Recommended object determination method, medium, device, and computing device

Abstract

Embodiments of the present disclosure provide a method, medium, device, and computing device for determining a recommended object. The method for determining a recommended object includes: determining candidate objects of common interest to at least two users based on historical behavior data of the at least two users for an object; then determining recalled objects based on the candidate objects, and determining recommended objects based on the recalled objects, the user portrait features of the at least two users, and multiple user behavior sequence features, where the user behavior sequence features characterize the features of a sequence composed of objects that the user has completely played within a preset time period. By determining recalled objects based on candidate objects of common interest to at least two users, and then combining the recalled objects with the user portrait features and user behavior sequence features of the at least two users to determine recommended objects of common interest to the at least two users, the recommendation accuracy of the recommended objects is improved, bringing a better experience to users.

Description

Technical Field

[0001] Embodiments of the present disclosure relate to the field of the Internet. More specifically, embodiments of the present disclosure relate to a method, medium, device, and computing device for determining a recommended object. Background Art

[0002] This section aims to provide background or context for the embodiments of the present disclosure recited in the claims. The description herein is not admitted to be prior art merely because it is included in this section.

[0003] Nowadays, with the continuous development of recommendation technologies, personalized recommendation technologies have been paid more and more attention. Among them, personalized recommended objects can include various audios, videos, advertisements, etc. Taking music as an example, songs and playlists of interest can be recommended to users, which can not only improve the user experience but also maximize the utilization of the music library resources. In addition, as a carrier of emotions, objects such as music are such that more and more users tend to share their favorite objects and play their favorite objects together with others, thus giving rise to the need for multi-user sharing of objects. However, how to generate recommended objects that are commonly liked by multiple users has become a challenge. Summary of the Invention

[0004] The present disclosure provides a method, medium, device, and computing device for determining a recommended object to solve the problem of poor recommendation accuracy when the current personalized recommendation technology generates recommended objects that are commonly liked by multiple users based on the interests and hobbies of a single user.

[0005] In a first aspect of the embodiments of the present disclosure, a method for determining a recommended object is provided, including:

[0006] Determining candidate objects that at least two users are commonly interested in according to historical behavior data of at least two users for an object;

[0007] Determining recalled objects according to the candidate objects;

[0008] Determining a recommended object according to the recalled objects, user portrait features of at least two users, and multiple user behavior sequence features, where the user behavior sequence features represent the features of a sequence composed of objects that a user has completely played within a preset time period.

[0009] In one embodiment of the present disclosure, determining a recommended object based on a recall object, user portrait features of at least two users, and multiple user behavior sequence features includes: determining corresponding user portrait feature vectors according to the user portrait features of at least two users; determining a corresponding recall object feature vector according to the features of the recall object; determining overall user portrait feature vectors corresponding to at least two users according to the user portrait feature vectors and the recall object feature vector; determining a user behavior sequence feature matrix corresponding to a user according to any user behavior sequence feature; determining an overall user interest feature vector according to the recall object feature vector and multiple user behavior sequence feature matrices, where the overall user interest feature vector represents the overall interest feature distribution of the overall user in different dimensions; and determining the recommended object according to the overall user portrait feature vector, the overall user interest feature vector, and the recall object feature vector.

[0010] In one embodiment of the present disclosure, determining overall user portrait feature vectors corresponding to at least two users according to the user portrait feature vectors and the recall object feature vector includes: performing attention weighting processing on the user portrait feature vectors and the recall object feature vector based on an attention mechanism to obtain the overall user portrait feature vector.

[0011] In one embodiment of the present disclosure, determining the overall user interest feature vector according to the recall object feature vector and multiple user behavior sequence feature matrices includes: respectively performing weighting processing on multiple user behavior sequence feature matrices based on a self-attention mechanism to determine user interest feature vectors corresponding to the users; and performing weighting processing on the vector sum of multiple user interest feature vectors in the same dimension and the recall object feature vector based on a multi-head attention mechanism to obtain the overall user interest feature vector.

[0012] In one embodiment of the present disclosure, determining the recommended object according to the overall user portrait feature vector, the overall user interest feature vector, and the recall object feature vector includes: combining the overall user portrait feature vector, the overall user interest feature vector, and the recall object feature vector to obtain a combined vector; extracting features from the combined vector and determining a score value of the recall object through a preset normalization function; and determining the recommended object according to the score value of the recall object.

[0013] In one embodiment of the present disclosure, it further includes at least one of the following: determining a corresponding user portrait feature vector according to the user portrait features of at least two users, including: determining the user portrait feature vector corresponding to the user based on the trained model according to the user portrait features of at least two users; determining a corresponding recall object feature vector according to the features of the recall object, including: determining the recall object feature vector corresponding to the recall object based on the trained model according to the features of the recall object; determining a corresponding user behavior sequence feature matrix according to any user behavior sequence feature, including: determining the user behavior sequence feature matrix corresponding to the user based on the trained model according to any user behavior sequence feature; wherein, the model is obtained by pre-constructing samples and training the initial model according to the samples.

[0014] In one embodiment of the present disclosure, determining a recall object according to a candidate object includes: determining the similarity between the candidate object and the first object; determining the overall interest scores of at least two users for the first object according to the preference degree and similarity of any one of the at least two users for the candidate object; sorting the multiple first objects based on the multiple overall interest scores to determine the recall object.

[0015] In one embodiment of the present disclosure, it further includes: generating an object list according to the recommended object; and / or, sending the recommended object to the user terminal to generate an object list according to the recommended object filtered by the user through the user terminal.

[0016] In the second aspect of the implementation manner of the present disclosure, there is provided a computer-readable storage medium, in which computer program instructions are stored, and when the computer program instructions are executed, the recommended object determination method described in any one of the first aspect is implemented.

[0017] In the third aspect of the implementation manner of the present disclosure, there is provided a recommended object determination device, including: a first determination module, configured to determine candidate objects of common interest to at least two users according to the historical behavior data of the at least two users for the object; a second determination module, configured to determine a recall object according to the candidate object; a third determination module, configured to determine a recommended object according to the recall object, the user portrait features of at least two users, and multiple user behavior sequence features, wherein the user behavior sequence feature characterizes the feature of the sequence composed of the objects completely played by the user within a preset time period.

[0018] In one embodiment of the present disclosure, the third determination module is specifically configured to: determine corresponding user portrait feature vectors according to the user portrait feature of at least two users; determine corresponding recall object feature vectors according to the features of the recall objects; determine overall user portrait feature vectors corresponding to at least two users according to the user portrait feature vectors and the recall object feature vectors; determine a user behavior sequence feature matrix corresponding to a user according to any user behavior sequence feature; determine an overall user interest feature vector according to the recall object feature vector and multiple user behavior sequence feature matrices, where the overall user interest feature vector represents the overall interest feature distribution of the overall user in different dimensions; determine a recommended object according to the overall user portrait feature vector, the overall user interest feature vector, and the recall object feature vector.

[0019] In one embodiment of the present disclosure, the third determination module is specifically configured to: perform attention weighting processing on the user portrait feature vector and the recall object feature vector based on the attention mechanism to obtain an overall user portrait feature vector.

[0020] In one embodiment of the present disclosure, the third determination module is specifically configured to: perform weighting processing on multiple user behavior sequence feature matrices respectively based on the self-attention mechanism to determine a user interest feature vector corresponding to the user; perform weighting processing on the vector sum of multiple user interest feature vectors in the same dimension and the recall object feature vector based on the multi-head attention mechanism to obtain an overall user interest feature vector.

[0021] In one embodiment of the present disclosure, the third determination module is specifically configured to: perform merging processing on the overall user portrait feature vector, the overall user interest feature vector, and the recall object feature vector to obtain a merged vector; perform feature extraction on the merged vector, and determine a score value of the recall object through a preset normalization function; determine a recommended object according to the score value of the recall object.

[0022] In one embodiment of the present disclosure, the third determination module is further configured to perform at least one of the following: determine a user portrait feature vector corresponding to a user according to the user portrait feature of at least two users based on the trained model; determine a recall object feature vector corresponding to the recall object according to the features of the recall object based on the trained model; determine a user behavior sequence feature matrix corresponding to a user according to any user behavior sequence feature based on the trained model; where the model is obtained by pre-constructing samples and training an initial model according to the samples.

[0023] In one embodiment of the present disclosure, the second determination module is specifically configured to: determine the similarity between the candidate object and the first object; determine the overall interest score of at least two users in the first object according to the preference degree and similarity of any user among the at least two users for the candidate object; and rank the multiple first objects based on the multiple overall interest scores to determine the recalled objects.

[0024] In one embodiment of the present disclosure, it further includes a generation module, configured to: generate an object list according to the recommended object; and / or send the recommended object to the user terminal to generate an object list according to the recommended object filtered by the user through the user terminal.

[0025] In the fourth aspect of the implementation manner of the present disclosure, a computing device is provided, including: a memory and a processor, where the memory is used to store program instructions; the processor is used to call the program instructions in the memory to execute the recommended object determination method as in the first aspect.

[0026] The present disclosure provides a recommended object determination method, device, medium and computing device. First, according to the historical behavior data of at least two users for the object, determine the candidate objects of common interest to at least two users; then, according to the candidate objects, determine the recalled objects, and determine the recommended objects according to the recalled objects, the user portrait features of at least two users and multiple user behavior sequence features, where the user behavior sequence features characterize the features of the sequence composed of the objects completely played by the user within a preset time period. The present disclosure determines the recalled objects based on the candidate objects of common interest to at least two users, and then combines the recalled objects with the user portrait features and user behavior sequence features of at least two users to determine the recommended objects of common interest to these at least two users, thereby improving the recommendation accuracy of the recommended objects. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] By referring to the accompanying drawings and reading the following detailed description, the above and other purposes, features and advantages of the exemplary embodiments of the present disclosure will become easily understood. In the drawings, several embodiments of the present disclosure are shown in an exemplary rather than restrictive manner, where:

[0028] Figure 1 Schematically shows an application scenario diagram according to an embodiment of the present disclosure;

[0029] Figure 2 Schematically shows a flowchart of a recommended object determination method according to an embodiment of the present disclosure;

[0030] Figure 3 Schematically shows a flowchart of a recommended object determination method according to another embodiment of the present disclosure;

[0031] Figure 4Schematically shows an application flowchart of a method for determining a recommended object according to an embodiment of the present disclosure;

[0032] Figure 5 Schematically shows a structural diagram of a neural network model in a multi - user music sorting module according to an embodiment of the present disclosure;

[0033] Figure 6 Schematically shows a program product diagram of a method for determining a recommended object provided by the present disclosure;

[0034] Figure 7 Schematically shows a structural diagram of a device for determining a recommended object according to an embodiment of the present disclosure;

[0035] Figure 8 Schematically shows a structural diagram of a computing device provided by the present disclosure.

[0036] In the drawings, the same or corresponding reference numerals represent the same or corresponding parts. Detailed Embodiments

[0037] The principles and spirit of the present disclosure will be described below with reference to several exemplary embodiments. It should be understood that these embodiments are given only to enable those skilled in the art to better understand and then implement the present disclosure, rather than limiting the scope of the present disclosure in any way. On the contrary, these embodiments are provided to make the present disclosure more thorough and complete, and to be able to fully convey the scope of the present disclosure to those skilled in the art.

[0038] Those skilled in the art know that the embodiments of the present disclosure can be implemented as a system, a device, an apparatus, a method, or a computer program product. Therefore, the present disclosure can be specifically implemented in the following forms: completely hardware, completely software (including firmware, resident software, microcode, etc.), or a combination of hardware and software. The data involved in the present disclosure can be data authorized by the user or fully authorized by all parties.

[0039] According to the embodiments of the present disclosure, a method, a medium, a device, and a computing device for determining a recommended object are provided.

[0040] The glossary of the present disclosure is as follows:

[0041] Recurrent Neural Network (RNN for short), a recurrent neural network is a type of recurrent neural network that takes sequence or vector data as input and recurs in the evolution direction of the sequence. Since the present disclosure needs to take vectors as input data, a recurrent neural network is required.

[0042] The multi-head self-attention mechanism (abbreviated as MHSF), compared with the attention mechanism, can be analogized to the relationship between the recurrent neural network and the neural network. The multi-head self-attention mechanism performs an attention mechanism on a sequence, that is, performs an attention mechanism operation on each dimension of a sequence simultaneously. The multi-head self-attention mechanism can be regarded as a high-dimensional mode of the attention mechanism.

[0043] In this disclosure, it should be understood that the number of any element in the drawings is for illustration rather than limitation, and any naming is only for distinction and does not have any limiting meaning.

[0044] Next, with reference to several representative embodiments of the present disclosure, the principles and spirits of the present disclosure will be elaborated in detail. Summary of the Invention

[0046] The inventor of the present invention has found that existing personalized recommendation technologies mainly generate recommended objects based on the interest preferences of individual users. Taking music as an example, although it is possible to calculate the preference scores of each user for each song, and then for each song among multiple songs, determine the average value of the preference scores of multiple users for that song, and sort the multiple songs according to the average value to obtain a list of songs with common preferences recommended to multiple users. This solution has the problem of poor recommendation accuracy.

[0047] Based on the above problems, the present disclosure first screens out the objects of interest to each user according to the historical behaviors of multiple users, and then based on the objects of interest to each user, screens out the objects of common interest to multiple users. Then, the objects similar to the objects of common interest to multiple users are used as recall objects. Further, it is also necessary to analyze and process the recall objects, and after the analysis and processing, sort the recall objects to generate a list of recommended objects, and the recommended objects can be obtained from the list.

[0048] After introducing the basic principles of the present disclosure, various non-limiting embodiments of the present disclosure will be specifically introduced below.

[0049] Overview of Application Scenarios

[0050] First, refer to Figure 1 , Figure 1 schematically shows an application scenario diagram according to an embodiment of the present disclosure, and the objects in this scenario are music. Figure 1It includes user 101, client 102 and server 103. Among them, server 103 may include a multi-user music recall module and a multi-user music sorting module. Server 103 records historical data related to interactions. For example, when user 101 listens to music using client 102, user 101 may first open the installed music application in client 102 and select the song to be played in the relevant interface of the music application displayed by client 102 for playback. Client 102 records historical data related to the interaction between user 101 and server 103 during this process, and server 103 records historical data related to the interaction with client 102 during this process. Optionally, the historical data at least includes historical data of at least one object. For example, the object can be music, advertisement or video, etc. Server 103 can obtain the historical data of at least one user through client 102 and process it; or, server 103 can obtain historical data locally and process it.

[0051] Exemplarily, the multi-user music recall module of server 103 is used to process the historical data obtained from client 102. The multi-user music recall module of server 103 will first count the historical interest music of multiple users to generate a historical interest music library. After that, in the historical interest music library, count the music that all users are interested in and use these musics as candidate musics. After the candidate musics are determined, this multi-user music recall module will re-determine the music with a higher similarity to the candidate musics among all the musics and use it as the recalled music. If a more accurate recommended music list needs to be generated, the recalled music needs to be further sorted by the multi-user music sorting module. Server 103 can generate a recommended music list based on the sorted recalled music and send it to client 102 for display to the user by client 102; in addition, server 103 can also directly send the sorted recalled music to client 102, and the user of client 102 can select the music by himself / herself and generate a music recommendation list.

[0052] Exemplary Method

[0053] The following combines Figure 1 the application scenario of Figure 2 to describe the method for generating a recommended object according to an exemplary embodiment of the present disclosure. It should be noted that the above application scenario is only shown for the convenience of understanding the spirit and principle of the present disclosure, and the embodiments of the present disclosure are not limited in this regard. On the contrary, the embodiments of the present disclosure can be applied to any applicable scenario.

[0054] Figure 2 Schematically shows a flowchart of a method for determining a recommended object according to an embodiment of the present disclosure. The method includes:

[0055] S201. Determine candidate objects that at least two users are jointly interested in based on the historical behavior data of the at least two users regarding the object.

[0056] Exemplarily, the historical behavior data includes the historical behavior data of at least one type of object, and in the historical behavior data, it may include at least one of the following relevant information: the degree of interest of the user in the object, the frequency of the user accessing the object within a period of time, the time when the user first accessed the object, etc.

[0057] In addition, the above object may be at least one of music, advertisement, video, etc.

[0058] Among them, the determination method may be to obtain all the objects of interest of all users at once, and further determine the objects that all users are jointly interested in among these objects, and use the determined objects as candidate objects; it may also be to separately count the objects of interest of each user first, and then determine the intersection of the objects of interest of each user, and further obtain the objects that all users are jointly interested in, that is, the candidate objects.

[0059] S202. Determine recalled objects based on the candidate objects.

[0060] Among them, the candidate objects represent the objects that have been determined to be jointly interested in by multiple users, while the recalled objects are the objects that multiple users may be jointly interested in. The meanings of the two are essentially different.

[0061] The method of determining the recalled objects may be to use the objects similar to the candidate objects as the recalled objects; it may also be to further determine the recalled objects of the multiple users by combining the candidate objects and analyzing the portrait features of the multiple users themselves.

[0062] S203. Determine recommended objects based on the recalled objects, the portrait features of at least two users, and the multiple user behavior sequence features, where the user behavior sequence features represent the features of the sequence composed of the objects that the user has completely played within a preset time period.

[0063] Among them, the user portrait features represent at least one of the user-related features such as the basic attribute features of the user, the user device features, and the user object preference features. Among them, the basic attribute features of the user may include at least one of the basic attribute features such as the age of the user, the gender of the user, and the city where the user is located. The user device features may include the operating system and the operator of the user device. Taking the object as music as an example, the user object preference features may include at least one of language preference, style preference, singer preference, or release year preference, etc.

[0064] Considering that determining recommended objects only based on user profile features and recalled objects may result in inaccurate results, therefore, the present disclosure introduces user behavior sequence features. The user behavior sequence features characterize the sequence of objects that the user has completely played within a period of time. Therefore, they can also be used to characterize the user's interest tendency in the next period of time. By extracting this interest tendency, more accurate recommended objects can be determined for the user.

[0065] Exemplarily, weights can also be set for the recalled objects. For different recalled objects, since the user's degree of interest in them is different, the weights are also different. After introducing the weights, combining the user profile features of multiple users and the user behavior sequence features of multiple users above, the recalled objects are sorted, and the corresponding recommended objects are selected through ascending or descending order.

[0066] In the embodiments of the present disclosure, first, according to the historical behavior data of at least two users for an object, candidate objects that at least two users are jointly interested in are determined; then, according to the candidate objects, recalled objects are determined, and according to the recalled objects, the user profile features of at least two users, and the user behavior sequence features of multiple users, recommended objects are determined, where the user behavior sequence features characterize the features of the sequence composed of the objects that the user has completely played within a preset time period. The embodiments of the present disclosure determine recalled objects based on candidate objects that at least two users are jointly interested in, and then combine the recalled objects with the user profile features and user behavior sequence features of at least two users to determine the recommended objects that at least two users are jointly interested in, thereby improving the recommendation accuracy of the recommended objects.

[0067] In some embodiments, it can be passed through Figure 3 to further show the determination process of the recommended objects. As Figure 3 shown, Figure 3 Schematically shows a flowchart of a method for determining recommended objects according to another embodiment of the present disclosure. The method includes:

[0068] S301. According to the historical behavior data of at least two users for an object, determine candidate objects that at least two users are jointly interested in.

[0069] This step is similar to the foregoing step S201 and will not be elaborated here.

[0070] S302. According to the candidate objects, determine recalled objects.

[0071] In some embodiments, determining recalled objects according to candidate objects may include: determining the similarity between the candidate objects and the first object; according to the preference degree and similarity of any one of the at least two users for the candidate objects, determining the overall interest score of the at least two users for the first object; based on multiple overall interest scores, sorting multiple first objects to determine recalled objects.

[0072] For example, it can be through the Item coordination filter (ItemCF for short). The specific operation method is as follows:

[0073] If the object is music, then: the above-mentioned candidate objects are the candidate music that users are interested in together; the above-mentioned first object is any music in the entire music library. The first step is: by calculating the similarity, select at least one first object similar to each candidate object from the entire music library and form a similar set. Among them, the selection method can be to first sort by similarity and then select a preset number of first objects to form a similar set; it can also be to select the first objects whose similarity is greater than the preset similarity threshold to form a similar set.

[0074] As an example, the calculation formula for similarity is as follows:

[0075]

[0076] Here, N(i) and N(j) respectively represent the number of people who like music i and music j, and w ij represents the similarity between music i and music j. The similarity between each candidate music and other music (the first object) in the music library can be determined through the above formula.

[0077] After determining the similar set from the music library, the interest score of the first object in the similar set is further determined in the following way:

[0078] Sim(u, j) = ∑ i∈S(j，K)∩N(u) W(j, i)R(u, i)

[0079] Among them, Sim(u, j) represents the interest score of user u in music j; N(u) is the set of music that user u is interested in, and this set can be determined when obtaining the historical behavior data of user u; S(j, K) is the set of the K music that is most similar to music j; W(j, i) is the similarity between music j and music i, and R(u, i) is the preference degree of user u for music i.

[0080] In the above formula, it can be understood that: music i represents the music that the user is interested in himself, that is, music i belongs to the above-mentioned candidate objects, while music j represents the first object music determined from the music library through similarity calculation. In addition, the preference degree is used to measure the preference degree of the user for music. Even if two pieces of music are both interested in by the user, however, the user's preference degree for these two pieces of music is also different. The influencing factors of the preference degree include the frequency of the user listening to this music, and the frequency of the user listening to music of the same type as this music, etc.

[0081] For any first object in a similar set, after calculating the interest scores of each user for the first object, the overall interest score of the first object can be determined by the method of weighted average, and the overall interest scores of all first objects can be determined by the same method as above. After sorting the overall interest scores of all first objects, a preset threshold number of first objects are selected as the recalled objects.

[0082] S303. Determine the corresponding user portrait feature vectors according to the user portrait features of at least two users.

[0083] In some embodiments, a neural network model can be used to convert the aforementioned user portrait features, user behavior sequence features, and recalled objects into the form of vectors or matrices, and then analyze and process the vectors or matrices to determine the recommended objects. Specifically as follows:

[0084] Before applying the neural network model, first, samples need to be constructed in advance and the neural network model is trained according to the samples.

[0085] Exemplarily, to construct a training sample, labels need to be set for the sample first: by obtaining the historical behavior data of a large number of users, and based on the historical behavior data, extracting data such as user IDs, object IDs, and play object labels, and using them as samples. Among them, the user ID is used as the identifier of the user and can represent the identity information of different users; the object ID is used as the identifier of the object and can characterize different object types; the play object label reflects the play information of the user and can correspond to the aforementioned user behavior sequence features.

[0086] After constructing the samples, the samples can be divided into a training set and a test set. Among them, the training set is used to train the neural network model, and the test set is used to test the neural network model. In addition, the ratio between the data included in the training set and the data included in the test set can be 8:2; it can also be 7:3, and the present disclosure does not limit it.

[0087] Based on the above description, for example: determining the corresponding user portrait feature vectors according to the user portrait features of at least two users may include: determining the user portrait feature vectors corresponding to the users based on the trained model according to the user portrait features of at least two users.

[0088] The user portrait feature vector can be represented by A. Among them, A can also be written as A = (aa, ab, ac, ad... az). Each dimension of this vector characterizes a certain feature of the user himself. For example, when aa is 0, the user is female; when aa is 1, the user is male. When ab is 24, it means the user's age is 24 years old. To a certain extent, ac characterizes the area where the user is located. For example, when ac is 2, it means the user is in a second-tier city; when ac is 1, it means the user is in a first-tier city.

[0089] S304. Determine the corresponding recall object feature vector according to the features of the recall object.

[0090] In some embodiments, determining the corresponding recall object feature vector according to the features of the recall object may include: determining the recall object feature vector corresponding to the recall object based on the trained model according to the features of the recall object.

[0091] Similarly, vectorize the recall object. Specifically, extract its features through the convolutional layer of the neural network.

[0092] The recall object feature vector obtained after vectorization can be represented as R, and R can also be written as: R = (ra, rb, rc, rd... rz).

[0093] Among them, each dimension of the recall object feature vector represents the features of the recall object in each dimension. For example, ra characterizes the language of the recall object. When ra is 1, the language is English; when ra is 2, the language is Chinese; when ra is 3, the language is Japanese. rb characterizes the type feature of the recall object. For example, when rb is 1, the recall object is of the entertainment type; when rb is 2, the recall object is of the sad type, etc. Further, when the object is music, rb can be characterized as: when rb is 1, it means the recalled music is folk music; when rb is 2, it characterizes the recalled music as pop music; when rb is 3, it characterizes the recalled music as rock.

[0094] S305. Determine the overall user portrait feature vector corresponding to at least two users according to the user portrait feature vector and the recall object feature vector.

[0095] Exemplarily, any one recall object feature vector and at least two user portrait feature vectors can be selected, and their feature fusion processing is performed to obtain the overall user portrait feature vector. Since the overall user portrait feature vector characterizes the user portrait features of multiple users, and at the same time, it also fuses the features of the recall object. Therefore, the overall user portrait feature vector can characterize the relationship between the portrait features of multiple users themselves and the features of the recall object.

[0096] S306. Determine the user behavior sequence feature matrix corresponding to the user according to any user behavior sequence feature.

[0097] Optionally, determining the user behavior sequence feature matrix corresponding to the user according to any user behavior sequence feature includes: determining the user behavior sequence feature matrix corresponding to the user based on the trained model according to any user behavior sequence feature.

[0098] As previously described, the user behavior sequence feature can characterize the user's interest tendency within a period of time. Therefore, the user behavior sequence feature can include at least one interest tendency. If each interest tendency is regarded as a vector, after vectorizing the user behavior sequence feature, multiple vectors, that is, a matrix, will be obtained. Therefore, each user behavior sequence feature corresponds to a user behavior sequence feature matrix, which can be represented by S. S can be further written as: S = (Sa, Sb, Sc, Sd... Sz). Wherein, Sa, Sb, Sc to Sz are all vectors, which represent the user's interest tendency in different dimensions. As an example, the larger the modulus value of the vector, the greater the user's interest tendency in this dimension. In addition, the user behavior sequence feature matrix corresponding to user 1 can be represented as S1, and S1 can be written as: S1 = (S11, S12, S13, S14... S1m); the user behavior sequence feature matrix corresponding to user 2 can be represented as S2, and S2 can be written as: S2 = (S21, S22, S23, S24... S2m), and so on.

[0099] S307. Determine the overall user interest feature vector according to the recall object feature vector and multiple user behavior sequence feature matrices, where the overall user interest feature vector characterizes the overall interest feature distribution of the overall user in different dimensions.

[0100] Combine the user behavior sequence feature matrix corresponding to each user and any recall object feature vector, and perform feature extraction on the vectors in different dimensions to obtain the overall user interest feature vector.

[0101] The overall user interest feature vector consists of multiple interest dimensions. Each interest dimension contains the features of multiple users in this interest dimension, and the features of the above-selected recall object in this interest dimension. Similarly, each interest dimension of the overall user interest feature vector is represented by a vector. Therefore, the overall user interest feature vector is actually a vector group.

[0102] S308. Determine the recommended object according to the overall user portrait feature vector, the overall user interest feature vector and the recall object feature vector.

[0103] Exemplarily, the above vectors can be fused, and then the fused vectors can be further dimension-reduced. Finally, after normalization, it is determined whether the recalled object is used as a recommended object.

[0104] In the embodiments of the present disclosure, by vectorizing the user portrait features of multiple users, the user behavior sequence features of multiple users, and the recalled object, and processing the vectors corresponding to the above multiple users through a trained neural network model, the recommended object is determined. By using the trained neural network model, the accuracy of the output recommended object can also be improved.

[0105] Based on the above embodiments, in some embodiments, an attention mechanism can be introduced: according to the user portrait feature vector and the recalled object feature vector, at least two overall user portrait feature vectors corresponding to users are determined, which may include: based on the attention mechanism, performing attention weighting processing on the user portrait feature vector and the recalled object feature vector to obtain the overall user portrait feature vector.

[0106] Exemplarily, if there are 3 users, their user portrait feature vectors can be represented as A1, A2, and A3. Among them, A1 = (a11, a12, a13); A2 = (a21, a22, a23); A3 = (a31, a32, a33). And the recalled object feature vector of one of the recalled objects is R1 = (r11, r12, r13). After the operation of the attention mechanism, the weight of vector A1 is 4, the weight of vector A2 is 5, and the weight of vector A3 is 2. Then the overall user portrait feature vector is: 4*A1 + 5*A2 + 2*A3.

[0107] In addition, the weight calculation of each user portrait feature vector can be obtained through a linear function or a non-linear function. The input of this function is the user portrait feature vector and the recalled object feature vector. Therefore, if the linear function is written as F(x), the weight of vector A1 can be written as: F(A1, R1), and the weight calculation of other vectors is the same and will not be elaborated here.

[0108] In some embodiments, the determination process of the overall user interest feature vector in the above embodiments can be further split into two steps as follows:

[0109] According to the recalled object feature vector and multiple user behavior sequence feature matrices, the overall user interest feature vector is determined, which may include: based on the self-attention mechanism, respectively performing weighting processing on multiple user behavior sequence feature matrices to determine the user interest feature vectors corresponding to the users; based on the multi-head attention mechanism, performing weighting processing on the vector sum of multiple user interest feature vectors in the same dimension and the recalled object feature vector to obtain the overall user interest feature vector.

[0110] It should be understood that the attention mechanism can be divided into self-attention mechanism and target-attention mechanism. Among them, the self-attention mechanism performs interactive operations between its own vectors without introducing external vectors; the target-attention mechanism is opposite to the self-attention mechanism, which performs interactive operations between internal vectors and external vectors when external vectors are introduced. At the same time, the attention mechanism also includes the multi-head attention mechanism. The so-called multi-head attention mechanism can be understood as performing attention mechanism operations in multiple dimensions.

[0111] Exemplarily, if there are 3 users, the user behavior sequence feature matrix of user 1 is S1, written as (S11, S12, S13); the user behavior sequence feature matrix of user 2 is S2, written as (S21, S22, S23); the user behavior sequence feature matrix of user 3 is S3, written as (S31, S32, S33). First, the self-attention mechanism is introduced, and the structural parameter of this self-attention mechanism is set to N, that is, the interest tendency of the user is divided into N dimensions, and each dimension represents the user's interest tendency in this interest dimension.

[0112] If the inputs are S1, S2, and S3, the outputs can be O1, O2, and O3 respectively. Among them, O1 can be written as O1 = (O11, O12, O13... O1N); similarly, O2 and O3 can be written as: O2 = (O21, O22, O23... O2N) and O3 = (O31, O32, O33... O3N). In this formula, O11 represents the interest tendency of user 1 in the first interest dimension, and O12 represents the interest tendency of user 1 in the second interest dimension. O21 represents the interest tendency of user 2 in the first interest dimension, while O31 represents the interest tendency of user 3 in the first interest dimension.

[0113] Taking music as an example of the object, the interest dimension can be divided into: retro music, modern music, electronic music, national style music, and so on. It can correspond one by one to each dimension in the aforementioned recall object feature vector, or can independently represent different features of music.

[0114] After dividing the interest tendency of each user into multiple interest dimensions, the multi-head attention mechanism is introduced to obtain the overall user interest feature vector containing multiple interest dimension features, as follows:

[0115] It should be understood that when the multi-head attention mechanism is processing, multiple dimensions are introduced, and in order to summarize the features of each interest dimension together, in each operation of the multi-head attention mechanism, a vector needs to be extracted from the same dimension in the aforementioned O1, O2, and O3 respectively.

[0116] Exemplarily, to determine the user interest feature vector of interest dimension 1, it is necessary to extract vector O11 from O1; extract vector O21 from O2; extract vector O31 from O3. At this time, since the recall object feature vector is also introduced, the multi-head attention mechanism also includes a target attention mechanism. Specifically, the weights of O11, O21, and O31 and the recall object feature vector are calculated respectively to obtain the weight vector W1 = (w11, w21, w31). This vector corresponds one-to-one with O11, O21, and O31. Among them, the calculation of the weights is the same as the foregoing method and will not be elaborated here. After the weight calculation is completed, the user interest feature vector of interest dimension 1 is obtained, denoted as: SPH1 (full name: sum-pooling head 1), where SPH1 = O11 * w11 + O21 * w21 + O31 * w31. Since a total of N dimensions of user interest feature vectors are introduced during the operation of the foregoing self-attention mechanism, the operation result of the final multi-head attention mechanism includes N interest dimension user interest feature vectors, namely SPH1, SPH2 to SPHN. Combining these N interest dimension user interest feature vectors can obtain the overall user interest feature vector.

[0117] After determining the overall user portrait feature vector and the overall user interest feature vector, in some embodiments, according to the overall user portrait feature vector, the overall user interest feature vector, and the recall object feature vector, determining the recommended object may include: merging the overall user portrait feature vector, the overall user interest feature vector, and the recall object feature vector to obtain a merged vector; extracting features from the merged vector, and determining the score value of the recall object through a preset normalization function; determining the recommended object according to the score value of the recall object.

[0118] Among them, the overall user portrait feature vector, the overall user interest feature vector, and the recall object feature vector can be merged together through the concat function to obtain a merged vector. And the features of the merged vector are extracted through a fully connected layer, thereby reducing the dimension of the merged vector. Exemplarily, the fully connected layer may include 3 layers, which are not limited in this disclosure.

[0119] After the output of the fully connected layer, it will also be normalized by the sigmoid function, and finally a one-dimensional vector is output. This vector can be used as the score value of multiple users for the recall object corresponding to a certain recall object feature vector. Among them, it can be understood that the higher the score value, the higher the degree of interest of the above users in the recall object; the lower the score value, the lower the degree of interest of the above users in the recall object. Specifically, the sigmoid function f(x) can be written as:

[0120]

[0121] After sorting the fractional values, the recall objects with fractional values higher than a certain threshold are selected as the recommended objects.

[0122] Exemplarily, after determining the recommended objects, at least one of the following methods can be used to determine the recommended object list: generating an object list based on the recommended objects; and / or, sending the recommended objects to the user terminal to generate an object list based on the recommended objects filtered by the user through the user terminal.

[0123] That is, the user can actively configure the recommended object list according to their own preferences based on the recommended objects; at the same time, the server can also automatically configure the recommended objects as the recommended object list and then send the recommended object list to the user terminal.

[0124] Next, Figure 4 Schematically shows an application flowchart of the method for determining recommended objects according to an embodiment of the present disclosure. It should be noted that Figure 4 and Figure 3 The two embodiments can be independent of each other or can be combined with each other. As Figure 4 shown, the method includes:

[0125] S401. Obtain the historical behavior data of at least two users and determine the music of interest to at least one user.

[0126] The historical behavior data includes the user's own music playing information within a period of time and the music of interest to the user. Among them, the music of interest to the user may include at least one of the following: user collection, like, heart, comment, share, and full play.

[0127] S402. Determine the music that all users are interested in together.

[0128] After obtaining the music that each user is interested in, through summarization and logical operations, the music that all users are interested in together can be determined.

[0129] S403. Determine the recall music list according to the music that all users are interested in together.

[0130] Based on the above similarity determination method and overall interest score determination method, recall music is determined from the music library and a recall music list is generated.

[0131] Exemplarily, if there are 3 pieces of music that all users are jointly interested in, and their types are slow-tempo English, Chinese pop music, and pure music respectively. When recalling similar music from the music library, 20 pieces of music with the highest similarity are determined for the first time, and these 20 pieces of music are used as the similar set. Next, the overall interest scores of these 20 pieces of music are determined, sorted, and the top 10 pieces of music with the overall interest scores after sorting are used as the recalled music to generate a recalled music list. Specifically, the recalled music can be "Yesterday Once More", "As Long As You Love Me", "Scarborough Fair", and so on.

[0132] S404. Determine the score value of each recalled music.

[0133] It should be understood that the generation of the recalled music described above can be generated by a multi-user music recall module. Further, the multi-user music sorting module can sort the recalled music list, and then determine the recommended music list.

[0134] In the multi-user music sorting module, there is a trained neural network model. The neural network model sorts the music list. The training process has been described in the foregoing embodiments and will not be elaborated here.

[0135] It can be passed through Figure 5 to intuitively represent the structure of the neural network model. Figure 5 Schematically shows the structural diagram of the neural network model in the multi-user music sorting module according to an embodiment of the present disclosure. As Figure 5 shown, the structure includes: an input layer, a vector layer, a multi-user interest extraction layer, and a multi-neural network layer.

[0136] Among them, the input of the input layer includes three parts: The first part is the user portrait features of at least two users. For example: the user 1 portrait feature of user 1, the user N portrait feature of user N. The user portrait feature reflects the basic features of the user, such as the user's age, gender, constellation, blood type, and the city where the user is located, etc. In addition, the user portrait feature can also include user device features and user music preference features. The user portrait feature can further include the device operating system and device operator of the user; the user music preference feature can further include language preference, style preference, release year preference, and so on.

[0137] The second part is the recalled music feature. The recalled music feature can include music attribute features and music portrait features. Among them, the music attribute features can further include music language, music style, release artist, playback sound quality, music duration, playback popularity, song comments, release version, and so on.

[0138] The third part is the user behavior sequence feature, which represents the user sequence composed of the music that the user has played completely within a period of time. It can reflect the user's interest tendency within a period of time.

[0139] After determining the above features, the neural network model will vectorize the above features in the vector layer and convert them into user portrait feature vectors, user behavior sequence feature matrices, and recalled music feature vectors. Among them, each time the recalled music input by the model is 1 piece, that is, the result output by the model each time is the score value corresponding to this music.

[0140] Exemplarily, the words in the above features can be converted into dense numbers through the look-up method, and then vectors or matrices are generated.

[0141] After vectorization is completed, each user's user portrait feature corresponds to a user portrait feature vector; each user's behavior sequence feature corresponds to a user behavior sequence feature matrix. At this time, in the multi-user interest extraction layer, an attention mechanism is introduced to fuse multiple user portrait feature vectors and recalled music feature vectors to obtain an overall user portrait feature vector. The specific formula is as follows:

[0142]

[0143] Among them, Vu represents the overall user portrait feature vector, N is the number of users, and the function g(x) can be linear or non-linear. In the function, V i represents the user portrait feature vector of user i, and V a represents the feature vector of the recalled music. The weight li of the user portrait feature vector of user i is output through g(x) i . After calculating the weights of each user, multiply them by the user portrait feature vectors and sum them to output the overall user portrait feature vectors corresponding to all users.

[0144] For the user behavior sequence feature matrix, two attention mechanism operations need to be performed here. The first attention mechanism operation is a self-attention mechanism operation, and the formula is the same as the above formula. After the self-attention mechanism operation, the user interest feature vectors of N users are obtained, which are: the user interest feature vector of user 1, the user interest feature vector of user 2... the user interest feature vector of user N. If the vector dimension set in the self-attention mechanism structure is n, the dimension of the user interest feature vector is also n. Each vector dimension reflects the characteristics of the user in this interest dimension.

[0145] The second attention mechanism operation is the multi-head attention mechanism operation. Its formula is still the same as the above formula. The difference is that the attention mechanism operations of the n interest dimensions are independent and parallel. That is, the first interest dimension of the user interest feature vectors from user 1 to user N is extracted and subjected to the attention mechanism operation with the recalled music feature vectors to obtain the interest feature vector 1. Similarly, the second interest dimension of the user interest feature vectors from user 1 to user N is extracted and subjected to the attention mechanism operation with the recalled music feature vectors to obtain the interest feature vector 2. And so on, until the interest feature vector N is obtained. Then, the interest feature vectors 1 to N are combined to obtain the overall user interest feature vector.

[0146] Finally, the overall user portrait feature vector, the overall user interest feature vector, and the recalled music feature vector are used as the inputs of the multi-neural network layer. The multi-neural network layer includes at least one convolutional layer and a fully connected layer. Specifically, the overall user portrait feature vector, the overall user interest feature vector, and the recalled music feature vector are combined with their cross features through the concat function, and the features are further extracted by the convolutional layer and the fully connected layer. After extraction, normalization is performed by the sigmoid function, and the score value is output. The range of the score value is from 0 to 1. According to the above method, the score value of each recalled music is output.

[0147] S405. Sort the recalled music according to the score value and determine the recommended music.

[0148] After the score value of each recalled music is determined, it is sorted according to the score value. And a threshold is set. If the score value is lower than the threshold, it will not be used as the recommended music. The recalled music with a score value higher than the threshold will be used as the recommended music and sent to the user.

[0149] S405. Send the recommended music to the client so that the user can select the recommended music on the client and form a recommended music list.

[0150] The user can select the recommended music matched by the server on the client by himself / herself to complete the configuration of the recommended music list.

[0151] In the embodiments of the present disclosure, the relevance between the song interest preferences of multiple users is fully considered, and a neural network model is used to solve the problem of music list generation. A multi-user common interest extraction layer is added to the neural network model to extract the interest preferences of multiple users in n interest dimensions from the user behavior sequence feature matrix, which improves the generation accuracy of the recommended music list. When extracting the interest preferences of the n interest dimensions of the user, first, the self-attention mechanism is used to extract the interest feature vectors of the n interest dimensions of each user, and then the multi-head attention mechanism is introduced to realize the weighted fusion of the n interest dimensions of the user. This method not only improves the preference degree of multiple users for the recommended music list, but also realizes the personalized matching recommendation of multiple users.

[0152] At the same time, the recommended music list conforms to the taste preferences of different users, improves the music listening experience of users listening to songs together, and thus drives the growth of the overall consumption of listening to songs together.

[0153] Exemplary Medium

[0154] After introducing the method of the exemplary embodiment of the present disclosure, next, reference is made to Figure 6 to describe the storage medium of the exemplary embodiment of the present disclosure.

[0155] Reference is made to Figure 6 As shown, a program product 60 for implementing the above method according to an embodiment of the present disclosure is described, which may be a portable compact disc read-only memory (CD-ROM) and includes program code, and may be run on a terminal device, such as a personal computer. However, the program product of the present disclosure is not limited thereto.

[0156] The program product may adopt any combination of one or more readable media. The readable media may be a readable signal medium or a readable storage medium. The readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination of the above. More specific examples (non-exhaustive list) of the readable storage medium include: an electrical connection having one or more wires, a portable disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above.

[0157] The readable signal medium may include a data signal propagated in a baseband or as part of a carrier wave, which carries the readable program code. Such a propagated data signal may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. The readable signal medium may also be any readable medium other than the readable storage medium.

[0158] Program code for performing the operations disclosed in this disclosure can be written in any combination of one or more programming languages, including object-oriented programming languages such as Java, C++, etc., and also including conventional procedural programming languages such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computing device, partially on the user's device, partially on a remote computing device, or entirely on a remote computing device or server. In cases involving a remote computing device, the remote computing device can be connected to the user's computing device through any type of network, including a local area network (LAN) or a wide area network (WAN).

[0159] Exemplary Apparatus

[0160] After introducing the media of the exemplary embodiments of this disclosure, next, reference is made to Figure 7 The recommended object determination device of the exemplary embodiment of this disclosure is described, which is used to implement the method in any of the above method embodiments. Its implementation principle and technical effects are similar and will not be elaborated here. Figure 7 The structural diagram of the recommended object determination device according to an embodiment of this disclosure is schematically shown. As Figure 7 shown, the recommended object determination device 700 includes:

[0161] A first determination module 701, configured to determine candidate objects that at least two users are jointly interested in according to the historical behavior data of at least two users for the object;

[0162] A second determination module 702, configured to determine recalled objects according to the candidate objects;

[0163] A third determination module 703, configured to determine recommended objects according to the recalled objects, the user portrait features of at least two users, and multiple user behavior sequence features, where the user behavior sequence features characterize the features of the sequence composed of the objects that the user has completely played within a preset time period.

[0164] In one embodiment of the present disclosure, the third determination module 703 may specifically be configured to: determine corresponding user portrait feature vectors according to the user portrait feature of at least two users; determine corresponding recall object feature vectors according to the features of the recall objects; determine overall user portrait feature vectors corresponding to at least two users according to the user portrait feature vectors and the recall object feature vectors; determine a user behavior sequence feature matrix corresponding to a user according to any user behavior sequence feature; determine an overall user interest feature vector according to the recall object feature vectors and multiple user behavior sequence feature matrices, where the overall user interest feature vector represents the overall interest feature distribution of the overall user in different dimensions; determine a recommended object according to the overall user portrait feature vector, the overall user interest feature vector, and the recall object feature vector.

[0165] In one embodiment of the present disclosure, the third determination module 703 may specifically be configured to: perform attention weighting processing on the user portrait feature vectors and the recall object feature vectors based on an attention mechanism to obtain overall user portrait feature vectors.

[0166] In one embodiment of the present disclosure, the third determination module 703 may specifically be configured to: perform weighting processing on multiple user behavior sequence feature matrices respectively based on a self-attention mechanism to determine user interest feature vectors corresponding to the users; perform weighting processing on the vector sums of multiple user interest feature vectors in the same dimension and the recall object feature vectors based on a multi-head attention mechanism to obtain an overall user interest feature vector.

[0167] In one embodiment of the present disclosure, the third determination module 703 may specifically be configured to: perform a merging process on the overall user portrait feature vector, the overall user interest feature vector, and the recall object feature vector to obtain a merged vector; perform feature extraction on the merged vector, and determine a score value of the recall object through a preset normalization function; determine a recommended object according to the score value of the recall object.

[0168] In one embodiment of the present disclosure, the third determination module 703 may also be used for at least one of the following:

[0169] 1) Determine user portrait feature vectors corresponding to users according to the user portrait features of at least two users based on a trained model;

[0170] 2) Determine recall object feature vectors corresponding to the recall objects according to the features of the recall objects based on a trained model;

[0171] 3) Determine user behavior sequence feature matrices corresponding to users according to any user behavior sequence feature based on a trained model.

[0172] Wherein, the model is obtained by pre-constructing samples and training an initial model according to the samples.

[0173] In one embodiment of the present disclosure, the second determination module 702 may be specifically configured to: determine the similarity between a candidate object and a first object; determine the overall interest score of at least two users for the first object according to the preference degree and similarity of any one of the at least two users for the candidate object; and based on multiple overall interest scores, sort multiple first objects to determine the recalled objects.

[0174] In one embodiment of the present disclosure, a generation module (not shown in the figure) is further included, which is configured to: generate an object list according to the recommended object; and / or send the recommended object to a user terminal to generate an object list according to the recommended object filtered by the user through the user terminal.

[0175] Exemplary Computing Device

[0176] After introducing the methods, media, and devices of the exemplary embodiments of the present disclosure, next, reference is made to Figure 8 describe the computing device of the exemplary embodiment of the present disclosure.

[0177] Figure 8 The computing device 80 shown is merely an example and should not impose any limitation on the functions and usage scope of the embodiments of the present disclosure.

[0178] As Figure 8 shown, the computing device 80 is presented in the form of a general-purpose computing device. The components of the computing device 80 may include, but are not limited to: at least one of the above-mentioned processing units 801, at least one of the above-mentioned storage units 802, and a bus 803 connecting different system components (including the processing unit 801 and the storage unit 802).

[0179] The bus 803 includes a data bus, a control bus, and an address bus.

[0180] The storage unit 802 may include a readable medium in the form of a volatile memory, such as a random access memory (RAM) 8021 and / or a cache memory 8022, and may further include a readable medium in the form of a non-volatile memory, such as a read-only memory (ROM) 8023.

[0181] The storage unit 802 may further include a program / utility 8025 having a set (at least one) of program modules 8024. Such program modules 8024 include, but are not limited to: an operating system, one or more application programs, other program modules, and program data. The implementation of a network environment may be included in each or some combination of these examples.

[0182] The computing device 80 may also communicate with one or more external devices 804 (such as a keyboard, a pointing device, etc.). Such communication may be carried out through the input / output (I / O) interface 805. Also, the computing device 80 may further communicate with one or more networks (such as a local area network (LAN), a wide area network (WAN), and / or a public network, such as the Internet) through the network adapter 806. As Figure 8 shown, the network adapter 806 communicates with other modules of the computing device 80 through the bus 803. It should be understood that, although not shown in the figures, other hardware and / or software modules may be used in conjunction with the computing device 80, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage systems, etc.

[0183] It should be noted that, although several units / modules or sub-units / modules of the recommendation object determination device are mentioned in the above detailed description, this division is merely exemplary and not mandatory. In fact, according to the embodiments of the present disclosure, the features and functions of two or more of the above-described units / modules may be embodied in one unit / module. Conversely, the features and functions of one unit / module described above may be further divided and embodied by multiple units / modules.

[0184] In addition, although the operations of the method of the present disclosure are described in a specific order in the drawings, this does not require or imply that these operations must be performed in that specific order, or that all of the shown operations must be performed to achieve the desired result. Additionally or alternatively, some steps may be omitted, multiple steps may be combined into one step for execution, and / or one step may be decomposed into multiple steps for execution.

[0185] Although the spirit and principles of the present disclosure have been described with reference to several specific embodiments, it should be understood that the present disclosure is not limited to the specific embodiments disclosed, and the division of each aspect does not mean that the features in these aspects cannot be combined for benefits. This division is only for the convenience of description. The present disclosure aims to cover various modifications and equivalent arrangements included within the spirit and scope of the appended claims.

Claims

1. A method for determining a recommended object, comprising: Determining candidate objects of common interest among at least two users based on historical behavior data of the at least two users for the object; Determining recalled objects based on the candidate objects; Determining corresponding user portrait feature vectors according to the user portrait feature of the at least two users; Determining corresponding recalled object feature vectors according to the features of the recalled objects; Determining the overall user portrait feature vectors corresponding to the at least two users according to the user portrait feature vectors and the recalled object feature vectors; Determining a user behavior sequence feature matrix corresponding to the user according to any user behavior sequence feature; Determining an overall user interest feature vector according to the recalled object feature vectors and a plurality of the user behavior sequence feature matrices, wherein the overall user interest feature vector characterizes the overall interest feature distribution of the overall user in different dimensions; Determining a recommended object according to the overall user portrait feature vector, the overall user interest feature vector and the recalled object feature vector; wherein the user behavior sequence feature characterizes the feature of a sequence composed of objects completely played by the user within a preset time period.

2. The method for determining a recommended object according to claim 1, wherein the determining the overall user portrait feature vectors corresponding to the at least two users according to the user portrait feature vectors and the recalled object feature vectors comprises: Performing attention weighting processing on the user portrait feature vectors and the recalled object feature vectors based on an attention mechanism to obtain the overall user portrait feature vectors.

3. The method for determining a recommended object according to claim 1, wherein the determining the overall user interest feature vector according to the recalled object feature vectors and a plurality of the user behavior sequence feature matrices comprises: Performing weighting processing on a plurality of the user behavior sequence feature matrices respectively based on a self-attention mechanism to determine user interest feature vectors corresponding to the users; Performing weighting processing on the vector sum of the plurality of user interest feature vectors in the same dimension and the recalled object feature vectors based on a multi-head attention mechanism to obtain the overall user interest feature vector.

4. The method for determining a recommended object according to claim 1, wherein the determining the recommended object according to the overall user portrait feature vector, the overall user interest feature vector and the recalled object feature vector comprises: Combining the overall user portrait feature vector, the overall user interest feature vector and the recalled object feature vector to obtain a combined vector; Performing feature extraction on the combined vector and determining a score value of the recalled object through a preset normalization function; Determining a recommended object according to the score value of the recalled object.

5. The method for determining a recommended object according to claim 1 further comprises at least one of the following: Determining a corresponding user portrait feature vector according to the user portrait feature of the at least two users includes: Determining user portrait feature vectors corresponding to the users based on a trained model according to the user portrait features of the at least two users; The determining the corresponding recalled object feature vectors according to the features of the recalled objects comprises: determining recalled object feature vectors corresponding to the recalled objects based on a trained model according to the features of the recalled objects; Determining the user behavior sequence feature matrix corresponding to the user according to any one of the user behavior sequence features includes: determining the user behavior sequence feature matrix corresponding to the user based on the trained model according to any one of the user behavior sequence features; Wherein, the model is obtained by pre-constructing samples and training an initial model according to the samples.

6. The method for determining a recommended object according to any one of claims 1 to 5, wherein determining the recalled object according to the candidate object includes: Determining the similarity between the candidate object and the first object; Determining the overall interest score of at least two users for the first object according to the preference degree of any one of at least two users for the candidate object and the similarity; Ranking the multiple first objects based on the multiple overall interest scores to determine the recalled object.

7. The method for determining a recommended object according to any one of claims 1 to 5 further includes: Generating an object list according to the recommended object; And / or, sending the recommended object to a user terminal to generate an object list according to the recommended object filtered by the user through the user terminal.

8. A recommended object determining device includes: A first determining module, configured to determine candidate objects of common interest to at least two users according to historical behavior data of the at least two users for objects; A second determining module, configured to determine a recalled object according to the candidate object; A third determining module, configured to determine a corresponding user portrait feature vector according to the user portrait feature of the at least two users; determine a corresponding recalled object feature vector according to the feature of the recalled object; determine a corresponding overall user portrait feature vector of the at least two users according to the user portrait feature vector and the recalled object feature vector; determine the user behavior sequence feature matrix corresponding to the user according to any one of the user behavior sequence features; determine an overall user interest feature vector according to the recalled object feature vector and multiple user behavior sequence feature matrices, where the overall user interest feature vector characterizes the overall interest feature distribution of the overall user in different dimensions; determine a recommended object according to the overall user portrait feature vector, the overall user interest feature vector and the recalled object feature vector, wherein the user behavior sequence feature characterizes the feature of a sequence composed of objects completely played by the user within a preset time period.

9. For the recommended object determining device according to claim 8, the third determining module is specifically configured to: Perform attention weighting processing on the user portrait feature vector and the recalled object feature vector based on an attention mechanism to obtain an overall user portrait feature vector.

10. For the recommended object determining device according to claim 8, the third determining module is specifically configured to: Perform weighting processing on multiple user behavior sequence feature matrices respectively based on a self-attention mechanism to determine the user interest feature vector corresponding to the user; Perform weighting processing on the vector sum of multiple user interest feature vectors in the same dimension and the recalled object feature vector based on a multi-head attention mechanism to obtain an overall user interest feature vector.

11. The recommended object determination device according to claim 8, wherein the third determination module is specifically configured to: Combine and process the overall user portrait feature vector, the overall user interest feature vector, and the recalled object feature vector to obtain a combined vector; Extract features from the combined vector, and determine the score value of the recalled object through a preset normalization function; Determine the recommended object according to the score value of the recalled object.

12. The recommended object determination device according to claim 8, wherein the third determination module is further configured to perform at least one of the following: Determining a corresponding user profile feature vector according to the user profile features of the at least two users includes: Based on the user portrait features of at least two users and a trained model, determine the user portrait feature vector corresponding to the user; The determining the recalled object feature vector corresponding to the recalled object according to the features of the recalled object includes: based on the features of the recalled object and a trained model, determining the recalled object feature vector corresponding to the recalled object; The determining the user behavior sequence feature matrix corresponding to the user according to any one of the user behavior sequence features includes: based on any one of the user behavior sequence features and a trained model, determining the user behavior sequence feature matrix corresponding to the user; Wherein, the model is obtained by pre-constructing a sample and training an initial model according to the sample.

13. The recommended object determination device according to any one of claims 8 to 12, wherein the second determination module is specifically configured to: Determine the similarity between the candidate object and the first object; According to the preference degree of any one of at least two users for the candidate object and the similarity, determine the overall interest score of at least two users for the first object; Based on multiple overall interest scores, sort multiple first objects to determine the recalled object.

14. The recommended object determination device according to any one of claims 8 to 12 further includes a generation module, configured to: Generate an object list according to the recommended object; And / or, send the recommended object to a user terminal to generate an object list according to the recommended object filtered by the user through the user terminal.

15. A computer-readable storage medium, including: Computer program instructions are stored in the computer-readable storage medium, and when the computer program instructions are executed, the recommended object determination method according to any one of claims 1 to 7 is implemented.

16. A computing device, comprising: A memory and a processor The memory is used to store program instructions; The processor is used to call the program instructions in the memory to execute the recommended object determination method according to any one of claims 1 to 7.

Images