Cross-domain action prediction
One or more computing devices, systems, and/or methods for cross-domain action prediction are provided. Action sequence embeddings are generated based upon a textual embedding and a graph embedding utilizing past user action sequences corresponding to sequences of past actions performed by users across a plurality of domains. An autoencoder is trained to utilize the action sequence embeddings to project the action sequence embeddings to obtain intent space vectors. A service switch classifier is trained using the intent space vectors. In response to the service switch classifier predicting that a current user will switch from a current domain to a next domain, the current user is provided with a recommendation of an action corresponding to the next domain.
1. A method, comprising:
executing, on a processor of a computing device, instructions that cause the computing device to perform operations, the operations comprising:
training a graph embedding utilizing past user action sequences corresponding to sequences of past actions performed by users across a plurality of domains;
generating action sequence embeddings based upon a textual embedding applied to the past user action sequences and based upon the graph embedding;
training an autoencoder utilizing the action sequence embeddings to project the action sequence embeddings to obtain intent space vectors, wherein a first intent space vector corresponds to a projection of a first action sequence embedding;
training a service switch classifier using the intent space vectors; and
utilizing the service switch classifier to predict whether users will switch between domains.
2. The method of claim 1 , wherein the utilizing comprises:
applying the action sequence embeddings to a current user action sequence of a current user to generate a current action sequence embedding for predicting whether the current user will switch from a current domain to a next domain.
3. The method of claim 2 , comprising:
projecting the current action sequence embedding to obtain a current intent space vector.
4. The method of claim 3 , comprising:
generating a prediction of whether the current user will switch from the current domain to the next domain based upon the current intent space vector.
5. The method of claim 3 , comprising:
utilizing the current intent space vector as input into a plurality of classifiers for determining a prediction of whether the current user will switch from the current domain to the next domain.
6. The method of claim 5 , comprising:
utilizing a voter mechanism to combine outputs from the plurality of classifiers to determine the prediction of whether the current user will switch from the current domain to the next domain.
7. The method of claim 6 , wherein a decision tree is utilized as a structure for performing the voter mechanism.
8. The method of claim 4 , comprising:
in response to the prediction corresponding to a probability of the current user switching to the next domain above a threshold, utilizing the current intent space vector and the intent space vectors to determine a next action to recommend to the current user.
9. The method of claim 8 , wherein the next action corresponds to an action that can be performed within the next domain.
10. The method of claim 8 , comprising:
identifying a threshold number of nearest intent space vectors in relation to the current intent space vector.
11. The method of claim 10 , comprising:
determining that a current intent of the current user corresponds to user action sequences of the threshold number of nearest intent space vectors.
12. The method of claim 11 , comprising:
recommending actions, of the user action sequences of the threshold number of nearest intent space vectors, to the current user.
13. A computing device comprising:
a processor; and
memory comprising processor-executable instructions that when executed by the processor cause performance of operations, the operations comprising:
generating action sequence embeddings based upon a textual embedding and a graph embedding utilizing past user action sequences corresponding to sequences of past actions performed by users across a plurality of domains;
training an autoencoder utilizing the action sequence embeddings to project the action sequence embeddings to obtain intent space vectors, wherein a first intent space vector corresponds to a projection of a first action sequence embedding;
training a service switch classifier using the intent space vectors; and
utilizing the service switch classifier to predict whether users will switch between domains.
14. The computing device of claim 13 , comprising:
applying the action sequence embeddings to a current user action sequence of a current user to generate a current action sequence embedding; and
utilizing the current action sequence embedding to determine a prediction as to whether the current user will switch from a current domain to a next domain.
15. The computing device of claim 14 , comprising:
in response to the prediction corresponding to a probability of the current user switching to the next domain above a threshold, utilizing a current intent space vector, derived from the current action sequence embedding, and the intent space vectors to determine a next action to recommend to the current user.
16. The computing device of claim 13 , comprising:
training the autoencoder based upon time spent metrics, wherein a time spent metric corresponds to a time difference between two consecutive actions within the past user action sequences.
17. The computing device of claim 13 , comprising:
training the autoencoder based upon times at which actions within the past user action sequences were performed.
18. The computing device of claim 13 , comprising:
training the autoencoder based upon content sizes of content items upon which actions within the past user action sequences were performed.
19. The computing device of claim 13 , comprising:
training the autoencoder based upon content types of content items upon which actions within the past user action sequences were performed.
20. A non-transitory machine readable medium having stored thereon processor-executable instructions that when executed cause performance of operations, the operations comprising:
generating action sequence embeddings based upon a textual embedding and a graph embedding utilizing past user action sequences corresponding to sequences of past actions performed by users across a plurality of domains;
training an autoencoder utilizing the action sequence embeddings to project the action sequence embeddings to obtain intent space vectors, wherein a first intent space vector corresponds to a projection of a first action sequence embedding;
training a service switch classifier using the intent space vectors; and
in response to the service switch classifier predicting that a current user will switch from a current domain to a next domain, providing the current user with a recommendation of an action corresponding to the next domain.