Material pushing method, device, equipment and medium

Abstract

The embodiment of the invention provides a material pushing method and device, equipment and a medium, and the method comprises the steps: obtaining a plurality of preferable materials and a plurality of target users associated with a target content publisher, and determining the preferable material corresponding to each target user based on the association relationship between the plurality of preferable materials and the plurality of target users; aiming at all the preferable materials corresponding to each target user, respectively determining a user interest degree score corresponding to each preferable material, and selecting a target material corresponding to the target user from all the preferable materials based on the user interest degree score; performing polymerization processing on all target materials corresponding to each target user to obtain a polymer material corresponding to each target user; and pushing the corresponding polymer material to each target user.

Description

Technical Field

[0001] The present invention relates to the field of content recommendation, and in particular to a material push method, device, equipment and medium. Background Art

[0002] With the rapid growth of online interactive resources, more and more social messages are posted on social platforms. How to notify designated users of designated messages through various channels has become more challenging. Push can be used to attract activity through external reminders when the phone is locked or in other phone interfaces. Push has also become a method for more and more companies to achieve user growth.

[0003] In the social push scenario, push is mainly based on the emerging massive blog information, through the user's social relationship and social interaction, to provide personalized push to users. Common personalized push usually filters information and predicts the information that users may be interested in before sending it. The information that can be displayed by personalized push is very limited, and frequent push will also cause a very bad user experience. Therefore, in most business scenarios, personalized push is mostly calculated offline for users and push content, and push is performed at fixed time periods to achieve more accurate recommendations.

[0004] In the Weibo push scenario, the most commonly used social push content is to build a model for users, materials, and cross-features, and then enter the multi-business mixed sorting for distribution. First, materials are recalled from the material pool. The material pool contains millions of candidate materials. It is necessary to recall the Weibo blog materials that users are interested in through different recall methods and store them in the recall pool. In the offline model training phase, features are extracted and samples are generated through the materials in the material pool. Then, samples are used for offline training. Model calculation and sorting are performed on the materials recalled every day, so as to sort the materials of each business. Finally, the materials to be distributed by each business are sorted together in the multi-business mixed sorting, and finally the best blog content is decided to be distributed.

[0005] In the process of realizing the present invention, the applicant found that how to screen high-quality materials suitable for users and push as many high-quality materials as possible to users under the condition of limited push times has become one of the technical problems to be solved urgently in the prior art. Summary of the invention

[0006] The embodiments of the present invention provide a material pushing method, device, equipment and medium, which can screen out high-quality materials suitable for users and push as many high-quality materials as possible to users under a limited number of push times.

[0007] To achieve the above-mentioned purpose, in a first aspect, an embodiment of the present invention provides a material pushing method, comprising:

[0008] Acquire multiple preferred materials and multiple target users associated with a target content publisher, and determine the preferred materials corresponding to each target user based on the association relationship between the multiple preferred materials and the multiple target users; the target content publisher is a content publisher that meets a preset blogger screening condition;

[0009] For all preferred materials corresponding to each target user, determine the user interest score corresponding to each preferred material, and select the target material corresponding to the target user from all preferred materials based on the user interest score;

[0010] Aggregate all target materials corresponding to each target user to obtain a piece of aggregated material corresponding to each target user;

[0011] Push the corresponding aggregated materials to each target user.

[0012] In a second aspect, an embodiment of the present invention provides a material pushing device, comprising:

[0013] A user material screening unit is used to obtain a plurality of preferred materials and a plurality of target users associated with a target content publisher, and determine a preferred material corresponding to each target user based on the association relationship between the plurality of preferred materials and the plurality of target users; the target content publisher is a content publisher that meets a preset blogger screening condition;

[0014] A material sorting and selecting unit, for determining, for all preferred materials corresponding to each target user, a user interest score corresponding to each preferred material, and selecting a target material corresponding to the target user from all preferred materials based on the user interest score;

[0015] The material aggregation unit is used to aggregate all target materials corresponding to each target user to obtain a piece of aggregated material corresponding to each target user;

[0016] The material pushing unit is used to push the corresponding polymer materials to each target user.

[0017] In a third aspect, an embodiment of the present invention provides a computer device, comprising at least one host, wherein the at least one host comprises a processor adapted to execute various instructions; and

[0018] A storage device is suitable for storing a plurality of instructions, wherein the instructions are suitable for being loaded by a processor and executing any of the above methods.

[0019] In a fourth aspect, an embodiment of the present invention provides a storage medium, in which a plurality of instructions are stored, and the instructions are suitable for being loaded by a processor and executing any of the methods described above.

[0020] The above technical solution has the following beneficial effects: through the attention relationship and publishing relationship between the content publisher (bloggers), materials (content published by bloggers), and users, the preferred materials corresponding to each target user are determined, and the user interest score of the preferred materials is predicted, and the target materials to be sent to the target users are determined according to the user interest score, so as to achieve the purpose of pushing high-quality materials to the target users. Before pushing, the target materials corresponding to the target users are aggregated into one aggregate material, so that multiple high-quality materials can be pushed to the target users in one push. Compared with the traditional push method, the overall computing system overhead of this application is relatively small, and more high-quality materials can be sent to users at one time to increase the amount of materials sent by social push; through the method of multi-material aggregate push, more high-quality materials can be distributed to users when more materials are recalled, thereby improving the quality of social sending; it can break through the push limit number of manufacturers such as client devices, and the multi-material automatic (aggregate) push method aggregates multiple materials in the intimacy materials that users pay attention to and sends them, breaking through the traditional limit on the number of sent items and avoiding the omission of high-quality materials; based on the social scenario, the method of constructing multi-material aggregate push is aimed at the problem of excessive and too frequent push commonly raised by users, and can well improve the user experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] 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 work.

[0022] Figure 1 This is a flow chart of a material pushing method according to one embodiment of the present invention;

[0023] Figure 2 This is a structural diagram of a material pushing device according to one embodiment of the present invention;

[0024] Figure 3 It is a flow chart of the material pushing method of the prior art;

[0025] Figure 4 This is a schematic diagram of a polymer material according to one embodiment of the present invention displayed on a user's client;

[0026] Figure 5 is another flow chart of a material pushing method according to one embodiment of the present invention;

[0027] Figure 6 This is another flow chart of a material pushing method according to one embodiment of the present invention. DETAILED DESCRIPTION

[0028] 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, not all of the embodiments. 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.

[0029] Figure 3 In the material push process under the existing technology, materials are recalled from the material pool and stored in the recall pool. During the offline model training phase, features are extracted and samples are generated through the materials in the material pool. Then, the samples are used for offline training, and the materials recalled every day are used for model calculation and sorting, so as to sort the materials of each business. Finally, the materials to be sent out by each business are sorted together in a multi-business mixed sorting process, and finally the best blog post content is decided to be sent out.

[0030] Through analysis of the prior art, the inventors found that the prior art has the following disadvantages: the number of material issuances is large, and multiple push notifications are displayed in one day, causing user dissatisfaction; there are too many materials competing, and some high-quality materials cannot compete and cannot reach users; during the sorting stage, many full-time or high-frequency users can recall many materials, and it is difficult to cover many high-quality materials; manufacturers limit the number of push notifications, and too many materials cannot be issued; when there are too many samples, the material sorting is not ideal, and the best quality materials may not be found. In view of the above shortcomings, the inventors provide various embodiments of the present invention to solve the following problems: in the social push scenario, for full attendance or high-frequency users, the user recall rate is very high, and there are many high-quality blog materials. Through aggregated push, more high-quality materials can be reached to users, thereby increasing the amount of materials sent by social push; in response to manufacturers' restrictions on the number of Push items, an automated aggregated Push method is proposed, which aggregates two or more intimacy materials that users are concerned about and sends them out, breaking through the traditional limit on the number of items sent and avoiding the omission of high-quality materials; a solution is proposed for using the DeepFM model to sort and select multiple blog materials, and model calculation and learning are performed using multi-dimensional features on the blogger side, the user side, the blogger to user U2U (user to user) side, and the blog material side, so that the model is more accurate and can recommend higher-quality materials.

[0031] On the one hand, if Figure 1 As shown, an embodiment of the present invention provides a material pushing method, including:

[0032] Step S10, obtaining multiple preferred materials and multiple target users associated with a target content publisher, and determining the preferred materials corresponding to each target user based on the association relationship between the multiple preferred materials and the multiple target users; the target content publisher is a content publisher that meets the preset blogger screening condition;

[0033] Step S11, for all preferred materials corresponding to each target user, respectively determine the user interest score corresponding to each preferred material, and select the target material corresponding to the target user from all preferred materials based on the user interest score;

[0034] Step S12, performing aggregation processing on all target materials corresponding to each target user to obtain a piece of aggregated material corresponding to each target user;

[0035] Step S13, pushing the corresponding aggregated material to each target user.

[0036] In some embodiments, the preferred material corresponding to the target user can be determined according to the user characteristics and material characteristics, and the matching relationship between the target user and the preferred material can also be established according to the attention relationship between the content publisher and the user and the publishing relationship between the material and the content publisher. The material can be selected according to certain preset preferred material conditions. The preset preferred material conditions can be set according to different specific business settings. For example, materials with high click-through rates can be selected as preferred materials, and the latest materials or materials with high forwarding rates or original materials can be selected as preferred materials. The target user can be all users of the platform, or users within a certain range obtained based on certain rules. Each target user may match many preferred materials. For each target user, these preferred materials can be further scored and sorted. There are many scoring methods, such as scoring according to one or more specified indicators, and then combining the scores of all indicators to obtain the (user interest level) score of the preferred material. The relevant characteristics of the material can also be obtained, and the (user interest level) score is predicted by the pre-trained model to obtain the (user interest level) score of the preferred material. The target material of each target user is further selected according to the (user interest level) score. The target material of each target user is the material that will be sent to the target user in the end. There may be many of them. In order to reduce the number of pushes, all the target materials of each target user can be aggregated to obtain one aggregate material for each target user, and the aggregate material can be pushed to the corresponding target user. One aggregate material can actually send multiple target materials to the user in one push.

[0037] The embodiments of the present invention have the following technical effects: compared with the traditional push method, the overall computing system overhead of the embodiments of the present invention is relatively small, and more high-quality materials can be sent at one time to reach users, thereby increasing the amount of materials sent through social push; through the method of multi-material aggregate push, more high-quality materials can be distributed to users when more materials are recalled, thereby improving the quality of social sending; it can break through the push limit number of manufacturers such as client devices, and the multi-material automatic (aggregate) push method aggregates 2 or more intimacy materials that users are concerned about and sends them, breaking through the traditional limit on the number of sent items and avoiding the omission of high-quality materials; based on social scenarios, a method for constructing a multi-material aggregate push is aimed at the problems of excessive and too frequent push commonly raised by users, and can greatly improve the user experience.

[0038] Furthermore, the acquiring of a plurality of preferred materials and a plurality of target users associated with a target content publisher, and determining a preferred material corresponding to each target user based on the association relationship between the plurality of preferred materials and the plurality of target users, includes:

[0039] Obtain multiple candidate materials and corresponding material information that meet preset material screening conditions in the material pool;

[0040] Determine the content publisher corresponding to each candidate material according to the material information corresponding to each candidate material;

[0041] Among all candidate materials, select the candidate material whose corresponding content publisher belongs to the target content publisher as the preferred material;

[0042] Determine the target user associated with each preferred material according to the target content publisher corresponding to each preferred material and the established correspondence between the target content publisher and the user who follows the target content publisher;

[0043] Based on the multiple preferred materials and the target users associated with each preferred material, a preferred material corresponding to each target user is determined.

[0044] In some embodiments, the materials in the material pool are updated regularly to maintain the timeliness and popularity of the materials in the material pool. The multiple candidate materials obtained from the material pool may be from one content publisher. Usually, the multiple candidate materials come from multiple different content publishers; the material information of the material records the content publisher corresponding to the material. The target content publisher may be all content publishers (all bloggers) or a content publisher is preferred according to a preset content publisher preference condition, such as the number of fans is greater than a preset fan number threshold, the number of content published is greater than a preset content number threshold, and / or the activity is greater than a preset activity threshold, etc. The attention relationship between all content publishers and users who follow the content publisher can be updated and saved periodically (so that the corresponding relationship between the target content publisher and the users who follow the target content publisher can be obtained), each preferred material corresponds to a target content publisher, and each target content publisher corresponds to one or more users who follow the target content publisher, so that one or more target users associated with each preferred material can be determined; using the association relationship between all preferred materials and one or more target users associated with the preferred material, one or more preferred materials associated with each target user can be determined.

[0045] Furthermore, the preset material screening conditions include at least one of the following:

[0046] Materials corresponding to original content released within the most recent preset time frame or new content generated by directly forwarding original content;

[0047] Materials with a click rate greater than or equal to the preset business requirement click rate. The click rate comes from the material information corresponding to the corresponding material; the click rate in the material information can be extracted from the log data.

[0048] In some embodiments, the materials are limited to those published within a recent preset time range, such as those published within the last day, to ensure that the materials pushed to the user are the latest materials; original content is selected for push to ensure the instructions for pushing materials and avoid pushing duplicate materials; it is generally believed that new content formed by directly forwarding original content, such as content that is 1-reposted for original content, is the original commentary content of the user who forwards the original content, and has higher content quality than 2-reposted and 3-reposted content. 1-repost content is defined as new content formed by directly forwarding original content, 2-repost content is defined as new content formed by forwarding 1-repost content, and 3-repost content is defined as new content formed by forwarding 2-repost content. According to specific business needs, the minimum preset business demand click-through rate corresponding to the business is determined. Before estimating and sorting the screened materials, the materials with click-through rates less than the preset business demand are filtered out to reduce invalid work in subsequent processing.

[0049] Furthermore, the acquiring of a plurality of preferred materials and a plurality of target users associated with a target content publisher, and determining a preferred material corresponding to each target user based on the association relationship between the plurality of preferred materials and the plurality of target users, includes:

[0050] Acquire at least one content publisher that meets a preset content publisher condition as a target content publisher;

[0051] For each target content publisher, obtain users who follow the target content publisher, and establish a corresponding relationship between the target content publisher and the user;

[0052] Obtain at least one candidate material and corresponding material information that meets preset preferred material conditions in the material pool;

[0053] For each candidate material, determining a content publisher of the candidate material according to the material information of the candidate material;

[0054] Selecting, from among all candidate materials, the candidate materials of the content publisher in the target content publisher as the preferred materials;

[0055] For each preferred material, according to the correspondence between the target content publisher and the user, the preferred material is allocated to the user corresponding to the target content publisher that published the preferred material, and the user to whom the preferred material is allocated is taken as the target user.

[0056] In some embodiments, in the scenario of sending push based on social relationship, the number of push items should be limited for manufacturers (such as mobile phone manufacturers or mobile phone systems or platform service providers, etc.). When there are sufficient materials, more and better materials will be aggregated and sent to users for personalized Push. At least one content publisher that meets the preset content publisher conditions is obtained as the target content publisher; the preset content publisher conditions can be all content publishers on the platform, or some content publishers, for example, content publishers that are active within the recent specified date range, or content publishers whose number of fans reaches a certain threshold, etc. Based on social relationships, according to the preset social relationship update cycle, such as daily, the blogger (target content publisher) recalls the user, that is, according to the user's attention relationship and intimacy relationship with the target content publisher, the corresponding relationship between the target content publisher and the user is determined, and the blogger side (content publisher side) and user side information are recalled, and the corresponding relationship between all target content publishers and users and the corresponding content publisher side of each target publisher and the user side information of each user are written into the database (such as Redis). Through each round of material mining, the preferred materials (high-quality materials) within one day are found, and at least one alternative material and corresponding material information that meet the preset preferred material conditions are obtained in the material pool; for each alternative material, the content publisher of the alternative material is determined according to the material information of the alternative material; among all the alternative materials, the alternative materials of the content publisher in the target content publisher are selected as the preferred materials; for each preferred material, according to the corresponding relationship between the target content publisher and the user, the preferred material is allocated to the user corresponding to the target content publisher who published the preferred material, and the user to whom the preferred material is allocated is used as the target user; preferably, the preset preferred material conditions include the latest one-day timeliness, originality and\or 1-reposted materials; the latest one-day timeliness indicates the materials published on the latest day, originality indicates the materials originally created by the content publisher himself, and 1-reposted materials are the materials forwarded for the first time. In an embodiment of the present invention, at least one content publisher that meets preset content publisher conditions can be obtained as a target content publisher according to a preset social relationship update cycle; for each target content publisher, users who follow the target content publisher are obtained, and a corresponding relationship between the target content publisher and the user is established. Two steps are independently deployed. The two steps of obtaining at least one candidate material that meets preset preferred material conditions and corresponding material information in a material pool for discovering materials and determining the content publisher of each candidate material according to the material information of the candidate material are independently deployed, so that the corresponding relationship between the target content publisher and the user and the mining of materials are independently decoupled and run, which is convenient for deployment and maintenance.The step of selecting the candidate materials of the content publisher in the target content publisher as the preferred materials from all the candidate materials, and the step of assigning the preferred materials to the users corresponding to the target content publisher who published the preferred materials according to the corresponding relationship between the target content publisher and the users for each preferred material, and taking the users assigned the preferred materials as the target users, and determining the list of preferred materials corresponding to each target user. Then, multiple target materials suitable for the target users are found through sorting, and the aggregated materials are automatically constructed to form the aggregated copywriting and other parameter information, and finally they are issued through transparent transmission, and multiple materials are displayed on the user's mobile phone.

[0057] The embodiment of the present invention has the following technical effects: Based on the social scenario, a method for constructing a multi-material aggregate push is proposed to break through the manufacturer's limit on the number of items. At the same time, the aggregation system can also solve the problem of excessive and too frequent pushes that users generally feedback. In each round of distribution, the latest time-level partitioned high-quality materials are obtained to quickly reach users, improve the quality of social distribution, and avoid the omission of high-quality materials. The proposed multi-material aggregate distribution is particularly effective for some high-frequency or full-time users. High-frequency users have a large amount of attention, many user behaviors, and more candidate materials. The user recall rate is high and the ranking competitiveness is strong. In each round of recall, users have a large number of optional materials to be distributed, and many high-quality materials cannot reach all the fans of the blogger (content publisher). Through the material push method of the embodiment of the present invention, more high-quality materials can be distributed at one time, with more user behavior data, and the overall coverage of materials in social Push can be improved. The aggregate push business is also applicable to the multi-material push processing in other company business scenarios. In other business scenarios, the multi-material automatic aggregate push method of the embodiment of the present invention can also be promoted to push when encountering the problem of multiple materials. The embodiment of the present invention is also suitable for performing targeted optimization of social push for key groups after segmenting users according to user activity, time, field, etc. If the amount of materials recalled by the user is small, it will be sent according to the traditional push. If there are many high-quality materials, multiple materials will be aggregated and sent.

[0058] In some other embodiments, the acquiring of multiple preferred materials and multiple target users associated with a target content publisher, and determining the preferred materials corresponding to each target user based on the association relationship between the multiple preferred materials and the multiple target users, includes:

[0059] Acquire at least one content publisher that meets a preset content publisher condition as a target content publisher;

[0060] For each target content publisher, obtain users who follow the target content publisher, and establish a corresponding relationship between the target content publisher and the user;

[0061] Obtain at least one candidate material and corresponding material information that meets preset preferred material conditions in the material pool;

[0062] According to the preset business demand conditions, the candidate materials are filtered to obtain the filtered candidate materials;

[0063] For each filtered candidate material, determining a content publisher of the filtered candidate material according to the material information of the filtered candidate material;

[0064] Selecting, from among all the filtered candidate materials, the filtered candidate materials of the content publisher among the target content publishers as the preferred materials;

[0065] For each preferred material, according to the correspondence between the target content publisher and the user, the preferred material is allocated to the user corresponding to the target content publisher that published the preferred material, and the user to whom the preferred material is allocated is taken as the target user.

[0066] In some embodiments, at least one candidate material and corresponding material information that meets the preset preferred material conditions are obtained in the material pool through the steps, the candidate materials are initially determined, and after the specific business module obtains the initially determined candidate materials, the candidate materials can be filtered and screened again according to the preset business demand conditions of the specific business module to obtain the candidate materials that meet the needs of the specific business module. For example, the candidate materials with a click count greater than a preset click value or a click rate greater than a preset click rate value are retained as the filtered candidate materials, for example, the user uid is used as the key, the characteristic information such as the number of issued items and the number of opened items of each material is obtained, and the materials are screened. When the material clicks on the day are greater than 5 or the click rate is greater than 1%, it is used as the filtered candidate material, and it is considered that such a material has a high click trend. The step of obtaining at least one content publisher that meets the preset content publisher conditions as the target content publisher, and the step of obtaining users who follow the target content publisher for each target content publisher, and establishing the corresponding relationship between the target content publisher and the user can be independently deployed to determine the corresponding relationship between the target content publisher and the user; the step of obtaining at least one candidate material and corresponding material information that meet the preset preferred material conditions in the material pool, and the step of filtering the candidate materials according to the preset business demand conditions to obtain the filtered candidate materials can be independently deployed to determine the initial candidate materials; but each business module may have its own needs and restrictions, and each business module can obtain the corresponding relationship data between the target content publisher and the user, the initial candidate material data, and then filter the candidate materials according to the specific needs of the business module to obtain the filtered candidate materials, and further optimize the preferred materials required by this business module. The establishment of the corresponding relationship between the target content publisher and the user, the selection of the initial candidate materials, and the final determination of the aggregated materials and push by each specific business module can all be independently deployed and maintained, which is conducive to the decoupling of the system.

[0067] The embodiments of the present invention have the following technical effects: it allows filtering of candidate materials according to the needs of specific business modules, and further determines preferred materials. Different preferred materials can be determined according to different businesses, and more suitable push notifications can be provided for users who use the response business.

[0068] In some other embodiments, the acquiring of multiple preferred materials and multiple target users associated with a target content publisher, and determining the preferred materials corresponding to each target user based on the association relationship between the multiple preferred materials and the multiple target users, includes:

[0069] Acquire at least one content publisher that meets a preset content publisher condition;

[0070] For each content publisher, at least one material published by the content publisher and meeting preset preferred material conditions is obtained as a preferred material, and the preferred material is allocated to users who follow the content publisher, and the users to whom the preferred material is allocated are taken as target users.

[0071] Furthermore, for all preferred materials corresponding to each target user, determining the user interest score corresponding to each preferred material, and selecting the target material corresponding to the target user from all preferred materials based on the user interest score, includes:

[0072] For each preferred material corresponding to any target user, obtain feature information associated with the preferred material, and vectorize the feature information to obtain a feature vector associated with the preferred material, input the feature vector associated with the preferred material into a pre-trained DeepFM model to predict the click-through rate of the preferred material, and obtain a user interest score corresponding to the preferred material output by the DeepFM model; the user interest score is used to characterize the click-through rate prediction value corresponding to the corresponding material; the DeepFM model is composed of a factor decomposition model and a deep neural network model, and can simultaneously learn low-order feature combinations and high-order feature combinations to predict the click-through rate;

[0073] A preset number of preferred materials whose user interest scores are greater than a preset score threshold and which are ranked high according to the user interest scores are selected as target materials corresponding to the target user.

[0074] In some embodiments, feature information includes user-side information (e.g., user gender, age, and address area), material-side information (e.g., material length, word count, domain label, etc.), and blogger-side information (e.g., number of fans, number of clicks on blogger's blog posts, etc.). Based on the generated target user data corresponding to each preferred material, one or more preferred materials corresponding to each target user are determined in turn. In the sorting stage, all preferred materials for each target user are sorted based on the target user. In the material sorting stage, the DeepFM model is used for sorting calculations. DeepFM is derived from the FM model and consists of two parts: the FM structure and the DNN structure. The bottom part shares the same feature Embedding input and learns low-order and high-order feature intersections at the same time. FM performs low-order combinations between features, and DNN performs high-order combinations between features. The two methods are combined in parallel so that the final architecture can train the model faster and more accurately. The parameters of all models can also be obtained through joint training of the loss function, as shown in formula (1):

[0075]

[0076] where yFM is the output of the FM part, y DNN is the output of the DNN part.

[0077] According to the material information, the blogger side information (content publisher side information), user information, content publisher and user relationship information, etc. are obtained, and the blogger side features (content publisher side features), user side features, material side features and blogger to user U2U features (content publisher to user U2U features), etc. are converted accordingly, and the 3245-dimensional features used in the material sorting of the aggregate push are formed through splicing and other operations. During the model training period, the embodiment of the present invention accumulates about 3 months of user click behavior data for the push service to generate samples, and offline training generates the model. If the user clicks on a push sent by the aggregate push, then the two or more blog posts contained in this aggregate push are counted as having one click, and the user may be interested in both blog posts. The feature combination part of the FM model can learn the interaction between different features through vectorized crossover, which is more advantageous when processing large-scale sparse features. In the embodiment of the present invention, some feature dimensions are also relatively sparse. In the data scenario of high coefficients, the part of the application of the FM model solves this problem better. The output of the FM part is an addition unit that reflects the first-order linear combination features, and the cumulative sum of many innerproduct units that use the inner product of the feature vector to extract the second-order feature combination, as shown in formula (2):

[0078]

[0079] Among them, w is the parameter to be learned, which is the first-order parameter feature; V i and V j It is a second-order combined feature. For feature x, there are combined features of i and j. j It represents the j-th dimension feature, and n represents the dimension of the feature.

[0080] The model calculates and sorts the user's scoring results for each material. Then, the materials that have been read and sent are filtered, and a Tree set structure is established to store and sort all the materials involved in the sorting for each user. The embodiment of the present invention screens the sorting and scoring results, sorts and selects the best two or more materials to aggregate and send to the user. The final output of this step is the sending user, the aggregated mid list, and the scoring list. According to the sorting score, the material click-through volume and click-through rate are limited, so as to screen out better target materials and enable high-quality target materials to quickly reach target users.

[0081] Furthermore, all target materials corresponding to each target user are aggregated to obtain a polymerized material corresponding to each target user, including:

[0082] For each target user, according to the content of each target material of the target user, a corresponding title is added to each target material;

[0083] All target materials and corresponding titles corresponding to the target user are spliced ​​together to obtain a polymer material corresponding to the target user.

[0084] In some embodiments, the embodiments of the present invention complete the automatic aggregation of multiple materials, as well as the splicing of material formats, recall and distribute to users according to materials, and process material titles from the material library. According to the amount of materials that users can recall and the quality of materials, materials are automatically aggregated. When the amount of materials that users can recall is large and the quality of materials is good, multiple suitable materials are automatically aggregated and distributed to users. When the number of materials is small, a single material is distributed to users. Different rounds of distribution attempts are made according to the frequency of user activity. For more active users such as full attendance, they may recall more related materials themselves. In the case of sufficient materials, more rounds of aggregate distribution can be performed within a day to achieve the purpose of distributing more materials to users. For less active users such as low-frequency and new users, the embodiments of the present invention will aggregate and distribute to these users during the time period of active blogging. During these time periods, blog materials are relatively sufficient, and it is more likely to meet the requirements of aggregate distribution, and break the number limit without affecting the distribution of materials by other businesses. Different from the traditional push distribution, the aggregate push needs to find two or more materials suitable for the user. It is necessary to ensure that the user has enough candidate materials, because the aggregate push is designed to solve the manufacturer's restrictions on the number of push items. If the user does not have particularly sufficient materials, the automated method of the embodiment of the present invention determines that the length of the material list recalled by the user list.size = 1, and then sends it in single mode; if the length of the material list recalled by the user list.size>1, then the Top2 or TopN suitable materials are intercepted according to the scenario for automatic aggregation. Compared with the traditional single push delivery, the aggregate push in the embodiment of the present invention is very different mainly in display. Specifically, all material tags of a single push are single material information, and the log data format is the tag of a single material, while in the aggregate push, it is necessary to record multiple blog material information at the same time to facilitate downstream and subsequent analysis of data processing. The format of recording the log data tag lfid is as follows:

[0085] lfid: lfid-prefix_${mid1},${mid2},${mid3}

[0086] Among them, for the log data tag lfid of the material, it is necessary to record the unique id corresponding to each material, that is, mid, in series into the aggregate material to form a format of a prefix plus multiple material ids (mid).

[0087] The scheme log parameter format also needs to be modified and added accordingly. All tags related to material information, such as mid list, blogger type list, blogger ID list, etc., need to add multiple material information. The text of the aggregate push is as follows:

[0088] Your multiple followers post

[0089] ①@Nickname 1: Weibo post…(picture)>>\n

[0090] ②@Nickname 2: Weibo post…(picture)>>

[0091] Figure 4 A schematic diagram of aggregate push display is given. Bloggers 1 and 2 followed by the user have posted blogs. "xxx" and "yyy" represent the main text of the Weibo posts of bloggers 1 and 2 respectively.

[0092] At the same time Figure 4 It is shown in the figure that the embodiment of the present invention also achieves corresponding optimization in the copywriting. The title shows your "your multiple followers post a blog", and the nicknames and contents of multiple materials are displayed in the copywriting. In order to display the copywriting, it is necessary to re-acquire the detailed information of the material, because in order to save computing resources in the calculation and sorting stages, the material title and other information are usually not obtained at the same time. When generating the copywriting, the material title and other information can be recalled again, and the obtained title and copywriting can be spliced ​​and stored.

[0093] After obtaining all aggregated materials and corresponding user-side information, the embodiment of the present invention performs the process of recalling aggregated materials before final delivery. Generate multiple aggregated pushes in the json format corresponding to the delivery user list, and deliver them through the downstream module. The json format contains the parameter information of the aggregated materials and the parameters of the corresponding delivery users. The data format is as follows:

[0094] {"msg":"

[0095] @Nickname 1: Weibo text...(Picture)>>\n②@Nickname 2: Weibo text...(Picture)>>",

[0096] "title":"Your multiple followers post a blog",

[0097] "toids":"uid1,uid2…………",

[0098] "url":"sinaweibo: / / xxxxxxxx"

[0099] ………………}

[0100] The aggregate push material is used as a unit, and each aggregate push material corresponds to multiple users. The embodiment of the present invention also applies split materials to avoid processing too many materials on the same node. For the nodes with a large number of users corresponding to the same aggregate material, slices are stored to facilitate the allocation of computing resources in the subsequent reduce stage, and the calculations are distributed on different nodes. And it is stored in a database such as Redis according to the above specific format, and merged with a large number of related materials every day to remove duplicates.

[0101] Furthermore, the pushing of the corresponding aggregated material to each target user includes:

[0102] In response to a click operation on the aggregated push by any target user, a polymerized material corresponding to the target user is sent to the target user's client, so that the client decomposes the polymerized material into all target materials corresponding to the target user, and displays all target materials in the target user's client.

[0103] In some embodiments, a traditional push is sent, and when the user clicks on the push, it will directly fall into the internal flow of the app page and display the blog post. Here, you can choose to fall into the blog post details page or the data stream for simple display according to the business form. The aggregate push of the embodiment of the present invention needs to display two or more blog post contents at the same time, so the details page that displays several blog post materials at the same time is obviously not suitable for the aggregated scenario. Based on the social scenario, the embodiment of the present invention drops the aggregated material into the relationship flow of the app. When the user clicks on the aggregate push, the downstream operation will be directly jumped to the relationship flow inside the app, and the corresponding two or more blog post materials will be directly and simply displayed in the flow of the user client, so that the user can directly see the multiple blog post contents of interest. Among them, the relationship flow is the content flow information display that can be scrolled and browsed based on the attention relationship generated on the microblog client. The final display stage requires in-stream display support. In other apps that apply the multi-material automatic aggregate push method of the embodiment of the present invention, when the push falls into the flow for display, it is necessary to support the display of multiple content, and match the corresponding flow display transformation top function to better apply the embodiment of the present invention.

[0104] Second, as Figure 2 As shown, an embodiment of the present invention provides a material pushing device, comprising:

[0105] The user material screening unit 200 is used to obtain multiple preferred materials and multiple target users associated with a target content publisher, and determine the preferred materials corresponding to each target user based on the association relationship between the multiple preferred materials and the multiple target users; the target content publisher is a content publisher that meets the preset blogger screening conditions;

[0106] The material sorting and selecting unit 201 is used to determine the user interest score corresponding to each preferred material for each target user, and select the target material corresponding to the target user from all the preferred materials based on the user interest score;

[0107] The material aggregation unit 202 is used to aggregate all target materials corresponding to each target user to obtain a piece of aggregated material corresponding to each target user;

[0108] The material pushing unit 203 is used to push the corresponding polymerized material to each target user.

[0109] Furthermore, the user material screening unit 200 includes:

[0110] The candidate material acquisition module is used to obtain multiple candidate materials and corresponding material information that meet the preset material screening conditions in the material pool;

[0111] A candidate material content publisher determination module is used to determine the content publisher corresponding to each candidate material according to the material information corresponding to each candidate material;

[0112] A preferred material selection module is used to select, from among all candidate materials, a candidate material whose corresponding content publisher belongs to the target content publisher as the preferred material;

[0113] A preferred material associated user determination module, used to determine the target user associated with each preferred material according to the target content publisher corresponding to each preferred material and the established correspondence between the target content publisher and the user who follows the target content publisher;

[0114] The target user preferred material determination module is used to determine the preferred material corresponding to each target user based on the multiple preferred materials and the target users associated with each preferred material.

[0115] Furthermore, the preset material screening conditions include at least one of the following:

[0116] Materials corresponding to original content released within the most recent preset time frame or new content generated by directly forwarding original content;

[0117] Materials with a click rate greater than or equal to the preset business requirement click rate. The click rate comes from the material information corresponding to the corresponding material; the click rate in the material information can be extracted from the log data.

[0118] Furthermore, the user material screening unit 200 includes:

[0119] A first publisher screening module, used to obtain at least one content publisher that meets a preset content publisher condition as a target content publisher;

[0120] A first publisher-user relationship determination module, configured to obtain, for each target content publisher, users who follow the target content publisher and establish a corresponding relationship between the target content publisher and the user;

[0121] A first material screening module, used to obtain at least one candidate material and corresponding material information that meets a preset preferred material condition in a material pool;

[0122] A first material publisher relationship determination module, configured to determine, for each candidate material, a content publisher of the candidate material according to the material information of the candidate material;

[0123] A first preferred material determination module is used to select, from among all candidate materials, the candidate material of the content publisher in the target content publisher as the preferred material;

[0124] The first target user and material matching module is used to assign each preferred material to the user corresponding to the target content publisher who published the preferred material according to the corresponding relationship between the target content publisher and the user, and to use the user to whom the preferred material is assigned as the target user.

[0125] Furthermore, the user material screening unit 200 includes:

[0126] A second publisher screening module, used to obtain at least one content publisher that meets a preset content publisher condition as a target content publisher;

[0127] A second publisher-user relationship determination module is used to obtain, for each target content publisher, users who follow the target content publisher and establish a corresponding relationship between the target content publisher and the user;

[0128] A second material screening module is used to obtain at least one candidate material and corresponding material information that meets the preset preferred material conditions in the material pool;

[0129] The business demand filtering module is used to filter the candidate materials according to the preset business demand conditions to obtain the filtered candidate materials;

[0130] A second material publisher relationship determination module is used to determine, for each filtered candidate material, a content publisher of the filtered candidate material according to the material information of the filtered candidate material;

[0131] A second preferred material determination module is used to select, from among all filtered candidate materials, the filtered candidate materials of the content publisher among the target content publishers as the preferred materials;

[0132] The second target user and material matching module is used to assign each preferred material to the user corresponding to the target content publisher who published the preferred material according to the corresponding relationship between the target content publisher and the user, and to use the user to whom the preferred material is assigned as the target user.

[0133] Furthermore, the user material screening unit 200 is configured to: obtain at least one content publisher that meets the preset content publisher conditions; for each content publisher, obtain at least one material published by the content publisher and meeting the preset preferred material conditions as a preferred material, and assign the preferred material to users who follow the content publisher, and use the users to whom the preferred material is assigned as target users.

[0134] Furthermore, the material sorting and selecting unit 201 includes:

[0135] The material score determination module is used to obtain the feature information associated with each preferred material corresponding to any target user, and vectorize the feature information to obtain the feature vector associated with the preferred material, input the feature vector associated with the preferred material into the pre-trained DeepFM model to predict the click-through rate of the preferred material, and obtain the user interest score corresponding to the preferred material output by the DeepFM model; the user interest score is used to characterize the click-through rate prediction value corresponding to the corresponding material; the DeepFM model is composed of a factor decomposition model and a deep neural network model, and can simultaneously learn low-order feature combinations and high-order feature combinations to predict the click-through rate;

[0136] The target material determination module is used to select a preset number of preferred materials whose user interest scores are greater than a preset score threshold and are ranked high according to the user interest scores as the target materials corresponding to the target user.

[0137] Furthermore, the material aggregation unit 202 includes:

[0138] A title adding module is used to add a corresponding title to each target material for each target user according to the content of each target material of the target user;

[0139] The aggregation module is used to splice all target materials and corresponding titles corresponding to the target user to obtain a piece of aggregated material corresponding to the target user.

[0140] Furthermore, the material push unit 203 is configured to: in response to a click operation on the aggregate push by any target user, send a polymerized material corresponding to the target user to the client of the target user, so that the client can decompose the polymerized material into all target materials corresponding to the target user, and display all target materials in the client of the target user.

[0141] The embodiment of the present invention is an embodiment of a material pushing device that corresponds one-to-one to the embodiment of the aforementioned material pushing method, and the embodiment of the present invention can be understood based on the embodiment of the aforementioned material pushing method.

[0142] In a third aspect, an embodiment of the present invention provides a computer device, comprising at least one host, wherein the at least one host comprises a processor adapted to execute various instructions; and

[0143] A storage device is suitable for storing a plurality of instructions, wherein the instructions are suitable for being loaded by a processor and executing any of the above methods.

[0144] The embodiment of the present invention is a device embodiment for storing and executing any of the material pushing methods described above. The embodiment of the present invention can be understood based on the embodiment of the aforementioned material pushing method.

[0145] In a fourth aspect, an embodiment of the present invention provides a storage medium, in which a plurality of instructions are stored, and the instructions are suitable for being loaded by a processor and executing any of the methods described above.

[0146] The embodiment of the present invention is an embodiment of a storage medium storing instructions of any of the aforementioned material pushing methods. The embodiment of the present invention can be understood based on the aforementioned embodiment of the material pushing method.

[0147] The above technical solution of the embodiment of the present invention is described in detail below in conjunction with specific application examples. For technical details not introduced during the implementation process, please refer to the relevant description in the previous text.

[0148] In the scenario of sending push based on social relationships, the embodiment of the present invention proposes a method for constructing an automatic aggregate push of multiple materials. The embodiment of the present invention aims to deal with manufacturers (such as mobile phone or mobile phone system manufacturers) limiting the number of push items, and when there are sufficient materials, more and better materials are aggregated and sent to users for personalized push.

[0149] First, the bloggers recall users on a daily basis under social relationships, and recall users based on the user's attention relationship and intimacy relationship with the blogger. The recalled information with the blogger and user side is stored in the Redis database. The relationship between the blogger (target content publisher) and the user is also stored in the Redis database. Then, through each round of material mining, high-quality materials within a day's validity period are found, and the DeepFM model is used to sort and predict the multi-dimensional features of the blogger side, the user side, the blogger's U2U side to the user, and the blog material side to find multiple materials suitable for the user. Then, on this basis, the aggregated materials are automatically constructed to form parameter information such as aggregated delivery copy, which is finally delivered through transparent transmission, and multiple materials are displayed on the user's mobile phone.

[0150] Based on the social scenario, the embodiment of the present invention proposes this method of constructing a multi-material aggregate push to break through the manufacturer's limit on the number of items. At the same time, the aggregation system of the embodiment of the present invention can also solve the problem of too many and too frequent pushes commonly reported by users. In each round of distribution, the latest time-level partitioned high-quality materials are obtained to quickly reach users, improve the quality of social distribution, and avoid the omission of high-quality materials.

[0151] Among them, the multi-material aggregation and distribution proposed in the embodiment of the present invention is particularly effective for some high-frequency or full-time users. High-frequency users have a large amount of attention, many user behaviors, more candidate materials, high user recall rate, and strong ranking competitiveness. In each round of recall, users have a large number of optional materials to distribute, and many high-quality materials cannot reach all the fans of the blogger. Through the material push method of the embodiment of the present invention, more high-quality materials can be distributed at one time, with more user behavior data, and the overall coverage of materials in social Push can be improved.

[0152] At the same time, the aggregate push service proposed in the embodiment of the present invention is also applicable to the multi-material push processing in other company business scenarios. In other business scenarios, the multi-material automated (aggregate) push method of the embodiment of the present invention can also be promoted to push when encountering multi-material problems. And the embodiment of the present invention is also suitable for targeted optimization of social Push for key groups after segmenting users according to dimensions such as user activity, time, and field. If the amount of materials recalled by the user is small, it will be issued according to the traditional push. If there are many high-quality materials, multiple materials will be aggregated and issued.

[0153] The embodiment of the present invention attempts to distribute materials in different rounds according to the frequency of users, and combines the distribution of the model calculation and sorting of CTR (Click-Through-Rate) of distributing multiple materials, the distribution of the model calculation and sorting of other services such as intimacy, sets the score threshold for aggregate distribution, and optimizes the display of copywriting. It is possible to increase the distribution of materials for users with different frequencies, screen high-quality materials with high interaction rates and large click volumes, and thus increase the distribution coverage of high-quality materials.

[0154] The system construction and application scheme of the present invention are further explained below in conjunction with the accompanying drawings, but are not limited to this. Any modification or equivalent replacement of the technical scheme of the present invention that does not depart from the spirit and scope of the technical scheme of the present invention should be included in the protection scope of the present invention.

[0155] The overall structure of the hotspot mining and distribution system based on the social push scenario is as follows: Figure 5 The specific steps are described below.

[0156] Daily recall of users by blogger: During the daily data preparation phase, bloggers are called back to users based on their following relationships. The purpose is to establish a corresponding relationship between bloggers and users who follow them. Bloggers can be selected from all bloggers or those that meet certain filtering conditions, such as bloggers with many fans and high activity.

[0157] Material mining: Before each round of calculation, pull the latest materials, recall users according to the material mid, and splice the materials. Select the latest materials (alternative materials) according to the preset preferred material conditions, establish a material library, and when selecting the latest materials, obtain the blogger side information from which the materials come. Based on the blogger side information and the correspondence between the blogger and the user, one or more users that the material may be suitable for can be determined. Finally, each user corresponds to one or more potentially suitable materials (preferred materials).

[0158] Material sorting: sort all materials (preferred materials) for each user, and the model calculates and sorts the sorting results of each material (preferred material) to determine the final aggregated and distributed materials (target materials). For each user, sort one or more materials suitable for the user, and determine the materials (target materials) that can be aggregated and distributed based on the sorting results.

[0159] Automatic material aggregation: Automatic material aggregation aggregates different numbers of materials (target materials) for different users, splices the material formats, and obtains aggregated materials. Aggregate the materials (target materials) that can be aggregated and distributed corresponding to the user (target user).

[0160] Aggregate material recall: After obtaining all materials to be aggregated (target materials) and corresponding user-side information, obtain parameter information from the material library to process the material title, process the material title from the material library, and finally splice and issue it. Generate material titles for the materials in each aggregated material of each user, format them, and issue them to the user.

[0161] Downstream processing display: Based on the aggregated materials mentioned in the above steps, the final operation of sending it to downstream users is completed in the system, and finally falls into the flow and is displayed through the relationship flow.

[0162] Daily Blogger Recall Instructions:

[0163] The embodiment of the present invention first performs a daily data preparation phase, and recalls users by blogger. In the social push scenario, due to the particularity of social networking, all materials issued are based on the attention relationship between users, where all recalled materials are relationship materials in mutual exchanges and communication, specifically referring to materials generated between all users with attention relationships, such as mutual friends, friend circles, classmates, relatives, colleagues, etc., and mining is performed through intimacy relationships. If a user follows a blogger, then all blog posts published by this blogger will enter the relationship material library of this user.

[0164] The embodiment of the present invention first screens the users that can be sent in the pushable library every day, and on this basis, the relationship pairs of the bloggers are recalled through the user relationships with attention relationships and intimacy indicators. Based on the relationship pairs, the user-side features and the blogger-side features are linked, and the U2U features of the blogger to the user are added to form a recall data format for the most users recalled by a blogger:

[0165] <blogUid,blogInfo,[uidInfoList 1,...,uidsInfoList n]> Among them, blogUid is the blogger's unique ID, blogInfo is the blogger-side feature information, and uidInfoList is the user-side feature information of each user. Since a recalled blogger corresponds to multiple users, a user information list is formed.

[0166] According to the above format, the daily recalled data is stored in Redis and read in each round of calculation on the same day. The step of calculating the corresponding user data of the recalled blogger is added, calculating how many users each blogger corresponds to, and slicing and storing bloggers with a large number of fans to facilitate the allocation of computing resources in the subsequent reduce phase. When a blogger corresponds to too many users, the calculation can be distributed on different nodes.

[0167] Material excavation instructions:

[0168] The embodiment of the present invention proposes to first pull the latest materials in each round of calculation and distribution, recall users according to the mid of the materials, and splice the materials. In the social push scenario, there are higher timeliness and quality requirements for the materials to be distributed, so the embodiment of the present invention sets up a screening mechanism with a timeliness of the latest day and limits the original and 1-repost blog materials to be distributed to users. Pull blog materials that meet the above requirements from Redis where the materials are stored, and splice the materials based on the users recalled by the blogger and the corresponding feature information, according to the blogger corresponding to the materials.

[0169] The embodiment of the present invention also splits the materials when sending them to users according to the material recall, so as to avoid the same node processing too many materials. Generate the material and blogger material side information, and the corresponding recalled user information format and store them in Redis:

[0170] <mid,midInfo,[uidInfoList 1,...,uidsInfoList n]>

[0171] Among them, mid is the unique ID of the material, midInfo is the material-side characteristic information, and uidInfoList is the user-side characteristic information of each user. Since one material to be issued corresponds to multiple users, a user information list is formed.

[0172] Material Sorting System Description:

[0173] The embodiment of the present invention proposes to sort all materials of each user in the material sorting stage to determine the materials to be finally aggregated and distributed. The overall flow chart of the sorting system is as follows: Figure 6 shown.

[0174] According to the user data corresponding to each material, the embodiment of the present invention recalls and sorts all materials of each user in the sorting stage on a user-by-user basis. The user uid is used as the key to obtain characteristic information such as the number of issued and opened items of each material, and preliminarily screen high-quality materials. Only materials with more than 5 clicks or a click-through rate of more than 1% on the same day can participate in the sorting, and it is considered that such blog materials have a high click trend.

[0175] The embodiment of the present invention uses the DeepFM model to perform sorting calculations in the material sorting stage. DeepFM is derived from the FM model and consists of two parts: the FM structure and the DNN structure. The bottom part shares the same feature Embedding input and learns low-order and high-order feature intersections at the same time. FM performs low-order combinations between features, and DNN performs high-order combinations between features. The two methods are combined in parallel, so that the final architecture can train the model faster and more accurately. The parameters of all models can also be obtained through joint training of the loss function, as shown in formula (1).

[0176] In the material sorting of aggregate push, the present invention applies 3245-dimensional features, including blogger-side features, user-side features, material-side features, and blogger-to-user U2U features. The present invention generates samples based on the click behavior data of users for push services for about 3 months, trains the generated model offline, and applies the model to the sorting calculation of the aggregate push service. Specifically, if a user clicks on a push sent by an aggregate push, then the two or more blog posts contained in this aggregate push are counted as having one click, and the user may be interested in both blog posts.

[0177] The feature combination part of the FM model can learn the interaction between different features through vectorized crossover, which is more advantageous when processing large-scale sparse features. In the present invention, some feature dimensions are also relatively sparse. In the data scenario with high coefficients, the application of the FM model solves this problem better. The output of the FM part is an addition unit that reflects the first-order linear combination feature, and the cumulative sum of many inner product units that use the inner product of the feature vector to extract the second-order feature combination, as shown in formula (2).

[0178] The model calculates and sorts the user's scoring results for each material. Then, the materials that have been read and sent are filtered, and a tree set structure is established to store and sort all the materials involved in the sorting for each user. The embodiment of the present invention screens the sorting and scoring results, sorts and selects the best two or more materials to aggregate and send to the user. The final output of this step is the sending user, the aggregated mid list, and the scoring list.

[0179] Automated Aggregation System Description

[0180] The embodiment of the present invention completes the automatic aggregation of multiple materials, splices the material format, sends the materials to the user according to the material recall, and processes the material title from the material library. According to the amount of materials that the user can recall and the quality of the materials, the embodiment of the present invention proposes automatic aggregation of materials. When the amount of materials that the user can recall is large and the quality of the materials is good, multiple suitable materials are automatically aggregated and sent to the user. When the number of materials is small, a single material is sent to the user.

[0181] Specifically, the embodiment of the present invention screens the model sorting ctr output on the premise of automatically aggregating multiple pushes, and attempts to distribute in different rounds according to the frequency of user activity. For more active users such as full attendance, they themselves may recall more related materials. In this case, if there are sufficient materials, more rounds of aggregation and distribution can be carried out within a day to achieve the purpose of distributing more materials to users. For less active users such as low-frequency and new users, the embodiment of the present invention will aggregate and distribute to these users during the time period of active blogging. During these time periods, blog materials are relatively sufficient and it is more likely that the requirements for aggregate distribution can be met, breaking the limit on the number of items without affecting the distribution of materials for other businesses.

[0182] The embodiment of the present invention combines the original sorting scores and CTR distribution of the two materials to set the CTR threshold for sending. According to the sorting score and CTR threshold, the material click sending volume and click rate are limited, thereby screening out better aggregate materials and allowing high-quality materials to quickly reach users.

[0183] Unlike the traditional one-push delivery, the aggregate push needs to find two or more materials suitable for the user. It is necessary to ensure that the user has enough candidate materials, because the aggregate push is designed to solve the manufacturer's restrictions on the number of push items. If the user does not have particularly sufficient materials, the automated method of the embodiment of the present invention determines that the length of the material list recalled by the user is list.size = 1 and sends it in single mode; if the length of the material list recalled by the user is list.size>1, the Top2 or TopN suitable materials are intercepted according to the scenario for automatic aggregation.

[0184] Compared with the traditional single push, the aggregate push in the embodiment of the present invention is very different in display. Specifically, all material tags of a single push are single material information, and the log data formats such as mid (the unique ID corresponding to each material) and lfid (log data tag of the material) are all tags of a single material. In the aggregate push, it is necessary to record multiple blog material information at the same time to facilitate downstream and subsequent analysis of data processing. The format of recording log data tag lfid is as follows:

[0185] lfid: lfid-prefix_${mid1},${mid2},${mid3}

[0186] The scheme log parameter format also needs to be modified accordingly. All tags related to material information, such as mid list, blogger type list, blogger ID list, etc., need to add multiple material information. The text of the aggregate push is as follows:

[0187] Your multiple followers post

[0188] @Nickname 1: Weibo post…(picture)>>\n

[0189] ②@Nickname 2: Weibo post…(picture)>>

[0190] Figure 4 This is a schematic diagram showing the actual aggregation push.

[0191] At the same time Figure 4 It is shown in the figure that the embodiment of the present invention also optimizes the copywriting accordingly. The title shows your "your multiple followers post a blog", and the nicknames and contents of multiple materials are displayed in the copywriting. In this step, it is necessary to re-acquire the detailed information of the material, because in order to save computing resources in the sorting stage of the previous step, the material title and other information are usually not brought down. After re-recalling the material title and other information, the midlist is processed to obtain the corresponding title and copywriting for splicing and storage.

[0192] Polymer Material Recall Instructions:

[0193] After obtaining all aggregated materials and corresponding user-side information, the embodiment of the present invention performs the process of recalling aggregated materials before final delivery. Specifically, in this step, multiple aggregated pushes corresponding to the delivery user list are generated in json format, and then delivered through the downstream. The json format contains the parameter information of the aggregated materials and the parameters of the corresponding delivery users, and the data format is as follows:

[0194] {"msg":"

[0195] @Nickname 1: Weibo text...(Picture)>>\n②@Nickname 2: Weibo text...(Picture)>>",

[0196] "title":"Your multiple followers post a blog",

[0197] "toids":"uid1,uid2…………",

[0198] "url":"sinaweibo: / / xxxxxxxx"

[0199] ………………}

[0200] The aggregate push material is used as a unit, and each aggregate push material corresponds to multiple users. Here, the embodiment of the present invention also applies splitting materials to avoid processing too many materials on the same node. The materials are stored in Redis according to the above specific format and merged with a large number of related materials every day to remove duplicates.

[0201] Description of the aggregation distribution system:

[0202] Based on the recall of aggregated materials mentioned in the above steps, the embodiment of the present invention proposes to complete the final operation of sending to the user's mobile phone in this system. Specifically, compared with the traditional push sending, when the user clicks on this push, it will directly fall into the internal flow of the app page and display this blog post. Here, it can be selected to fall into the blog post details page or the data flow for simple display according to the business form.

[0203] Aggregate push needs to display two or more blog posts at the same time, so the detail page that displays several blog posts at the same time is obviously not suitable for the aggregation scenario. Based on the social scenario, the embodiment of the present invention drops the aggregated material into the relationship flow of the app. When the user clicks on the aggregate push, we will directly jump to the relationship flow inside the app in the downstream operation, and directly and simply display the corresponding two or more blog posts in the user client flow, so that the user can directly see the multiple blog posts of interest.

[0204] The final display stage requires support for in-stream display. Therefore, in other apps that apply the multi-material automatic (aggregation) push method of the embodiment of the present invention, when the push falls into the stream for display, it is necessary to be able to support the display of multiple contents and match the corresponding in-stream display transformation and top-pinning function to better apply the embodiment of the present invention.

[0205] The embodiments of the present invention have the following technical effects: compared with the traditional push method, the overall computing system overhead of the embodiments of the present invention is relatively small, and more high-quality materials can be sent at one time to reach users, thereby increasing the amount of materials sent by social Push; the embodiments of the present invention can distribute more high-quality blog materials to users when more materials are recalled through the method of multi-material aggregation push, thereby improving the quality of social delivery; compared with the system of distribution by traditional model calculation, the embodiments of the present invention proposes a solution of using DeepFM model application to sort and select multiple blog materials, and uses the blogger side, user side, and blogger to perform model calculation learning on the multi-dimensional features of the user U2U side and the blog material side, so that the model is more accurate and can recommend higher-quality materials; the embodiments of the present invention can break through the manufacturer's limit on the number of items, and the multi-material automatic (aggregation) push method aggregates 2 or more intimacy materials that the user is concerned about and sends them, breaking through the traditional limit on the number of items sent and avoiding the omission of high-quality materials; based on the social scenario, the present invention proposes this method of constructing multi-material aggregation push to address the problems of excessive and too frequent push commonly raised by users, which can well improve the user experience.

[0206] In other embodiments, the present invention provides a method for constructing a multi-material automatic aggregate push based on a social push scenario, including:

[0207] Recall users by blogger every day: During the daily data preparation stage, bloggers recall users according to their following relationships.

[0208] Material mining: Pull the latest materials before each round of calculation, recall users according to the material mid, and splice the materials.

[0209] Material sorting: All materials of each user are sorted. The model calculates and sorts the sorting results of each material to determine the final aggregated and distributed materials.

[0210] Automatic material aggregation: Automatic material aggregation, aggregation of different numbers of materials for different users, and splicing of material formats.

[0211] Aggregate material recall: After obtaining all aggregate materials and corresponding user-side information, obtain parameter information from the material library to process material titles, recall aggregate materials and send them to users, process material titles from the material library, and finally assemble and send them.

[0212] Downstream processing display: Based on the aggregate material call mentioned in the above steps, the final operation of sending it to downstream users is completed in the system, and finally falls into the flow and the relationship flow is displayed.

[0213] In the social scenario, the blogger recalls the user relationship direction, and proposes that the blogger first recalls the user to determine the range of materials that the user will recall every day, and proposes the material storage form and rules, and the calculation formula for the click ranking. The rules for building material aggregation are proposed, which can break through the manufacturer's limit on the number of items, break through the traditional limit on the number of items issued, and avoid the omission of high-quality materials. The proposed DeepFM model calculation method proposes to use the blogger side, the user side, and the blogger to calculate and learn the model on the user U2U side and the blog material side of the multi-dimensional features, so that the model is more accurate and can recommend higher-quality materials. Including the material storage form, rules, and calculation formulas described in the present invention. The proposed automated aggregation and aggregation copywriting rules specifically propose that it can automatically determine whether to issue a single or multiple aggregated materials and the number of aggregated materials by recalling the number and quality of materials. And proposed material storage forms and rules, automated identification methods, display copywriting formats, flow display rules, etc. The proposed automated aggregate push mobile client display transformation proposes that the aggregate push will eventually be displayed within the stream, which requires the application to support the display of multiple blog post contents, including support methods within the stream, multiple material display rules, and the application of the top-level function within the stream.

[0214] It should be understood that the specific order or hierarchy of steps in the disclosed process is an example of an exemplary method. Based on design preferences, it should be understood that the specific order or hierarchy of steps in the process can be rearranged without departing from the scope of protection of the present disclosure. The attached method claims present the elements of the various steps in an exemplary order and are not intended to be limited to the specific order or hierarchy described.

[0215] In the above detailed description, various features are grouped together in a single embodiment to simplify the disclosure. This method of disclosure should not be interpreted as reflecting an intention that the embodiments of the claimed subject matter require more features than are clearly stated in each claim. On the contrary, as reflected in the appended claims, the invention is in a state of having less than all the features of the disclosed individual embodiments. Therefore, the appended claims are hereby expressly incorporated into the detailed description, with each claim standing on its own as a separate preferred embodiment of the invention.

[0216] The disclosed embodiments are described above to enable any person skilled in the art to implement or use the present invention. Various modifications of these embodiments are obvious to those skilled in the art, and the general principles defined herein may also be applied to other embodiments without departing from the spirit and scope of the present disclosure. Therefore, the present disclosure is not limited to the embodiments given herein, but is consistent with the broadest scope of the principles and novel features disclosed in this application.

[0217] The above description includes examples of one or more embodiments. Of course, it is impossible to describe all possible combinations of components or methods for the purpose of describing the above embodiments, but it should be recognized by those skilled in the art that the various embodiments may be further combined and arranged. Therefore, the embodiments described herein are intended to cover all such changes, modifications and variations that fall within the scope of protection of the appended claims. In addition, with respect to the term "comprising" used in the specification or claims, the word is covered in a manner similar to the term "including". In addition, any term "or" used in the specification of the claims is intended to mean "non-exclusive or".

[0218] Those skilled in the art may also understand that the various illustrative logical blocks, units, and steps listed in the embodiments of the present invention may be implemented by electronic hardware, computer software, or a combination of the two. In order to clearly demonstrate the interchangeability of hardware and software, the various illustrative components, units, and steps described above have generally described their functions. Whether such functions are implemented by hardware or software depends on the specific application and the design requirements of the entire system. Those skilled in the art may use various methods to implement the described functions for each specific application, but such implementation should not be understood as exceeding the scope of protection of the embodiments of the present invention.

[0219] The various illustrative logic blocks or units described in the embodiments of the present invention can be implemented or operated by a general-purpose processor, a digital signal processor, an application-specific integrated circuit (ASIC), a field programmable gate array or other programmable logic device, a discrete gate or transistor logic, a discrete hardware component, or any combination of the above. The general-purpose processor can be a microprocessor, and optionally, the general-purpose processor can also be any conventional processor, controller, microcontroller or state machine. The processor can also be implemented by a combination of computing devices, such as a digital signal processor and a microprocessor, a plurality of microprocessors, one or more microprocessors combined with a digital signal processor core, or any other similar configuration.

[0220] The steps of the method or algorithm described in the embodiments of the present invention can be directly embedded in hardware, a software module executed by a processor, or a combination of the two. The software module can be stored in a RAM memory, a flash memory, a ROM memory, an EPROM memory, an EEPROM memory, a register, a hard disk, a removable disk, a CD-ROM, or other storage media of any form in the art. Exemplarily, the storage medium can be connected to the processor so that the processor can read information from the storage medium and can write information to the storage medium. Optionally, the storage medium can also be integrated into the processor. The processor and the storage medium can be arranged in an ASIC, and the ASIC can be arranged in a user terminal. Optionally, the processor and the storage medium can also be arranged in different components in the user terminal.

[0221] In one or more exemplary designs, the above functions described in the embodiments of the present invention can be implemented in hardware, software, firmware or any combination of the three. If implemented in software, these functions can be stored on a computer-readable medium, or transmitted in the form of one or more instructions or codes on a computer-readable medium. Computer-readable media include computer storage media and communication media that facilitate the transfer of computer programs from one place to another. The storage medium can be any available medium that can be accessed by any general or special computer. For example, such computer-readable media can include but are not limited to RAM, ROM, EEPROM, CD-ROM or other optical disk storage, disk storage or other magnetic storage devices, or any other medium that can be used to carry or store program codes in the form of instructions or data structures and other forms that can be read by general or special computers, or general or special processors. In addition, any connection can be appropriately defined as a computer-readable medium, for example, if the software is transmitted from a website site, server or other remote resource through a coaxial cable, fiber optic cable, twisted pair, digital subscriber line (DSL) or wirelessly, such as infrared, wireless and microwave, it is also included in the defined computer-readable medium. The disk and disc include compact disk, laser disk, optical disk, DVD, floppy disk and blue-ray disk. Disks usually copy data magnetically, while discs usually copy data optically with lasers. The above combination can also be included in computer readable media.

[0222] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A material pushing method, characterized in that: include: Acquire multiple preferred materials and multiple target users associated with a target content publisher, and determine the preferred materials corresponding to each target user based on the association relationship between the multiple preferred materials and the multiple target users; the target content publisher is a content publisher that meets a preset blogger screening condition; For all preferred materials corresponding to each target user, determine the user interest score corresponding to each preferred material, and select the target material corresponding to the target user from all the preferred materials based on the user interest score; Aggregate all target materials corresponding to each target user to obtain a piece of aggregated material corresponding to each target user; Push the corresponding aggregated materials to each target user.

2. The material pushing method according to claim 1, characterized in that: The step of acquiring a plurality of preferred materials and a plurality of target users associated with a target content publisher, and determining a preferred material corresponding to each target user based on the association relationship between the plurality of preferred materials and the plurality of target users, includes: Obtain multiple candidate materials and corresponding material information that meet preset material screening conditions in the material pool; Determine the content publisher corresponding to each candidate material according to the material information corresponding to each candidate material; Among all candidate materials, select the candidate material whose corresponding content publisher belongs to the target content publisher as the preferred material; Determine the target user associated with each preferred material according to the target content publisher corresponding to each preferred material and the established correspondence between the target content publisher and the user who follows the target content publisher; Based on the multiple preferred materials and the target users associated with each preferred material, a preferred material corresponding to each target user is determined.

3. The material pushing method according to claim 2, characterized in that: The preset material screening conditions include at least one of the following: Materials corresponding to original content released within the most recent preset time frame or new content generated by directly forwarding original content; Materials whose click-through rate is greater than or equal to the preset business requirement click-through rate.

4. The material pushing method according to claim 1, characterized in that: The step of determining the user interest score corresponding to each preferred material for each target user, and selecting the target material corresponding to the target user from all the preferred materials based on the user interest score, includes: For each preferred material corresponding to any target user, obtain feature information associated with the preferred material, and vectorize the feature information to obtain a feature vector associated with the preferred material, input the feature vector associated with the preferred material into a pre-trained DeepFM model to predict the click-through rate of the preferred material, and obtain a user interest score corresponding to the preferred material output by the DeepFM model; the user interest score is used to characterize the click-through rate prediction value corresponding to the corresponding material; A preset number of preferred materials whose user interest scores are greater than a preset score threshold and which are ranked high according to the user interest scores are selected as target materials corresponding to the target user.

5. The material pushing method according to claim 1, characterized in that: All target materials corresponding to each target user are aggregated to obtain a polymerized material corresponding to each target user, including: For each target user, according to the content of each target material of the target user, a corresponding title is added to each target material; All target materials and corresponding titles corresponding to the target user are spliced ​​together to obtain a polymer material corresponding to the target user.

6. The material pushing method according to claim 1, characterized in that: Push the corresponding aggregated materials to each target user, including: In response to a click operation on the aggregated push by any target user, a polymerized material corresponding to the target user is sent to the target user's client, so that the client decomposes the polymerized material into all target materials corresponding to the target user, and displays all target materials in the target user's client.

7. A material pushing device, characterized in that: include: A user material screening unit is used to obtain a plurality of preferred materials and a plurality of target users associated with a target content publisher, and determine a preferred material corresponding to each target user based on the association relationship between the plurality of preferred materials and the plurality of target users; the target content publisher is a content publisher that meets a preset blogger screening condition; A material sorting and selecting unit, for determining, for all preferred materials corresponding to each target user, a user interest score corresponding to each preferred material, and selecting a target material corresponding to the target user from all preferred materials based on the user interest score; The material aggregation unit is used to aggregate all target materials corresponding to each target user to obtain a piece of aggregated material corresponding to each target user; The material pushing unit is used to push the corresponding polymer materials to each target user.

8. The material pushing device according to claim 7, characterized in that: The material sorting and selecting unit comprises: The material score determination module is used to obtain the feature information of each preferred material corresponding to any target user, and vectorize the feature information to obtain a feature vector associated with the preferred material, input the feature vector associated with the preferred material into a pre-trained DeepFM model to predict the click-through rate of the preferred material, and obtain the user interest score corresponding to the preferred material output by the DeepFM model; the user interest score is used to characterize the click-through rate prediction value corresponding to the corresponding material; The target material determination module is used to select preferred materials whose user interest score is greater than a preset score threshold and whose preset number of materials are ranked high according to the user interest score as the target material corresponding to the target user.

9. A computer device, characterized in that: The device comprises at least one host, wherein the at least one host comprises a processor adapted to execute various instructions; and A storage device, adapted to store a plurality of instructions, wherein the instructions are adapted to be loaded by a processor and execute the method according to any one of claims 1 to 6.

10. A storage medium, characterized in that: A plurality of instructions are stored therein, and the instructions are suitable for being loaded by a processor and executing the method according to any one of claims 1 to 6.

Images