Systems and methods for multi-market browse facet mapping and ranking using machine learning
Systems and methods including one or more processors and one or more non-transitory storage devices storing computing instructions configured to run on the one or more processors and perform: receiving a first set of facet information corresponding to a first marketplace; receiving a second set of facet information corresponding to a second marketplace; analyzing the first set of facet information to reduce a number of attributes to process; building a machine learning model to map candidate facets from the first set of facet information to facets from the second set of facet information; training the machine learning model based on the first set of facet information, the second set of facet information, and the mapping of candidate facets from the first set of facet information to facets from the second set of facet information; and displaying a selected number of re-ranked facets based on an output from the machine learning model. Other embodiments are disclosed herein.
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, perform:
receiving a first set of facet information corresponding to a first online marketplace, the first set of facet information comprising first facets of the first online marketplace;
receiving a second set of facet information corresponding to a second online marketplace, the second set of facet information comprising second facets for the second online marketplace;
analyzing the first set of facet information to reduce a number of the first facets as candidate facets;
building a machine learning model to map candidate facets to the second facets from the second set of facet information by determining, for each candidate facet, a facet candidate score indicative of a similarity between the candidate facets and the second facets, wherein the facet candidate scores are determined using one or more similarity functions comprising one or more of a n-gram Jaccard similarity based on tokenizing the candidate and second facets into tokens and constructing n-grams by splitting the tokens, an edit-distance based measure, a Monge-Elkan similarity that computes an average similarity between similar token pairs using an inter-token similarity measure, and a cosine similarity of word embeddings constructed from the candidate and second facets;
training the machine learning model to perform facet ranking on the candidate facets and second facets using the first set of facet information, the second set of facet information, and the mapping of the candidate facets to the second facets to a curated set of factors for the first online marketplace; and
displaying, on a navigation panel of a graphical user interface, a selected number of ranked facets as a curated set of facets for the first online marketplace, the navigation panel comprising one or more of the first facets and one or more of the second facets that are selected based on the facet ranking limited number of facet display positions.
2 . The system of claim 1 , wherein the first set of facet information comprises: a product catalog for the first marketplace, engagement data for the first marketplace, and facet rules for the first marketplace.
3 . The system of claim 1 , wherein the second set of facet information comprises a whitelist of facets for the second marketplace.
4 . The system of claim 1 , wherein analyzing the first set of facet information further comprises:
determining a respective stock keeping unit (SKU) coverage score for each facet in the first set of facet information; and
determining a respective cardinality value for each facet in the first set of facet information; and
determining a respective facet value dispersion for each facet in the first set of facet information.
5 . The system of claim 4 , wherein determining the respective SKU coverage score includes using an equation comprising:
score
coverage
(
f
)
=
number
of
SKUs
with
facet
f
total
number
of
SKUs
in
category
L
.
6 . The system of claim 4 , wherein determining the respective cardinality value includes using an equation comprising:
score
valueCard
(
f
)
=
{
0
if
c
f
<
min
Card
e
c
f
-
optCard
θ
2
if
min
Card
≤
c
f
≤
optCard
s
1
+
(
c
f
-
optCard
)
if
c
f
>
optCard
where c f corresponds to a number of facet values, optCard corresponds to an optimal number of facet values, minCard corresponds to a minimum number of facet values, and s corresponds to a scale constant that controls a slope of decay after optCard.
7 . The system of claim 4 , wherein determining the respective facet value dispersion includes using an equation comprising:
score
sizeDispersion
(
f
)
=
1
1
+
sizeVar
(
f
)
sizeVar
(
f
)
=
1
N
×
∑
i
=
1
N
(
x
i
-
x
_
)
2
x
_
where N corresponds to the number of facet values, x; corresponds to the i th facet value size (i.e., the number of products under the i th facet value) and x corresponds to the average facet value size.
8 . The system of claim 1 , wherein building the machine learning model further comprises:
determining the n-gram Jaccard similarity;
determining the edit-distance based measures;
determining the Monge-Elkan similarity;
determining the cosine similarity of word embeddings; and
determining the facet candidate score.
9 . The system of claim 8 , wherein the n-gram Jaccard similarity comprises a Bigram Jaccard similarity, and wherein determining the Bigram Jaccard similarity includes using an equation comprising:
sim
jaccard
(
S
a
,
S
b
)
=
❘
"\[LeftBracketingBar]"
S
a
⋂
S
b
❘
"\[RightBracketingBar]"
❘
"\[LeftBracketingBar]"
S
a
⋃
S
b
❘
"\[RightBracketingBar]"
=
❘
"\[LeftBracketingBar]"
S
a
⋂
S
b
❘
"\[RightBracketingBar]"
❘
"\[LeftBracketingBar]"
S
a
❘
"\[RightBracketingBar]"
+
❘
"\[LeftBracketingBar]"
S
b
❘
"\[RightBracketingBar]"
-
❘
"\[LeftBracketingBar]"
S
a
⋂
S
b
❘
"\[RightBracketingBar]"
where a corresponds to a facet in the first set of facet information, b corresponds to a facet in the second set of facet information, S a corresponds to a set of bigrams constructed from string a, and S b corresponds to a set of bigrams constructed from string b.
10 . The system of claim 8 , wherein determining the Monge-Elkan similarity includes using an equation comprising:
sim
MongeElkan
(
a
,
b
)
=
1
❘
"\[LeftBracketingBar]"
a
❘
"\[RightBracketingBar]"
∑
i
=
1
❘
"\[LeftBracketingBar]"
a
❘
"\[RightBracketingBar]"
max
j
=
1
,
…
,
❘
"\[LeftBracketingBar]"
b
❘
"\[RightBracketingBar]"
{
sim
inter
(
a
i
,
b
j
)
}
where |a| and |b| are the number of tokens contained in string a and b respectively, ai corresponds to the i th token in a, and b j corresponds the j th token in b.
11 . A method implemented via execution of computing instructions configured to run at one or more processors and configured to be stored at non-transitory computer-readable media, the method comprising:
receiving a first set of facet information corresponding to a first online marketplace, the first set of facet information comprising first facets of the first online marketplace;
receiving a second set of facet information corresponding to a second online marketplace, the second set of facet information comprising second facets for the second online marketplace;
analyzing the first set of facet information to reduce a number of the first facets as candidate facets;
building a machine learning model to map candidate facets to the second facets from the second set of facet information by determining, for each candidate facet, a facet candidate score indicative of a similarity between the candidate facets and the second facets, wherein the facet candidate scores are determined using one or more similarity functions comprising one or more of a n-gram Jaccard similarity based on tokenizing the candidate and second facets into tokens and constructing n-grams by splitting the tokens, an edit-distance based measure, a Monge-Elkan similarity that computes an average similarity between similar token pairs using an inter-token similarity measure, and a cosine similarity of word embeddings constructed from the candidate and second facets;
training the machine learning model to perform facet ranking on the candidate facets and second facets using the first set of facet information, the second set of facet information, and the mapping of the candidate facets to the second facets to a curated set of factors for the first online marketplace; and
displaying, on a navigation panel of a graphical user interface, a selected number of ranked facets as a curated set of facets for the first online marketplace, the navigation panel comprising one or more of the first facets and one or more of the second facets that are selected based on the facet ranking limited number of facet display positions.
12 . The method of claim 11 , wherein the first set of facet information comprises: a product catalog for the first marketplace, engagement data for the first marketplace, and facet rules for the first marketplace, and wherein the second set of facet information comprises a whitelist of facets for the second marketplace.
13 . The method of claim 11 , wherein analyzing the first set of facet information further comprises:
determining a respective stock keeping unit (SKU) coverage score for each facet in the first set of facet information; and
determining a respective cardinality value for each facet in the first set of facet information; and
determining a respective facet value dispersion for each facet in the first set of facet information.
14 . The method of claim 13 , wherein determining the respective SKU coverage score includes using an equation comprising:
score
coverage
(
f
)
=
number
of
SKUs
with
facet
f
total
number
of
SKUs
in
category
L
.
15 . The method of claim 13 , wherein determining the respective cardinality value includes using an equation comprising:
score
valueCard
(
f
)
=
{
0
if
c
f
<
min
Card
e
c
f
-
optCard
θ
2
if
min
Card
≤
c
f
≤
optCard
s
1
+
(
c
f
-
optCard
)
if
c
f
>
optCard
where c f corresponds to a number of facet values, optCard corresponds to an optimal number of facet values, minCard corresponds to a minimum number of facet values, and s corresponds to a scale constant that controls a slope of decay after optCard.
16 . The method of claim 13 , wherein determining the respective facet value dispersion includes using an equation comprising:
score
sizeDispersion
(
f
)
=
1
1
+
sizeVar
(
f
)
sizeVar
(
f
)
=
1
N
×
∑
i
=
1
N
(
x
i
-
x
_
)
2
x
_
where N corresponds to the number of facet values, x i corresponds to the i th facet value size (i.e., the number of products under the i th facet value) and x corresponds to the average facet value size.
17 . The method of claim 11 , wherein building the machine learning model further comprises:
determining the n-gram Jaccard similarity;
determining the edit-distance based measures;
determining the Monge-Elkan similarity;
determining the cosine similarity of word embeddings; and
determining the facet candidate score.
18 . The method of claim 17 , wherein the n-gram Jaccard similarity comprises a Bigram Jaccard similarity, and wherein determining the Bigram Jaccard similarity includes using an equation comprising:
sim
jaccard
(
S
a
,
S
b
)
=
❘
"\[LeftBracketingBar]"
S
a
⋂
S
b
❘
"\[RightBracketingBar]"
❘
"\[LeftBracketingBar]"
S
a
⋃
S
b
❘
"\[RightBracketingBar]"
=
❘
"\[LeftBracketingBar]"
S
a
⋂
S
b
❘
"\[RightBracketingBar]"
❘
"\[LeftBracketingBar]"
S
a
❘
"\[RightBracketingBar]"
+
❘
"\[LeftBracketingBar]"
S
b
❘
"\[RightBracketingBar]"
-
❘
"\[LeftBracketingBar]"
S
a
⋂
S
b
❘
"\[RightBracketingBar]"
where a corresponds to a facet in the first set of facet information, b corresponds to a facet in the second set of facet information, S a corresponds to a set of bigrams constructed from string a, and S b corresponds to a set of bigrams constructed from string b.
19 . The method of claim 17 , wherein determining the Monge-Elkan similarity includes using an equation comprising:
sim
MongeElkan
(
a
,
b
)
=
1
❘
"\[LeftBracketingBar]"
a
❘
"\[RightBracketingBar]"
∑
i
=
1
❘
"\[LeftBracketingBar]"
a
❘
"\[RightBracketingBar]"
max
j
=
1
,
…
,
❘
"\[LeftBracketingBar]"
b
❘
"\[RightBracketingBar]"
{
sim
inter
(
a
i
,
b
j
)
}
where |a| and |b| are the number of tokens contained in string a and b respectively, ai corresponds to the i th token in a, and b j corresponds the j th token in b.
20 . The system of claim 1 , wherein the machine learning model comprises one or more Sparse Group Lasso models trained on a utility score that is based on user engagement data and the one or more similarity functions.