AUTOMATICALLY DETERMINING USER INTENT BY SEQUENCE CLASSIFICATION BASED ON NON-TIME-SERIES-BASED MACHINE LEARNING

1 . A system comprising:

one or more processors; and

one or more non-transitory computer-readable media storing computing instructions that, when executed on the one or more processors, cause the one or more processors to perform:

receiving, via a computer network, an intent prediction request from a frontend system;

obtaining, from a database, one or more events in a lookback period associated with one or more items ordered by a user for the intent prediction request;

determining a time-based feature encoding for the one or more events for the user by:

determining a feature encoding for the one or more events;

determining a positional encoding for the one or more events, wherein:

the positional encoding comprises one or more positional vectors associated with a temporal sequence of the one or more events; and

determining the time-based feature encoding based at least in part on the feature encoding, the positional encoding, and a decay function; and

determining, in real-time via a machine learning model, a user intent for the user based on the time-based feature encoding.

2 . The system in claim 1 , wherein:

the feature encoding comprises one or more multi-dimensional feature vectors for the one or more events.

3 . The system in claim 2 , wherein:

each of the one or more multi-dimensional feature vectors comprises one or more of:

an embedding for a respective event of the one or more events;

an item quantity of a respective order for the respective event;

an amount of the respective order; or

a time difference between the respective event and a current time.

4 . The system in claim 1 , wherein:

the positional encoding is sinusoidal.

5 . The system in claim 4 , wherein:

the positional encoding comprises:

[

v

(

0

)

⋮

v

(

k

-

1

)

]

,

wherein:

k is a quantity of the one or more events;

n is a length of each of one or more feature vectors of the feature encoding;

v (i) is a positional vector of the one or more positional vectors for an i th event of the one or more events, 0≤i<k;

v (i) (q), a q th element of v (i) , 0≤q<n, is one of:

if (q mod 2)=0, then cos(ω q x i ), else sin(ω q x i ); or

if (q mod 2)=1, then cos(ω q x j ), else sin(ω q x j );

ω j is a frequency for a j th element of each positional vector of the one or more positional vectors, 0≤j<n; and

x j is a position of the j th element of each positional vector of the one or more positional vectors.

6 . The system in claim 1 , wherein:

the decay function is configured to determine a respective weightage for each event of the one or more events in the time-based feature encoding; and

the respective weightage for a first event of the one or more events, as determined by the decay function, is greater than the respective weightage for a second event of the one or more events, as determined by the decay function, when the first event is closer in time to a current time than the second event.

7 . The system in claim 6 , wherein:

the decay function comprises:

λ

⁢

T

-

Δ

⁢

t

i

T

,

wherein:

λ is a domain-specific constant, 0<λ≤1;

T is a time period of the lookback period; and

Δt i is a respective time difference between an i th event of the one or more events and the current time.

8 . The system in claim 1 , wherein:

the time-based feature encoding comprises one or more time-based feature vectors for the one or more events; and

each of the one or more time-based feature vectors is determined based on:

( v feature (i) +v position (i) )* f d ( i ), wherein:

 v feature (i) is a feature vector of one or more feature vectors of the feature encoding for an i th event of the one or more events;

v position (i) is a positional vector of the one or more positional vectors of the positional encoding for the i th event;

f d (i) is the decay function for the i th event; and

0≤i<a quantity of the one or more events.

9 . The system in claim 1 , wherein:

the machine learning model is pre-trained based on historical time-based feature encodings for historical events for one or more users and historical output intent data; and

the one or more users comprise the user.

10 . The system in claim 1 , wherein:

the machine learning model comprises a classification algorithm.

11 . A method being implemented via execution of computing instructions configured to run at one or more processors and stored at one or more non-transitory computer-readable media, the method comprising:

receiving, via a computer network, an intent prediction request from a frontend system;

obtaining, from a database, one or more events in a lookback period associated with one or more items ordered by a user for the intent prediction request;

determining a time-based feature encoding for the one or more events for the user by:

determining a feature encoding for the one or more events;

determining a positional encoding for the one or more events, wherein:

the positional encoding comprises one or more positional vectors associated with a temporal sequence of the one or more events; and

determining the time-based feature encoding based at least in part on the feature encoding, the positional encoding, and a decay function; and

determining, in real-time via a machine learning model, a user intent for the user based on the time-based feature encoding.

12 . The method in claim 11 , wherein:

the feature encoding comprises one or more multi-dimensional feature vectors for the one or more events.

13 . The method in claim 12 , wherein:

each of the one or more multi-dimensional feature vectors comprises one or more of:

an embedding for a respective event of the one or more events;

an item quantity of a respective order for the respective event;

an amount of the respective order; or

a time difference between the respective event and a current time.

14 . The method in claim 11 , wherein:

the positional encoding is sinusoidal.

15 . The method in claim 14 , wherein:

the positional encoding comprises:

[

v

(

0

)

⋮

v

(

k

-

1

)

]

,

wherein:

k is a quantity of the one or more events;

n is a length of each of one or more feature vectors of the feature encoding;

v (i) is a positional vector of the one or more positional vectors for an i th event of the one or more events, 0≤i<k;

v (i) (q), a q th element of v (i) , 0≤q<n, is one of:

if (q mod 2)=0, then cos(ω q x i ), else sin(ω q x i ); or

if (q mod 2)=1, then cos(ω q x j ), else sin(ω q x j );

ω j is a frequency for a j th element of each positional vector of the one or more positional vectors, 0≤j<n; and

x j is a position of the j th element of each positional vector of the one or more positional vectors.

16 . The method in claim 11 , wherein:

the decay function is configured to determine a respective weightage for each event of the one or more events in the time-based feature encoding; and

the respective weightage for a first event of the one or more events, as determined by the decay function, is greater than the respective weightage for a second event of the one or more events, as determined by the decay function, when the first event is closer in time to a current time than the second event.

17 . The method in claim 16 , wherein:

the decay function comprises:

λ

⁢

T

-

Δ

⁢

t

i

T

,

wherein:

λ is a domain-specific constant, 0<λ≤1;

T is a time period of the lookback period; and

Δt i is a respective time difference between an i th event of the one or more events and the current time.

18 . The method in claim 11 , wherein:

the time-based feature encoding comprises one or more time-based feature vectors for the one or more events; and

each of the one or more time-based feature vectors is determined based on:

( v feature (i) +v position (i) )* f d ( i ), wherein:

 v feature (i) is a feature vector of one or more feature vectors of the feature encoding for an i th event of the one or more events;

v position (i) is a positional vector of the one or more positional vectors of the positional encoding for the i th event;

f d (i) is the decay function for the i th event; and

0≤i<a quantity of the one or more events.

19 . The method in claim 11 , wherein:

the machine learning model is pre-trained based on historical time-based feature encodings for historical events for one or more users and historical output intent data; and

the one or more users comprise the user.

20 . The method in claim 11 , wherein:

the machine learning model comprises a classification algorithm.