Content recommendation method, device and storage medium for large model cold start

Abstract

The present application discloses a content recommendation method, device and storage medium for cold start of a large model, including: collecting target behavior data corresponding to a target user, and inputting the target behavior data into a first model and a second model to generate a first feature vector; based on the first feature vector, generating multiple distance functions corresponding to each center point; according to the multiple distance functions, determining a target optimization function for maximizing the difference between the first distance and the second distance, and generating a second feature vector corresponding to the target user; based on the second feature vector, determining a first behavior data sample; inputting the target behavior data and the first behavior data sample into the large model together, and outputting an interest topic category corresponding to the target user; and when the interest topic category corresponding to the target user is determined, generating recommended content corresponding to the target user, and returning it to the terminal device.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and in particular to a content recommendation method, device and storage medium for cold start of a large model. Background Art

[0002] With the continuous development of big models, the use of recommendation systems based on big models is becoming more and more widespread. However, in the above recommendation systems, there is a very common and unavoidable problem, namely the cold start problem. The cold start problem refers to the situation where when the recommendation system is just started or when a new user uses the recommendation system, the recommendation system lacks sufficient user behavior data, so it is difficult for the recommendation system to provide users with accurate personalized recommendations.

[0003] For example, a user downloads a video player software and registers as a new user. When the page of the video player software jumps from the new user registration page to the homepage, the video player software needs to generate personalized recommendation content. The personalized recommendation content needs to be consistent with the user's interests in order to attract the user's use. Therefore, it can be seen from the above that it is very necessary for the recommendation system to provide accurate personalized recommendation content to attract users.

[0004] In response to the above-mentioned problems, some solutions have been proposed, such as generating personalized recommendation content based on the registration information of new users, generating personalized recommendation content based on the registration environment of new users, or generating personalized recommendation content based on the age group and gender of new users. However, no matter which of the above solutions is used, it is only a coarse-grained recommendation and cannot make personalized recommendations for the user itself. In other words, the above solutions all have the problem of low recommendation efficiency and low recommendation accuracy.

[0005] Currently, no effective solution has been proposed to the technical problem in the above-mentioned prior art that, due to the lack of sufficient user behavior data, the recommendation system based on the large model is difficult to provide accurate personalized recommendation content to the target users. Summary of the invention

[0006] The embodiments of the present disclosure provide a content recommendation method, device and storage medium for large model cold start, so as to at least solve the technical problem existing in the prior art that the recommendation system based on the large model has difficulty in providing accurate personalized recommended content to the target users due to the lack of sufficient user behavior data.

[0007] According to one aspect of an embodiment of the present disclosure, a content recommendation method for a large model cold start is provided, comprising: collecting target behavior data corresponding to a target user, and inputting the target behavior data into a pre-trained first model and a second model, thereby generating a first feature vector corresponding to the target user; based on the first feature vector, generating a plurality of distance functions corresponding to each center point, wherein each center point corresponds to a different sample user category; based on the plurality of distance functions, determining a target optimization function for indicating the difference between the first distance and the second distance, and generating a second feature vector corresponding to the target user when the target optimization function is maximized, wherein the first distance is used to indicate the difference between the first feature vector and each center point. a second minimum distance between the first feature vector and each center point; based on the second feature vector, determining a first behavior data sample, wherein the first behavior data sample is used to indicate a second behavior data sample corresponding to a similar user of the target user in the target sample user category, and wherein the second behavior data sample is used to indicate behavior data collected in advance and different from the target user; inputting the target behavior data and the first behavior data sample into the big model together, and using the big model to output the interest topic category corresponding to the target user; and in the case of determining the interest topic category corresponding to the target user, generating recommended content corresponding to the target user, and returning the recommended content to the terminal device corresponding to the target user.

[0008] According to another aspect of an embodiment of the present disclosure, a storage medium is further provided, the storage medium including a stored program, wherein when the program is running, a processor executes any one of the methods described above.

[0009] According to another aspect of the embodiment of the present disclosure, there is also provided a content recommendation device for large-model cold start, including: a first data acquisition module, used to acquire target behavior data corresponding to a target user, and input the target behavior data into a pre-trained first model and a second model, thereby generating a first feature vector corresponding to the target user; a distance function generation module, used to generate a plurality of distance functions corresponding to each center point based on the first feature vector, wherein each center point corresponds to a different sample user category; a second feature vector determination module, used to determine a target optimization function indicating the difference between the first distance and the second distance based on the plurality of distance functions, and to generate a second feature vector corresponding to the target user when the target optimization function is maximized, wherein the first distance is used to indicate the difference between the first feature vector and each center point. a second minimum distance between the first feature vector and each center point, the second distance is used to indicate the minimum distance between the first feature vector and each center point; a target sample user category determination module, used to determine the first behavior data sample based on the second feature vector, wherein the first behavior data sample is used to indicate the second behavior data sample corresponding to a similar user of the target user in the target sample user category, and wherein the second behavior data sample is used to indicate behavior data collected in advance and different from the target user; an interest topic category output module, used to input the target behavior data and the first behavior data sample into the big model together, and use the big model to output the interest topic category corresponding to the target user; and a recommended content generation module, used to generate recommended content corresponding to the target user when the interest topic category corresponding to the target user is determined, and return the recommended content to the terminal device corresponding to the target user.

[0010] According to another aspect of the embodiment of the present disclosure, there is also provided a content recommendation device for large model cold start, comprising: a processor; and a memory connected to the processor, for providing the processor with instructions for processing the following processing steps: collecting target behavior data corresponding to a target user, and inputting the target behavior data into a pre-trained first model and a second model, thereby generating a first feature vector corresponding to the target user; based on the first feature vector, generating a plurality of distance functions corresponding to each center point respectively, wherein each center point corresponds to a different sample user category; based on the plurality of distance functions, determining a target optimization function for indicating the difference between the first distance and the second distance, and generating a second feature vector corresponding to the target user when the target optimization function is maximized, wherein the first feature vector A distance is used to indicate the second minimum distance between the first feature vector and each center point, and a second distance is used to indicate the minimum distance between the first feature vector and each center point; based on the second feature vector, a first behavior data sample is determined, wherein the first behavior data sample is used to indicate a second behavior data sample corresponding to a similar user of the target user in the target sample user category, and wherein the second behavior data sample is used to indicate behavior data collected in advance and different from the target user; the target behavior data and the first behavior data sample are input into the big model together, and the big model is used to output the interest topic category corresponding to the target user; and when the interest topic category corresponding to the target user is determined, recommended content corresponding to the target user is generated, and the recommended content is returned to the terminal device corresponding to the target user.

[0011] The present application provides a content recommendation method for cold start of a large model. First, an application server collects target behavior data corresponding to a target user, and inputs the target behavior data into a first model and a second model. Then, the generation module generates a plurality of distance functions corresponding to each center point based on the first feature vector. Then, the optimization module determines a target optimization function for maximizing the difference between the first distance and the second distance based on the plurality of distance functions. Further, the optimization module determines a target optimization function based on the plurality of distance functions, and generates a second feature vector corresponding to the target user when the target optimization function is maximized. Then, the generation module determines a first behavior data sample based on the second feature vector. Then, the generation module inputs the target behavior data and the first behavior data sample into the large model together, and uses the large model to output the interest topic category corresponding to the target user. Finally, when the interest topic category corresponding to the target user is determined, the generation module generates recommended content corresponding to the target user, and the application server returns the recommended content to the terminal device corresponding to the target user through the user interface module.

[0012] With reference to the above contents, it can be seen that the present application predetermines a first feature vector corresponding to a target user, and based on a plurality of distance functions corresponding to each center point, determines a target optimization function for indicating the maximization of the difference between the first distance and the second distance. Thus, when the target optimization function is maximized (i.e., when the difference between the first distance and the second distance is maximized), the variable value in the target optimization function and the target sample user category can be determined, and a second feature vector corresponding to the target user is generated based on the variable value.

[0013] In addition, since the target sample user category corresponding to the target user has been determined based on the above optimization algorithm, the first behavior data sample corresponding to the target sample user category can be used as supplementary data for the target behavior data and input into the large model together with the target behavior data. Thus, the large model is used to output the interest topic category corresponding to the target user. In other words, although there is a lack of user behavior data corresponding to the target user, when a similar user corresponding to the target user is determined, the first behavior data sample corresponding to the similar user can be used as supplementary data, and the target behavior data and the first behavior data sample can be input into the large model together, thereby making up for the lack of sufficient user behavior data.

[0014] Thus, even in the case of a cold start, the technical effect of being able to provide accurate personalized recommendation content to the target user based on a small amount of target behavior data corresponding to the target user and a large model is achieved. This solves the technical problem in the prior art that the recommendation system based on a large model is difficult to provide accurate personalized recommendation content to the target user due to the lack of sufficient user behavior data. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] The drawings described herein are used to provide a further understanding of the present disclosure and constitute a part of the present application. The illustrative embodiments of the present disclosure and their descriptions are used to explain the present disclosure and do not constitute an improper limitation on the present disclosure. In the drawings:

[0016] Figure 1 is a hardware structure block diagram of a computing device for implementing the method according to Embodiment 1 of the present disclosure;

[0017] Figure 2 is a schematic diagram of a content recommendation system for cold start of a large model according to Example 1 of the present application;

[0018] Figure 3 is a modular schematic diagram of the application server according to Embodiment 1 of the present application;

[0019] Figure 4 is a flowchart of the content recommendation method for large model cold start according to Example 1 of the present application;

[0020] Figure 5 is a schematic diagram of outputting a first feature vector using a first model and a second model according to Embodiment 1 of the present application;

[0021] Figure 6 is a schematic diagram of the first characteristic vector according to Embodiment 1 of the present application;

[0022] Figure 7 is a schematic diagram of the model architecture of the third model according to Embodiment 1 of the present application; and

[0023] Figure 8 is a schematic diagram of a content recommendation device for large model cold start according to Embodiment 2 of the present application; and

[0024] Fig. 9 It is a schematic diagram of a content recommendation device for cold start of a large model according to Example 3 of the present application. DETAILED DESCRIPTION

[0025] In order to enable those skilled in the art to better understand the technical solutions of the present disclosure, the technical solutions in the embodiments of the present disclosure will be clearly and completely described below in conjunction with the drawings in the embodiments of the present disclosure. Obviously, the described embodiments are only embodiments of a part of the present disclosure, not all of the embodiments. Based on the embodiments in the present disclosure, all other embodiments obtained by ordinary technicians in this field without making creative work should fall within the scope of protection of the present disclosure.

[0026] It should be noted that the terms "first", "second", etc. in the specification and claims of the present disclosure and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present disclosure described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products, or devices.

[0027] Example 1

[0028] According to this embodiment, a method embodiment of content recommendation for large model cold start is provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and although a logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0029] The method embodiment provided in this embodiment can be executed in a mobile terminal, a computer terminal, a server or a similar computing device. Figure 1 FIG. 1 shows a hardware structure block diagram of a computing device for implementing a content recommendation method based on a large model cold start. Figure 1 As shown, the computing device may include one or more processors (the processor may include but is not limited to a processing device such as a microprocessor MCU or a programmable logic device FPGA), a memory for storing data, a transmission device for communication functions, and an input / output interface. The memory, the transmission device, and the input / output interface are connected to the processor via a bus. In addition, it may also include: a display, a keyboard, and a cursor control device connected to the input / output interface. A person skilled in the art can understand that Figure 1 The structure shown is only for illustration and does not limit the structure of the above electronic device. Figure 1 More or fewer components as shown, or with Figure 1 Different configurations are shown.

[0030] It should be noted that the one or more processors and / or other data processing circuits described above may generally be referred to herein as "data processing circuitry". The data processing circuitry may be embodied in whole or in part as software, hardware, firmware, or any other combination thereof. In addition, the data processing circuitry may be a single independent processing module, or may be incorporated in whole or in part into any of the other components of the computing device. As involved in the embodiments of the present disclosure, the data processing circuitry acts as a processor control (e.g., selection of a variable resistor terminal path connected to an interface).

[0031] The memory can be used to store software programs and modules of application software, such as the program instructions / data storage device corresponding to the content recommendation method for large model cold start in the embodiment of the present disclosure. The processor executes various functional applications and data processing by running the software programs and modules stored in the memory, that is, the content recommendation method for large model cold start of the above-mentioned application is realized. The memory may include a high-speed random access memory, and may also include a non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory may further include a memory remotely arranged relative to the processor, and these remote memories can be connected to the computing device via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and a combination thereof.

[0032] The transmission device is used to receive or send data via a network. The specific example of the above network may include a wireless network provided by a communication provider of the computing device. In one example, the transmission device includes a network adapter (Network Interface Controller, NIC), which can be connected to other network devices through a base station so as to communicate with the Internet. In one example, the transmission device can be a radio frequency (Radio Frequency, RF) module, which is used to communicate with the Internet wirelessly.

[0033] The display may be, for example, a touch screen liquid crystal display (LCD) that may enable a user to interact with a user interface of the computing device.

[0034] It should be noted that, in some optional embodiments, the above Figure 1 The computing device shown may include hardware elements (including circuits), software elements (including computer code stored on a computer-readable medium), or a combination of both hardware elements and software elements. Figure 1 This is merely one example of a particular embodiment and is intended to illustrate the types of components that may be present in the computing devices described above.

[0035] Figure 2 is a schematic diagram of a content recommendation system for a large model cold start according to this embodiment. Figure 2 As shown, the system includes: a terminal device 100, an application server 200, a large model server 300 running a recommendation system, and a database 400. The terminal device 100 is connected to the application server 200 through a network, and the application server 200 is connected to the large model server 300 through a network. In addition, the database 400 also stores pre-collected second behavior data samples corresponding to multiple users. And the second behavior data samples include the first behavior data samples.

[0036] Figure 3 2 is a modular schematic diagram of the application server 200 according to an embodiment of the present application. Figure 3 As shown, the application server 200 includes a user interface module, a generation module, a clustering module, a model training module and an optimization module.

[0037] The generation module and the clustering module are used to interact with the database 400, to perform clustering operations on the second behavior data samples in the database 400, and to generate center points corresponding to each sample user category. In addition, the generation module is also used to generate a first feature vector corresponding to the target user based on the target behavior data corresponding to the target user, and to generate multiple distance functions corresponding to each center point based on the first feature vector. The optimization module is used to determine the target optimization function according to multiple distance functions, and to generate a second feature vector corresponding to the target user when the target optimization function is maximized. The generation module is also used to determine the first behavior data sample (i.e., the second behavior data sample corresponding to a similar user of the target user in the target sample user category) based on the second feature vector. The generation module is also used to interact with the large model server 300 when determining the target sample user category, and to generate interest topic categories and personalized recommended content corresponding to the target user. The user interface module interacts with the terminal device 100, and the application server 200 returns the personalized recommended content corresponding to the target user to the terminal device 100 through the user interface module.

[0038] In addition, the model training module is used to construct the first model and the second model, and train the first model and the second model.

[0039] It should be noted that the terminal device 100, application server 200, large model server 300 and database 400 in the system can all be applied to the hardware structure described above.

[0040] Under the above operating environment, according to the first aspect of this embodiment, a content recommendation method for a large model cold start is provided. The method comprises: Figure 2 and Figure 3 The application server 200 shown in FIG.

[0041] Figure 4 A schematic diagram showing the process of the method is shown in FIG. Figure 4 As shown, the method includes:

[0042] S402: collecting target behavior data corresponding to the target user, and inputting the target behavior data into a pre-trained first model and a second model, thereby generating a first feature vector corresponding to the target user;

[0043] S404: Based on the first feature vector, generate a plurality of distance functions corresponding to each center point respectively, wherein each center point corresponds to a different sample user category respectively;

[0044] S406: determining, according to the plurality of distance functions, a target optimization function for indicating a difference between the first distance and the second distance, and generating a second feature vector corresponding to the target user when the target optimization function is maximized, wherein the first distance is used to indicate a second minimum distance between the first feature vector and each center point, and the second distance is used to indicate a minimum distance between the first feature vector and each center point;

[0045] S408: Determine a first behavior data sample based on the second feature vector, wherein the first behavior data sample is used to indicate a second behavior data sample corresponding to a user similar to the target user in the target sample user category, and wherein the second behavior data sample is used to indicate behavior data collected in advance and different from the target user;

[0046] S410: inputting the target behavior data and the first behavior data sample into the big model together, and using the big model to output the interest topic category corresponding to the target user; and

[0047] S412: When the interest topic category corresponding to the target user is determined, recommended content corresponding to the target user is generated, and the recommended content is returned to the terminal device corresponding to the target user.

[0048] Specifically, first, when the target user registers as a new user using the terminal device 100 or just starts the recommendation system, the application server 200 responds to the target user P q The operation behavior of the target user P is collected through the user interface module q The corresponding target behavior data S q Among them, the target behavior data S q For example, the application server 200 can be in the target user P q Behavior data collected when operating using the terminal device 100. For example, the target user P q Like or give a negative review to an article on a web page; for example, the target user P q The behavior of chatting through instant messaging tools, etc. However, since the target user Pq is a new user, the application server 200 can only collect the target user P q Therefore, the sparse behavior data of the target user P q The corresponding target behavior data S q For example, it only includes a small amount of data, and this small amount of data is not sufficient to support a large model-based recommendation system to accurately generate interest topics corresponding to the target users.

[0049] In addition, the application server 200 collects the target user P q The corresponding target behavior data S q In the case of q Divided into a plurality of first behavior indicator data I 1 ~I m The preset behavior indicators include, but are not limited to, liked articles, badly rated articles, commented articles, watched videos, liked videos, videos with a stay time of more than 5 seconds, and purchased goods, etc. Those skilled in the art should note that the above are only examples of behavior indicators, and the actual situation is not limited to this.

[0050] However, it is worth noting that since the above preset behavior indicators are obtained by the generation module based on a large amount of historical behavior data, there can be multiple different behavior indicators. However, since the target user obtained is a single individual, in the case of cold start, it is different from the target user P q The corresponding first behavior is indicator data I 1 ~I m The data may include some behavioral indicator data or may not include some behavioral indicator data. q There are no “bad review articles” or “purchased products”, so the target user P q There is no first behavior indicator data corresponding to “bad review article” and no first behavior indicator data corresponding to “purchased product”.

[0051] And in the target user P q The corresponding first behavior is indicator data I 1 ~I m In the case where certain behavior indicator data is not included, the corresponding behavior indicator data is "empty". q The corresponding first behavior indicator data I 1 ~I m For example, it may include the first behavior indicator data I corresponding to "liking the article" 1 (For example, it can be the text content of the article that the user has liked), and the first behavior indicator data I corresponding to the “bad review article” 2 is "empty", and the first line corresponding to "comment article" is indicator data I 3 , the first behavior indicator data I corresponding to the “watched video” 4 (For example, it can be text information related to the video content watched by the user), and the first behavior indicator data I corresponding to the “liked video” 5 ,......, the first behavior corresponding to "video with a stay time of more than 5s" is indicator data Im-1 , the first behavior corresponding to "purchased goods" is the indicator data I m is "empty".

[0052] Figure 5 is a schematic diagram of outputting a first feature vector using a first model and a second model according to an embodiment of the present application. Figure 5 As shown, the generation module uses the pre-trained first model to generate a plurality of first behavior indicator data I 1 ~I m Vectorize and generate corresponding multiple first-line indicator vectors Q 1 ~Q m The generation module uses the pre-trained second model to generate multiple first indicator behavior vectors Q 1 ~Q m Perform feature extraction to generate the target user P q The corresponding first eigenvector Q x1 (S402). The first model includes, for example, a One-hot encoder, a Word2vec model, and a Transformer-based encoding model, and the second model includes, for example, a plurality of convolutional layers. Thus, the first model can process the plurality of first behavior indicator data I 1 ~I m The second model can encode and vectorize the multiple first indicator behavior vectors Q generated by the first model. 1 ~Q m Extract features. q The corresponding first behavior indicator data I 1 ~I m When input into the first model and the second model, the target user P q The missing first line is the first eigenvector Q corresponding to the indicator data x1 The position is "empty". Figure 6 is a schematic diagram of a first eigenvector according to an embodiment of the present application. Figure 6 As shown, for example, in the absence of the first behavior indicator data corresponding to “bad review articles” and the first behavior indicator data corresponding to “purchased goods”, the first feature vector can be expressed as Q x1 =[q 1 , 0, q 2 ,q 3 ,q 4 ,q 5 , ..., q m-1 ,0] TIt should be noted that the applicant uses 0 to represent "empty" in the first feature vector, but the actual situation is not limited to this. The above content will be described in detail later, so it will not be repeated here.

[0053] Then, the generation module generates a plurality of distance functions corresponding to each center point based on the first feature vector (S404). Specifically, before the generation module generates a plurality of distance functions corresponding to each center point based on the first feature vector, the clustering module needs to predetermine a plurality of center points corresponding to different sample user categories. First, the clustering module determines a plurality of different sample user categories using a clustering algorithm and a plurality of pre-collected second behavior data samples. Thereafter, the clustering algorithm determines a plurality of center points corresponding to each sample user category based on the determined plurality of different sample user categories. The content of the clustering module determining a plurality of center points corresponding to each sample user category based on the clustering algorithm will be described in detail later, so it will not be described in detail here.

[0054] In the case where the clustering module determines multiple center points corresponding to different sample user categories, the generation module determines multiple distance functions. Specifically, the first feature vector can be expressed as Q x1 =[q 1 , 0, q 2 ,q 3 ,q 4 ,q 5 , ..., q m-1 ,0] T , and then the generation module sets the position of the first eigenvector "0" to y 1 ~y l , and determine the first eigenvector Q x1 The distance d from each center point 1 ~d k , so based on the vector distance calculation formula, we can determine the 1 ~y l Multiple distance functions D 1 (y 1 ~y l )~D t (y 1 ~y l ). Among them, l is less than m.

[0055] For example, the first eigenvector Q x1 The distance d from the first center point 1 It can be expressed as So with the distance d 1 The corresponding distance function can be expressed as D 1 (y 1 ~yl ). Among them, a 1,1 represents the coordinates corresponding to the first line index of the first center point, a 1,2 represents the coordinates corresponding to the second line index of the first center point, ..., a 1,m-1 represents the coordinates corresponding to the index of the m-1th row of the first center point, ..., a 1,m Represents the coordinates corresponding to the mth row of indices of the first center point.

[0056] And so on.

[0057] For example, the first eigenvector Q x1 The distance d from the tth center point t It can be expressed as So with the distance d t The corresponding distance function can be expressed as D t (y 1 ~y l ). Among them, a t,1 represents the coordinates corresponding to the first row of indices of the t-th center point, a t,2 represents the coordinates corresponding to the index of the second line of the t-th center point, ..., a t,m-1 represents the coordinates corresponding to the index of the m-1th row of the tth center point, ..., a t,m Represents the coordinates corresponding to the mth row of indices at the tth center point.

[0058] Thus, through the above operation, the generation module can determine the first feature vector Q x1 Multiple distance functions D corresponding to each center point 1 (y 1 ~y l )~D t (y 1 ~y l ).

[0059] Further, when the generation module determines a plurality of distance functions corresponding to the respective center points, the optimization module determines a target optimization function indicating the difference between the first distance and the second distance, and generates a second feature vector corresponding to the target user when the target optimization function is maximized (S406). min1 It is used to indicate the minimum distance between the target optimization function and each center point, the second distance D min2 It is used to indicate the second smallest distance between the target optimization function and each center point. The target optimization function can be calculated by the following formula:

[0060] F(y 1 ~y l )=Dmin1 -D min2

[0061] Among them, F(y 1 ~y l ) represents the target optimization function, D min1 Indicates the first distance, D min2 Indicates the second distance.

[0062] And wherein the optimization module can use, for example, a particle swarm optimization algorithm to determine the target optimization function F(y 1 ~y l ) maximizes y 1 ~y l That is, at the first distance D min1 The second distance D min2 The maximum difference between 1 ~y l The value of .

[0063] The following is the algorithm flow of the particle swarm optimization algorithm: First, the position and speed of each particle are initialized. Then, for the current position of each particle, its fitness is calculated. Furthermore, if the fitness of the particle position at the current moment is better than the fitness of the optimal position it has searched, the optimal position searched by the particle is replaced with the current position of the particle. Then, the optimal position searched by the particle is replaced with the optimal position searched by the particle swarm. And based on the optimization formula of the particle swarm algorithm, the speed and position of the particle are updated. The above process is iterated repeatedly to finally determine y 1 ~y l The value of .

[0064] Thus, through the above method, the optimization module can be at the first distance D min1 and the second distance D min2 When the difference between 1 ~y l And in the optimization module, determine y 1 ~y l In the case of 1 ~y l The value of is filled into the corresponding "empty" position in the first eigenvector, thereby generating the second eigenvector Q x2 For example, the first eigenvector is Q x1 =[q 1 , 0, q 2 ,q 3 ,q 4 ,q 5 , ..., q m-1 ,0] T, so that y 1 ~y l The value of is filled into the corresponding "empty" position, thereby generating the second eigenvector Q x2 =[q 1 ,y 1 ,q 2 ,q 3 ,q 4 ,q 5 , ..., q m-1 ,y l ] T .

[0065] It is worth noting that at the first distance D min1 and the second distance D min2 When the difference between 1 ~y l ), the second eigenvector Q x2 Compared with the first distance D min1 The corresponding center point is closer to the second distance D min2 The corresponding center point. The center point is calculated by the clustering module through the clustering algorithm, and each center point represents a different sample user category. Therefore, based on the above method, the sample user category that best fits the target user (ie, the target sample user category) can be determined.

[0066] Further, when the optimization module determines the second feature vector, the generation module can determine the first behavior data sample based on the second feature vector (S408). The first behavior data sample is used to indicate the second behavior data sample corresponding to the similar user of the target user in the target sample user category. And the second behavior data sample is used to indicate the behavior data collected in advance and different from the target user. From the above content, it can be seen that the clustering module determines the sample user categories different from the multiple second behavior data samples corresponding to the multiple sample users and uses the clustering algorithm, so each sample user category must be composed of multiple second behavior data samples. Thus, when the optimization module determines the target sample user category, the generation module can determine the second behavior data sample (i.e., the first behavior data sample) corresponding to the similar user of the target user in the target sample user category. Thus, when the generation module subsequently inputs the determined first behavior data sample and the target behavior data into the large model together, the accuracy of the output result can be greatly improved.

[0067] Furthermore, the generation module inputs the target behavior data and the first behavior data sample into the big model together, and uses the big model to output the interest topic category corresponding to the target user (S410). Specifically, when the optimization module determines the first behavior data sample, the generation module uses the first behavior data sample as supplementary data for the target behavior data, and inputs it into the big model together with the target behavior data, and uses the big model to output the interest topic category corresponding to the target user. Thus, the operation of inputting the target behavior data and the first behavior data sample into the big model together can make up for the defect that the personalized recommendation content corresponding to the target user cannot be output due to the lack of target behavior data.

[0068] In addition, since the first behavior data sample used as supplementary data in this embodiment is very consistent with the target user, the interest topic category generated based on the first behavior data sample and the target behavior data and using a large model has a higher accuracy rate than traditional coarse-grained recommendations.

[0069] Finally, when the generation module determines the interest topic category corresponding to the target user, the web page title corresponding to the interest topic category of the target user is obtained through online search, and then the large model is used and based on the web page title to generate the recommended content corresponding to the target user. Finally, the application server 200 returns the recommended content to the terminal device 100 corresponding to the target user through the user interface module (S412).

[0070] As described in the background technology, some solutions have been proposed for the cold start problem of large model recommendation systems, such as generating personalized recommendation content based on the registration information of new users, generating personalized recommendation content based on the registration environment of new users, or generating personalized recommendation content based on the age group and gender of new users. However, no matter which of the above solutions is used, it is only a coarse-grained recommendation and cannot make personalized recommendations for the user itself. In other words, the above solutions all have the problem of low recommendation efficiency and low recommendation accuracy.

[0071] In view of this, the present application predetermines a first feature vector corresponding to the target user, and based on a plurality of distance functions corresponding to each center point, determines a target optimization function for indicating the maximization of the difference between the first distance and the second distance. Thus, when the target optimization function is maximized (i.e., when the difference between the first distance and the second distance is maximized), the variable value in the target optimization function can be determined, and a second feature vector corresponding to the target user is generated based on the variable value.

[0072] Furthermore, the generation module can determine the distance between the second feature vector and each center point based on the second feature vector. When the distance between the second feature vector and each center point is determined, the center point corresponding to the minimum distance can be further determined, and the sample user category corresponding to the target user and the center point is more closely matched than other sample user categories.

[0073] In addition, since the target sample user category corresponding to the target user has been determined based on the above optimization algorithm, the first behavior data sample can be used as supplementary data for the target behavior data and input into the large model together with the target behavior data. Thus, the large model is used to output the interest topic category corresponding to the target user. In other words, although there is a lack of user behavior data corresponding to the target user, the first behavior data sample can be used as supplementary data when the target sample user category corresponding to the target user is determined, and the target behavior data and the first behavior data sample can be input into the large model together, thereby making up for the lack of sufficient user behavior data.

[0074] Thus, even in the case of a cold start, the technical effect of being able to provide accurate personalized recommendation content to the target user based on a small amount of target behavior data corresponding to the target user and a large model is achieved. This solves the technical problem in the prior art that the recommendation system based on a large model is difficult to provide accurate personalized recommendation content to the target user due to the lack of sufficient user behavior data.

[0075] Optionally, when the interest topic category corresponding to the target user is determined, recommended content corresponding to the target user is generated, and the recommended content is returned to the terminal device corresponding to the target user, including: when the interest topic category corresponding to the target user is determined, obtaining the web page title corresponding to the interest topic category of the target user through an online search; and using a large model and based on the web page title, generating recommended content corresponding to the target user, and returning the recommended content to the terminal device corresponding to the target user.

[0076] Specifically, the generation module determines the target user P q In the case of the corresponding interest topic category, the web page title corresponding to the interest topic category is searched through the network search method. Then, the generation module creates content based on the web page title corresponding to the interest topic category and uses the big model to generate content related to the target user P. q The generation module may form a 4-tuple of the target user's registered name, target interest topic, matching title, and text content, and use the 4-tuple as the personalized recommendation content. Finally, the application server 200 returns the personalized recommendation content generated by the generation module to the target user P through the user interface module.q The corresponding terminal device 100 is provided to the target user P q Watch.

[0077] Thus, by generating recommended content corresponding to the target user based on the interest topic category corresponding to the target user and through online search, the technical effect of being able to display personalized recommended content to the target user and increasing the target user's interest in use is achieved.

[0078] Optionally, the target behavior data is input into a pre-trained first model and a second model to generate an operation of a first feature vector corresponding to the target user: according to preset behavior indicators, the target behavior data corresponding to the target user is divided into corresponding multiple first behavior indicator data; the multiple first behavior indicator data are vectorized using the first model, and a corresponding multiple first behavior indicator vectors are generated; the multiple first behavior indicator vectors are combined to generate a first behavior indicator matrix; and the first behavior indicator matrix is ​​input into the second model to generate multiple third feature vectors; and the multiple third feature vectors are added and averaged to generate the first feature vector corresponding to the target user.

[0079] Specifically, first, the application server collects information related to the target user P q The corresponding target behavior data S q In addition, as mentioned above, the target behavior data S q It can also be divided into a plurality of first behavior indicator data I based on the preset behavior indicator 1 ~I m .

[0080] And the generation module in the application server 200 receives the target user P q The corresponding first behavior indicator data I 1 ~I m In the case of multiple first behavior indicator data I 1 ~I m The first model is used to train a plurality of first behavior indicator data. 1 ~I m In this embodiment, the first model includes, for example, a One-hot encoding layer, a Word2vec layer (or an embedding layer), and a Bert model. First, the first model uses the One-hot encoding layer to encode the received plurality of first behavior indicator data I 1 ~I m Then, the first model uses the Word2vec model (or embedding layer) to encode the encoded multiple first behavior indicator data I 1 ~Im Vectorize and use the BERT model to understand the context and semantic relationship, thereby generating multiple first behavior indicator vectors Q corresponding to multiple first behavior indicator data 1 ~Q m . The dimension of each first line index vector can be, for example, R w×1 .

[0081] In the generating module, the first module is used and based on a plurality of first behavior indicator data I 1 ~I m Generate multiple first-line indicator vectors Q 1 ~Q m Afterwards, the generation module converts the first rows into the index vector Q 1 ~Q m Combination, thus generating the first line of the indicator matrix Q M Among them, the first line is the indicator matrix Q M The dimension can be, for example, R w×m .

[0082] The generation module then converts the first line into the indicator matrix Q M Input to the pre-trained second model, and use the second model to calculate the first behavior indicator matrix Q M Feature extraction is performed, and multiple third feature vectors are output. The second model, for example, includes multiple convolutional layers, and the convolutional layer includes u convolutional kernels, each of which is arranged in the form of w rows and 1 column (ie, R w×1 ). The first line of each convolution kernel is the indicator matrix Q M After convolution, an m-dimensional second eigenvector is obtained, so that u convolution kernels can obtain u third eigenvectors. Finally, the generation module takes the average value of the u third eigenvectors to obtain the index data I corresponding to the first behavior. 1 ~I m The corresponding first eigenvector Q x1 .

[0083] Thus, by using the pre-trained first model and the second model to vectorize and extract features of multiple first behavior indicator data corresponding to the target user, and generating a first feature vector corresponding to the multiple first behavior indicator data, the necessary basis for subsequent operations is provided, thereby ensuring the technical effect of being able to determine the sample user category that best matches the target user.

[0084] Optionally, it also includes: predetermining the center point corresponding to each sample user category, wherein the operation of predetermining the center point corresponding to each sample user category includes: collecting second behavior data samples corresponding to multiple sample users, and inputting the multiple second behavior data samples into the first model and the second model, thereby outputting the fourth feature vector corresponding to each sample user; and clustering the fourth feature vector corresponding to each sample user, and determining the center point corresponding to each sample user category. Further optionally, the operation of collecting second behavior data samples corresponding to multiple sample users, and inputting the multiple second behavior data samples into the first model and the second model, thereby outputting the fourth feature vector corresponding to each sample user, includes: according to the preset behavior indicator, dividing the second behavior data samples corresponding to the multiple sample users into multiple second behavior indicator data corresponding to each sample user; using the first model to vectorize the multiple second behavior indicator data corresponding to each sample user, and generating multiple second behavior indicator vectors corresponding to each sample user; combining the multiple second behavior indicator vectors corresponding to each sample user, and generating multiple second behavior indicator matrices; and inputting the multiple second behavior indicator matrices into the second model respectively, and generating the fourth feature vector corresponding to each sample user. Further optionally, clustering the fourth eigenvectors corresponding to the respective sample users and determining the center points corresponding to the respective sample user categories includes: determining the respective sample user categories based on the fourth eigenvectors corresponding to the respective sample users using a clustering algorithm; and determining the corresponding multiple center points based on the respective sample user categories.

[0085] Specifically, referring to the above content, it can be known that the clustering module needs to predetermine the center points corresponding to each sample user category. In this embodiment, the application server 200 uses the clustering module and a clustering algorithm to determine multiple center points corresponding to each sample user category.

[0086] First, the application server 200 collects data from multiple sample users P 1 ~P n The corresponding second line data sample S 1 ~S n The application server 200 collects data from multiple sample users P 1 ~P n The corresponding second line data sample S 1 ~S n The method includes but is not limited to crawling relevant data from web pages, or historical data pre-stored in the database 400. And with multiple sample users P 1 ~P n The corresponding second line data sample S 1 ~S nIt can also be divided into different groups according to the preset behavior indicators. 1 ~P n Corresponding multiple second behavior indicator data. For example, with the first sample user P 1 The corresponding plurality of second behavior indicator data can be expressed as I 1,1 ~I 1,m , and with the first sample user P 1 The corresponding plurality of second behavior indicator data I 1,1 ~I 1,m Including the second behavior indicator data I corresponding to "like article" 1,1 , the second behavior indicator data corresponding to "bad review article" I 1,2 , the second behavior corresponding to "comment article" is indicator data I 1,3 , the second behavior indicator data I corresponding to the “watched video” 1,4 , and the second behavior indicator data I corresponding to the “liked video” 1,5 , and the second behavior indicator data I corresponding to the "video with a stay time of more than 5 seconds" 1,6 , ..., the second behavior corresponding to "purchased goods" is the indicator data I 1,m .

[0087] For another example, with the second sample user P 2 The corresponding plurality of second behavior indicator data can be expressed as I 2,1 ~I 2,m , and with the second sample user P 2 The corresponding plurality of second behavior indicator data I 2,1 ~I 2,m Including the second behavior indicator data I corresponding to "like article" 2,1 , the second behavior indicator data I corresponding to the “bad review article” 2,2 , the second behavior corresponding to "comment article" is indicator data I 2,3 , and the second behavior indicator data I corresponding to the “watched video” 2,4 , and the second behavior indicator data I corresponding to the “liked video” 2,5 , and the second behavior indicator data I corresponding to the “video with a stay time of more than 5 seconds” 2,6 , ..., the second behavior corresponding to "purchased goods" is the indicator data I 2,m .

[0088] And so on.

[0089] For example, with the nth sample user P n The corresponding plurality of second behavior indicator data can be expressed as I n,1 ~I n,m , and with the nth sample user Pn The corresponding plurality of second behavior indicator data I n,1 ~I n,m Including the second behavior indicator data I corresponding to "like article" n,1 , the second behavior indicator data corresponding to "bad review article" I n,2 , the second behavior corresponding to "comment article" is indicator data I n,3 , the second behavior indicator data I corresponding to the “watched video” n,4 , and the second behavior indicator data I corresponding to the “liked video” n,5 , and the second behavior indicator data I corresponding to the "video with a stay time of more than 5 seconds" n,6 , ..., the second behavior corresponding to "purchased goods" is the indicator data I n,m .

[0090] After that, the generation module will be connected with each sample user P 1 ~P n The corresponding plurality of second behavior indicator data are input into the pre-trained first model, and the first model is used to train the first model for each sample user P. 1 ~P n The corresponding multiple second behavior indicator data are encoded and vectorized to generate the corresponding second behavior indicator data of each sample user P. 1 ~P n The corresponding multiple second behaviors are index vectors. The dimension of each second behavior index vector can be, for example, R m×1 For example, with the first sample user P 1 The corresponding plurality of second behavior indicator vectors include Q 1,1 ~Q 1,m , and the second sample user P 2 The corresponding plurality of second behavior indicator vectors include Q 2,1 ~Q 2,m , and so on, the multiple second behavior indicator vectors corresponding to the nth sample user include Q n,1 ~Q n,m The above content has described the model architecture of the first model in detail, so it will not be repeated here.

[0091] Then, the generation module will be combined with each sample user P 1 ~P n The corresponding multiple second behavior indicator vectors are combined to generate the corresponding second behavior indicator vectors for each sample user P 1 ~P n The corresponding multiple second lines are indicator matrices Q 1,M ~Q n,M . The dimension of each second line indicator matrix can be, for example, R w×m .

[0092] Furthermore, the generation module will be associated with each sample user P 1 ~P n The corresponding second line is the indicator matrix Q 1,M ~Q n,M Input to the pre-trained second model, and use the second model to compare with each sample user P 1 ~P n The corresponding second line is the indicator matrix Q 1,M ~Q n,M Perform feature extraction to generate features related to each sample user P 1 ~P n The corresponding fourth eigenvector Q 1,x ~Q n,x For example, with the first sample user P 1 The corresponding fourth eigenvector is denoted as Q 1,x , and the second sample user P 2 The corresponding fourth eigenvector is denoted as Q 2,x , and so on, with the nth sample user P n The corresponding fourth eigenvector is denoted as Q n,x .

[0093] After that, the clustering module is used to identify each sample user P 1 ~P n The corresponding fourth eigenvector Q 1,x ~Q n,x Perform clustering operations and determine multiple different sample user categories. In this embodiment, the clustering module can use the k-mans mean clustering method to cluster the sample users P 1 ~P n The corresponding fourth eigenvector Q 1,x ~Q n,x Perform clustering operations. And in the clustering module, use the k-means mean clustering method to cluster the samples of users P 1 ~P n The corresponding fourth eigenvector Q 1,x ~Q n,x In the case of clustering, based on the characteristics of the k-means clustering method, the center point corresponding to each sample user category can be further determined.

[0094] Thus, the application server 200 outputs the fourth feature vector corresponding to each sample user by utilizing the generation module and based on the second behavior data samples corresponding to multiple sample users, and utilizes the clustering module to perform clustering operations on the fourth feature vector corresponding to each sample user, thereby achieving the technical effect of providing the necessary basis for the subsequent determination of the distance function and the target optimization function, thereby ensuring the ability to determine the sample user category that best matches the target user.

[0095] Optionally, it also includes: pre-training the first model and the second model, wherein the operation of pre-training the first model and the second model includes: constructing a third model, wherein the third model includes the first model, the second model, a fully connected layer connected to the second model, and a softmax classifier connected to the fully connected layer; classifying the second behavior data samples corresponding to multiple sample users according to preset personality emotion categories, and generating multiple third behavior data samples corresponding to each personality emotion category; inputting the multiple third behavior data samples corresponding to each personality emotion category into the third model, and training the third model; and when the third model is trained, removing the fully connected layer and the softmax classifier, and retaining the trained first model and second model.

[0096] Specifically, before the generation module generates the first feature vector (or the fourth feature vector) using the first model and the second model, the application server 200 also needs to pre-train the first model and the second model through the model training module.

[0097] Figure 7 is a schematic diagram of the model architecture of the third model according to an embodiment of the present application. Figure 7 Said, first, the model building module builds a third model. The third model consists of the first model, the second model, a fully connected layer connected to the second model, and a softmax classifier connected to the fully connected layer. And the first model includes a One-hot encoding layer, a Word2vec model (or an embedding layer), and a Transformer-based encoding model. The second model includes multiple convolutional layers.

[0098] Afterwards, the model building module classifies the multiple second behavior data samples corresponding to each sample user collected by the application server 200 according to the preset personality and emotion categories, and generates multiple third behavior data samples corresponding to each personality and emotion category. Among them, the personality and emotion categories include but are not limited to MBTI personality types, etc. However, it is worth noting that although the model building module classifies the multiple second behavior data samples according to the preset personality and emotion categories, the sample user categories determined by the clustering module and the clustering algorithm may be more than the number of personality and emotion categories.

[0099] Then, the model building module inputs the plurality of third behavior data samples into the third model, and trains the third model using the plurality of third behavior data samples. Finally, when the model building module has completed training the third model, the fully connected layer and the softmax classifier connected to the second model are removed, and the trained first model and the second model are retained.

[0100] Thus, by using the above-mentioned operation of training the first model and the second model using multiple second behavior data samples corresponding to each sample user collected in advance, the technical effect of providing the necessary basis for the subsequent generation module to generate the first feature vector corresponding to the target user using the first model and the second model is achieved.

[0101] Thus, according to the first aspect of this embodiment, a technical effect is achieved that, even in the case of a cold start, accurate personalized recommendation content can be provided to the target user based on a small amount of target behavior data corresponding to the target user and a large model.

[0102] In addition, reference Figure 1 As shown, according to the second aspect of this embodiment, a storage medium is provided, wherein the storage medium includes a stored program, wherein when the program is run, a processor executes any one of the above methods.

[0103] Thus, according to this embodiment, even in the case of a cold start, the technical effect of being able to provide accurate personalized recommendation content for the target user based on a small amount of target behavior data corresponding to the target user and a large model is achieved.

[0104] It should be noted that, for the above-mentioned method embodiments, for the sake of simplicity, they are all described as a series of action combinations, but those skilled in the art should know that the present invention is not limited by the described action sequence, because according to the present invention, certain steps can be performed in other sequences or simultaneously. Secondly, those skilled in the art should also know that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required by the present invention.

[0105] Through the description of the above implementation methods, those skilled in the art can clearly understand that the method according to the above embodiment can be implemented by means of software plus a necessary general hardware platform, and of course can also be implemented by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, a magnetic disk, or an optical disk), and includes a number of instructions for enabling a terminal device (which can be a mobile phone, a computer, a server, or a network device, etc.) to execute the methods described in each embodiment of the present invention.

[0106] Example 2

[0107] Figure 8 FIG. 8 shows a content recommendation device 800 for cold start of a large model according to this embodiment, and the device 800 corresponds to the method according to embodiment 1. Figure 8As shown, the device 800 includes: a first data collection module 810, which is used to collect target behavior data corresponding to the target user, and input the target behavior data into a pre-trained first model and a second model, so as to generate a first feature vector corresponding to the target user; a distance function generation module 820, which is used to generate a plurality of distance functions corresponding to each center point based on the first feature vector, wherein each center point corresponds to a different sample user category; a second feature vector determination module 830, which is used to determine a target optimization function indicating a difference between the first distance and the second distance based on the plurality of distance functions, and generate a second feature vector corresponding to the target user when the target optimization function is maximized, wherein the first distance is used to indicate a second minimum distance between the first feature vector and each center point, and the second ... Indicates the minimum distance between the first feature vector and each center point; a first behavior data sample determination module 840 is used to determine the first behavior data sample based on the second feature vector, wherein the first behavior data sample is used to indicate the second behavior data sample corresponding to a similar user of the target user in the target sample user category, and wherein the second behavior data sample is used to indicate behavior data collected in advance and different from the target user; an interest topic category output module 850 is used to input the target behavior data and the first behavior data sample into the big model together, and use the big model to output the interest topic category corresponding to the target user; and a recommended content generation module 860 is used to generate recommended content corresponding to the target user when the interest topic category corresponding to the target user is determined, and return the recommended content to the terminal device corresponding to the target user.

[0108] Optionally, the recommended content generation module 860 includes: a web page title acquisition module, which is used to obtain the web page title corresponding to the target user's interest topic category through online search after determining the interest topic category corresponding to the target user; and a content return module, which is used to generate recommended content corresponding to the target user based on the large model and the web page title, and return the recommended content to the terminal device corresponding to the target user.

[0109] Optionally, the first data acquisition module 810 includes: a first data division module, which is used to divide the target behavior data corresponding to the target user into corresponding multiple first behavior indicator data according to preset behavior indicators; a first behavior indicator vector generation module, which is used to vectorize multiple first behavior indicator data using the first model and generate corresponding multiple first behavior indicator vectors; a first behavior indicator matrix generation module, which is used to combine multiple first behavior indicator vectors and generate a first behavior indicator matrix; a third eigenvector generation module, which is used to input the first behavior indicator matrix into the second model and generate multiple third eigenvectors; and a first eigenvector generation module, which is used to add and average multiple third eigenvectors to generate a first eigenvector corresponding to the target user.

[0110] Optionally, the device 800 also includes: a center point determination module, used to predetermine the center point corresponding to each sample user category, wherein the center point determination module includes: a fourth feature vector output module, used to collect second behavior data samples corresponding to multiple sample users, and input the multiple second behavior data samples into the first model and the second model, thereby outputting fourth feature vectors corresponding to each sample user; and a center point determination submodule, used to cluster the fourth feature vectors corresponding to each sample user, and determine the center point corresponding to each sample user category.

[0111] Optionally, the fourth feature vector output module includes: a second data division module, which is used to divide the second behavior data samples corresponding to multiple sample users into multiple second behavior indicator data corresponding to each sample user according to preset behavior indicators; a second behavior indicator vector generation module, which is used to vectorize the multiple second behavior indicator data corresponding to each sample user using the first model, and generate multiple second behavior indicator vectors corresponding to each sample user; a second behavior indicator matrix generation module, which is used to combine the multiple second behavior indicator vectors corresponding to each sample user, and generate multiple second behavior indicator matrices; and a second feature vector generation module, which is used to input the multiple second behavior indicator matrices into the second model respectively, and generate second feature vectors corresponding to each sample user.

[0112] Optionally, the center point determination submodule includes: a sample user category determination module, which is used to determine each sample user category by using a clustering algorithm and based on a fourth feature vector corresponding to each sample user; and a center point generation module, which is used to determine multiple corresponding center points based on each sample user category.

[0113] Optionally, the device 800 also includes: a model training module for pre-training the first model and the second model, wherein the model training module includes: a model construction module for constructing a third model, wherein the third model includes the first model, the second model, a fully connected layer connected to the second model, and a softmax classifier connected to the fully connected layer; a third behavior data sample generation module for classifying the second behavior data samples corresponding to multiple sample users according to preset personality emotion categories, and generating multiple third behavior data samples corresponding to each personality emotion category; a data input module for inputting multiple third behavior data samples corresponding to each personality emotion category into the third model and training the third model; and a elimination module for eliminating the fully connected layer and the softmax classifier when the third model is trained, and retaining the trained first model and second model.

[0114] Thus, according to this embodiment, even in the case of a cold start, the technical effect of being able to provide accurate personalized recommendation content for the target user based on a small amount of target behavior data corresponding to the target user and a large model is achieved.

[0115] Example 3

[0116] Fig. 9 FIG. 9 shows a content recommendation device 900 for cold start of a large model according to this embodiment, and the device 900 corresponds to the method according to embodiment 1. Fig. 9As shown, the device 900 includes: a processor 910; and a memory 920, connected to the processor 910, for providing the processor 910 with instructions for processing the following processing steps: collecting target behavior data corresponding to the target user, and inputting the target behavior data into a pre-trained first model and a second model, thereby generating a first feature vector corresponding to the target user; based on the first feature vector, generating a plurality of distance functions corresponding to each center point, wherein each center point corresponds to a different sample user category; according to the plurality of distance functions, determining a target optimization function for indicating the difference between the first distance and the second distance, and generating a second feature vector corresponding to the target user when the target optimization function is maximized, wherein the first distance is used to indicate the difference between the first distance and the second distance. a second minimum distance between a feature vector and each center point, the second distance being used to indicate the minimum distance between the first feature vector and each center point; determining a first behavior data sample based on the second feature vector, wherein the first behavior data sample is used to indicate a second behavior data sample corresponding to a similar user of the target user in the target sample user category, and wherein the second behavior data sample is used to indicate behavior data collected in advance and different from the target user; inputting the target behavior data and the first behavior data sample into the big model together, and using the big model to output the interest topic category corresponding to the target user; and in the case of determining the interest topic category corresponding to the target user, generating recommended content corresponding to the target user, and returning the recommended content to the terminal device corresponding to the target user.

[0117] Optionally, when the interest topic category corresponding to the target user is determined, recommended content corresponding to the target user is generated, and the recommended content is returned to the terminal device corresponding to the target user, including: when the interest topic category corresponding to the target user is determined, obtaining the web page title corresponding to the interest topic category of the target user through an online search; and using a large model and based on the web page title, generating recommended content corresponding to the target user, and returning the recommended content to the terminal device corresponding to the target user.

[0118] Optionally, the target behavior data is input into a pre-trained first model and a second model to generate an operation of a first feature vector corresponding to the target user: according to preset behavior indicators, the target behavior data corresponding to the target user is divided into corresponding multiple first behavior indicator data; the multiple first behavior indicator data are vectorized using the first model, and a corresponding multiple first behavior indicator vectors are generated; the multiple first behavior indicator vectors are combined to generate a first behavior indicator matrix; and the first behavior indicator matrix is ​​input into the second model to generate multiple third feature vectors; and the multiple third feature vectors are added and averaged to generate the first feature vector corresponding to the target user.

[0119] Optionally, the device 900 also includes: pre-determining the center point corresponding to each sample user category, wherein the operation of pre-determining the center point corresponding to each sample user category includes: collecting second behavior data samples corresponding to multiple sample users, and inputting the multiple second behavior data samples into the first model and the second model, thereby outputting fourth feature vectors corresponding to each sample user; and clustering the fourth feature vectors corresponding to each sample user, and determining the center point corresponding to each sample user category.

[0120] Optionally, the operation of collecting second behavior data samples corresponding to multiple sample users, and inputting the multiple second behavior data samples into the first model and the second model, thereby outputting a fourth feature vector corresponding to each sample user, includes: dividing the second behavior data samples corresponding to the multiple sample users into multiple second behavior indicator data corresponding to each sample user according to preset behavior indicators; vectorizing the multiple second behavior indicator data corresponding to each sample user using the first model, and generating multiple second behavior indicator vectors corresponding to each sample user; combining the multiple second behavior indicator vectors corresponding to each sample user, and generating multiple second behavior indicator matrices; and inputting the multiple second behavior indicator matrices into the second model respectively, and generating fourth feature vectors corresponding to each sample user.

[0121] Optionally, the operation of clustering the fourth eigenvectors corresponding to each sample user and determining the center points corresponding to each sample user category includes: determining each sample user category based on the fourth eigenvectors corresponding to each sample user using a clustering algorithm; and determining the corresponding multiple center points based on each sample user category.

[0122] Optionally, it also includes: pre-training the first model and the second model, wherein the operation of pre-training the first model and the second model includes: constructing a third model, wherein the third model includes the first model, the second model, a fully connected layer connected to the second model, and a softmax classifier connected to the fully connected layer; classifying the second behavior data samples corresponding to multiple sample users according to preset personality emotion categories, and generating multiple third behavior data samples corresponding to each personality emotion category; inputting the multiple third behavior data samples corresponding to each personality emotion category into the third model, and training the third model; and when the third model is trained, removing the fully connected layer and the softmax classifier, and retaining the trained first model and second model.

[0123] Thus, according to this embodiment, even in the case of a cold start, the technical effect of being able to provide accurate personalized recommendation content for the target user based on a small amount of target behavior data corresponding to the target user and a large model is achieved.

[0124] The serial numbers of the above embodiments of the present invention are only for description and do not represent the advantages or disadvantages of the embodiments.

[0125] In the above embodiments of the present invention, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0126] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. Among them, the device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of units or modules, which can be electrical or other forms.

[0127] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0128] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0129] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for a computer device (which can be a personal computer, a server or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, RandomAccess Memory), mobile hard disk, magnetic disk or optical disk and other media that can store program codes.

[0130] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A content recommendation method for large model cold start, characterized in that: include: Collecting target behavior data corresponding to a target user, and inputting the target behavior data into a pre-trained first model and a second model, thereby generating a first feature vector corresponding to the target user; Based on the first feature vector, generating a plurality of distance functions corresponding to respective center points, wherein the respective center points correspond to different sample user categories; Determine, according to the multiple distance functions, a target optimization function for indicating a difference between a first distance and a second distance, and generate a second feature vector corresponding to the target user when the target optimization function is maximized, wherein the first distance is used to indicate a second minimum distance between the first feature vector and each of the center points, and the second distance is used to indicate a minimum distance between the first feature vector and each of the center points; Based on the second feature vector, determining a first behavior data sample, wherein the first behavior data sample is used to indicate a second behavior data sample corresponding to a user similar to the target user in the target sample user category, and wherein the second behavior data sample is used to indicate behavior data collected in advance and different from the target user; Inputting the target behavior data and the first behavior data sample into a large model, and using the large model to output an interest topic category corresponding to the target user; as well as When the interest topic category corresponding to the target user is determined, recommended content corresponding to the target user is generated, and the recommended content is returned to the terminal device corresponding to the target user.

2. The method according to claim 1, characterized in that In the case of determining the interest topic category corresponding to the target user, generating recommended content corresponding to the target user, and returning the recommended content to the terminal device corresponding to the target user, including: When the interest subject category corresponding to the target user is determined, obtaining the web page title corresponding to the interest subject category of the target user through online search; and The large model is used to generate recommended content corresponding to the target user based on the web page title, and the recommended content is returned to a terminal device corresponding to the target user.

3. The method according to claim 1, characterized in that The operation of inputting the target behavior data into a pre-trained first model and a second model to generate a first feature vector corresponding to the target user includes: According to the preset behavior indicators, the target behavior data corresponding to the target user is divided into corresponding multiple first behavior indicator data; vectorizing the plurality of first behavior indicator data using the first model, and generating a plurality of corresponding first behavior indicator vectors; Combining the plurality of first behavior indicator vectors and generating a first behavior indicator matrix; Inputting the first behavior indicator matrix into the second model and generating a plurality of third eigenvectors; and The plurality of third feature vectors are added and averaged to generate a first feature vector corresponding to the target user.

4. The method according to claim 1, characterized in that: Also includes: Predetermine the central points corresponding to the respective sample user categories, wherein the operation of predetermining the central points corresponding to the respective sample user categories includes: Collecting second behavior data samples corresponding to a plurality of sample users, and inputting the plurality of second behavior data samples into the first model and the second model, thereby outputting a fourth feature vector corresponding to each sample user; as well as Clustering is performed on the fourth feature vectors corresponding to the respective sample users, and a center point corresponding to the respective sample user category is determined.

5. The method according to claim 4, characterized in that The operation of collecting second behavior data samples corresponding to a plurality of sample users and inputting the plurality of second behavior data samples into the first model and the second model, thereby outputting a fourth feature vector corresponding to each sample user, includes: According to the preset behavior indicators, the second behavior data samples corresponding to the multiple sample users are divided into multiple second behavior indicator data corresponding to the respective sample users; Vectorizing a plurality of second behavior indicator data corresponding to each of the sample users by using the first model, and generating a plurality of second behavior indicator vectors corresponding to each of the sample users; Combining a plurality of second behavior indicator vectors corresponding to the respective sample users to generate a plurality of second behavior indicator matrices; and The plurality of second behavior indicator matrices are respectively input into the second model, and a fourth feature vector corresponding to each of the sample users is generated.

6. The method according to claim 5, characterized in that The operation of clustering the fourth feature vectors corresponding to the respective sample users and determining the center points corresponding to the respective sample user categories includes: Determining the category of each sample user by using a clustering algorithm and based on the fourth feature vector corresponding to each sample user; and Based on the respective sample user categories, a corresponding plurality of center points are determined.

7. The method according to claim 6, characterized in that Also includes: Pre-training the first model and the second model, wherein the operation of pre-training the first model and the second model comprises: Constructing a third model, wherein the third model includes the first model, the second model, a fully connected layer connected to the second model, and a softmax classifier connected to the fully connected layer; Classifying the second behavior data samples corresponding to the plurality of sample users according to preset personality and emotion categories, and generating a plurality of third behavior data samples corresponding to the respective personality and emotion categories; Inputting a plurality of third behavior data samples corresponding to the respective personality emotion categories into the third model, and training the third model; and When the third model is trained, the fully connected layer and the softmax classifier are removed, and the trained first model and the second model are retained.

8. A storage medium, characterized in that: The storage medium includes a stored program, wherein when the program is run, the processor executes the method according to any one of claims 1 to 7.

9. A content recommendation device for large model cold start, characterized in that: include: A first data collection module, used for collecting target behavior data corresponding to a target user, and inputting the target behavior data into a pre-trained first model and a second model, so as to generate a first feature vector corresponding to the target user; A distance function generating module, configured to generate a plurality of distance functions corresponding to respective center points based on the first feature vector, wherein the respective center points correspond to different sample user categories; a second feature vector determination module, configured to determine, based on the multiple distance functions, a target optimization function indicating a difference between a first distance and a second distance, and generate a second feature vector corresponding to the target user when the target optimization function is maximized, wherein the first distance is used to indicate a second minimum distance between the first feature vector and each of the center points, and the second distance is used to indicate a minimum distance between the first feature vector and each of the center points; a target sample user category determination module, configured to determine a first behavior data sample based on the second feature vector, wherein the first behavior data sample is used to indicate a second behavior data sample corresponding to a user similar to the target user in the target sample user category, and wherein the second behavior data sample is used to indicate behavior data collected in advance and different from the target user; An interest topic category output module, used to input the target behavior data and the first behavior data sample into a large model, and use the large model to output an interest topic category corresponding to the target user; as well as The recommended content generation module is used to generate recommended content corresponding to the target user when the interest topic category corresponding to the target user is determined, and return the recommended content to the terminal device corresponding to the target user.

10. A content recommendation device for large model cold start, characterized in that: include: processor; as well as A memory, connected to the processor, configured to provide the processor with instructions for processing the following processing steps: Collecting target behavior data corresponding to a target user, and inputting the target behavior data into a pre-trained first model and a second model, thereby generating a first feature vector corresponding to the target user; Based on the first feature vector, generating a plurality of distance functions corresponding to respective center points, wherein the respective center points correspond to different sample user categories; Determine, according to the multiple distance functions, a target optimization function for indicating a difference between a first distance and a second distance, and generate a second feature vector corresponding to the target user when the target optimization function is maximized, wherein the first distance is used to indicate a second minimum distance between the first feature vector and each of the center points, and the second distance is used to indicate a minimum distance between the first feature vector and each of the center points; Based on the second feature vector, determining a first behavior data sample, wherein the first behavior data sample is used to indicate a second behavior data sample corresponding to a user similar to the target user in the target sample user category, and wherein the second behavior data sample is used to indicate behavior data collected in advance and different from the target user; Inputting the target behavior data and the first behavior data sample into a large model, and using the large model to output an interest topic category corresponding to the target user; as well as When the interest topic category corresponding to the target user is determined, recommended content corresponding to the target user is generated, and the recommended content is returned to the terminal device corresponding to the target user.

Images