similar preferences or behaviors for other items. It identifies the commonalities among users based on the users' historical ratings, and then generates new recommendations according to the preferences of other users who are similar to the target user's interests. The item-based CF method, on the other hand, assumes that when an item is similar to the items that a user has liked in the past, the user will also like that item. However, no matter which type of CF method, it faces the cold start problem of new items.

[0004] Currently, there have been many studies on new item recommendations. There are mainly two ways. One is to use content-based filtering methods, and the other is to mix collaborative filtering with content-based methods. However, the current new item recommendation methods all start from the user's perspective and only consider the matching degree between new items and the user's interests. But for enterprises or new item publishers, they hope that new items can be promoted as soon as possible and gain more attention and popularity. Obviously, the current new item recommendation methods have a low promotion effect. Summary of the Invention

[0005] (1) Technical Problems to be Solved

[0006] Aiming at the deficiencies of the prior art, the present invention provides a personalized recommendation method and system for promoting new projects, and solves the technical problem that the existing new item recommendation methods have a low promotion effect.

[0007] (2) Technical Solutions

[0008] To achieve the above objectives, the present invention is realized through the following technical solutions:

[0009] In a first aspect, the present invention provides a personalized recommendation method for promoting new projects, and the method includes:

[0010] S1. Obtain candidate user log data and data of new projects to be promoted published by publishers;

[0011] S2. Based on a pre-constructed multi-task model, the candidate user log data, and the data of new projects to be promoted, obtain the probability that candidate users share the new projects, and select candidate users with the top-N sharing probabilities as recommended users.

[0012] Preferably, the candidate user log data includes: users, projects with which the users have interaction relationships, the time of the interaction relationships, and specific interaction relationships; the specific interaction relationships include click, like, and share behaviors.

[0013] Preferably, the construction process of the pre-constructed multi-task model is as follows:

[0014] A1. Obtain log data and project data that interact with the users in the log data;

[0015] A2. Preprocess the log data and project data to obtain training data, test data, and validation data. The training data, test data, and validation data all include the historical click data of users and new project data, and the new project data is the project generated within the corresponding time range of the training data, test data, and validation data;

[0016] A3. Obtain an initial multi-task model;

[0017] A4. Train, validate, and test the initial multi-task model based on the training data, validation data, and test data to obtain a multi-task model.

[0018] Preferably, A2 includes:

[0019] A201. Divide the log data according to time. The click behaviors that occurred before the preset time point are regarded as the historical click data of users, and the data after the preset time point is divided into training data, validation data, and test data according to the ratio of days x:y:z;

[0020] A202. Obtain new project data in the training data, validation data, and test data respectively;

[0021] A203. Process the training data to obtain training log data in the form of new project ID, user ID, whether clicked, whether liked, and whether shared; process the validation data and test data to obtain validation log data and test log data in the form of new projects and the user set that has shared the project.

[0022] Preferably, the A3 includes:

[0023] Based on the attention mechanism, fuse different attribute features of the new item to obtain the new item feature vector, and based on the long short-term memory network, learn the user's interest feature vector from the user's historical click data; splice the new item feature vector and the user's interest feature vector, and input them into four multi-layer perceptron layers to respectively predict the probability of exposure → click, the probability of click → like, the probability of like → share, and the probability of other cases → share; calculate the probabilities of exposure → click → like and exposure → click → like / other cases → share based on the above four probabilities.

[0024] Preferably, the method of fusing different attribute features of the new item based on the attention mechanism to obtain the new item feature vector includes:

[0025] Given the attribute a = [a1, a2,..., a M of the new item i;

[0026] Convert the attribute of the new item into an embedding vector through the embedding layer, and the attribute embedding matrix of the new item is represented as

[0027]

[0028] where: d is the dimension of the embedding vector, M is the number of attributes of the new item, is the embedding vector of the attribute a m ;

[0029] Use the attention mechanism to perform weighted fusion on different attribute features of the new item to obtain the new item feature vector;

[0030] The method of learning the user's interest feature vector from the user's historical click data based on the long short-term memory network includes:

[0031] Given the item sequence {i1, i2,..., i N} of the user's historical clicks;

[0032] Convert the item sequence into an embedding vector through the embedding layer as

[0033] Convert the embedding vector into the user's interest feature vector e u .

[0034] Preferably, the A4 includes:

[0035] When training a multi-task model, cross-entropy loss functions are constructed respectively according to exposure→click, exposure→click→like, exposure→click→like / other situations→share, and then weighted summation is performed on them to form the joint loss function of the multi-task model;

[0036]

[0037] where: w c 、w z 、w s are the weights of click, like, and share behaviors respectively, and they satisfy w c +w z +w s = 1; is the loss function corresponding to the click behavior; is the loss function corresponding to the like behavior; is the loss function corresponding to the share behavior;

[0038] The Adam optimizer is used to train the model, and the hyperparameters of the multi-task model are optimized according to the performance of the trained multi-task model on the validation data.

[0039] In a second aspect, the invention provides a personalized recommendation system for promoting new projects, and the system includes:

[0040] A data acquisition module for acquiring candidate user log data and new project data to be promoted published by the publisher;

[0041] A recommendation module for obtaining the probability that a candidate user shares the new project based on a pre-constructed multi-task model, the candidate user log data, and the new project data to be promoted, and selecting candidate users with a sharing probability of top-N as recommended users.

[0042] In a third aspect, the invention provides a computer-readable storage medium, which stores a computer program for personalized recommendation for promoting new projects, wherein the computer program enables a computer to execute the personalized recommendation method for promoting new projects as described above.

[0043] In a fourth aspect, the invention provides an electronic device, including:

[0044] One or more processors;

[0045] A memory; and

[0046] One or more programs, wherein the one or more programs are stored in the memory and are configured to be executed by the one or more processors, and the programs include those for executing the personalized recommendation method for promoting new projects as described above.

[0047] (3) Beneficial effects

[0048] The present invention provides a personalized recommendation method and system for promoting new projects. Compared with the prior art, the following beneficial effects are achieved:

[0049] Based on a pre-constructed multi-task model, the candidate user log data, and the new project data to be promoted, the present invention obtains the probability that a candidate user shares the new project, and selects the candidate users with the top-N sharing probabilities as the recommended users. The present invention comprehensively considers the interests of users and project publishers, recommends new projects to users who may take sharing actions, not only can recommend new projects that users are interested in, but also can improve the promotion effect of new projects, enhance the popularity and activity of project publishers, and promote the healthy development of social media websites and online communities. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0051] Figure 1 It is a block diagram of a personalized recommendation method for promoting new projects in an embodiment of the present invention;

[0052] Figure 2 It is a schematic diagram of a multi-task model in an embodiment of the invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0053] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the protection scope of the present invention.

[0054] By providing a personalized recommendation method and system for promoting new projects, the embodiments of the present application solve the technical problem of the low promotion effect of existing new project recommendation methods, and achieve the improvement of the promotion effect and the enhancement of the popularity of new project publishers.

[0055] The overall idea of the technical solutions in the embodiments of the present application to solve the above technical problems is as follows:

[0056] Currently, there have been many studies on the recommendation for new projects. There are mainly two methods. One is to use content-based filtering methods, and the other is to mix collaborative filtering with content-based methods. However, current new project recommendation methods all start from the perspective of users and only consider the matching degree between new projects and users' interests. But for enterprises or new project publishers, they hope that new projects can be promoted as soon as possible and gain more attention and popularity. Therefore, when implementing new project recommendation, it is also a valuable issue to promote new projects as much as possible. Existing research shows that the mutual connection between people on social websites can strengthen the process of information dissemination and expand the influence of this information. Interactions such as sharing can directly or indirectly affect others in the social network and accelerate the spread of information. Therefore, by using users' sharing behaviors, the promotion of new projects can be effectively achieved. However, current research on personalized recommendation for new project promotion is very scarce, and users' sharing behaviors have not been fully utilized. How to promote new projects while solving the cold start problem of new projects remains to be further studied. In view of the above problems, the embodiments of the present invention propose a personalized recommendation method for new project promotion. This method takes into account the interests of both users and project publishers, recommends new projects to users who may take sharing behaviors, not only can recommend new projects that users are interested in, but also can bring a certain promotion effect to new projects, improve the popularity and activity of project publishers, and promote the healthy development of social media websites and online communities. The present invention uses an attention mechanism to fuse different attribute features, can assign different weights to different attributes, and can obtain a more accurate new project feature representation to solve the cold start problem of new projects.

[0057] To better understand the above technical solution, the above technical solution will be described in detail below in combination with the accompanying drawings of the specification and specific embodiments.

[0058] The embodiments of the present invention provide a personalized recommendation method for new project promotion, as Figure 1 shown, this method includes steps S1 to S2:

[0059] S1. Obtain candidate user log data and new project data to be promoted published by the publisher;

[0060] S2. Based on a pre-constructed multi-task model, the candidate user log data, and the new project data to be promoted, obtain the probability that the candidate user shares the new project, and select the candidate users with the top-N sharing probabilities as the recommended users.

[0061] The embodiments of the present invention comprehensively consider the interests of users and project publishers, recommend new projects to users who may take sharing actions, not only can recommend new projects that users are interested in, but also can improve the promotion effect of new projects, enhance the popularity and activity of project publishers, and promote the healthy development of social media websites and online communities.

[0062] The implementation process of the embodiments of the present invention will be described in detail below:

[0063] In step S1, candidate user log data and data of new projects to be promoted published by publishers are obtained. The specific implementation process is as follows:

[0064] The candidate user log data includes: users, projects with which the users have interaction relationships, the time of the interaction relationships, and specific interaction relationships, and the specific interaction relationships include click, like, and share behaviors.

[0065] The data of new projects to be promoted published by publishers includes: the publishers of the projects, the release time, the release content, etc.

[0066] In step S2, based on a pre-constructed multi-task model, the candidate user log data, and the data of new projects to be promoted, the probability that a candidate user shares the new project is obtained, and candidate users with the top-N sharing probabilities are selected as recommended users. The specific implementation process is as follows:

[0067] It should be noted that in the embodiments of the present invention, the construction process of the pre-constructed multi-task model is as follows:

[0068] A1. Obtain log data and project data of projects that interact with the users in the log data.

[0069] Specifically:

[0070] The log data includes: users, projects that interact with the users, timestamps, and labels indicating whether click, like, and share behaviors occur;

[0071] The project data includes: the publishers of the projects, the release time, the content, etc.

[0072] A2. Preprocess the log data and the project data to obtain training data, test data, and validation data. The training data, test data, and validation data all include the historical click data of users and new project data, and the new project data is the project generated within the corresponding time range of the training data, test data, and validation data. Specifically:

[0073] A201. Divide the log data by time. The click behaviors that occurred before a certain time point are regarded as the historical click data of the user, and the data after this time point is divided into training data, validation data, and test data according to the ratio of 6:2:2 by days.

[0074] A202. Obtain the relevant log records of new items in the modeling process involved in the training, validation, and test data respectively. Here, a new item refers to an item generated within the time range corresponding to each data. For example, if the training data is the log data generated during the period from November 1, 2020 to November 18, 2020, then the new items in the training set refer to the items generated during the period from November 1, 2020 to November 18, 2020.

[0075] A203. Process the training data to obtain training log data in the form of new item ID, user ID, whether clicked, whether liked, and whether shared; process the validation data and test data to obtain validation log data and test log data in the form of new items and the set of users who have shared the item.

[0076] The historical click data of the user is the user behavior data, that is, the sequence of items that the user has clicked in the past.

[0077] A3. Obtain the initial multi-task model. Specifically, it includes:

[0078] First, fuse different attribute features of new items based on the attention mechanism to obtain new item feature vectors, and at the same time learn the user's interest feature vectors from the user's click history based on the long short-term memory network (LSTM). Then concatenate the new item feature vectors with the user's interest feature vectors and input them into four multi-layer perceptron (MLP) layers to predict the probabilities of "exposure → click", "click → like", "like → share", and "other cases → share" respectively. Based on these four probabilities, further calculate the probabilities of "exposure → click → like" and "exposure → click → like / other cases → share".

[0079] Among them:

[0080] The process of fusing different attribute features of new items based on the attention mechanism to obtain new item feature vectors is achieved through new item feature modeling. Specifically:

[0081] In social media websites and online communities, items often have different attribute information. Although these attributes are all descriptions of the items, the attributes in different aspects often have different impacts on users. The attention model can be used to learn the different importance of different aspects of item attributes. Given the attribute a = [a1, a2,..., a M of the new item i, the specific modeling process is as follows:

[0082] First, the attributes of the new item need to be converted into an embedding vector through the embedding layer. This embedding vector is randomly initialized and will be continuously learned according to the training of the model. The embedding matrix of the new item's attributes can be expressed as

[0083]

[0084] where: d is the dimension of the embedding vector, M is the number of attributes of the new item, is the embedding vector of attribute a m .

[0085] Then, the attention mechanism is used to weight and fuse the different attribute features of the new item to obtain the new item feature vector. The specific formula is as follows:

[0086]

[0087]

[0088]

[0089] where: are the parameters of the attention network, d1 is the dimension of the attention network parameters; s(x) is the Leaky ReLU activation function. The obtained α m is normalized to obtain the importance α′ of the attributes in different aspects m . Then, weighted summation is performed to obtain the new item feature vector e i .

[0090] Based on the long short-term memory network (LSTM), the user's interest feature vector is learned from the user's click history through user feature modeling. Specifically:

[0091] The user's click history can well reflect the user's preferences. Here, the LSTM network is used for learning. Given the sequence of items {i1, i2,..., i N} that the user has clicked in the past, after passing through the embedding layer, the corresponding embedding vectors of the items can be obtained as After passing through the LSTM network, the user's interest feature vector e u can finally be obtained. The LSTM network includes an input gate, a forget gate, and an output gate. The specific calculation formulas are as follows:

[0092]

[0093] i t = σ(W i ·[h t-1 , x t +bi )

[0094] f t = σ(W f · [h t-1 ,x t + b f )

[0095] o t = σ(W o · [h t-1 ,x t + b o )

[0096]

[0097] h t = o t * tanh(c t )

[0098] Where: W c , W i , W f , W o are weight matrices, b c , b i , b f , b o are bias vectors; h t-1 and h t represent the hidden states at time t-1 and time t respectively; x t is the input vector at time t; is the candidate value, which is obtained through the tanh layer; c t-1 and c t represent the cell states at time t-1 and time t respectively; i t , f t , o t represent the input gate, forget gate and output gate respectively.

[0099] Concatenate the new item feature vector e i and the user's interest feature vector e u , and input them into four prediction networks respectively. Here, MLP is used as the prediction network, and the probabilities of "exposure → click", "click → like", "like → share", and "other cases → share" are predicted respectively. The four MLP layer structures are the same, but the parameters are not shared. The formula of the MLP layer is as follows:

[0100] z1 = ReLU(W1[e i , e u + b1)

[0101] ……

[0102] z L = ReLU(W L z L-1 + b L )

[0103] y i = σ(h T z L + b)

[0104] where: L represents the number of hidden layers in the multi-layer perceptron, W x , b x are the weight matrix and bias vector of the x-th hidden layer. The activation function used in the hidden layer is the ReLU function, which is beneficial for building a deep model. [e i , e u represents concatenating the new item feature vector e i and the user's interest feature vector e u . h and b are the weight vector and bias of the output layer. The activation function used in the output layer is the sigmoid function, and the predicted probability y i can be obtained. Through four prediction network modules, the predicted probabilities of "exposure → click", "click → like", "like → share", "other cases → share"

[0105] Then, predictions of click, like, and share behaviors are made:

[0106]

[0107]

[0108]

[0109] where: are respectively the predicted probability of "exposure → click", the predicted probability of "exposure → click → like", and the predicted probability of "exposure → click → like / other cases → share".

[0110] where: The calculation formulas for "exposure → click", "click → like", "like → share", "click → other cases", and "other cases → share" are as follows:

[0111] The user's behavior path can be decomposed into 5 basic paths, including "exposure → click", "click → like", "like → share", "click → other cases", and "other cases → share". The conditional probability calculation formula for each path is as follows:

[0112] (1) The probability of the "exposure → click" path describes the conditional probability that a user clicks on an item given that the item has been exposed to the user's view. It is expressed by the following formula:

[0113]

[0114] Where: c i ∈ {0, 1} defines whether item i is clicked, i ∈ [1, N], where N is the total number of items; v i ∈ {0, 1} indicates whether item i is seen by the user; y 1i is the symbol representing the conditional probability of "exposure → click".

[0115] (2) The probability of the "click → like" path describes the probability that a user likes an item given that the item has been clicked by the user. It is expressed by the following formula:

[0116]

[0117] Where: z i ∈ {0, 1} defines whether item i is liked, y 2i is the symbol representing the conditional probability of "click → like".

[0118] (3) The probability of the "like → share" path describes the probability that a user shares an item given that the item has been liked by the user. It is expressed by the following formula:

[0119]

[0120] Where: s i ∈ {0, 1} defines whether item i is shared, y 3i is the symbol representing the conditional probability of "like → share".

[0121] (4) The probability of the "click → other cases" path describes the probability that a user does not like an item given that the item has been clicked by the user. It is expressed by the following formula:

[0122] P(z i =0|v i =1, c i =1)=1 - y 2i

[0123] (5) The probability of the "other cases → share" path describes the probability that a user shares an item given that the item has been clicked by the user but not liked. It is expressed by the following formula:

[0124]

[0125] where: y 4i is the indicator of the conditional probability of "other cases → sharing".

[0126] Based on the probabilities of these 5 basic paths, the probability of "exposure → click → like" can be further calculated, and its calculation formula is:

[0127]

[0128]

[0129] The calculation formula of "exposure → click → like / other cases → sharing" is as follows:

[0130]

[0131] A4. Train, validate, and test the initial multi-task model based on the training data, validation data, and test data to obtain the final multi-task model.

[0132] When training the model, construct cross-entropy loss functions according to "exposure → click", "exposure → click → like", and "exposure → click → like / other cases → sharing" respectively, and then perform weighted summation on them to form the joint loss function of the model.

[0133] Specifically, the loss functions of the three parts of training include:

[0134] (1) The loss function corresponding to the click behavior is expressed as follows:

[0135]

[0136] where: C + and C - are the positive and negative sample sets of the click behavior.

[0137] (2) The loss function corresponding to the like behavior is expressed as follows:

[0138]

[0139] where: Z + and Z - are the positive and negative sample sets of the like behavior.

[0140] (3) The loss function corresponding to the sharing behavior is expressed as follows:

[0141]

[0142] where: S + and S - are the positive and negative sample sets of the sharing behavior.

[0143] The final training objective is to minimize the following combined loss function:

[0144]

[0145] where: w c 、w z 、w s are the weights of click, like, and share behaviors respectively, which satisfy w c +w z +w s = 1, and control the influence of different types of behaviors in the combined training.

[0146] Then use the Adam optimizer to train the model, and optimize the hyperparameters of the model according to the performance of the trained model on the validation data. The hyperparameters that can be optimized include the dimension d of the embedding vector; the dimension d1 of the attention network parameters; the dimension of the hidden layer of the long short-term memory network; the number L of hidden layers in the multi-layer perceptron; the dimension of the hidden layer in the multi-layer perceptron; the weights w c 、w z 、w s of click, like, and share behaviors in the loss function; the learning rate; the batch size.

[0147] During validation and testing, for each specific new project, the probability that all users share this new project is predicted by the multi-task model, and then sorted from large to small according to the probability values. The top N users are used as the recommendation list. Based on the feedback of the users, that is, the user set that has shared the project, relevant metrics such as accuracy and recall are further calculated to optimize the hyperparameters of the multi-task model and obtain the multi-task model.

[0148] As Figure 2 shown, the attributes of the historical data of the candidate users and the data of the new project to be promoted are input into the multi-task model to obtain the sharing probability of the candidate users, and the candidate users with the sharing probability of top-N are selected as the recommended users.

[0149] The embodiment of the present invention also provides a personalized recommendation system for new project promotion, and the system includes:

[0150] A data acquisition module for acquiring candidate user log data and the data of the new project to be promoted published by the publisher;

[0151] A recommendation module for obtaining the probability that a candidate user shares the new project based on a pre-constructed multi-task model, the candidate user log data, and the data of the new project to be promoted, and selecting the candidate users with the sharing probability of top-N as the recommended users.

[0152] It is understood that the personalized recommendation system for new project promotion provided by the embodiments of the present invention corresponds to the above-mentioned personalized recommendation method for new project promotion. For the explanations, examples, beneficial effects, etc. of the relevant content, reference can be made to the corresponding content in the personalized recommendation method for new project promotion, which will not be elaborated here.

[0153] The embodiments of the present invention also provide a computer-readable storage medium, which stores a computer program for personalized recommendation for new project promotion. Wherein, the computer program enables a computer to execute the above-mentioned personalized recommendation method for new project promotion.

[0154] The embodiments of the present invention also provide an electronic device, including: one or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and are configured to be executed by the one or more processors, and the programs include those for executing the above-mentioned personalized recommendation method for new project promotion.

[0155] In summary, compared with the prior art, the following beneficial effects are achieved:

[0156] 1. The embodiments of the present invention comprehensively consider the interests of users and project publishers, and recommend new projects to users who may take sharing actions. It can not only recommend new projects that users are interested in, but also improve the promotion effect of new projects, enhance the popularity and activity of project publishers, and promote the healthy development of social media websites and online communities.

[0157] 2. Since the sharing behavior is very sparse compared with click and like behaviors. Therefore, the embodiments of the present invention construct a multi-task model, which will not only use the sharing behavior, but also take into account the click and like behaviors, alleviating the problem of sparse sharing behavior, providing better support for the parameters in the training model, and improving the accuracy of the recommendation results.

[0158] 3. For new projects, due to the lack of interaction information and the cold start problem, usually only the attributes of the project can be used to learn the features of the new project. However, there are many attributes of the new project, and different attributes reflect different aspects of the project's features, so their importance is different. The present invention uses an attention mechanism to fuse different attribute features, can assign different weights to different attributes, and can obtain a more accurate feature representation of the new project.

[0159] It should be noted that through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a necessary general hardware platform. Based on such an understanding, the essence of the above technical solution, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0160] In this article, relational terms such as "first" and "second" are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements not only includes those elements, but also includes other elements not expressly listed, or further includes elements inherent to such process, method, article or device. Without further limitation, an element defined by the statement "including one..." does not exclude the existence of additional identical elements in the process, method, article or device including the said element.

[0161] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A personalized recommendation method for new project promotion, characterized in that The method includes: S1. Obtain candidate user log data and new project data to be promoted published by the publisher; S2. Based on a pre-constructed multi-task model, the candidate user log data, and the new project data to be promoted, obtain the probability that a candidate user shares the new project, and select the top-N candidate users with the sharing probability as the recommended users; Among them, the construction process of the pre-constructed multi-task model is as follows: A1. Obtain log data and project data that interact with the users in the log data; A2. Preprocess the log data and project data to obtain training data, test data, and validation data. The training data, test data, and validation data all include the user's historical click data and new project data. The new project data is the project generated within the corresponding time range in the training data, test data, and validation data; A3. Obtain an initial multi-task model; A4. Train, validate, and test the initial multi-task model based on the training data, validation data, and test data to obtain a multi-task model; The A3 includes: Fuse different attribute features of the new project based on the attention mechanism to obtain a new project feature vector, and learn the user's interest feature vector from the user's historical click data based on the long short-term memory network; splice the new project feature vector and the user's interest feature vector, and input them into four multi-layer perceptron layers to predict the probability of exposure → click, the probability of click → like, the probability of like → share, and the probability of other situations → share; calculate the probability of exposure → click → like and exposure → click → like / other situations → share based on the above four probabilities; The A4 includes: When training the multi-task model, construct cross-entropy loss functions according to exposure → click, exposure → click → like, and exposure → click → like / other situations → share respectively, and then perform weighted summation on them to form the joint loss function of the multi-task model; Use the Adam optimizer to train the model, and optimize the hyperparameters of the multi-task model according to the performance of the trained multi-task model on the validation data.

2. The personalized recommendation method for new project promotion according to claim 1, wherein The candidate user log data includes: users, projects that have an interaction relationship with the users, the time of the interaction relationship, and the specific interaction relationship; the specific interaction relationship includes click, like, and share behaviors.

3. The personalized recommendation method for new project promotion according to claim 1, wherein, The A2 includes: A201. Divide the log data according to time. The click behavior that occurred before the preset time point is regarded as the user's historical click data, and the data after the preset time point is divided into training data, validation data, and test data according to the ratio of days x:y:z; A202. Obtain the new project data in the training data, validation data, and test data respectively; A203. Process the training data to obtain training log data in the form of new project ID, user ID, whether clicked, whether liked, and whether shared; process the validation data and test data to obtain validation log data and test log data in the form of new projects and the user set that has shared the project.

4. The personalized recommendation method for new project promotion according to claim 1, characterized in that, The fusing different attribute features of the new project based on the attention mechanism to obtain a new project feature vector includes: Given the attributes a = [a1, a2, …, a of the new project i M ; Convert the attributes of the new item into an embedding vector through the embedding layer. The attribute embedding matrix of the new item is represented as where: d is the dimension of the embedding vector, and M is the number of new item attributes, is the attribute a m 's embedding vector; Use the attention mechanism to weight and fuse different attribute features of the new item to obtain the new item feature vector; The learning of the user's interest feature vector from the user's historical click data based on the long short-term memory network includes: Given the sequence of items {i1, i2, …, i N} clicked by the user; Convert the item sequence into an embedding vector through the embedding layer as The embedded vector is transformed into the user's interest feature vector e through the LSTM network u .

5. The personalized recommendation method for new project promotion according to claim 1, wherein, The joint loss function of the multi-task model is as follows; where: w c , w z , w s are the weights of click, like, and share behaviors respectively, and they satisfy w c + w z + w s = 1; is the loss function corresponding to the click behavior; is the loss function corresponding to the like behavior; is the loss function corresponding to the share behavior.

6. A personalized recommendation system for promoting new projects, characterized in that, The system includes: A data acquisition module for acquiring candidate user log data and new item data to be promoted published by the publisher; A recommendation module for obtaining the probability that a candidate user shares the new item based on a pre-constructed multi-task model, the candidate user log data, and the new item data to be promoted, and selecting the candidate users with the top-N sharing probabilities as the recommended users; Among them, the construction process of the pre-constructed multi-task model is as follows: A1. Obtain log data and item data that interact with the users in the log data; A2. Preprocess the log data and item data to obtain training data, test data, and validation data. The training data, test data, and validation data all include the user's historical click data and new item data. The new item data is the item generated within the corresponding time range in the training data, test data, and validation data; A3. Obtain an initial multi-task model; A4. Train, validate, and test the initial multi-task model based on the training data, validation data, and test data to obtain a multi-task model; The A3 includes: Fuse different attribute features of the new item based on the attention mechanism to obtain the new item feature vector, and learn the user's interest feature vector from the user's historical click data based on the long short-term memory network; concatenate the new item feature vector with the user's interest feature vector and input it into four multi-layer perceptron layers to respectively predict the probability of exposure → click, click → like, like → share, and other cases → share; calculate the probabilities of exposure → click → like and exposure → click → like / other cases → share based on the above four probabilities; The A4 includes: When training the multi-task model, construct cross-entropy loss functions according to exposure → click, exposure → click → like, and exposure → click → like / other cases → share respectively, and then perform weighted summation on them to form the joint loss function of the multi-task model; Use the Adam optimizer to train the model, and optimize the hyperparameters of the multi-task model according to the performance of the trained multi-task model on the validation data.

7. A computer-readable storage medium, characterized in that, It stores a computer program for personalized recommendation for new item promotion, wherein the computer program enables a computer to execute the personalized recommendation method for new item promotion as described in any one of claims 1 to 5.

8. An electronic device, characterized in that, Including: One or more processors; A memory; And One or more programs, wherein the one or more programs are stored in the memory and are configured to be executed by the one or more processors. The programs include the personalized recommendation method for new item promotion as described in any one of claims 1 to 5.

Images