Noise reduction and smart ticketing for social media-based communication systems
A noise reduction and smart ticketing application for social media-based communication systems may identify social media-based communications from users who are attempting to engage with a brand or entity on a social media platform as actionable, and distinguish other communications as noise. The noise reduction and smart ticketing system may use machine learning to determine which social media communications are actionable for a given company or other organization, and generates tickets for actionable communications. Actionable communications may include, but are not limited to, technical support issues, inquiries about a product release date, grievances, incidents, suggestions to improve service, critiques of company policies, etc. Non-actionable communications (i.e., “noise”) may include, but are not limited to, suggestions to other users, promotions, coupons, offers, marketing campaigns, affiliate marketing, statements that a user is attending an event, etc.
1. A computer-implemented method, comprising:
calculating, by a computing system, averaged local and global prediction scores for a mathematical representation of a plurality of words in a social media communication;
combining, by the computing system, the local and global prediction scores to obtain a single combined value S;
calculating, by the computing system, a probability P using:
P
=
1
1
+
e
-
(
a
+
bS
)
where a and b are both 1; and
predicting, by the computing system, a class or label of the social media communication using the combined value.
2. The computer-implemented method of claim 1 , wherein when the predicted class or label is incorrect, the method further comprises:
updating a local model, by the computing system; and
recalculating, by the computing system, the averaged local and global prediction scores using the updated local model and a global model.
3. The computer-implemented method of claim 2 , further comprising:
updating, by the computing system, the local model and recalculate the averaged local and global prediction scores using the updated local model and the global model until the predicted class or label is correct.
4. The computer-implemented method of claim 2 , wherein the local model is updated by updating each weight of the local model using one of the following equations:
w
i
=
w
i
+
f
i
sign
(
E
)
(
max
(
E
,
0
)
)
f
2
+
1
/
2
C
w
i
=
w
i
+
f
i
(
E
f
2
)
w
i
=
w
i
+
f
i
(
min
(
C
,
E
)
f
2
)
w
i
=
w
i
+
f
i
(
E
f
2
+
1
2
C
)
where f i is an i th value of a final text vector, C is a parameter, ∥f∥ is the norm of a final text vector, and E is an error between P and an actual label.
5. The computer-implemented method of 1 , further comprising:
calculating, by the computing system, the natural logarithm of P÷(1−P); and
comparing, by the computing system, the calculated natural logarithm with a threshold to predict the class or label.
6. A computer-implemented method, comprising:
calculating, by a computing system, averaged local and global prediction scores for a mathematical representation of a plurality of words in a social media communication;
combining, by the computing system, the local and global prediction scores to obtain a single combined value S;
calculating, by the computing system, a probability P using:
P
=
1
1
+
e
-
(
a
+
bS
)
where a and b are both real numbers;
calculating, by the computing system, the natural logarithm of P÷(1−P); and
comparing, by the computing system, the calculated natural logarithm with a threshold to predict the class or label of the social media communication.
7. A computer program embodied on a non-transitory computer-readable medium, the program configured to cause at least one processor to:
process a social media communication comprising a plurality of words to convert each of the plurality of words of the communication into a mathematical representation;
calculate local and global prediction scores for the mathematical representation of the plurality of words in the social media communication;
combine the local and global prediction scores to obtain a single combined value;
predict a class or label of the social media communication using the combined value; and
when the predicted class or label is incorrect:
update a local model, and
recalculate the averaged local and global prediction scores using the updated local model and a global model, wherein
the local model is updated by updating each weight of the local model using one of the following equations:
w
i
=
w
i
+
f
i
sign
(
E
)
(
max
(
E
,
0
)
)
f
2
+
1
/
2
C
w
i
=
w
i
+
f
i
(
E
f
2
)
w
i
=
w
i
+
f
i
(
min
(
C
,
E
)
f
2
)
w
i
=
w
i
+
f
i
(
E
f
2
+
1
2
C
)
where f i is an i th value of a final text vector, C is a parameter, ∥f∥ is the norm of a final text vector, and E is an error between P and an actual label.
8. The computer program of claim 7 , wherein the program is further configured to cause the at least one processor to update the local model and recalculate the averaged local and global prediction scores using the updated local model and the global model until the predicted class or label is correct.
9. The computer-implemented method of claim 7 , wherein equations of the local model are implemented directly in code of the local model, reducing a size of the local model.
10. The computer program of claim 7 , wherein the prediction of the class or label further comprises:
assigning a value of 0 to the class or label when the class or label is determined to be noise; and
assigning a value of 1 to the class or label when the class or label is determined to be actionable.
11. The computer program of claim 7 , wherein the program is further configured to cause the at least one processor to:
determine whether the social media communication comprises an override word or phrase; and
when the override word or phrase is present in the social media communication, set the class or label of the social media communication to a class or label associated with the override word or phrase.
12. The computer program of claim 7 , wherein the local and global prediction scores are calculated by adding vectors for each word to obtain a sum vector and dividing the sum vector by a total number of the words.
13. The computer program of claim 7 , wherein when a mathematical representation of a word is not available in a dictionary, the computer program is further configured to cause the at least one processor to generate a random representation for the word.
14. The computer program of claim 7 , wherein the global prediction score and local prediction score are combined by summing their values to obtain a single real number S.
15. The computer program of claim 14 , the program further configured to cause the at least one processor to:
calculate a probability P using:
P
=
1
1
+
e
-
(
a
+
bS
)
where a and b are real numbers.
16. The computer program of claim 15 , wherein a and b are both 1.
17. The computer program of claim 15 , wherein the program is further configured to cause the at least one processor to:
calculate the natural logarithm of P÷(1−P); and
compare the calculated natural logarithm with a threshold to predict the class or label.