Relevance group suggestions

Patent metadata

Field	Value
Publication number	US-10042961-B2
Application number	US-201514811397-A
Country	US
Kind code	B2
Filing date	Jul 28, 2015
Priority date	Apr 28, 2015
Publication date	Aug 7, 2018
Grant date	Aug 7, 2018

Abstract

Techniques for providing a people recommendation system for predicting and recommending relevant people (or other entities) to include in a conversation. In an exemplary embodiment, a plurality of conversation boxes associated with communications between a user and target recipients, or between other users and recipients, are collected and stored as user history. During a training phase, the user history is used to train encoder and decoder blocks in a de-noising auto-encoder model. During a prediction phase, the trained encoder and decoder are used to predict one or more recipients for a current conversation box composed by the user, based on contextual and other signals extracted from the current conversation box. The predicted recipients are ranked using a scoring function, and the top-ranked individuals or entities may be recommended to the user.

First claim

The invention claimed is:

1. A method for execution by a computer processor coupled to a memory, the method comprising:
using an encoding function, encoding a vector comprising a pre-specified recipient group, a source user, and at least one context signal derived from a current communications item;
using a decoding function, decoding the encoded vector to generate a relevance group;
generating a recommendation comprising a member of the relevance group not in the pre-specified recipient group; and
training said encoding and decoding functions using a de-noising auto-encoder neural network model applied to a plurality of historical communications items, each of said plurality of historical communications items being associated with a data sample comprising a recipient group, a source user, and at least one context signal from the item;
said training comprising, for each data sample:
applying a corrupting function to the recipient group of the data sample to corrupt the recipient group of the data sample and generate a corrupted vector;
encoding the corrupted vector to generate an encoded training vector to extract features of the corrupted vector;
decoding the encoded training vector to generate a target recipient group corresponding to an estimate of the recipient group;
computing a loss function for the data sample based on the difference between the target recipient group and the recipient group; and
updating parameters of the encoding and decoding functions using the computed loss function for the data sample.

2. The method of claim 1 , the training comprising: updating one or more weights in the encoding function to minimize a ranking based reconstruction loss function comprising a log-likelihood computation.

3. The method of claim 1 , the training further comprising: updating one or more weights in the decoding function to minimize a ranking based reconstruction loss function comprising a log-likelihood computation.

4. The method of claim 1 , the training comprising: extracting a recipient vector and a context vector corresponding to each of the plurality of communications items; and corrupting at least one element of the recipient vector.

5. The method of claim 1 , wherein the pre-specified recipient group is represented using a sparse binary vector, and wherein the at least one context signal is represented using a multi-dimensional vector comprising a plurality of dimensions corresponding to occurrences of N-grams.

6. The method of claim 1 , wherein the at least one context signal is represented using a multi-dimensional vector comprising a plurality of dimensions corresponding to semantic symbols or extracted topics.

7. The method of claim 1 , further comprising assigning scores to members of the relevance group using a scoring function, the generating the recommendation comprising selecting a member of the relevance group having a highest score.

8. The method of claim 1 , the recommendation comprising a recommendation to add a recipient to an email.

9. An apparatus comprising a processor and a memory coupled to the processor, the memory storing instructions for causing the processor for executing:
a training block configured to train an encoding function and a decoding function using a plurality of historical communications items; and
a prediction block configured to generate a recipient recommendation using the encoding function to encode a vector comprising a pre-specified recipient group, a source user, and at least one context signal derived from a current communications item, and further using the decoding function to decode the encoded vector to generate a relevance group, wherein the recipient recommendation comprises a member of the relevance group not in the pre-specified recipient group;
the training block configured to train the encoding and decoding functions using a de-noising auto-encoder neural network model applied to the plurality of historical communications items, each of said plurality of historical communications items being associated with a data sample comprising a recipient group, a source user, and at least one context signal from the item, the training block configured to, for each data sample:
apply a corrupting function to the recipient group of the data sample to corrupt the recipient group of the data sample and generate a corrupted vector;
using the encoding function, encode the corrupted vector to generate an encoded training vector to extract features of the corrupted vector;
using the decoding function, decode the encoded training vector to generate a target recipient group corresponding to an estimate of the recipient group;
compute a loss function for the data sample based on the difference between the target recipient group and the recipient group; and
update parameters of the encoding and decoding functions using the computed loss function for the data sample.

10. The apparatus of claim 9 , wherein the vector comprises a representation of the pre-specified recipient group as a sparse binary vector, and a representation of the at least one context signal as a multi-dimensional N-gram vector.

11. The apparatus of claim 9 , wherein the vector comprises a representation of the at least one context signal as a multi-dimensional bag of words vector or as a multi-dimensional semantic symbol vector.

12. The apparatus of claim 9 , further comprising a recommendation block configured to:
apply a scoring function to generate a score for each member of the relevance group not in the pre-specified recipient group;
generate the recipient recommendation as the scored member having the highest score.

13. The apparatus of claim 12 , the recommendation block further configured to receive an acceptance or rejection by a user of the recipient recommendation, the prediction block further configured to generate the recipient recommendation using the encoding function to encode a second vector comprising a second pre-specified recipient group based on the acceptance or rejection, the source user, and the at least one context signal.

14. The apparatus of claim 9 , the loss function comprising a squared-loss function.

15. The apparatus of claim 9 , the loss function comprising a ranking-based function.

16. An apparatus comprising computer hardware configurable to execute:
means for, using an encoding function, encoding a vector comprising a pre-specified recipient group, a source user, and at least one context signal derived from a communications item;
means for, using a decoding function, decoding the encoded vector to generate a relevance group;
means for generating a recommendation comprising a member of the relevance group not in the pre-specified recipient group; and
means for training the encoding and decoding functions using a de-noising auto-encoder neural network model applied to a plurality of historical communications items, each of said plurality of historical communications items being associated with a data sample comprising a recipient group, a source user, and at least one context signal from the item;
said means for training comprising:
means for, for each data sample, applying a corrupting function to the recipient group of the data sample to corrupt the recipient group of the data sample and generate a corrupted vector;
means for encoding the corrupted vector to generate an encoded training vector to extract features of the corrupted vector;
means for decoding the encoded training vector to generate a target recipient group corresponding to an estimate of the recipient group;
means for computing a loss function for the data sa

Assignees

• Microsoft Technology Licensing Llc

Inventors

• Shen Yelong
• Song Xinying
• Gao Jianfeng
• Guo Chenlei
• Byun Byungki
• Wang Ye-Yi
• Remick Brian D
• Thiele Edward
• Ali Mohammed Aatif
• Gois Marcus
• Zou Yang
• Stepp Mariana
• Jetley Divya
• Friesen Stephen

Classifications

• G06F16/24578

  using ranking · CPC title

• G06F16/285

  Clustering or classification · CPC title

• G06Q10/06315

  Needs-based resource requirements planning or analysis · CPC title

• G06F16/335

  Filtering based on additional data, e.g. user or group profiles (filtering in web context G06F16/9535, G06F16/9536) · CPC title

• G06Q10/06313

  Resource planning in a project environment · CPC title