User targeted content generation using multimodal embeddings
A method, apparatus and system for determining user-content associations for determining and providing user-preferred content using multimodal embeddings include creating an embedding space for multimodal content by creating a first modality vector representation of the multimodal content having a first modality, creating a second modality vector representation of the multimodal content having a second modality, creating a user vector representation, as a third modality, for each user associated with at least a portion of the multimodal content, and embedding the first and the second modality vector representations and the user vector representations in the common embedding space using at least a mixture of loss functions for each modality pair of the first, the at least second and the third modalities that pushes closer co-occurring pairs of multimodal content. Embodiments can further include generating content using determined attributes of a message to be conveyed and features of the user-preferred content.
1. A method for training a neural network for creating a common embedding space for multimodal content for determining user-content associations, comprising:
for each of a plurality of content of the multimodal content, creating respective first modality vector representations of the content of the multimodal content having a first modality using a machine learning model;
creating a first training set comprising the first modality vector representations;
training the neural network in a first stage using the first training set by embedding the first modality vector representations in a common embedding space;
for each of a plurality of content of the multimodal content, creating at least respective second modality vector representations of content of the multimodal content having at least a second modality using a machine learning model;
creating a second training set comprising the second modality vector representations;
training the neural network in a second stage using the second training set by embedding the second modality vector representations in the common embedding space;
creating a respective user vector representation, as a third modality, for each user associated with at least a portion of the multimodal content using information regarding a respective portion of the multimodal content associated with each user;
creating a third training set comprising the third modality vector representations; and
training the neural network in a third stage using the third training set by embedding the third modality vector representations in the common embedding space;
wherein the first modality vector representations, the at least second modality vector representations and the user vector representations are embedded in the common embedding space using at least a combined loss function which comprises a ranking loss component for each modality pair of the first, the at least second and the third modalities that pushes closer co-occurring pairs of multimodal content.
2. The method of claim 1 , wherein at least three mixture loss functions are used to embed the first modality vector representations, the at least second modality vector representations and the user vector representations in the common embedding space.
3. The method of claim 1 , wherein the first modality comprises text and the at least second modality comprises at least images.
4. The method of claim 3 , wherein the machine learning model for creating the first modality vector representation comprises a text-based convolutional neural network (CNN) with fully connected (FC) layers and the machine learning model for creating the at least second modality vector representation comprises a deep image encoder implementing a convolutional neural network (CNN) with fully connected (FC) layers.
5. The method of claim 1 , comprising:
for each of a plurality of content of the multimodal content, creating at least a respective fourth modality vector representation of content of the multimodal content having at least a fourth modality using a machine learning model; and
embedding the first modality vector representations, the at least second modality vector representations, the user vector representations and the at least fourth modality vector representation in a common embedding space using at least a mixture of loss functions for each modality pair of the first, the at least second, the third, and the fourth modalities that pushes closer co-occurring pairs of multimodal content.
6. The method of claim 5 , wherein the at least fourth modality comprises audio and wherein at least four mixture loss functions are used to embed the first modality vector representations, the at least second modality vector representations, the user vector representations and the at least fourth modality vector representations in the common embedding space.
7. The method of claim 1 , wherein the user vector representation of a user is created as an arithmetic average of respective first modality vector representations and respective at least second modality vector representations of multimodal content associated with the user.
8. The method of claim 1 , further comprising:
determining a similarity between the embedded user vector representation and the embedded first modality vector representations and the at least second modality vector representations using a distance function to identify user multimodal content preferences.
9. A method for predicting user content preferences of multimodal content in a multimodal content database using a trained embedding space having embedded therein a plurality of user vector representations and multimodal content including a first modality and at least a second modality, comprising:
identifying a user for which to predict user content preferences of the multimodal content;
locating an embedded user vector representation of the identified user in the trained embedding space, wherein the user vector representation was embedded in the trained embedding space using at least a combined loss function which comprises a ranking loss component for each modality pair of the first, the at least second and the user vector representation that pushes closer co-occurring pairs of multimodal content;
determining a similarity between the embedded user vector representation and embedded vector representations of multimodal content in the trained embedding space using a distance function to identify user multimodal content preferences; and
comparing the identified user multimodal content preferences to multimodal content in the multimodal content dataset to predict user content preferences of multimodal content in the multimodal content dataset;
wherein the embedding space is trained using a method comprising:
for each of a plurality of content of the multimodal content, creating respective first modality vector representations of the content of the multimodal content having a first modality using a machine learning model:
creating a first training set comprising the first modality vector representations;
training the embedding space in a first stage using the first training set by embedding the first modality vector representations in a common embedding space;
for each of a plurality of content of the multimodal content, creating at least respective second modality vector representations of content of the multimodal content having at least a second modality using a machine learning model:
creating a second training set comprising the second modality vector representations;
training the embedding space in a second stage using the second training set by embedding the second modality vector representations in the common embedding space;
creating a respective user vector representation, as a third modality, for each user associated with at least a portion of the multimodal content using information regarding a respective portion of the multimodal content associated with each user;
creating a third training set comprising the third modality vector representations; and
training the embedding space in a third stage using the third training set by embedding the third modality vector representations in the common embedding space:
wherein the first modality vector representations, the at least second modality vector representations and the user vector representations are embedded in the common embedding space using at least a combined loss function which includes a ranking loss component for each modality pair of the first, the at least second and the third modalities that pushes closer co-occurring pairs of multimodal content.
10. The method of claim 9 , wherein user content preferences are inferred based on a proximity of the user vector representations to at least one other feature vector embedded in the embedding space.
11. A method for using determined user preferred content, determined using the method of claim 9 , for conveying message intents, further comprising:
determining attributes, including at least one of meaning, color, and appearance of a message to be conveyed by the user preferred content, using a first pre-trained embedding space;
determining features of the user preferred content for conveying the message using a second pre-trained embedding space; and
generating content using the determined attributes of the message to be conveyed and the determined features of the user preferred content for conveying the message.
12. The method of claim 11 , wherein the generating content comprises at least one of modifying existing content or generating new content.
13. The method of claim 11 , wherein the attributes of the message determine an intent of the message to be conveyed by the user preferred content.
14. The method of claim 11 , wherein the message comprises at least one of an advertisement, a social media post, content used for marketing, content used for educational purposes, and content used for entertainment.
15. The method of claim 11 , wherein the content is generated using an adversarial relationship between a discriminator and a generator.
16. An apparatus for training a neural network for creating a common embedding space for multimodal content for determining user-content associations, comprising:
a processor; and
a memory accessible to the processor, the memory having stored therein at least one of programs or instructions executable by the processor to configure the apparatus to:
create a respective first modality vector representation of the content of the multimodal content having a first modality using a machine learning model for each of a plurality of content of the multimodal content;
create a first training set comprising the first modality vector representations;
train the neural network in a first stage using the first training set by embedding the first modality vector representations in a common embedding space;
create at least a respective second modality vector representation of content of the multimodal content having at least a second modality using a machine learning model for each of a plurality of content of the multimodal content;
create a second training set comprising the second modality vector representations;
train the neural network in a second stage using the second training set by embedding the second modality vector representations in the common embedding space;
create a respective user vector representation, as a third modality, for each user associated with at least a portion of the multimodal content using information regarding a respective portion of the multimodal content associated with each user;
create a third training set comprising the third modality vector representations; and
train the neural network in a third stage using the third training set by embedding the third modality vector representations in the common embedding space;
wherein the first modality vector representations, the at least second modality vector representations and the user vector representations are embedded in the common embedding space using at least a combined loss function which comprises a ranking loss component for each modality pair of the first, the at least second and the third modalities that pushes closer co-occurring pairs of multimodal content.
17. The apparatus of claim 16 , wherein, upon identification of a user for which to predict user content preferences of the multimodal content, the apparatus is further configured to:
locate an embedded user vector representation of the identified user in the embedding space;
determine a similarity between the located, embedded user vector representation and embedded vector representations of the first and at least the second modalities in the embedding space using a distance function to identify user multimodal content preferences; and
compare the identified user multimodal content preferences to multimodal content in a multimodal content dataset to predict user content preferences of multimodal content in the multimodal content dataset.
18. The apparatus of claim 17 , wherein the multimodal content dataset comprises at least one of advertising content, marketing content, entertainment content, educational content, or social media content.
19. The apparatus of claim 16 , wherein the apparatus is further configured to:
determine attributes including at least one of meaning, color, and appearance of a message to be conveyed by the user preferred content using a first pre-trained embedding space, and determining features of the user preferred content for conveying the message using a second pre-trained embedding space; and
generate content generator content using the determined attributes of the message to be conveyed and the determined features of the user preferred content for conveying the desired message.
20. The apparatus of claim 16 , wherein the user vector representation of a user is created as an arithmetic average of respective first modality vector representations and respective at least second modality vector representations of multimodal content associated with the user.