Predicting Reactions to Short-Text Posts

• US 20150052087A1
• Filed: 08/14/2013
• Published: 02/19/2015
• Est. Priority Date: 08/14/2013
• Status: Active Grant

• Current Assignee: Adobe Inc.
• Original Assignee: Adobe Systems Incorporated (Adobe Inc.)
• Inventors: Srinivasan, Balaji Vasan, Natarajan, Anandhavelu, Gupta, Vineet, Revankar, Shriram V., Sinha, Ritwik, Ravindran, Balaraman
• Granted Patent: US 9,256,826 B2
• Time in Patent Office: Days
• US Class Current: 706/12
• CPC Class Codes: G06N 20/00 Machine learning
• CPC Class Codes: G06N 20/10 using kernel methods, e.g. support vector machines [SVM]
• CPC Class Codes: G06N 20/20 Ensemble learning
• CPC Class Codes: G06N 5/02 Knowledge representation; Symbolic representation

Abstract

This document describes techniques for predicting reactions to short-text posts. In one or more implementations, a prediction model for short-text posts is generated from previous posts to a social network and responses to the posts by the social network community. Subsequently, the prediction model can be used to predict the social network community'"'"'s reaction to a proposed post prior to the proposed post being posted to the social network.

Claims

1. One or more computer-readable storage media comprising instructions stored thereon that, responsive to execution by a computing device, cause the computing device to perform operations comprising:

collecting posts to a social network and responses to each post by a social network community;

binning the posts into different post classes based on a number of the responses to each post;

extracting features from the posts, the extracted features including at least one non-textual feature; and

forming a prediction model by applying a learning model to the different classes of posts and the extracted features.

2. The one or more computer-readable storage media of claim 1, wherein the different post classes include at least a popular post class and an unpopular post class, wherein the popular post class includes a first subset of posts that received a high number of responses, and wherein the unpopular post class includes a second subset of posts that received a low number of responses.

4. The one or more computer-readable storage media of claim 1, wherein the binning is performed by applying Otsu thresholding techniques to a histogram of the number of the responses to each post.

5. The one or more computer-readable storage media of claim 6, wherein prior to applying the Otsu thresholding techniques, the instructions, responsive to execution by the computing device, cause the computing device to perform operations further comprising normalizing the number of the responses to each post by dividing the number of the responses to each post by a total number of subscribers of the social network community.

6. The one or more computer-readable storage media of claim 1, wherein the at least one non-textual feature comprises one of:

a length of the post, a time of day or week of the post, a duration of time since a last post, a sentiment of the post, a hotness of topic of the post, a newness of topic of the post, or a type of the post.

7. The one or more computer-readable storage media of claim 1, wherein the extracted features further include at least one textual feature.

8. The one or more computer-readable storage media of claim 1, wherein the extracted features further include a differential term frequency-inverse document frequency (Diff-TFIDF) of each post.

9. The one or more computer-readable storage media of claim 1, wherein the learning model comprises one of a support vector machine (SVM) learning model or a Random Forest learning model.

10. The one or more computer-readable storage media of claim 1, wherein the posts comprise short-text posts.

11. The one or more computer-readable storage media of claim 1, wherein the responses to each post comprise one or more of comments, likes, shares, retweets, or replies.

3. The one or more computer-readable storage media of claim 3, wherein the popular post class includes approximately 10% of the collected posts.

12. A system comprising:

one or modules implemented at least partially in hardware, the one or more modules configured to perform operations comprising;

receiving a request to predict a reaction, by a social network community, to a proposed post prior to the proposed post being posted;

extracting one or more features from the proposed post;

comparing the one or more extracted features of the proposed post to features associated with different post classes of a prediction model;

selecting a post class of the different post classes that includes features that are most similar to the extracted features of the proposed post; and

predicting the reaction by the social network community to the proposed post based on a number of responses received by posts in the selected post class.

13. The system of claim 12, wherein the predicted reaction indicates a predicted number of responses that the proposed post will receive, the predicted number of responses being similar to the number of responses received by posts in the selected post class.

14. The system of claim 13, wherein the predicted number of responses that the proposed post will receive is based on one or more of a predicted number of comments the proposed post will receive, a predicted number of likes the proposed post will receive, a predicted number of shares the proposed post will receive, a predicted number of tweets the proposed post will receive, or a predicted number of replies the proposed post will receive.

15. The system of claim 12, wherein the predicted reaction indicates whether the post will be popular or unpopular.

16. A method comprising:

collecting posts to a social network and responses to each post;

automatically binning the posts into different post classes based on a number of the responses to each post;

extracting features from the posts;

generating a prediction model by applying a learning model to the different post classes and the extracted features;

receiving a request to predict a reaction to a proposed post prior to the proposed post being posted to the social network; and

predict the reaction to the proposed post using the learning model.

17. The method of claim 16, wherein the extracted features include at least one non-textual feature.

18. The method of claim 16, wherein the extracted features include at least one of a length of the post, a time of day or week of the post, a duration of time since a last post, a sentiment of the post, a type of the post, a hotness of topic of the post, a newness of topic of the post, or a differential term frequency-inverse document frequency (Diff-TFIDF) of each post.

19. The method of claim 16, wherein the binning is performed by applying Otsu thresholding techniques to a histogram of the number of the responses to each post.

20. The method of claim 16, wherein the posts comprise short-text posts.