Flexible multi-task neutral network for content ranking

1. A computer-implemented method performed by an online system, the method comprising:

receiving a plurality of content items associated with a viewing user;

deriving a feature vector for each content item;

predicting a likelihood of interacting with each content item using a prediction model associated with a plurality of tasks, wherein a first task of the plurality of tasks is associated with a first domain and a second task of the plurality of tasks is associated with a second domain, the prediction model comprising:

a first independent bottom layer associated with the first domain, the first independent bottom layer configured to extract features that are specific to the first domain and not shared with the second domain;

a second independent bottom layer configured to extract features that are shared across the first domain and the second domain;

a plurality of shared layers configured to extract common features that are shared across the plurality of tasks, the plurality of shared layers using the extracted features from the first independent bottom layer and the second independent bottom layer as input;

a plurality of separate layers, each separate layer configured to predict likelihood of the viewing user performing a task associated with the separate layer based on the features extracted from the first independent bottom layer, the second independent bottom layer, and the plurality of shared layers; and

ranking the plurality of content items based on the predicting.

2. The computer-implemented method of claim 1 , wherein the feature vector comprises characteristics of a corresponding content item, characteristics of a poster who posted the corresponding content item, characteristics of the viewing user, and relationships among the corresponding content item, the poster and the viewing user.

3. The computer-implemented method of claim 2 , wherein the relationships comprises: prior interactions between the corresponding content item and the viewing user, affinities between the poster and/or poster's friends and the viewing user, a time period when the corresponding content item is to be presented to the viewing user, and a recurring frequency of the corresponding content item that was presented to the viewing user.

4. The computer-implemented method of claim 1 , wherein the content item comprises stories, photos, videos, invitations, and advertisements.

5. The computer-implemented method of claim 1 , wherein the prediction model is not trained by training data for at least one of the plurality of tasks.

6. The computer-implemented method of claim 5 , wherein the prediction model comprises the plurality of shared layers trained by a first group of the plurality of tasks and at least one of the plurality of separate layers not trained by the first group of the plurality of tasks.

7. The computer-implemented method of claim 1 , wherein the plurality of independent layers comprises a plurality of independent bottom layers configured to extract features as inputs for the plurality of shared layers, and a plurality of independent low layers configured to extract features as inputs for the plurality of separate layers.

8. The computer-implemented method of claim 7 , wherein the plurality of independent layers further comprises a plurality of independent middle layers configured to combine common features extracted from the plurality of shared layers and features extracted from the plurality of independent low layers to generate combined features as inputs for the plurality of separate layers.

9. The computer-implemented method of claim 1 , wherein the plurality of independent layers and the plurality of separate layers are trained jointly.

10. The computer-implemented method of claim 1 , wherein the plurality of shared layers and the plurality of independent layers are trained individually.

11. A non-transitory computer-readable medium comprising computer program instructions that when executed by a computer processor of an online system causes the processor to perform steps comprising:

receiving a plurality of content items associated with a viewing user;

deriving a feature vector for each content item;

predicting a likelihood of interacting with each content item using a prediction model associated with a plurality of tasks, wherein a first task of the plurality of tasks is associated with a first domain and a second task of the plurality of tasks is associated with a second domain, the prediction model comprising:

a first independent bottom layer associated with the first domain, the first independent bottom layer configured to extract features that are specific to the first domain and not shared with the second domain;

a second independent bottom layer configured to extract features that are shared across the first domain and the second domain;

a plurality of shared layers configured to extract common features that are shared across the plurality of tasks, the plurality of shared layers using the extracted features from the first independent bottom layer and the second independent bottom layer as input;

a plurality of separate layers, each separate layer configured to predict likelihood of the viewing user performing a task associated with the separate layer based on the features extracted from the first independent bottom layer, the second independent bottom layer, and the plurality of shared layers; and

ranking the plurality of content items based on the predicting.

12. The non-transitory computer-readable medium of claim 11 , wherein the feature vector comprises characteristics of a corresponding content item, characteristics of a poster who posted the corresponding content item, characteristics of the viewing user, and relationships among the corresponding content item, the poster and the viewing user.

13. The non-transitory computer-readable medium of claim 12 , wherein the relationships comprises: prior interactions between the corresponding content item and the viewing user, affinities between the poster and/or poster's friends and the viewing user, a time period when the corresponding content item is to be presented to the viewing user, and a recurring frequency of the corresponding content item that was presented to the viewing user.

14. The non-transitory computer-readable medium of claim 11 , wherein the content item comprises stories, photos, videos, invitations, and advertisements.

15. The non-transitory computer-readable medium of claim 11 , wherein the prediction model is not trained by training data for at least one of the plurality of tasks.

16. The non-transitory computer-readable medium of claim 15 , wherein the prediction model comprises the plurality of shared layers trained by a first group of the plurality of tasks and at least one of the plurality of separate layers not trained by the first group of the plurality of tasks.

17. The non-transitory computer-readable medium of claim 11 , wherein the plurality of independent layers comprises a plurality of independent bottom layers configured to extract features as inputs for the plurality of shared layers, and a plurality of independent low layers configured to extract features as inputs for the plurality of separate layers.

18. The non-transitory computer-readable medium of claim 17 , wherein the plurality of independent layers further comprises a plurality of independent middle layers configured to combine common features extracted from the plurality of shared layers and features extracted from the plurality of independent low layers to generate combined features as inputs for the plurality of separate layers.

19. The non-transitory computer-readable medium of claim 11 , wherein the plurality of independent layers and the plurality of separate layers are trained jointly.

20. The non-transitory computer-readable medium of claim 11 , wherein the plurality of shared layers and the plurality of independent layers are trained individually.