Detailed Action
Notice of Pre-AIA or AIA status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Continued Examination under 37 C.F.R. § 1.114
A request for continued examination under 37 C.F.R. § 1.114, including the fee set forth in 37 C.F.R. § 1.17(e), was filed in this application after final rejection. Since this application is eligible for continued examination under 37 C.F.R. § 1.114, and the fee set forth in 37 C.F.R. § 1.17(e) has been timely paid, the finality of the previous Office action has been withdrawn pursuant to 37 C.F.R. § 1.114. Applicant’s submission filed on June 26, 2026 has been entered.
Response to Amendment
This Non-Final Office action is responsive to the Request for Continued Examination filed on June 26, 2026 (hereafter “Response”). The amendments to the claims are acknowledged and have been entered.
Claims 1, 8, and 9 are now amended.
Claims 7 and 15–17 are now canceled.
Claims 1–6 and 8–14 are pending in the application.
Response to Arguments
Claims 1–15 were previously rejected under 35 U.S.C. § 103 as being unpatentable over U.S. Patent Application Publication No. 2016/0205509 A1 (hereafter “Hopcraft”) in view of Haoyu Han et al., STGCN: A Spatial-Temporal Aware Graph Learning Method for POI Recommendation, 2020 IEEE International Conference on Data Mining (ICDM) (Nov. 17, 2020), available at https://doi.org/10.1109/ICDM50108.2020.00124 (hereafter “Han”), and claims 16 and 17 were further rejected under 35 U.S.C. § 103 over the combination of Hopcraft and Han with Chang Liu et al., Improving Location Recommendation with Urban Knowledge Graph (Nov. 1, 2021) https://doi.org/10.48550/arXiv.2111.01013 (hereafter “Liu”).
In order to comply with the rejection, the Applicant incorporated a narrower version of the subject matter of claim 7 into the independent claims, while traversing its substance. The Examiner does not fully agree with every point in the traversal, but in view of the additional time provided with this request for continued examination, along with principles of compact prosecution, the previous ground of rejection is hereby withdrawn in favor of a new ground of rejection under 35 U.S.C. § 103 that replaces one or more of the prior art references.
Since some of the references that the Applicant challenges in its response persist in this new ground of rejection, the Applicant’s arguments that remain relevant to those references will now be addressed.
Starting with pages 10–11 of the Response, neither Hopcraft nor Han are relied upon in the new ground of rejection to for what the Applicant contests, and therefore, those arguments are no longer relevant to the current ground of rejection. This also includes the “teaching away” argument for Han on page 12, as Han is only relied-upon in the rejections of the dependent claims to show why its teaching of the movement layer of an analogous graph renders the claims obvious in view of what Wu and Liu already teach with the same data. “The test for obviousness is not whether the features of a secondary reference may be bodily incorporated into the structure of the primary reference,” In re Keller, 642 F.2d 413, 425 (CCPA 1981), and thus, there is no need to show how a person of ordinary skill in the art would have incorporated the whole graph structure from Liu.
The Applicant’s argument concerning the Liu reference remains applicable to the current grounds of rejection asserted herein, so it must be addressed. As a reminder, both the previous rejection and the one here today rely on Liu to show why it was obvious to give “a greater weight to a movement between two nodes of the region mobility graph having a greater physical distance” in the model, before the effective filing date of the claimed invention.
The Applicant seems to admit that Liu teaches “that, when a user actually interacts with a distant POI rather than a nearby POI, such interaction may be evaluated as being more meaningful.” (Response 12). However, the Applicant argues that “the weight/score of Liu is closer to a bias-correction value in user-POI recommendation, and does not correspond to the weight assigned to movement between two nodes in the region mobility graph of the claimed invention.” (Response 12).
The Examiner respectfully disagrees. Liu’s counterfactual learning technique—which the Applicant correctly identifies as treating interactions with distant POIs as more meaningful—mathematically acts to upweight long-distance check-ins when user intent overcame spatial friction, by subtracting out the confounding effect of locality to check-ins. The fact that this technique helps with “bias-correction” is irrelevant to whether Liu teaches the limitation, and if anything, is evidence of a motivation to combine the references independent from hindsight.
Accordingly, Liu stands as part of the rejection, and since all of the claims are rejected, the Applicant’s request for a notice of allowance (Response 13) is respectfully denied.
Claim Rejections – 35 U.S.C. § 103
The following is a quotation of 35 U.S.C. § 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102 of this title, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
This application currently names joint inventors. In considering patentability of the claims the examiner presumes that the subject matter of the various claims was commonly owned at the time any inventions covered therein were effectively filed absent any evidence to the contrary. Applicant is advised of the obligation under 37 C.F.R. § 1.56 to point out the inventor and effective filing dates of each claim that was not commonly owned at the time a later invention was effectively filed in order for the examiner to consider the applicability of 35 U.S.C. § 102(b)(2)(C) for any potential 35 U.S.C. § 102(a)(2) prior art against the later invention.
I. Wu and Liu teach claims 1 and 9.
Claims 1 and 9 are rejected under 35 U.S.C. § 103 as being unpatentable over Ning Wu et al., Learning Effective Road Network Representations with Hierarchical Graph Neural Networks, KDD ‘20: Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining (August 20, 2020), available at https://doi.org/10.1145/3394486.3403043 (hereafter “Wu”) in view of Chang Liu et al., Improving Location Recommendation with Urban Knowledge Graph (Nov. 1, 2021), available at https://doi.org/10.48550/arXiv.2111.01013 (hereafter “Liu”).
Claim 1
Wu teaches a method for controlling a server, the method comprising:
obtaining road information in a region having a predetermined range and information on a plurality of places in the region;
Wu begins with a “road network” that “is characterized as a directed graph
G
=
S
,
A
S
⟩
, where
S
is a vertex set of
k
S
road segments and
A
S
∈
R
k
S
×
k
S
is the adjacency matrix.” Wu 8. Each of the road segments are “usually associated with some side features (e.g., longitude and latitude, segment type, and length).” Wu 8. Note that while some embodiments of Wu’s scheme define road segments as the vertices, others “define locations (e.g., POI or location cell) as vertices)” instead. Wu 8.
In practice, Wu populates the road network
G
by obtaining “corresponding road network information from open street map.” Wu 11.
obtaining a region mobility graph corresponding to the region,
Wu proposes creating a supplemented version of the road network called a “Hierarchical Road Network” (shown within the dashed line in Figure 1) and formalized as “
H
” in the equations. Wu 8.
based on movement information of at least one user between the plurality of places, the region mobility graph including a plurality of nodes corresponding to the plurality of places and an edge connecting the plurality of nodes;
The nodes of the Hierarchical Road Network are
V
=
S
∪
R
∪
Z
, and the edges include
E
=
A
S
,
A
R
,
A
Z
,
A
S
R
,
A
R
Z
. Per Definition 2, and as discussed above,
S
and
A
S
come from the road network graph
G
, with
S
providing the nodes for the POIs (or intersections, depending on the embodiment), and
A
S
providing the adjacency matrix (i.e., edges) that represent the road segments linking those vertices. Wu 8.
The Hierarchical Road Network is “based on” movement information of at least one user between the vertices of the as claimed, because
H
is supplemented with “real trajectory data for capturing functional characteristics.” Wu 10. Specifically, Wu collects “trajectory sequence data of real users, which is a time-ordered road segment sequence visited by a user,” and then Wu “construct[s] a road segment transition matrix
T
λ
∈
R
k
S
×
k
S
, in which entry
T
λ
s
i
,
s
j
indicates the frequency that si has reached sj with a step length λ in all trajectory sequences.” Wu 10.
wherein the edge connecting the plurality of nodes is identified based on at least one of information on a number of times of movement of the at least one user between the plurality of nodes
As mentioned above, Wu uses “real trajectory data for capturing functional characteristics.” Wu 10. “We collect trajectory sequence data of real users, which is a time-ordered road segment sequence visited by a user. In order to utilize the trajectory data, we construct a road segment transition matrix
T
λ
∈
R
k
S
×
k
S
, in which entry
T
λ
s
i
,
s
j
indicates the frequency that si has reached sj with a step length λ in all trajectory sequences.” Wu 10.
or information on a physical distance between the plurality of nodes;
Additionally, each road segment
s
∈
S
“is usually associated with some side features,” one of which is “length.” Wu 7–8.
generating a learned region mobility graph, by using a graph convolutional network (GCN) model, the GCN model being configured to learn the region mobility graph based on the plurality of nodes, feature information included in the plurality of nodes, and edges connecting the plurality of nodes,
Before discussing how Wu uses a GCN to generate a learned version of the Hierarchical Road Network, it is necessary to discuss the other nodes in the Hierarchical Road Network, namely,
R
and
Z
. As shown in Figure 1 and Definitions 3 and 4, the Hierarchical Road Network organizes the road segments s into respective regions
r
∈
R
, and organizes each of those regions r into respective functional zones
z
∈
Z
. Wu 8. To represent this hierarchy as a flat graph, Wu provides edges
A
S
R
connecting each road segment s to its respective structural region r, and likewise provides edges
A
R
Z
connecting each of the regions r to a respective functional zone z.
With this structure in place Wu uses “a standard Graph Convolutional Network (GCN) to update the zone embeddings
N
Z
t
+
1
=
GCN
N
Z
t
,
A
Z
.” Wu 10. “Then, it sends the zone embeddings to the next level for updating region embeddings,” which updates its own embedding representations with the GCN, according to
N
S
t
+
1
=
GCN
N
~
S
t
,
A
S
. Wu 10.
and predict a relationship between the plurality of nodes in the region mobility graph;
The embeddings learned using the GCN are then used to predict downstream relationships. Wu 10 (section 4.5).
and providing the learned region mobility graph to an external apparatus,
Once the learned hierarchical road network is created, it may be provided to a computer responsible for performing one of the five tasks discussed on page 11 of Wu’s disclosure.
wherein the generating the learned region mobility graph comprises obtaining an embedding vector for predicting a number of times of movement of the at least one user to a node corresponding to a specific category of a place between the plurality of nodes by using the GCN model,
As discussed above, Wu groups nodes into “functional zones” (specific categories of places), and the model obtains (among other things), a matrix of embedding vectors NZ with representations for each of those functional zones. See Wu 9. Wu then uses NZ to fit the segment representations based on the zone representations via equation (18), i.e.,
N
^
S
=
A
S
R
A
R
Z
N
Z
, and finally, use the segment representations to predict the connectivity matrix with
C
^
=
N
^
S
N
^
S
⊤
. Wu 10. Notice that
C
^
predicts the number of times of movement because it is a predicted version of the original connectivity matrix C, which “considers the connectivity in terms of both road network structure and human moving behaviors.” Wu 10. More precisely, per Equation (17), C has the sum of each T(j) from 1 to
λ
, where each “entry T(λ)[si,sj] indicates the frequency that si has reached sj with a step length λ in all trajectory sequences.” Wu 10.
To summarize, NZ is an embedding vector that is used to make a prediction
C
^
of C, and C is a matrix of tallied trajectories between the different nodes of the graph.
and determining an edge between the plurality of nodes based on the obtained embedding vector,
Determining
C
^
also determines an edge between nodes based on the embedding vector (i.e., NZ as discussed above), because
C
^
is a prediction of C, which is defined as
C
∈
R
k
S
×
k
S
. Wu 10.
Wu does not explicitly disclose “giving a greater weight to a movement between two nodes of the region mobility graph having a greater physical distance.”
Liu, however, teaches a method that includes obtaining an embedding vector from a graph (which, as shown in Figure 2, similarly includes nodes for places of interest, nodes for categories of places of interest, and edges connecting the nodes), using a GCN model, and, wherein the obtaining the embedding vector comprises:
obtaining the embedding vector by using the GCN model configured to learn the region mobility graph by giving a greater weight to a movement between two nodes of the region mobility graph having a greater physical distance.
As shown in Figure 2 (page 4), Liu proposes obtaining embeddings
u
i
from a graph describing both the geography and functions of POIs in the real world, but then weighting them using a counterfactual learning model, to “eliminate the geographical bias” in the initial embeddings that result from users biasing POIs that are nearby, even if those POIs would not ordinarily be their first choice. Liu 5. This has the effect of giving greater weight to long-range movements, because users who travel a greater physical distance to reach a specific functional zone yield a higher score in the model as a bonus for their effort.
Specifically, using equation (16), Liu calculates a debiased interaction prediction score between user
u
i
and POI
p
j
as the difference between the total effect (“TE”) (expressed as
Y
u
i
,
p
j
,
g
j
) and the natural direct effect (“NDE”) (expressed as
Y
u
i
,
p
*
j
,
g
j
) of the geography of a POI on the probability that a user will interact with a POI. By subtracting the second part, the model effectively removes the “bonus points” a POI gets just for being close to the user, such that a very distant POI will have a very small value for
Y
u
i
,
p
*
j
,
g
j
, and yield a relatively higher final score if the user actually interacts with it, as opposed to a closer POI of the same nature.
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to improve Hopcraft and Han’s combined GCN model with Liu’s counterfactual learning model, so as to give greater weight to POIs that are further from a user in the embeddings. There would have been a reasonable expectation of success in the combination because Liu explicitly instructs the person of ordinary skill to apply the counterfactual learning model to a GCN layer. And such a person would have been motivated to combine Liu with Hopcraft because the close proximity of POIs in a model sometimes “serves as a confounder” of the users’ true intentions or desires, which “will not only make it difficult to accurately match users’ interests but drastically damage the performance of recommender systems.” Liu 1.
Claim 9
Claim 9 recites a server with a memory and processor that performs exactly the same server control method set forth in claim 1. Therefore, claim 9 is rejected over the same findings and rationale as provided above for those claims.
II. Wu, Liu, and Han teach claims 2–6, 8, and 10–14.
Claims 2–6, 8, and 10–14 are rejected under 35 U.S.C. § 103 as being unpatentable over Wu in view of Liu as applied to claims 1 and 9 above, and further in view of Haoyu Han et al., STGCN: A Spatial-Temporal Aware Graph Learning Method for POI Recommendation, 2020 IEEE International Conference on Data Mining (ICDM) (Nov. 17, 2020), available at https://doi.org/10.1109/ICDM50108.2020.00124 (hereafter “Han”).
Claim 2
Wu teaches the method according to claim 1, wherein obtaining the region mobility graph comprises:
obtaining a place mobility graph corresponding to the region, the place mobility graph including a plurality of first nodes corresponding to the plurality of places, respectively, and a first edge connecting the plurality of nodes,
The hierarchical road network further includes functional zones
z
∈
Z
, the nodes z of which describe respective places “e.g., shopping areas and transportation hubs,” and are connected to one another with edges
A
z
∈
R
k
z
×
k
z
. Wu 8. “A functional zone is constructed on top of functionally related structural regions.” Wu 9.
In Wu, “movement information of the at least one user between the plurality of places” is eventually added to the functional zone graph, but only for the learned region mobility graph. See Wu 10 (section 4.3.2). In contrast, claim 2 requires the place mobility graph to have the movement information when it is obtained for the original region mobility graph. (It could also be argued that claim 2 differs from Wu if the “places” are required to be points of interest rather than regions, but the language of claim 2 is not yet that specific).
and obtaining a road network graph including a plurality of second nodes corresponding to intersections and a second edge connecting the plurality of second nodes, based on the road information in the region.
Meanwhile, Wu begins with a “road network” that “is characterized as a directed graph
G
=
S
,
A
S
⟩
, where
S
is a vertex set,” and where”
A
S
∈
R
k
S
×
k
S
is the adjacency matrix.” Wu 8. “Each entry
A
S
s
i
,
s
j
is a binary value indicating whether there exists a directed link from road segment si to road segment sj.” Wu 8.
Wu thus lacks the place mobility graph identical to the one described in claim 2 prior to the learning phase of the claim. Liu similarly lacks such a graph.
Han, however, teaches a method comprising:
obtaining a place mobility graph corresponding to the region, the place mobility graph including a plurality of first nodes corresponding to the plurality of places, respectively, and a first edge connecting the plurality of nodes, the place mobility graph further including the movement information of the at least one user between the plurality of places;
“To fuse the spatial-temporal context information, we propose the user record multigraph, as shown in Figure 3,” Han 1053, which includes a plurality of POI nodes
P
(e.g.,
p
i
and
p
j
), and a plurality of edges connecting the POI nodes, e.g., “edge
e
i
,
j
connects
p
i
and
p
j
if users visit POI
p
i
and POI
p
j
during a time period.” Han 1054.
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to include Han’s POI-POI graph in Wu’s model, either as its own hierarchical level, or encoded in the node and edge information of Wu’s existing road network graph
G
—the latter of which Wu already acknowledged was possible by teaching that the road segments are “usually associated with some side features.” Wu 8. One would have been motivated to enhance Wu’s model with this data from Han because “different users may prefer the same POI in different time periods.” Han Abstract.
Claim 3
Wu, Liu, and Han teach the method according to claim 2, and Wu further teaches that the obtaining the region mobility graph comprises:
identifying, as a node of the region mobility graph, an area defined by the plurality of second nodes and the second edge of the road network graph;
Wu’s hierarchical road network further includes structural regions
r
∈
R
, whose adjacency matrix ASR associates a plurality of road segments from
S
that each have edges connecting one another. Wu 8.
and identifying, as an edge of the region mobility graph,
The edges of the hierarchical road network include the whole set of edges in the road network graph, AS, and those edges may be “associated with some side features (e.g., longitude and latitude, segment type, and length).” Wu 8.
As discussed in the rejections of the prior claims, Wu further teaches identifying movement information between the nodes, but this information is only used in the learned version of the region mobility graph, not the pre-learned version.
Han, however, teaches obtaining an original, pre-learned graph via the following steps:
identifying, as a node of the region mobility graph, an area defined by the plurality of second nodes and the second edge of the road network graph;
In Han’s graph, a POI-Region layer provides a sub-graph with edges connecting POI nodes to Region nodes when a POI is in a given Region. Han 1054.
and identifying, as an edge of the region mobility graph, movement information of the at least one user between a plurality of identified nodes and information on a physical distance between the plurality of identified nodes.
The graph further includes a POI-POI layer that provides a subgraph where an In the multigraph, an “edge
e
i
,
j
connect[s] POI nodes
p
i
and
p
j
if users visit POI
p
i
. The POI-POI edges point the proximity between the POIs.” Han 1054.
Claim 4
Wu, Liu, and Han teach the method according to claim 3, wherein,
in the region mobility graph, each of the plurality of identified nodes comprises information on a place, with respect to at least one place located in an area corresponding to each of the plurality of nodes, as the feature information, and wherein the information on the place comprises location information of the place and category information corresponding to the place.
Each of the POIs in
P
include both information about the POI itself, as well as its location. Han 1053. In some instances—e.g., when using the Gowalla dataset—the POIs include data about “POI, time, and POI location,” as well as “seven categories of POIs.” Han 1055.
Claim 5
Claim 5 is a broader version of claim 3 in which all of the limitations are the same, but without the requirement of identifying “an area defined by” the plurality of second nodes; it simply requires us to identify the second nodes. Therefore, claim 5 is rejected over all of the same findings and rationale as provided in the rejection of claim 3, above.
Claim 6
Wu, Liu, and Han teach the method according to claim 3,
wherein in the region mobility graph, each of the plurality of identified nodes comprises information on a place, with respect to at least one place related to a location corresponding to each of the plurality of identified nodes, as the feature information.
Each of the POIs in
P
include both information about the POI itself, as well as its location. Han 1053. In some instances—e.g., when using the Gowalla dataset—the POIs include data about “POI, time, and POI location,” as well as “seven categories of POIs.” Han 1055.
Claim 8
Wu, Liu, and Han teach the method according to claim 1, and in Wu, physical distance is available in the region mobility graph, but not necessarily part of the process for obtaining the embedding vectors.
Liu, however, teaches a method for learning a region mobility graph that involves obtaining embedding vectors for the graph,
wherein the obtaining the embedding vector comprises obtaining the embedding vector by using the GCN model configured to learn the region mobility graph by giving a weight to the physical distance between the plurality of nodes of the region mobility graph.
As shown in Figure 2 (page 4), Liu proposes obtaining embeddings
u
i
from a graph describing both the geography and functions of POIs in the real world, but then weighting them using a counterfactual learning model, to “eliminate the geographical bias” in the initial embeddings that result from users biasing POIs that are nearby, even if those POIs would not ordinarily be their first choice. Liu 5. This has the effect of giving greater weight to long-range movements, because users who travel a greater physical distance to reach a specific functional zone yield a higher score in the model as a bonus for their effort.
Specifically, using equation (16), Liu calculates a debiased interaction prediction score between user
u
i
and POI
p
j
as the difference between the total effect (“TE”) (expressed as
Y
u
i
,
p
j
,
g
j
) and the natural direct effect (“NDE”) (expressed as
Y
u
i
,
p
*
j
,
g
j
) of the geography of a POI on the probability that a user will interact with a POI. By subtracting the second part, the model effectively removes the “bonus points” a POI gets just for being close to the user, such that a very distant POI will have a very small value for
Y
u
i
,
p
*
j
,
g
j
, and yield a relatively higher final score if the user actually interacts with it, as opposed to a closer POI of the same nature.
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to improve Hopcraft and Han’s combined GCN model with Liu’s counterfactual learning model, so as to give greater weight to POIs that are further from a user in the embeddings. There would have been a reasonable expectation of success in the combination because Liu explicitly instructs the person of ordinary skill to apply the counterfactual learning model to a GCN layer. And such a person would have been motivated to combine Liu with Hopcraft because the close proximity of POIs in a model sometimes “serves as a confounder” of the users’ true intentions or desires, which “will not only make it difficult to accurately match users’ interests but drastically damage the performance of recommender systems.” Liu 1.
Claims 10–14
Claims 10–14 recite a server with a memory and processor that performs exactly the same server control method set forth in corresponding claims 2–6 and 8. Therefore, claims 10–14 are rejected over the same findings and rationale as provided above for those claims.
Conclusion
Any inquiry concerning this communication or earlier communications from the examiner should be directed to Justin R. Blaufeld whose telephone number is (571)272-4372. The examiner can normally be reached M-F 9:00am - 4:00pm ET.
Examiner interviews are available via telephone, in-person, and video conferencing using a USPTO supplied web-based collaboration tool. To schedule an interview, applicant is encouraged to use the USPTO Automated Interview Request (AIR) at http://www.uspto.gov/interviewpractice.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, James K Trujillo can be reached at (571) 272-3677. The fax phone number for the organization where this application or proceeding is assigned is 571-273-8300.
Information regarding the status of published or unpublished applications may be obtained from Patent Center. Unpublished application information in Patent Center is available to registered users. To file and manage patent submissions in Patent Center, visit: https://patentcenter.uspto.gov. Visit https://www.uspto.gov/patents/apply/patent-center for more information about Patent Center and https://www.uspto.gov/patents/docx for information about filing in DOCX format. For additional questions, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free). If you would like assistance from a USPTO Customer Service Representative, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000.
Justin R. Blaufeld
Primary Examiner
Art Unit 2151
/Justin R. Blaufeld/Primary Examiner, Art Unit 2151
Prosecution Insights