System and method for personalization via user embeddings and applications thereof

1 . A method implemented on at least one processor, a memory, and a communication platform for personalized online content service, comprising:

obtaining information related to a user, including a user profile identifying multiple interests of the user;

training, via machine learning and based on training data including online textual information defining each of the multiple interests and additional information related to each of the multiple interests, semantic embeddings of interest embeddings associated with at least some interest that is selected from the multiple interests based on a score associated with each of the multiple interests;

training, via machine learning and based on training data including online content within a sliding window in time, temporal embeddings of the interest embeddings associated with the at least some interest;

retraining, based on newly collected training data, the semantic embeddings and the temporal embeddings, wherein the temporal embeddings are retrained at a higher frequency than the sematic embeddings thereby maintaining an updated storage of the interest embeddings;

scaling attributes of semantic embeddings for each of the at least some interest based on the score to derive scaled semantic embeddings of the interest with scaled attributes;

combining the scaled attributes across the semantic embeddings for the at least some interest to obtain aggregated semantic embeddings having aggregated scaled attributes therein;

dynamically constructing, based on the aggregated semantic embeddings and the temporal embeddings associated with the at last some interest, user embeddings with respect to the user, wherein the user embeddings capture semantics of the at least some interest, similarity between the semantics of the at least some interest and semantics of other of the multiple interests, and additional interests that are not yet included in the user profile but appear together with the at least some interest in the sliding window in time;

in response to receipt of user interaction at a user interface, identifying, based on the user embeddings, online content relating to the at least some interest and the additional interests not yet known to be within the multiple interests of the user; and

providing, via the user interface, the online content to the user to personalize the online content service and to explore the additional interests of the user.

2 . The method of claim 1 , wherein each of the multiple interests is represented in the user profile based on a description of the interest and the score indicating a level of the interest with respect to the user.

3 . The method of claim 2 , wherein:

the semantic embeddings represent semantics of each of a plurality of interests; and

the temporal embeddings capture at least one additional interest that co-occurs with any of the plurality of interests within the sliding window in time.

4 . The method of claim 3 , further comprising obtaining the temporal embeddings by:

obtaining online articles within the sliding window in time;

extracting multiple entities appearing in each of the online articles;

creating temporal training data based on ranked entity lists, each of which corresponds to one of the online articles with multiple entities of the article ranked based on a pre-determined criterion; and

deriving, via machine learning based on the temporal training data, the temporal embeddings to capture temporal relationship among the plurality of interests within the sliding window in time.

5 . A machine readable and non-transitory medium having information recorded therein for personalized online content service, wherein the information, when read by the machine, causes the machine to perform the following steps:

obtaining information related to a user, including a user profile identifying multiple interests of the user;

training, via machine learning and based on training data including online textual information defining each of the multiple interests and additional information related to each of the multiple interests, semantic embeddings of interest embeddings associated with at least some interest that is selected from the multiple interests based on a score associated with each of the multiple interests;

training, via machine learning and based on training data including online content within a sliding window in time, temporal embeddings of the interest embeddings associated with the at least some interest;

retraining, based on newly collected training data, the semantic embeddings and the temporal embeddings, wherein the temporal embeddings are retrained at a higher frequency than the sematic embeddings thereby maintaining an updated storage of the interest embeddings;

scaling attributes of semantic embeddings for each of the at least some interest based on the score to derive scaled semantic embeddings of the interest with scaled attributes;

combining the scaled attributes across the semantic embeddings for the at least some interest to obtain aggregated semantic embeddings having aggregated scaled attributes therein;

dynamically constructing, based on the aggregated semantic embeddings and the temporal embeddings associated with the at last some interest, user embeddings with respect to the user, wherein the user embeddings capture semantics of the at least some interest, similarity between the semantics of the at least some interest and semantics of other of the multiple interests, and additional interests that are not yet included in the user profile but appear together with the at least some interest in the sliding window in time;

in response to receipt of user interaction at a user interface, identifying, based on the user embeddings, online content relating to the at least some interest and the additional interests not yet known to be within the multiple interests of the user; and

providing, via the user interface, the online content to the user to personalize the online content service and to explore the additional interests of the user.

6 . The medium of claim 5 , wherein each of the multiple interests is represented in the user profile based on a description of the interest and the score indicating a level of the interest with respect to the user.

7 . The medium of claim 6 , wherein:

semantic embeddings represent semantics of each of a plurality of interests; and

temporal embeddings capture at least one additional interest that co-occurs with any of the plurality of interests within the sliding window in time.

8 . The medium of claim 7 , further comprising obtaining the temporal embeddings by:

obtaining online articles within the sliding window in time;

extracting multiple entities appearing in each of the online articles;

creating temporal training data based on ranked entity lists, each of which corresponds to one of the online articles with multiple entities of the article ranked based on a pre-determined criterion; and

deriving, via machine learning based on the temporal training data, the temporal embeddings to capture temporal relationship among the plurality of interests within the sliding window in time.

9 . A system for personalized online content service, comprising:

a user profile retriever implemented by a processor and configured for obtaining information related to a user, including a user profile identifying multiple interests of the user;

a user embedding generator implemented by a processor and configured for

training, via machine learning and based on training data including online textual information defining each of the multiple interests and additional information related to each of the multiple interests, semantic embeddings of interest embeddings associated with at least some interest that is selected from the multiple interests based on a score associated with each of the multiple interests,

training, via machine learning and based on training data including online content within a sliding window in time, temporal embeddings of the interest embeddings associated with the at least some interest,

retraining, based on newly collected training data, the semantic embeddings and the temporal embeddings, wherein the temporal embeddings are retrained at a higher frequency than the sematic embeddings thereby maintaining an updated storage of the interest embeddings,

scaling attributes of semantic embeddings for each of the at least some interest based on the score to derive scaled semantic embeddings of the interest with scaled attributes,

combining the scaled attributes across the semantic embeddings for the at least some interest to obtain aggregated semantic embeddings having aggregated scaled attributes therein, and

dynamically constructing, based on the aggregated semantic embeddings and the temporal embeddings associated with the at last some interest, user embeddings with respect to the user, wherein the user embeddings capture semantics of the at least some interest, similarity between the semantics of the at least some interest and semantics of other of the multiple interests, and additional interests that are not yet included in the user profile but appear together with the at least some interest in the sliding window in time;

a content search/recommendation engine implemented by a processor and configured for in response to receipt of user interaction at a user interface, identifying, based on the user embeddings, online content relating to the at least some interest and the additional interests not yet known to be within the multiple interests of the user; and

the user interface implemented by the processor and configured for providing the online content to the user to personalize the online content service and to explore the additional interests of the user.

10 . The system of claim 9 , wherein each of the multiple interests is represented in the user profile based on a description of the interest and the score indicating a level of the interest with respect to the user.

11 . The system of claim 10 , wherein:

semantic embeddings represent semantics of each of a plurality of interests; and

temporal embeddings capture at least one additional interest that co-occurs with any of the plurality of interests within the sliding window in time.

12 . The system of claim 11 , further comprising obtaining the temporal embeddings by:

obtaining online articles within the sliding window in time;

extracting multiple entities appearing in each of the online articles;

creating temporal training data based on ranked entity lists, each of which corresponds to one of the online articles with multiple entities of the article ranked based on a pre-determined criterion; and

deriving, via machine learning based on the temporal training data, the temporal embeddings to capture temporal relationship among the plurality of interests within the sliding window in time.