Content representation determination method and device, equipment, readable storage medium and program product

Abstract

The invention relates to a content representation determination method and device, computer equipment, a computer readable storage medium and a computer program product. The method comprises the following steps: for to-be-processed content, obtaining object representation of a seed object interested in the content; performing enhancement processing on the projection of each object representation on the to-be-solved unit vector to obtain a target function; taking the maximized objective function as an optimization objective, and solving a solving result of a unit vector when the optimization objective is achieved; a content characterization of the content is determined based on a solution of the unit vector. In this way, the content representation determination accuracy is improved, and the content pushing effect is improved.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and in particular to a method, apparatus, computer device, computer-readable storage medium, and computer program product for determining content representation. Background Art

[0002] With the rapid development of computer technology and the popularization of Internet technology, a variety of content is usually pushed to users by push systems. In the link of the recommendation system, the recall link is a key step in determining the recalled content set, and the crowd diffusion method is one of the most important methods in the recall link. The crowd diffusion method first needs to align the user representation and content representation in space, and then push the content.

[0003] In the related art, during the representation alignment process, for a certain content, the user representations of users who are directly interested in the content are averaged, and the obtained average is used as the content representation of the content.

[0004] However, with the rapid development of community business and the rapid expansion of user scale, the concentration of the population will inevitably decline gradually, which increases the complexity of locating the target population. The representation alignment operation of related technologies is difficult to accurately describe the content representation corresponding to the interests of complex populations, that is, the accuracy of content representation determination is not high, resulting in poor content push effect. Summary of the invention

[0005] Based on this, it is necessary to provide a content representation determination method, apparatus, computer device, computer-readable storage medium and computer program product that can improve the accuracy of content representation determination to improve the content push effect in response to the above technical problems.

[0006] In a first aspect, the present application provides a method for determining content representation, comprising:

[0007] For the content to be processed, obtaining an object representation of a seed object of interest in the content;

[0008] The projection of each object representation on the unit vector to be solved is enhanced to obtain the objective function;

[0009] Taking maximizing the objective function as the optimization goal, solving the solution of the unit vector when the optimization goal is achieved;

[0010] A content representation of the content is determined based on a solution result of the unit vector.

[0011] In a second aspect, the present application further provides a content representation determination device, comprising:

[0012] An object representation acquisition module, used for acquiring, for the content to be processed, an object representation of a seed object of interest to the content;

[0013] A function acquisition module is used to enhance the projection of each object representation on the unit vector to be solved to obtain the target function;

[0014] A solution module, used for taking maximization of the objective function as an optimization goal and solving a solution result of the unit vector when the optimization goal is achieved;

[0015] The content representation determination module is used to determine the content representation of the content based on the solution result of the unit vector.

[0016] In a third aspect, the present application further provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the above-mentioned content representation determination method when executing the computer program.

[0017] In a fourth aspect, the present application further provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the steps of the above-mentioned content representation determination method are implemented.

[0018] In a fifth aspect, the present application also provides a computer program product, including a computer program, which implements the steps of the above-mentioned content characterization determination method when executed by a processor.

[0019] The above-mentioned content representation determination method, apparatus, computer device, computer-readable storage medium and computer program product obtain the object representation of the seed objects that are interested in the content to be processed, so as to accurately reflect the object information of each seed object in the crowd that is interested in the content. Then, the projection of each object representation on the unit vector to be solved is determined, and the projection is enhanced, that is, the projection of the object representation of the object that is farther away from the crowd is smaller, and the projection of the object representation of the seed object that is closer to the crowd is larger. Then, by taking the maximization of the objective function as the optimization goal, the influence of discrete seed objects far away from the crowd can be effectively reduced. In this way, regardless of whether the concentration of the crowd decreases, the solution result of the unit vector when the optimization goal is achieved can accurately describe the content representation to improve the content push effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the technical solutions in the embodiments of the present application or related technologies, the drawings required for use in the embodiments of the present application or related technical descriptions will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

[0021] Figure 1 A diagram of an application environment of a content representation determination method in an embodiment;

[0022] Figure 2 is a flow chart of a method for determining content representation in one embodiment;

[0023] Figure 3 is a schematic diagram of an object representation query process in one embodiment;

[0024] Figure 4 is a schematic diagram representing an updating step in one embodiment;

[0025] Figure 5 A schematic diagram of object representation comparison in one embodiment;

[0026] Figure 6 is a schematic diagram of a specific application environment in an embodiment;

[0027] Figure 7 A schematic diagram of a note characterization determination process in one embodiment;

[0028] Figure 8 is a structural block diagram of a content representation determination device in one embodiment;

[0029] Fig. 9 FIG. 4 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0030] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0031] Before introducing this application, the relevant terms are explained:

[0032] Recommendation system (push system): A recommendation system is an information filtering system that is used to predict the "rating" or "preference" that a user may have for an item or product. This system is widely used in different Internet applications, such as movie, music, news recommendations, product recommendations, etc. The main goal of the recommendation system is to provide users with personalized content or product recommendations by learning the user's historical behavior and interest preferences, thereby improving user experience and satisfaction, and increasing user stickiness and activity.

[0033] Crowd diffusion: Spread the user's interests to friends, followers, and other groups with explicit or implicit connections through the user's social network relationships. Crowd diffusion can effectively utilize interest representations in social networks and accurately spread crowd interests, thereby improving the accuracy and diversity of recommendations.

[0034] Object: used to refer to a user. For example, the object is the account of the user who logs in to the push application.

[0035] Content: The specific information to be pushed to the target, which can be in the form of pictures, articles, graphics, videos, objects, music, audio and video, etc.

[0036] Representation mapping: The multi-dimensional features of an object or content are mapped to a low-dimensional space through a function, so that similar objects or content are closer in this low-dimensional space. This low-dimensional feature vector after mapping is usually called the "representation" or "embedding" of the object or content.

[0037] Object representation (also called object interest representation): refers to the representation content obtained by mapping the representation of the object.

[0038] Content representation (also called content interest representation): the representation content obtained by mapping the representation of a certain content.

[0039] Representation alignment: The process of mapping object representation and content representation into the same semantic space, so that the similarity between object representation and content representation in this space can be measured by calculating the similarity distance between the two representations.

[0040] Recall: refers to recalling content for rough ranking from the candidate pool through targeted strategies or intelligent means.

[0041] Interactive behavior: The behavior of an object on content in a push scenario or a search scenario. For example, interactive behaviors include but are not limited to clicks, likes, forwarding, purchases, etc.

[0042] In the related art, before the recall phase, a representation alignment operation needs to be completed. The mathematical expression of the representation alignment operation is that the vector distance between the object representation Eu (which can be regarded as vector Eu) and the content representation Ev (which can be regarded as vector Ev) is as small as possible, as shown in the following formula (1):

[0043] (1)

[0044] In the above formula, the modulus of the object representation Eu and the content representation Ev are both unit values ​​1, argmin(.) is the minimum point function, <.> vector inner product, is the vector distance.

[0045] From formula (1), we can see that in order to minimize the vector distance, the inner product between vector Eu and vector Ev needs to be maximized, that is, .

[0046] In the push scenario, for a certain content, obtain the object representations of multiple objects that are interested in the content, for example, , …, At this time, to achieve representation alignment, it is necessary to calculate the sum of the vector distances between each object representation and the content representation, and ensure that the sum of the vector distances is minimized. In other words, it is necessary to ensure that the sum of the inner products is maximized, that is:

[0047] (2)

[0048] In the above formula (2), The modulus of the content representation is 1. Representation of object The inner product with the content representation Ev.

[0049] After equation conversion, the following formula (3) is obtained:

[0050] (3)

[0051] In formula (3) is the mean of all object representations, is the modulus of the mean.

[0052] Based on the above formulas (1)-(3), it can be seen that in the related art, the representation alignment scheme is achieved by optimizing the inner product sum, and optimizing the inner product sum is equivalent to normalizing and averaging the object representations of all objects. In the related art, the mean of the object representation is used as the content representation of the content. In other words, for a certain content, the object representations of all objects among the people who are interested in the content are aligned indiscriminately.

[0053] However, with the rapid development of community services and the rapid expansion of user scale, the concentration of the crowd will inevitably decline, which increases the complexity of locating the target population. In other words, there is at least one object in the crowd whose object representation is relatively scattered and far away from the crowd. At this time, the object can be regarded as a discrete object, and it is necessary to weaken the influence of the object representation of the discrete object in the content representation determination process.

[0054] However, as mentioned above, the representation alignment of related technologies is indiscriminate alignment, that is, the influence of discrete objects and non-discrete objects in the process of determining content representation is equal. Obviously, when the concentration of a population decreases, the use of related technologies cannot effectively describe the content representation corresponding to the population, that is, the mapping between object representation and content representation is inaccurate, and it is difficult to accurately find the right population to diffuse and distribute the content, which seriously restricts the recall efficiency and ultimately affects the push quality, that is, the push effect is not good.

[0055] The embodiment of the present application provides a method for determining content representation, which obtains the object representation of the seed objects that are interested in the content to be processed, so as to accurately reflect the object information of each seed object in the crowd that is interested in the content. Then, the projection of each object representation on the unit vector to be solved is determined, and the projection is enhanced, that is, the projection of the object representation of the object that is farther away from the crowd is smaller, and the projection of the object representation of the seed object that is closer to the crowd is larger. Then, by taking the maximization of the objective function as the optimization goal, the influence of discrete seed objects far away from the crowd can be effectively reduced. In this way, regardless of whether the concentration of the crowd decreases, the solution result of the unit vector when the optimization goal is achieved can accurately describe the content representation to improve the content push effect.

[0056] The content representation determination method provided in the embodiment of the present application can be applied to Figure 1 In the application environment shown, the terminal 102 communicates with the server 104 through a network. The data storage system can store data that the server 104 needs to process. The data storage system can be integrated on the server 104, or placed on the cloud or other network servers.

[0057] In some embodiments, for each content to be processed, the server 104 obtains the object representation of the seed object of interest in the content; enhances the projection of each object representation on the unit vector to be solved to obtain the objective function; takes maximizing the objective function as the optimization goal, and solves the solution result of the unit vector when the optimization goal is achieved; and determines the content representation of the content based on the solution result of the unit vector.

[0058] Subsequently, the terminal 102 where any object is located sends a push request to the server 104. The server 104 determines the corresponding target object representation based on the object identifier carried in the push request, and recalls multiple recalled contents based on the content representation of each content and the target object representation. The recalled contents are used to determine the pushed contents and are returned to the terminal 102 to expose the pushed contents in the terminal 102.

[0059] The terminal 102 may be, but is not limited to, various personal computers, laptops, smart phones, tablet computers, IoT devices, and portable wearable devices. The IoT devices may be smart speakers, smart TVs, smart air conditioners, smart car devices, projection devices, etc. Portable wearable devices may be smart watches, smart bracelets, head-mounted devices, etc. Head-mounted devices may be virtual reality (VR) devices, augmented reality (AR) devices, smart glasses, etc. The server 104 may be an independent physical server, a server cluster or distributed system consisting of multiple physical servers, or a cloud server that provides cloud computing services.

[0060] In an exemplary embodiment, Figure 2 As shown, a content representation determination method is provided, and the method is applied to Figure 1 The server 104 in the example is used as an example to illustrate the method, which includes the following steps S202 to S208. Among them:

[0061] Step S202: for the content to be processed, obtain the object representation of the seed object of interest in the content.

[0062] Among them, the content published on the target application is regarded as the content to be processed, and the target application is an application with a content push function, including but not limited to e-commerce applications, social applications, and multimedia applications.

[0063] A seed object refers to an object that is interested in the content. For example, an object that has performed at least one interactive behavior on the content is used as a seed object. For example, an object that has performed continuous interactive behaviors on the content (for example, clicking, liking, collecting, and forwarding the content in sequence) is used as a seed object.

[0064] Exemplarily, the server determines candidate objects that have performed interactive behaviors on the content, and screens out candidate objects with high interactive behavior frequencies from the candidate objects, and uses the screened candidate objects as seed objects. For example, for each candidate object, the interactive behavior frequency for the content is determined, and if the interactive behavior frequency reaches a preset frequency, the candidate object is used as a seed object, and if it does not reach the preset frequency, the candidate object is not used as a seed object.

[0065] Optionally, the server obtains the content to be processed from the candidate pool, and obtains object statistics data of the content, where the object statistics data includes a plurality of objects that have implemented interactive behaviors on the content.

[0066] The server treats each object in the object statistics as a seed object of interest in the content.

[0067] Alternatively, the object statistics also include the frequency of interaction behaviors of each object with respect to the content. The process of determining the seed object based on the frequency of interaction behaviors can be referred to in the previous text, and will not be described in detail here.

[0068] The server then determines an object representation for each seed object.

[0069] Exemplarily, the server randomly selects a content in an exposed state from the candidate pool, and uses the selected content as the content to be processed.

[0070] Exemplarily, after determining the seed object, the server queries the object representation of each seed object from local storage or other servers for storing object representations based on the object identifier of each seed object.

[0071] In one embodiment, for the content to be processed, before obtaining the object representation of the seed object of interest in the content, the method includes: if a content representation update condition is met, taking any content in an exposed state as the content to be processed.

[0072] The content representation update condition refers to a condition for updating the content representation. Optionally, the content representation update condition for each content may be the same or different.

[0073] Optionally, the server checks whether the content representation update condition is satisfied at the current moment, and if so, obtains any content in the exposure state as the content to be processed. If not, does not obtain the content to be processed.

[0074] Exemplarily, the server determines the last update time of the last content representation update, and calculates the duration between the current time and the last update time, and if the duration is equal to a first preset duration, it is determined that the current time satisfies the content representation update condition. If the first preset duration is not satisfied, it is determined that the current time does not satisfy the content representation update condition.

[0075] In this way, by checking whether the time between the current time and the last update time satisfies the first preset time, the content representation is periodically updated.

[0076] Furthermore, the corresponding first preset time length can be adaptively adjusted in combination with the content type. For example, for some first content types with high timeliness requirements, such as news types, the corresponding first preset time length is the first candidate time length, and for some second content types with low timeliness requirements, such as popular science types, the corresponding first preset time length is the second candidate time length, and the first candidate time length is smaller than the second candidate time length.

[0077] For example, for any candidate content in the candidate pool that is in an exposed state, after the server determines the duration of the candidate content (the duration between the current moment and the last update moment of the last content representation update), it selects the candidate duration corresponding to the content type to which the candidate content belongs from the candidate durations corresponding to each content type, and uses the selected candidate duration as the first preset duration. The server verifies whether the duration is equal to the first candidate duration. If it is equal, it is determined that the content representation update condition is met, and the candidate content is used as the content to be processed. If it is not equal, it is determined that the content representation update condition is not met. Among them, if the content type is the first content type, the first preset duration is the first candidate duration, and if the content type is the second content type, the second preset duration is the first candidate duration.

[0078] In the above two examples, the first preset duration can be regarded as a period for updating the content representation, such as one day, one week, one month, etc.

[0079] In this way, in combination with the timeliness requirements of the content type and the current moment, the content representation of each content can be adaptively updated, and more accurate and effective content push can be performed subsequently with the updated content representation, further improving the content push effect.

[0080] Exemplarily, the server counts the time period in which the maximum number of active objects is within a preset period, sets the time period before and adjacent to the time period as the content representation update period, and checks whether the current moment is within the content representation update period within the current preset period. If so, it is determined that the current moment meets the content representation update condition. The number of active objects within each time period refers to the number of refresh operations performed on the target application within the time period, or the number of login operations performed on the target application, etc.

[0081] For example, the preset period can be a day, a week, a month, a quarter, etc. The time period can be understood as a basic unit within the preset period. For example, if the preset period is a day, there are 24 time periods within the preset period, and the duration of each time period is one hour; for another example, if the preset period is a week, there are 7 time periods within the preset period, and the duration of each time period is 24 hours.

[0082] The following explanation is based on the preset cycle of a week. The day with the largest daily activity in the past week, such as Wednesday, is determined. The content representation update period is Tuesday, that is, from 0:00 to 24:00 on Tuesday. If the current time is 8:00 on Tuesday of the current week, it is determined that the current time meets the content representation update conditions.

[0083] In this way, the content representation of all content can be updated in advance before the period when the number of objects is the largest, and the updated content representation can be used for more accurate, comprehensive and effective content push later, further improving the content push effect.

[0084] In this embodiment, before obtaining the object representation, it is pre-checked whether the content representation update condition is met. If so, the step of obtaining the object representation is triggered to update the content representation. Subsequently, content can be pushed based on the latest updated content representation to ensure the validity of the content push.

[0085] In some embodiments, obtaining an object representation of a seed object that is interested in the content includes: obtaining multiple objects that have performed interactive behaviors on the content within a first preset time period, and using the multiple objects as seed objects about the content; and determining the object representation of each seed object respectively.

[0086] The first preset duration refers to the duration between the current time and the last update time of the last content representation update.

[0087] Exemplarily, for each content to be processed, the server queries the object statistical data of the content from the respective object statistical data of a plurality of contents stored locally according to the content identifier of the content.

[0088] The server searches out sub-object statistical data generated within a first preset time period from the object statistical data, wherein the sub-object statistical data includes a plurality of objects that have implemented interactive behaviors on the content within the first preset time period.

[0089] The server treats each object in the sub-object statistics as a seed object.

[0090] Exemplarily, after the seed objects are determined, the server may directly obtain the object representation of each seed object from the object representation storage space according to the object identifier of each seed object.

[0091] For example, Figure 3 The figure is a schematic diagram of an object representation query process in an embodiment. The server queries n seed objects, and their respective object identifiers are: object identifier uid1, ..., object identifier uidn. The server sequentially queries the corresponding object representations from the local object representation storage space or the object representation storage space of other servers used to store object representations according to uid1, ..., uidn, that is, object representations ,…,Object Representation .

[0092] Alternatively, when determining that the object representation update condition is met at the current moment, the server updates the object representation of the seed object to obtain the object representation of the seed object.

[0093] In this embodiment, multiple objects that have performed at least one interactive behavior on the content within the first preset time period are used as seed objects of the content to query the group of people who are interested in the content. In this way, the object representation belonging to the group of people can be obtained in a targeted manner, ensuring that the interests of the group of people can be accurately described later, thereby ensuring the accuracy of the content representation.

[0094] In some embodiments, the object representation of each seed object is determined separately, including: for each seed object, when an object representation update condition of the seed object is met, calling a representation extraction model to extract the representation of the interaction behavior data of the seed object within a second preset time length to obtain the object representation of the seed object.

[0095] The object representation update condition refers to a condition for updating the object representation. Optionally, the object representation update condition of each seed object may be the same or different.

[0096] The representation extraction model is a model built based on a neural network and is used to extract object representations.

[0097] The second preset duration refers to the duration between the current time and the last update time of the last object representation update.

[0098] Optionally, the server checks whether the object representation update condition is met at the current moment. If so, it determines that the object representation of the seed object needs to be updated at the current moment, and obtains the interaction behavior data of each seed object within a second preset time length. The interaction behavior data records the content identifiers of multiple contents that the seed object has interacted with within the second preset time length.

[0099] The server calls the trained representation extraction model, inputs the interaction behavior data of each seed object into the representation extraction model, and obtains the object representation of the seed object.

[0100] If not satisfied, it is determined that the object representation of the seed object does not need to be updated at the current moment, and the object representation of each seed object obtained in the previous update is used as the object representation of the seed object at the current moment.

[0101] Exemplarily, the server determines the last update time of the object representation update and calculates the duration between the current time and the last update time. If the duration is equal to the second preset duration, it is determined that the current time meets the object representation update condition, and returns to call the trained representation extraction model step to continue execution.

[0102] If it is not equal to the second preset time length, it is determined that the current moment does not meet the object representation update condition, and the process returns to the step of determining that the object representation of the seed object does not need to be updated at the current moment and continues to execute.

[0103] In the above example, the second preset duration can be understood as a period for updating the object representation.

[0104] Exemplarily, the first preset duration may be different from or the same as the second preset duration;

[0105] If they are the same, and the first preset duration and the second preset duration are respectively regarded as the content representation update cycle and the object representation update cycle, it means that the object representation update and the content representation update are synchronously updated regularly, such as Figure 4 , which is a schematic diagram of a representation update step in an embodiment. In the case where the object representation and the content representation are updated synchronously, the server determines the duration between the current time and the last update time of the last update. If the duration is greater than or equal to the preset duration (the first preset duration or the second preset duration), it is determined that the update condition is met, and the server needs to synchronously update the content representation of each content and each seed object representation, as follows:

[0106] The server regards each content in the exposed state in the candidate pool as content to be processed, and determines a seed object for each content.

[0107] For any seed object i, the server returns to call the representation extraction model step to continue execution and determine the object representation of seed object i , the object identifier uidi and the object representation Send to other servers (for storing object representations), such as the server SU in the figure, so as to be stored in SU accordingly. For any content j, the server returns to step S202 to continue execution and obtains the corresponding content representation , and the content identifier nidj and the content representation The content is sent to another server, such as the server SN (used to store content representation) in the figure, so as to be stored accordingly in SN.

[0108] In some other examples, the method further includes: for each seed object, the server counts the most active time period of the seed object within a preset period (i.e., the frequency of interactive behaviors in the time period is the highest), and the previous time period before the time period and adjacent to the time period is used as the object representation update period. The server determines whether the current moment is in the object representation update period. If so, it is determined that the object representation update condition corresponding to the seed object is met, and returns to the step of calling the representation extraction model to continue execution. If not, it is determined that it is not met, and it is determined not to update the object representation.

[0109] For example, the most active time period of seed object 1 is from 12:00 to 13:00, and the corresponding object representation update time period is from 11:00 to 12:00. Then the object representation update condition 1 corresponding to seed object 1 is the time period between 11:00 and 12:00 at the current time.

[0110] The most active period of seed object 2 is from 9 o'clock to 10 o'clock, and the object representation update period is from 8 o'clock to 9 o'clock. Then the object representation update condition 2 corresponding to seed object 2: the current time is between 8 o'clock and 9 o'clock.

[0111] If the current time is 11:30, it is obvious that seed object 1 needs to update its object representation, while seed object 2 does not need to update its object representation.

[0112] In this way, based on the interaction habits of the seed objects themselves, more matching object representation update conditions are set for the seed objects, ensuring that the object representation of each seed object is more accurate.

[0113] In this embodiment, for each seed object, it is pre-checked whether the object representation update condition of the seed object is met to determine whether to update the corresponding object representation, ensuring that the content representation can be determined more accurately based on the more real-time object representation.

[0114] Step S204: Enhance the projection of each object representation on the unit vector to be solved to obtain the target function.

[0115] The unit vector to be solved refers to a unit vector representing the content to be solved for the content, and the unit vector refers to a vector whose modulus is the unit value 1. It can be understood that the essence of the unit vector to be solved is to solve the direction of the unit vector.

[0116] The projection of each object representation on the unit vector to be solved actually refers to the inner product of the object representation and the unit vector to be solved, which is equal to the product of the modulus of each object representation, the modulus of the unit vector, and the corresponding cosine value of each object representation (the cosine value of the angle between the object representation and the unit vector).

[0117] Since the modulus of a unit vector is always 1 and the modulus of each object representation is known, the projection of each object representation onto the unit vector to be solved is only related to the cosine value of the angle between each object representation and the unit vector to be solved.

[0118] It should be noted that the object set of seed objects that are interested in the content is regarded as the population of the content. Then, after a unit vector is determined, the larger the angle, the smaller the cosine value, the smaller the corresponding projection, the less similar the object representation is to the unit vector to be solved, and the seed object corresponding to the object representation belongs to a discrete seed object. The smaller the angle, the larger the cosine value, the larger the corresponding projection, the more similar the object representation is to the unit vector to be solved, and the seed object corresponding to the object representation belongs to a non-discrete seed object.

[0119] The enhancement processing of the projection refers to the adaptive scaling processing of the projection, that is, the smaller the projection is, the smaller the projection is, the smaller the projection is, and the larger the projection is, the larger the projection is.

[0120] It should be noted that, by enhancing the projection corresponding to each object representation, the large projection becomes a larger projection, and the small projection becomes a smaller projection, so that the gap between the large projection and the small projection can be magnified.

[0121] Exemplarily, the projection enhancement process may be performed by calculating the product of the projection and a large positive integer, for example, the large positive integer is 10,000.

[0122] Exemplarily, the enhancement processing for the projection may be to calculate the power of the projection, for example, to calculate the second power, the third power, the fourth power, etc. of the projection.

[0123] Exemplarily, the projection enhancement processing may also be performed by calling a preset function for calculation, where the preset function is a direct proportional function, such as the preset function is: f(x)=1 / (1-x).

[0124] The objective function is a function whose independent variable is the unit vector to be solved, and the dependent variable is the sum of the powers after each enhancement. After the content is determined, the object representation of the corresponding seed object is a known quantity in the target object.

[0125] Optionally, for each object representation, the server determines an expression of the projection of the independent variable and the object representation, and performs corresponding enhancement processing on the expression to obtain an enhanced expression corresponding to the object representation, and superimposes the enhanced expressions corresponding to each object representation to obtain the objective function.

[0126] For example, let the unit vector to be solved be x, then the objective function The formula (4) is:

[0127] (4)

[0128] in, is the enhanced expression corresponding to the seed object i, which is specifically expressed as follows:

[0129] (5)

[0130] in, is the projection of the object representation of the seed object i on the unit vector to be solved, and f(.) is the enhancement processing.

[0131] In order to further distinguish the changes in the projection before and after enhancement, each object representation is normalized to obtain the target object representation (that is, the modulus of the target object representation is the unit value 1); at this time, the projection of each target object representation on the unit vector to be solved is enhanced to obtain the objective function.

[0132] In some embodiments, the projection of each object representation on the unit vector to be solved is enhanced to obtain a target function, including: for each object representation, the projection of the object representation on the unit vector to be solved is squared to obtain the power corresponding to the object representation; and the power corresponding to each object representation is superimposed to obtain the target function.

[0133] As mentioned above, the exponentiation process can be quadratic, cubic, etc.

[0134] The following is an example of quadratic equation. The expression of maximizing the objective function is shown in the following formula (6):

[0135] (6)

[0136] in, is the unit vector to be solved corresponding to content j.

[0137] Of course, in other examples, each object representation can be normalized separately to obtain the corresponding target object representation. For each target object representation, the projection of the target object representation on the unit vector to be solved is squared to obtain the corresponding power of the object representation; the corresponding power of each target object representation is superimposed to obtain the objective function.

[0138] The following is a comparison of the differences between the object representations determined in this application and the related art with specific examples, such as Figure 5 As shown, it is a schematic diagram of object representation comparison in an embodiment. Figure 5 It is a schematic diagram under the scenario where the concentration of the crowd decreases.

[0139] Figure 5 The right part is the content representation 1 determined by this embodiment, and the left part is the content representation 2 determined by the related technology. In the process of determining the content representation of a certain content by the related technology, the weight of the object representation of each corresponding seed object is equal, that is, the object representation weights of the three discrete seed objects in the area 11 where the discrete seed objects are located in the left part and the object representation weights of the eight non-discrete seed objects in the area 12 where the non-discrete seed objects are located are the same, and the content representation 1 determined by it cannot accurately describe the interests of the group of people.

[0140] The object representation weights of the three discrete seed objects in the area 21 where the outlier seed objects are located in the right part are smaller than the object representation weights of the eight discrete seed objects in the area 22 where the non-outlier seed objects are located. That is, the influence of the outlier seed objects is reduced by quadratic processing, and the determined content representation 2 is closer to the object representation of the non-discrete objects, that is, it can more accurately describe the interests of the group of people.

[0141] In this embodiment, by performing a power processing on the projection corresponding to each object representation, the projection of the object representation of the object farther away from the crowd can be made smaller, and the projection of the object representation of the seed object closer to the crowd can be made larger.

[0142] Step S206, taking the maximization of the objective function as the optimization goal, and solving the unit vector solution when the optimization goal is achieved.

[0143] It should be noted that the essence of maximizing the objective function is to weaken the influence of discrete seed objects and increase the influence of close seed objects (seed objects concentrated in the crowd).

[0144] The solution result refers to solving a unit vector in the target direction, and the sum of the projections of all object representations corresponding to the content after enhancement in the target direction is greater than the sum of the projections of all object representations in other directions after enhancement.

[0145] The following continues to introduce the process of solving the unit vector to be solved when the exponentiation is quadratic. In some embodiments, the optimization goal is to maximize the objective function, and the solution result of the unit vector when the optimization goal is achieved is solved, including: for each object representation, the object representation is transposed to obtain the corresponding transposed representation, and the inner product of the object representation and the corresponding transposed representation is calculated; the inner products corresponding to each object representation are superimposed to obtain the target matrix; at least one eigenvalue of the target matrix is ​​determined, and the eigenvector corresponding to the maximum eigenvalue in at least one eigenvalue is determined, and the eigenvector is used as the solution result of the unit vector.

[0146] Exemplarily, for content j, the server determines the corresponding n object representations, namely , …, , and then determine the corresponding transposed representation, that is, , …, ,for ,calculate and The inner product of The corresponding product value; ...; for ,calculate and The inner product of The corresponding product value determines the target matrix based on the sum of n product values.

[0147] Exemplarily, if the sum is a square matrix, the sum is directly used as the target matrix. Alternatively, if the sum is not a square matrix, the server uses the sum as the matrix to be processed, and uses principal component analysis to determine the covariance matrix of the matrix to be processed, and uses the covariance matrix as the target matrix.

[0148] The server performs eigenvalue decomposition on the target matrix to obtain multiple eigenvalues, selects the maximum eigenvalue, and uses the maximum eigenvalue and the target matrix to solve the corresponding eigenvector, and uses the solved eigenvector as the solution result of the unit vector.

[0149] The following is a specific example to illustrate the correctness of the above solution process:

[0150] After obtaining formula (6), we transpose it and change the equation to get formula (7):

[0151] (7)

[0152] Among them, based on , determine the target matrix A, and obtain the following formula (8):

[0153] (8)

[0154] Among them, the target matrix A is a square matrix. Based on formulas (6)-(8), solving the maximum value of the target function is equivalent to solving The maximum value of .

[0155] Next, let According to formula (8), the actual solution is The maximum value of is a unit vector, then, , which is equivalent to calculating the Rayleigh quotient ,in, According to the Rayleigh quotient theorem, for any unit vector, ,in is the largest eigenvalue of the matrix A.

[0156] therefore, Taking the maximum value, the solution must be the eigenvector corresponding to the maximum eigenvalue, that is, the eigenvector corresponding to the maximum eigenvalue is the solution result of the unit vector.

[0157] That is to say, when the power is quadratic, the solution is the eigenvector of the maximum eigenvalue corresponding to the target matrix.

[0158] In this embodiment, by calculating the inner product of each object representation and the corresponding transposed representation, which is equivalent to calculating the power of the object representation in the quadratic, and according to the inner product corresponding to each object representation, the representation is converted into the problem of solving the eigenvector of the matrix. In this way, by solving the eigenvector corresponding to the maximum eigenvalue of the target matrix, the solution process is completed more simply and accurately, thereby improving the efficiency and accuracy of content representation solution.

[0159] Step S208: determining the content representation of the content based on the solution result of the unit vector.

[0160] Exemplarily, the server directly uses the unit vector in the target direction in the solution result of the unit vector as the content representation of the content.

[0161] Exemplarily, the server denormalizes the unit vector in the target direction in the solution result to obtain a content representation of the content, where the modulus of the content representation is equal to the modulus of the object representation.

[0162] It should be noted that normalization refers to changing the modulus of a vector to 1. It can be understood that denormalization is the inverse operation of normalization, which is to scale the values ​​of each element in a unit vector with a modulus of 1 proportionally so that the modulus of the vector after denormalization is not 1.

[0163] Exemplarily, after the content representation of the content to be processed is determined, the content identifier and the content representation of the content are stored in a content representation storage space in correspondence. The content representation storage space may be a local storage space or another server for storing content representations, which is not specifically limited.

[0164] In the above content representation determination method, the object representation of the seed objects that are interested in the content to be processed is obtained to accurately reflect the object information of each seed object in the crowd that is interested in the content. Then, the projection of each object representation on the unit vector to be solved is determined, and the projection is enhanced, that is, the projection of the object representation of the object farther away from the crowd is smaller, and the projection of the object representation of the seed object closer to the crowd is larger. Then, taking the maximization of the objective function as the optimization goal, the influence of discrete seed objects far away from the crowd can be effectively reduced. In this way, regardless of whether the concentration of the crowd decreases, the solution result of the unit vector when the optimization goal is achieved can accurately describe the content representation to improve the content push effect.

[0165] In some embodiments, the method also includes: in response to a push request, obtaining a corresponding target object representation based on an object identifier carried in the push request; obtaining content representations of multiple contents to be screened; determining recalled content based on the similarity between each content representation and the target object representation; and the recalled content is used to determine the push content for the push request.

[0166] The push request is a request sent by any terminal, and is used to request content push. The push request carries the object identifier of the object initiating the request.

[0167] Optionally, the target application on any terminal is logged in with the object identifier, and the terminal generates a push request carrying the object identifier in response to a trigger operation on the interface of the target application, and sends it to the server. The trigger operation is an operation for requesting a push, such as a refresh operation on the interface, a query operation on the search box in the interface, etc.

[0168] The server parses the received push request and obtains the object identifier. Based on the object identifier, the server queries the corresponding target object representation from the object representation storage space. The server uses the content of the candidate pool as the content to be screened, and obtains the corresponding content representation from the content representation storage space based on the content identifier of each content to be screened.

[0169] Calculate the similarity between the target object representation and each content representation, and sort them from high to low according to the similarity, and determine the target number of contents with the front sequence number as the recalled content. According to the target number of recalled contents, perform rough sorting, fine sorting and mixed sorting in turn to obtain the pushed content, and return it to the terminal for exposure.

[0170] Exemplarily, the server uses the time when the push request is received as the current time, and determines whether the content representation update condition is met based on the current time;

[0171] If satisfied, refer to the previous text, update the content representation of each content, calculate the similarity between the target object representation and the updated content representation, and return to the step of determining the recalled content to continue execution.

[0172] If not, the step of returning to the content identifier based on the selected content and acquiring the corresponding content representation from the content representation storage space is continued.

[0173] Exemplarily, if the triggering operation is a query operation, the terminal obtains the input term entered in the search box, generates a push request according to the input term and the object identifier, and sends it to the server.

[0174] The server obtains the input term and the object identifier by parsing the push request, and searches for the corresponding target object representation from the object representation storage space based on the object identifier. The server performs keyword recall based on the input term to obtain the first amount of content.

[0175] The server returns to the above step of using the content in the candidate pool as the content to be screened, and continues to execute until the similarities corresponding to the respective content representations are obtained, and the content is sorted from high to low according to the similarities, and the second number of content with the highest sequence number is screened out. The first number of content and the second number of content are both used as recalled content.

[0176] The keyword recall process may include: obtaining key text of each content in the candidate pool, such as the title and summary of the content, and filtering out a first number of contents with the highest similarity based on the similarity between the input terms and each key text of the content.

[0177] In this example, the recall based on object representation is personalized recall. Therefore, by combining keyword recall and personalized recall, multi-channel recall is achieved to avoid missing relevant content in the recall process and ensure the effectiveness and accuracy of the recall.

[0178] In this embodiment, in response to a push request, the corresponding target object representation is determined based on the object identifier carried in the push request. Since the content representation that has been aligned is obtained by reducing the influence of discrete seed objects, its content representation is more accurate. Therefore, based on the more accurate content representation and target object representation, more appropriate push content can be pushed to the object to which the object identifier belongs, thereby improving the content push effect.

[0179] The present application also provides an application scenario, which applies the above-mentioned content representation determination method. The application of the content representation determination method in this application scenario is as follows: in the note push scenario (notes are diversified content carriers created and published by creators, which can integrate at least one form of expression such as text, pictures, videos, etc.), in order to be able to push notes more accurately, it can be achieved through the following steps: for the notes to be processed, the server obtains the object representation of the seed object of interest to the note; enhances the projection of each object representation on the unit vector to be solved to obtain the objective function; takes maximizing the objective function as the optimization goal, and solves the solution result of the unit vector when the optimization goal is achieved; determines the note representation of the note based on the solution result of the unit vector. In this way, after receiving the note push request sent by any terminal, multiple recalled notes are accurately recalled according to the similarity between the target object representation of the object identifier carried and each note representation, so as to achieve more accurate note push.

[0180] Of course, it is not limited to this. The content representation determination method provided in this application can also be applied to other scenarios, for example, in the scenario of product push, the corresponding content is the product, the content representation is the product representation, and the pushed content is the pushed product. For another example, in the scenario of video push, the corresponding content is the video, the content representation is the video representation, and the pushed content is the pushed video.

[0181] In a specific embodiment, Figure 6 As shown, it is a schematic diagram of a specific application environment in an embodiment. This embodiment is described by taking the note push scenario as an example. Figure 6 This embodiment involves the interaction between the terminal and the server cluster. Specifically, it involves the client of the target application on the terminal. The server cluster includes a target server ST for determining the note representation, a server SU for storing the object representation, and a server SN for storing the note representation. ST communicates with SU and SN respectively.

[0182] like Figure 7 FIG. 1 is a schematic diagram of a process for determining a note characterization in one embodiment. Figure 7 For relevant instructions:

[0183] Step 1: After receiving the push request sent by the terminal at the current time, the server ST checks whether the note representation update condition is met based on the current time. Then proceed to step 2.

[0184] The specific verification process can refer to the previous text, based on whether the content representation update conditions are met at the current moment, and the content is replaced with notes to perform a similar process.

[0185] Step 2: When the note representation update condition is satisfied at the current time, the server ST takes any note in the exposure state as a note to be processed. Then proceed to step 3, which is described by taking note j as an example.

[0186] If it is verified that the note representation update conditions are not met at the current moment, the note representation of each note obtained in the previous update is directly obtained from the server SN, and the process jumps to step 8 to continue.

[0187] Step 3: The server ST obtains n objects that have interacted with the note j within the first preset time period, and uses the n objects as seed objects of the note j, wherein the object identifiers of the seed objects are uid1, ..., uidn in sequence. Then proceed to step 4.

[0188] Step 4: The server ST verifies whether the object representation update condition is met based on the current moment. The specific verification process can refer to the previous text. If the object representation update condition is met based on the current moment, the note is replaced with the note to perform a similar process. Continue to step 5.

[0189] Step 5: When the object representation update condition of the seed object is met at the current moment, for each seed object, the server ST calls the representation extraction model to extract the representation of the interactive behavior data of the seed object within the second preset time length to obtain the object representation of the seed object.

[0190] The server ST sends the object identifiers of the n seed objects and the determined object representations to the server SU. The server SU queries the object representation obtained by the last object representation update according to each object identifier in turn, and replaces the object representation obtained by the last object representation update with the currently received object representation. Then, the process jumps to step 7 to continue.

[0191] Step 6: If the object representation update condition of the seed object is not met at the current moment, the server ST sends all n object identifiers to the server SU for object representation query. The server SU queries the n object representations obtained in the previous update according to each object identifier, namely , …, , and returns it to the server ST. Continue to step 7.

[0192] Step 7: After obtaining the object representations of each seed object, the server ST performs a representation alignment operation. The specific process is as follows:

[0193] For each object representation, the server ST squares the projection of the object representation on the unit vector to be solved to obtain the square corresponding to the object representation; superimposes the corresponding power of each object representation to obtain the objective function. The server ST solves the unit vector to be solved with maximizing the objective function as the optimization goal. The specific solution process is as follows:

[0194] For each object representation, the server ST transposes the object representation to obtain the corresponding transposed representation, and calculates the inner product of the object representation and the corresponding transposed representation; superimposes the inner products corresponding to each object representation to obtain the target matrix; determines at least one eigenvalue of the target matrix, and determines the eigenvector corresponding to the maximum eigenvalue in at least one eigenvalue, and uses the eigenvector as the solution result of the unit vector.

[0195] The server ST determines the note representation of note j based on the solution result. , and the note identifier and the determined note representation of note j The message is sent to the server SN. The server SN queries the note representation obtained from the last note representation update based on the note identifier, and replaces the note representation obtained from the last note representation update with the currently received note representation. , and store. Continue to step 8.

[0196] Step 8: The server ST determines the object identifier carried in the push request, and sends the object identifier to the server SU. The server SU returns the target object representation corresponding to the object identifier to the server ST.

[0197] The server ST uses the content in the candidate pool as the content to be screened, determines the content identifier of each content to be screened, and sends a note representation acquisition request carrying each content identifier to the server SN. The server SN returns the note representations of multiple notes to be screened to the server ST. Continue to step 9.

[0198] Step 9: The server ST selects the TopK notes as recalled notes based on the similarity between each note representation and the target object representation, in descending order of similarity; the recalled notes are used to determine the pushed notes for the push request.

[0199] It should be noted that in order to verify the note push effect, an online comparative experiment was conducted in the note recommendation system, and the experimental group was set up using this embodiment. Compared with the control group, the experimental group greatly improved the new note recall index (cold start completion rate increased by +8.6%) while the user experience was basically intact. In other words, the embodiment of the present application can more accurately describe the interests of the crowd and achieve the established goal of accurately recalling and spreading to the interested groups.

[0200] In this embodiment, the object representation of the seed object that is interested in the note to be processed is obtained to accurately reflect the object information of each seed object in the crowd that is interested in the note. Then, the projection of each object representation on the unit vector to be solved is determined, and the projection is enhanced, that is, the projection of the object representation of the object farther away from the crowd is smaller, and the projection of the object representation of the seed object closer to the crowd is larger. Then, by maximizing the objective function as the optimization goal, the influence of discrete seed objects far away from the crowd, that is, the calculation offset of the effective discrete seed objects, can be effectively reduced. In this way, regardless of whether the concentration of the crowd decreases, the solution result of the unit vector when the optimization goal is achieved can accurately describe the note representation, thereby realizing the precise diffusion of the crowd's interests, and ultimately improving the quality of note distribution to improve the note push effect.

[0201] It should be understood that, although the various steps in the flowcharts involved in the above-mentioned embodiments are displayed in sequence according to the indication of the arrows, these steps are not necessarily executed in sequence according to the order indicated by the arrows. Unless there is a clear explanation in this article, the execution of these steps does not have a strict order restriction, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowcharts involved in the above-mentioned embodiments can include multiple steps or multiple stages, and these steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a part of the steps or stages in other steps.

[0202] Based on the same inventive concept, the embodiment of the present application also provides a content representation determination device for implementing the content representation determination method involved above. The implementation solution provided by the device to solve the problem is similar to the implementation solution recorded in the above method, so the specific limitations in one or more content representation determination device embodiments provided below can refer to the limitations of the content representation determination method above, and will not be repeated here.

[0203] In an exemplary embodiment, Figure 8 As shown, a content representation determination device 800 is provided, comprising: an object representation acquisition module 802, a function acquisition module 804, a solution module 806 and a content representation determination module 808, wherein:

[0204] An object representation acquisition module 802 is used to acquire object representations of seed objects of interest in the content to be processed;

[0205] A function acquisition module 804 is used to enhance the projection of each object representation on the unit vector to be solved to obtain the target function;

[0206] A solution module 806 is used to take the maximization of the objective function as the optimization goal and solve the solution result of the unit vector when the optimization goal is achieved;

[0207] The content representation determination module 808 is used to determine the content representation of the content based on the solution result of the unit vector.

[0208] In some embodiments, the object representation acquisition module 802 is further configured to, when a content representation update condition is met, take any content in an exposed state as content to be processed.

[0209] In some embodiments, the object representation acquisition module 802 is further used to acquire multiple objects that have implemented interactive behaviors on the content within a first preset time period, and use the multiple objects as seed objects about the content; and determine the object representation of each seed object respectively.

[0210] In some embodiments, the object representation acquisition module 802 is also used to call the representation extraction model for each seed object, extract the representation of the interactive behavior data of the seed object within a second preset time period, and obtain the object representation of the seed object when the object representation update condition of the seed object is met.

[0211] In some embodiments, the function acquisition module 804 is used to, for each object representation, square the projection of the object representation on the unit vector to be solved to obtain the power corresponding to the object representation; and superimpose the power corresponding to each object representation to obtain the target function.

[0212] In some embodiments, the function acquisition module 804 is used to transpose each object representation to obtain a corresponding transposed representation, and calculate the inner product of the object representation and the corresponding transposed representation; superimpose the inner products corresponding to each object representation to obtain a target matrix; determine at least one eigenvalue of the target matrix, and determine the eigenvector corresponding to the maximum eigenvalue in at least one eigenvalue, and use the eigenvector as the solution result of the unit vector.

[0213] In some embodiments, the device also includes a recall module, which is used to respond to a push request and obtain a corresponding target object representation based on an object identifier carried in the push request; obtain content representations of multiple contents to be screened; determine the recalled content based on the similarity between each content representation and the target object representation; the recalled content is used to determine the push content for the push request.

[0214] Each module in the above content representation determination device can be implemented in whole or in part by software, hardware, or a combination thereof. Each module can be embedded in or independent of a processor in a computer device in the form of hardware, or can be stored in a memory in a computer device in the form of software, so that the processor can call and execute operations corresponding to each module.

[0215] In an exemplary embodiment, a computer device is provided. The computer device may be a server, and its internal structure diagram may be as shown in FIG. Fig. 9 As shown. The computer device includes a processor, a memory, an input / output interface (Input / Output, referred to as I / O) and a communication interface. The processor, the memory and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a content representation determination method is implemented.

[0216] Those skilled in the art will understand that Fig. 9 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0217] In one embodiment, a computer device is further provided, including a memory and a processor, wherein a computer program is stored in the memory, and the processor implements the steps in the above method embodiments when executing the computer program.

[0218] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in the above-mentioned method embodiments are implemented.

[0219] In one embodiment, a computer program product is provided, including a computer program, which implements the steps in the above method embodiments when executed by a processor.

[0220] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant regulations.

[0221] A person of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment method can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided in the present application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. Non-relational databases may include distributed databases based on blockchains, etc., but are not limited to this. The processor involved in each embodiment provided in this application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, an artificial intelligence (AI) processor, etc., but are not limited to this.

[0222] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0223] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

Claims

1. A method for determining content representation, characterized in that: The method comprises: For the content to be processed, obtaining an object representation of a seed object of interest in the content; The projection of each object representation on the unit vector to be solved is enhanced to obtain the objective function; Taking maximizing the objective function as the optimization goal, solving the solution of the unit vector when the optimization goal is achieved; A content representation of the content is determined based on a solution result of the unit vector.

2. The method according to claim 1, characterized in that Before obtaining the object representation of the seed object of interest in the content to be processed, the method includes: When the content representation update condition is met, any content in the exposure state is used as content to be processed.

3. The method according to claim 1, characterized in that The obtaining of the object representation of the seed object of interest to the content includes: Acquire multiple objects that have implemented interactive behaviors on the content within a first preset time period, and use the multiple objects as seed objects related to the content; The object representation of each seed object is determined separately.

4. The method according to claim 3, characterized in that: The determining of the object representation of each seed object respectively includes: For each seed object, when an object representation update condition of the seed object is met, a representation extraction model is called to extract the representation of the interaction behavior data of the seed object within a second preset time period to obtain the object representation of the seed object.

5. The method according to claim 1, characterized in that: The projection of each object representation on the unit vector to be solved is enhanced to obtain the objective function, including: For each object representation, performing a square process on the projection of the object representation on the unit vector to be solved to obtain the square corresponding to the object representation; The objective function is obtained by superimposing the corresponding powers of each object representation.

6. The method according to claim 5, characterized in that The step of taking maximizing the objective function as the optimization goal and solving the solution result of the unit vector when the optimization goal is achieved includes: For each object representation, transpose the object representation to obtain a corresponding transposed representation, and calculate an inner product between the object representation and the corresponding transposed representation; Superimpose the inner products corresponding to each object representation to obtain the target matrix; At least one eigenvalue of the target matrix is ​​determined, and an eigenvector corresponding to a maximum eigenvalue among the at least one eigenvalue is determined, and the eigenvector is used as a solution result of the unit vector.

7. The method according to any one of claims 1 to 6, characterized in that: The method further comprises: In response to the push request, obtaining a corresponding target object representation based on the object identifier carried in the push request; Obtaining content representations of multiple contents to be filtered; The recalled content is determined according to the similarity between each content representation and the target object representation; the recalled content is used to determine the push content for the push request.

8. A content representation determination device, characterized in that: The device comprises: An object representation acquisition module, used for acquiring, for the content to be processed, an object representation of a seed object of interest to the content; A function acquisition module is used to enhance the projection of each object representation on the unit vector to be solved to obtain the target function; A solution module, used for taking maximization of the objective function as an optimization goal and solving a solution result of the unit vector when the optimization goal is achieved; The content representation determination module is used to determine the content representation of the content based on the solution result of the unit vector.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

11. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images