Method for obtaining user portrait and related apparatus

1 . A method for training a user portrait model, performed by a computing device, the method comprising:

obtaining the user portrait model by using a to-be-trained user portrait model and a training sample, to obtain the user portrait model, the training sample including sample multimedia content and user features of a sample user, wherein the user portrait model is obtained by:

extracting feature vectors of the user features of the sample user and tag feature vectors of content tags of the sample multimedia content;

performing hierarchical embedding processing on the feature vectors of the user features of the sample user, to obtain a user feature vector of the sample user, wherein the performing the hierarchical embedding processing on the feature vectors of the user features comprises: for each of a plurality of feature fields, fusing the feature vectors of the user features in the feature field to obtain an intra-field feature vector by summation or attention-weighted summation; and subsequent to obtaining the intra-field feature vector for each of the plurality of feature fields, fusing the intra-field feature vectors across the plurality of feature fields to obtain the user feature vector by summation or attention-weighted summation;

performing hierarchical embedding processing on the tag feature vectors, to obtain a content feature vector of the sample multimedia content;

adjusting a parameter of the to-be-trained user portrait model based on a degree of association between the user feature vector and the content feature vector, wherein the adjusting the parameter comprises minimizing a cross-entropy loss using adaptive moment estimation until a preset stop condition is met; and

storing parameters of the user portrait model after the adjusting in a database of a server that provides services for a target application for subsequent use by the target application.

2 . The method according to claim 1 , wherein extracting the feature vectors comprises:

determining user features of the sample user in a feature field, and extracting feature vectors of user features in the feature field; and

determining content tags of the sample multimedia content in each of a plurality of tag fields, and extracting tag feature vectors of content tags in each of the plurality of tag fields.

3 . The method according to claim 2 , wherein performing the hierarchical embedding processing on the tag feature vectors of the content tags comprises:

fusing the tag feature vectors of the content tags in each of the plurality of tag fields, to obtain an intra-field tag vector of each of the plurality of tag fields, the fusing the tag feature vectors comprising performing summation; and

fusing intra-field tag vectors of the plurality of tag fields, to obtain the content feature vector of the sample multimedia content, the fusing the intra-field tag feature vectors comprising performing summation.

4 . A method for obtaining a user portrait, performed by a computing device, the method comprising:

determining a user feature vector of a target user according to attribute information and historical behavior data of the target user, wherein feature vectors of user features of the target user in a plurality of feature fields are embedded in the user feature vector of the target user by intra-field fusion followed by inter-field fusion;

determining content tags of each piece of multimedia content in the multimedia content in a plurality of tag fields;

extracting tag feature vectors of content tags in each tag field through a user portrait model, the user portrait model being obtained by training based on a degree of association between a user feature vector of a sample user and a content feature vector of sample multimedia content, a cross-entropy loss being minimized using adaptive moment estimation until a preset stop condition is met during the training, the content feature vector of the sample multimedia content being obtained by performing hierarchical embedding processing on tag feature vectors of content tags of the sample multimedia content, the user feature vector of the sample user being obtained by performing hierarchical embedding processing on feature vectors of user features of the sample user, wherein the feature vectors of the user features of the sample user in the plurality of feature fields are embedded in the user feature vector of the sample user by intra-field fusion followed by inter-field fusion;

determining an alternative tag of the target user from the content tags of the multimedia content according to similarities between the user feature vector and the tag feature vectors;

determining a user portrait of the target user based on the alternative tag of the target user; and

in response to determining that the target user triggers content recommendation in a target application installed in a terminal device, transmitting, from a server, multimedia content recommended to the target user according to the user portrait and causing the terminal device to display the multimedia content in the target application.

5 . The method according to claim 4 , wherein the content feature vector of the sample multimedia content is obtained further by:

determining content tags of the sample multimedia content in the plurality of tag fields, and extracting tag feature vectors of content tags in each tag field;

fusing the tag feature vectors of the content tags in each tag field, to obtain an intra-field tag vector of each tag field, the fusing the tag feature vectors comprising performing summation; and

fusing intra-field tag vectors of the plurality of tag fields, to obtain the content feature vector of the sample multimedia content, the fusing the intra-field tag vectors comprising performing summation.

6 . The method according to claim 4 , wherein the user feature vector of the sample user is obtained further by:

determining user features of the sample user in the plurality of feature fields, and extracting feature vectors of user features in each feature field;

fusing the feature vectors of the user features in the feature field, to obtain an intra-field feature vector of the feature field, the fusing the feature vectors comprising performing summation or weighted summation; and

fusing intra-field feature vectors of the plurality of feature fields, to obtain the user feature vector of the sample user, the fusing the intra-field feature vectors comprising performing summation or weighted summation.

7 . The method according to claim 6 , wherein the plurality of feature fields includes at least a first feature field that does not include historical behavior data associated with any application, and a second feature field that includes historical behavior data associated with an application.

8 . The method according to claim 4 , wherein determining the user feature vector comprises:

determining user features of the target user in the plurality of feature fields according to the attribute information and the historical behavior data of the target user; and

extracting feature vectors of user features in each feature field and performing hierarchical embedding processing on the feature vectors of the user features in the plurality of feature fields through the user portrait model, to determine the user feature vector of the target user.

9 . The method according to claim 8 , wherein performing the hierarchical embedding processing on the feature vectors of the user features comprises:

fusing the feature vectors of the user features in each of the plurality of feature fields, to obtain an intra-field feature vector of each of the plurality of feature fields, the fusing the feature vectors comprising performing summation or weighted summation; and

fusing intra-field feature vectors of plurality of feature fields, to obtain the user feature vector of the target user, the fusing the intra-field feature vectors comprising performing summation or weighted summation.

10 . The method according to claim 4 , wherein determining the alternative tag comprises:

determining similarities between the user feature vector and the tag feature vectors of the content tags in the plurality of tag fields; and

determining a content tag in the content tags of the multimedia content in the plurality of tag fields whose similarity meets a preset condition as the alternative tag of the target user.

11 . The method according to claim 4 , further comprising:

performing hierarchical embedding processing on tag feature vectors of content tags of each piece of multimedia content in the plurality of tag fields through the user portrait model, to obtain a content feature vector of each piece of multimedia content; and

determining the multimedia content recommended to the target user from the multimedia content according to a degree of association between the user feature vector of the target user and the content feature vector of each piece of multimedia content.

12 . The method according to claim 11 , wherein performing the hierarchical embedding processing on the tag feature vectors of the content tags comprises:

fusing the tag feature vectors of the content tags of the multimedia content in each tag field, to obtain an intra-field tag vector of each tag field, the fusing the tag feature vectors comprising performing summation; and

fusing intra-field tag vectors of the plurality of tag fields, to obtain the content feature vector of the multimedia content, the fusing the intra-field tag feature vectors comprising performing summation.

13 . An apparatus for obtaining user portrait, the apparatus comprising: a memory storing computer program instructions; and a processor coupled to the memory and configured to execute the computer program instructions and perform:

determining a user feature vector of a target user according to attribute information and historical behavior data of the target user, wherein feature vectors of user features of the target user in a plurality of feature fields are embedded in the user feature vector of the target user by intra-field fusion followed by inter-field fusion;

determining content tags of each piece of multimedia content in the multimedia content in a plurality of tag fields;

extracting tag feature vectors of content tags in each tag field through a user portrait model, the user portrait model being obtained by training based on a degree of association between a user feature vector of a sample user and a content feature vector of sample multimedia content, a cross-entropy loss being minimized using adaptive moment estimation until a preset stop condition is met during the training, the content feature vector of the sample multimedia content being obtained by performing hierarchical embedding processing on tag feature vectors of content tags of the sample multimedia content, the user feature vector of the sample user being obtained by performing hierarchical embedding processing on feature vectors of user features of the sample user, wherein the feature vectors of the user features of the sample user in the plurality of feature fields are embedded in the user feature vector of the sample user by intra-field fusion followed by inter-field fusion;

determining an alternative tag of the target user from the content tags of the multimedia content according to similarities between the user feature vector and the tag feature vectors;

determining a user portrait of the target user based on the alternative tag of the target user; and

in response to determining that the target user triggers content recommendation in a target application installed in a terminal device, transmitting, from a server, multimedia content recommended to the target user according to the user portrait and causing the terminal device to display the multimedia content in the target application.

14 . The apparatus according to claim 13 , wherein the content feature vector of the sample multimedia content is obtained further by:

determining content tags of the sample multimedia content in the plurality of tag fields, and extracting tag feature vectors of content tags in each tag field;

fusing the tag feature vectors of the content tags in each tag field, to obtain an intra-field tag vector of each tag field, the fusing the tag feature vectors comprising performing summation; and

fusing intra-field tag vectors of the plurality of tag fields, to obtain the content feature vector of the sample multimedia content, the fusing the intra-field tag vectors comprising performing summation.

15 . The apparatus according to claim 13 , wherein the user feature vector of the sample user is obtained further by:

determining user features of the sample user in the plurality of feature fields, and extracting feature vectors of user features in each feature field;

fusing the feature vectors of the user features in each of the plurality of feature fields, to obtain an intra-field feature vector of each of the plurality of feature fields, the fusing the feature vectors comprising performing summation or weighted summation; and

fusing intra-field feature vectors of the plurality of feature fields, to obtain the user feature vector of the sample user, the fusing the intra-field feature vectors comprising performing summation or weighted summation.

16 . The apparatus according to claim 13 , wherein determining the user feature vector comprises:

determining user features of the target user in the plurality of feature fields according to the attribute information and the historical behavior data of the target user; and

extracting feature vectors of user features in each feature field and performing hierarchical embedding processing on the feature vectors of the user features in the plurality of feature fields through the user portrait model, to determine the user feature vector of the target user.

17 . The apparatus according to claim 16 , wherein performing the hierarchical embedding processing on the feature vectors of the user features comprises:

fusing the feature vectors of the user features in each of the plurality of feature fields, to obtain an intra-field feature vector of each of the plurality of feature fields, the fusing the feature vectors comprising performing summation or weighted summation; and

fusing intra-field feature vectors of the plurality of feature fields, to obtain the user feature vector of the target user, the fusing the intra-field feature vectors comprising performing summation or weighted summation.

18 . The apparatus according to claim 13 , wherein determining the alternative tag comprises:

determining similarities between the user feature vector and the tag feature vectors of the content tags in the plurality of tag fields; and

determining a content tag in the content tags of the multimedia content in the plurality of tag fields whose similarity meets a preset condition as the alternative tag of the target user.