CN117573986A - Interest point recommendation method based on sequential and ground understanding coupling characterization - Google Patents
Interest point recommendation method based on sequential and ground understanding coupling characterization Download PDFInfo
- Publication number
- CN117573986A CN117573986A CN202410061576.2A CN202410061576A CN117573986A CN 117573986 A CN117573986 A CN 117573986A CN 202410061576 A CN202410061576 A CN 202410061576A CN 117573986 A CN117573986 A CN 117573986A
- Authority
- CN
- China
- Prior art keywords
- interest
- representation
- geographic
- user
- points
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 238000000034 method Methods 0.000 title claims abstract description 65
- 238000012512 characterization method Methods 0.000 title claims abstract description 33
- 230000008878 coupling Effects 0.000 title claims abstract description 12
- 238000010168 coupling process Methods 0.000 title claims abstract description 12
- 238000005859 coupling reaction Methods 0.000 title claims abstract description 12
- 238000010586 diagram Methods 0.000 claims abstract description 54
- 230000007246 mechanism Effects 0.000 claims abstract description 8
- 230000006870 function Effects 0.000 claims description 32
- 238000013528 artificial neural network Methods 0.000 claims description 22
- 230000000694 effects Effects 0.000 claims description 13
- 230000003993 interaction Effects 0.000 claims description 11
- 239000003795 chemical substances by application Substances 0.000 claims description 9
- 239000013598 vector Substances 0.000 claims description 9
- 238000010276 construction Methods 0.000 claims description 7
- 230000009466 transformation Effects 0.000 claims description 7
- 230000002776 aggregation Effects 0.000 claims description 6
- 238000004220 aggregation Methods 0.000 claims description 6
- 239000011159 matrix material Substances 0.000 claims description 6
- 230000008569 process Effects 0.000 claims description 6
- 230000006399 behavior Effects 0.000 claims description 5
- 238000011176 pooling Methods 0.000 claims description 4
- 238000012549 training Methods 0.000 claims description 4
- 230000004931 aggregating effect Effects 0.000 claims description 3
- 238000004364 calculation method Methods 0.000 claims description 3
- 230000000644 propagated effect Effects 0.000 claims description 3
- 238000013461 design Methods 0.000 claims description 2
- 238000004458 analytical method Methods 0.000 abstract 1
- 238000011161 development Methods 0.000 description 2
- 238000012935 Averaging Methods 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 230000008901 benefit Effects 0.000 description 1
- 235000013410 fast food Nutrition 0.000 description 1
- 235000013305 food Nutrition 0.000 description 1
- 230000007774 longterm Effects 0.000 description 1
- 238000000844 transformation Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/90—Details of database functions independent of the retrieved data types
- G06F16/95—Retrieval from the web
- G06F16/953—Querying, e.g. by the use of web search engines
- G06F16/9535—Search customisation based on user profiles and personalisation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/90—Details of database functions independent of the retrieved data types
- G06F16/95—Retrieval from the web
- G06F16/953—Querying, e.g. by the use of web search engines
- G06F16/9536—Search customisation based on social or collaborative filtering
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/044—Recurrent networks, e.g. Hopfield networks
- G06N3/0442—Recurrent networks, e.g. Hopfield networks characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/048—Activation functions
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q30/00—Commerce
- G06Q30/02—Marketing; Price estimation or determination; Fundraising
- G06Q30/0241—Advertisements
- G06Q30/0251—Targeted advertisements
- G06Q30/0259—Targeted advertisements based on store location
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q50/00—Systems or methods specially adapted for specific business sectors, e.g. utilities or tourism
- G06Q50/10—Services
- G06Q50/12—Hotels or restaurants
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q50/00—Systems or methods specially adapted for specific business sectors, e.g. utilities or tourism
- G06Q50/10—Services
- G06Q50/14—Travel agencies
Abstract
The invention provides an interest point recommendation method based on sequential and ground understanding coupling characterization, which aims at the problems that the prior interest point recommendation method is used for carrying out algorithm modeling by coupling the sequence influence and the geographic influence of an interest point, and the independent modeling analysis on the dominant influence is absent, so that the recommendation result is suboptimal and poor in interpretation. Specifically, the method and the system respectively construct the sequence diagram and the geographic diagram according to the sign-in sequence of the user to carry out decoupling characterization, customize the propagation scheme of the sequence diagram and the geographic diagram to capture the sequence influence and the geographic influence respectively, introduce a soft attention mechanism, and finally construct the interest point click prediction model, so that the problems of cold start and recommendation accuracy of the interest point recommendation are better solved, meanwhile, the algorithm interpretability is improved, and the optimal interest point is recommended to the user.
Description
Technical Field
The invention relates to prediction of a travel recommendation system field based on sequence and geographic interest point recommendation, which solves the problem of a field recommendation industry prediction model through sequence and geographic dependence.
Background
The rapid development of location-based social networks has prompted the development of point-of-interest recommendations, which are central to a variety of location-based services, including location-based advertising and online food distribution, which are the mainstay of numerous popular applications such as Foursquare, gowalla and meiuan, which can be restaurants, hotels, attractions, or any item with a location tag. In point of interest recommendations, location is not just a common feature dimension: it constructs a network structure between points of interest, interconnecting them by their interaction history sequence and geographic affinity.
Recent point of interest recommendation efforts have focused on combining geographical impact with sequential transformations of user access to point of interest data, sequence-based methods like LSTPM have combined long-term interests of users with short-term interests of geographical expansion, using geographical affinities as an aid to sequence behavior modeling, geoIE explicitly uses physical distances to model user-specific geographical impact, GTAG performs preference propagation based on physical distances and slot affinities, although these point of interest recommenders have versatility, few methods can explicitly reveal collaboration signals, complex interactions between users and points of interest can be easily depicted graphically, and thus one promising approach is to utilize rich graph structures to utilize higher order connections between points of interest.
Graph neural networks have been widely deployed for recommendations to better capture similarities between higher-order neighbors, NGCF and LightGCN collect collaborative signals on user-point of interest bipartite graphs, SR-GNN propagates along the user's interaction sequence graph, while these paradigms provide valuable insight into organizing interaction data into graph structures of point of interest recommendations, they do not take advantage of geographical relationships, which is a prominent feature of point of interest recommendations.
In fact, the sequential impact and geographic impact should be considered equivalent counterparts, which are the two primary driving forces recommended by points of interest, and user access to items is affected by the geography and neighboring nodes in the sequence diagram. In one aspect, users tend to revisit familiar points of interest that have appeared in the interaction sequence. For example, frequent visits to clothing stores may reveal user preferences for fashion, indicating that similar clothing stores are more likely to be points of interest for the next visit, and on the other hand, nearby points of interest are more likely to be visited. For example, a user who has just come out of a mall shopping mall is more likely to eat lunch at a nearby fast food restaurant than his/her favorite rural restaurant. These two effects affect user behavior in a substantially different manner, and previous point of interest recommendation methods have not attempted to distinguish, since these methods do not impose explicit supervision on learned user preferences, there is no guarantee that sequential effects and geographic effects in user preferences are explicitly captured.
Disclosure of Invention
Aiming at the limitations and challenges of the existing solution, the invention provides an interest point recommendation method based on sequential and ground understanding coupling characterization. Specifically, the method and the system respectively construct a sequence diagram and a geographic diagram according to the sign-in sequence of the user to carry out decoupling characterization, customize a propagation scheme of the sequence diagram and the geographic diagram to capture sequence influence and geographic influence respectively, introduce a soft attention mechanism, and finally construct an interest point click prediction model.
The invention discloses a method for recommending interest points based on sequential and ground understanding coupling characterization, which comprises the following steps:
s1: based on the user and interest point interaction history record, respectively constructing a sequence chart and a geographic chart to recommend the interest points, specifically, using geographic information in the user and interest point interaction history record to represent the user set asThe set of points of interest is denoted +.>Where U represents each user, |U| represents the number of users, V represents each point of interest, |V| represents the number of points of interest; every interest point->Through its coordinates%) The tuples are geocoded, i.e.)>For each user->The access sequence of the interest points is expressed as +.>Given user +.>And target interest point->The goal of the point of interest recommendation is to predict +.>Access->Is expressed as +.>Wherein->Is a learning function parameterized as +.>；
Constructing a sequence chart: given userIs>Then construct the directed sequence diagram +.>For each edge->Indicating that the user is +.>Middle visit->After which access is continued +.>Thus, the sequence diagram->Collecting sequence information from a history record of a user access interest point to perform decoupling characterization of sequence influence;
constructing a geographic diagram: given the point of interest location information, then construct an undirected geographic mapUndirected edgeRepresentation->And->The distance between them is at a specific distance threshold +.>In, edge weight matrix->Is interest point->And->Geographic distance between, thus, geographic map ∈ ->Performing decoupling characterization of geographic influence from the point of interest location information;
s2: customizing the propagation process of the sequence diagram and the geographic diagram, constructing a diagram propagation layer to respectively carry out decoupling characterization on the sequence diagram and the geographic diagram, and outputting two groups of interest point representations with sequence and geographic information so as to better utilize the inherent characteristics of each diagram;
s3: after decoupling characterization is carried out on the sequence diagram and the geographic diagram according to the previous interest point access history of the user, the soft attention mechanism is introduced, and a sequence embedded representation and a geographic embedded representation of the interest points are respectively generated;
s4: the self-supervision decoupling process comprises the steps of firstly, acting on a graph propagation layer through two reading functions to obtain output, and then designing a projection head structure contrast loss for each decoupling characterization;
s5: and constructing a click rate prediction model based on the interest points, and performing model prediction and loss function construction.
According to a specific implementation manner of the embodiment of the present invention, the specific steps of S2 are:
s21, a sequence diagram propagation layer: given target userAnd sequence diagram of construction->，/>From the connection matrix->Determining, applying a gated neural network +.>And decoupling characterization embedding +.>Is updated as:
1) The interest points represent:the method comprises the steps of carrying out a first treatment on the surface of the Wherein->The representation is decoupled from the token embedding,
2) The point of interest representation is then built as an input state:
；
wherein the method comprises the steps ofThe connection matrix, b is the parameter,
3) And then input it to the update gate for controlling the last memory information to continue to hold the amount of data to the current time:
；
wherein:is a parameter that can be trained and is,
4) After updating the gate and then determining that the network state at the previous moment is multiplied by the output of the reset gate through the reset gate, the network state is used as a parameter to calculate the candidate state at the current moment:
；
wherein the method comprises the steps ofIs a parameter that can be trained and is,
5) Candidate states are calculated after the update gate and the reset gate, and the formula is:
；
wherein:is a reset gate->The parameters may be a function of the training parameters,
6) Finally, the final state of the current time step is obtained:
；
wherein:is update gate->A current memory content;
designing a gating neural network to sequentially perform on all nodes of the sequence diagram, wherein hidden states are updated by the updating gatesAnd reset gate->Control, the sequential effect behind the user behavior is propagated by gating +.>Simulating to obtain the output of a hidden layer as the embedded representation of the sequential decoupling representation of the interest point:
；
s22, a geographical map propagation layer: 1) Message construction function: given a location-based graphConstructing a geographical map propagation layer using a messaging scheme of a map neural network, and then constructing a message, the +.>A pair of adjacent interest points on the layer->And->The message is expressed as:
；
wherein the method comprises the steps ofIs a message coding function, obtains interest point representation from the graph neural network of the previous layer>I.e. the input of the first layer is initialized with the embedded X of interest points:
；
message functionRepresenting the influence of the distance between the points of interest +.>Is realized by:
；
wherein the method comprises the steps ofIs a trainable weight matrix, as a lower layer representation a linear transformation taking into account distance effects; distance core->Along with->Exponentially decaying->And->The distance between them increases, using the graph Laplace norm of the graph roll-up neural network, where +.>And->Respectively indicate->Go up->And->Is the first hop neighbor of (a);
2) Message aggregation function: generating a node representation of each graph neural network layer by defining a message aggregation function:
；
wherein the method comprises the steps ofBy->Is derived from the representation of (a);
after propagation along the L-layer graph neural network, the hidden representation of the L-th layer is adopted as the geocoding of all interest points in V:
；
for the userSign-in history of->By aggregating users->Higher order geographic neighbors of points of interest in the point of interest access history to capture geographic preferences, user's geocode +.>Is->The geographical decoupling representation of all points of interest appearing in the list is embedded.
According to a specific implementation manner of the embodiment of the present invention, the specific steps of S3 are:
s31, after decoupling characterization is carried out on the sequence diagram and the geographic diagram according to the previous interest point access history of the user, a soft attention mechanism is introduced, and according to the current interest pointAggregate coding, coding according to a given geographical map +.>Will->Hidden representation +.>Generating a geo-embedded representation as a query vector>：
；
；
Wherein the method comprises the steps ofRepresenting sigmoid function->Representing attention vectors, queries and key matrices +.>Is a trainable parameter for +.>Will->Is +.>Generating a sequential embedded representation as a query vector>：
；
。
According to a specific implementation manner of the embodiment of the present invention, the specific steps of S4 are:
s41 willAnd->The two representation forms are separated from each other, the order effect and the geographic effect are fully utilized to predict the click rate of interest points, the two embedded agents are obtained, the average pooling is selected as a reading function, and the average pooling is acted on a graph propagation layer to obtain output so as to generate ∈>And->Is->And->：
；
；
Wherein the method comprises the steps ofAnd->Representing a geographical map and a sequence map, < >>Carrying geographical preference information about points of interest of the user,carrying sequential preference information about user points of interest, < >>Representing a set of one-hop geographical neighbors,>an access order representation set representing points of interest thereof;
s42, before sending the embedding and its proxy to generate the self-supervising signals, designing a projection head for each decoupling token, the projection representation of which is calculated as follows:
；
；
wherein the method comprises the steps ofIs a linear transformation with trainable parameters, the projection head projects the embedding into another potential space, obtaining information from different aspects;
representing the projection poolAnd->Regarded as->And->I.e. the two agents act as positive samples of the respective embeddings, while acting as negative samples of each other, the design contrast loss function is:
；
wherein the method comprises the steps ofThe Bayesian personalized ranking loss is represented by the following formula:
；
wherein the method comprises the steps ofRepresenting the inner product of two given representations.
According to a specific implementation manner of the embodiment of the present invention, the specific steps of S5 are:
s51, model prediction: at the time of obtaining user access historyAnd->After embedding the representation of the decoupling representation, embedding the decoupling representation with the target +.>Is embedded->And geographical characterization->In connection, a 2-layer multi-layer perceptron is applied to predict +.>Click rate of->，/>The calculation formula of (2) is as follows:
；
s52, objective function and course learning: given a label y, carrying out supervised interest point click rate prediction by adopting binary cross entropy loss, wherein the formula is as follows:
；
after adding decoupling to the model, the overall model loss is obtained:
；
wherein the method comprises the steps ofIs a decoupled weight, training the model by dynamically increasing the contrast loss weight pre-heating program:
；
wherein the method comprises the steps ofAnd->Is a superparameter->Representing the current course number.
Drawings
Fig. 1 is a flow chart of the method.
Detailed Description
The present invention will be described in further detail below with reference to examples and drawings in order to facilitate the understanding and practice of the present invention by those of ordinary skill in the art.
As shown in fig. 1, a method for recommending interest points based on sequential and well-understood coupling characterization comprises the following steps:
step 1:
based on the interaction history of the user and the interest points, respectively constructing a sequence diagram and a geographic diagram to recommend the interest points, specificallyIn other words, the user set is represented as the geographic information in the interaction history of the user and the interest pointThe set of points of interest is denoted +.>Wherein each interest point->Through its coordinates%) The tuples are geocoded, i.e.)>For each user->The access sequence of the interest points is expressed as +.>Given user +.>And target interest point->The goal of the point of interest recommendation is to predict +.>Access->Is expressed as +.>Wherein->Is a learning function parameterized as +.>；
Constructing a sequence chart: given userIs>Then construct the directed sequence diagram +.>For each edge->Indicating that the user is +.>Middle visit->After which access is continued +.>Thus, the sequence diagram->Collecting sequence information from a history record of a user access interest point to perform decoupling characterization of sequence influence;
constructing a geographic diagram: given the point of interest location information, then construct an undirected geographic mapUndirected edgeRepresentation->And->The distance between them is at a specific distance threshold +.>In, edge weight matrix->Is interest point->And->Geographic distance between, thus, geographic map ∈ ->Performing decoupling characterization of geographic influence from the point of interest location information;
step 2: customizing the propagation process of the sequence diagram and the geographic diagram, respectively carrying out decoupling characterization on the sequence diagram and the geographic diagram by a construction diagram propagation layer, outputting two groups of interest point representations with sequence and geographic information, and better utilizing the inherent characteristics of each diagram:
sequence diagram propagation layer: given target userAnd sequence diagram of construction->，/>From the connection matrix->Determining, applying a gated neural network +.>And decoupling characterization embedding +.>Is updated as:
1) The interest points represent:the method comprises the steps of carrying out a first treatment on the surface of the Wherein->Representing the sign of the embedded condition,
2) The point of interest representation is then built as an input state:
；
wherein the method comprises the steps ofThe connection matrix, b is the parameter,
3) It is then input to an update gate to control how much history state remains in the output state at the current time:
；
wherein:is a parameter that can be trained and is,
4) After updating the gate and then determining that the network state at the previous moment is multiplied by the output of the reset gate through the reset gate, the network state is used as a parameter to calculate the candidate state at the current moment:
；
wherein the method comprises the steps ofIs a parameter that can be trained and is,
5) Candidate states are calculated after the update gate and the reset gate, and the formula is:
；
wherein:is a reset gate which is configured to reset the gate,
6) Finally, the final state of the current time step is obtained:
；
wherein:is update gate->A current memory content; a current memory content;
inspired by a gate circulation unit, the gate control neural network sequentially performs on all nodes of the sequence diagram, and the hidden state is updated by a gateAnd reset gate->Control, the sequential effect behind the user behavior is propagated by gating +.>Simulating to obtain the output of a hidden layer as the embedded representation of the sequential decoupling representation of the interest point:
；
s22, a geographical map propagation layer: 1) Message construction function: given a location-based graphConstructing a geographical map propagation layer using a messaging scheme of a map neural network, and then constructing a message, the +.>A pair of adjacent interest points on the layer->And->The message is expressed as:
；
wherein the method comprises the steps ofIs a message coding function, obtains interest point representation from the graph neural network of the previous layer>I.e. the input of the first layer is initialized with the embedded X of interest points:
；
message functionRepresenting the influence of the distance between points of interest, in particular, points of interest that are closer tend to have more similarities, formally, ++>Is realized by:；
wherein the method comprises the steps ofIs a trainable weight matrix, as a lower layer representation a linear transformation taking into account distance effects; distance core->Along with->Exponentially decaying->And->The distance between them increases, using the graph Laplace norm of the graph roll-up neural network, where +.>And->Respectively indicate->Go up->And->Is the first hop neighbor of (a);
2) Message aggregation function: to generate a node representation of each graph neural network layer, the message aggregation function is defined as:
wherein->By->Is derived from the representation of (a);
after propagation along the L-layer graph neural network, the hidden representation of the L-th layer is adopted as the geocoding of all interest points in V:
；
for the userSign-in history of->By aggregating users->Higher order geographic neighbors of points of interest in the point of interest access history to capture geographic preferences, user's geocode +.>Is->The geographical decoupling representation of all points of interest appearing in the list is embedded.
Step 3: after decoupling characterization is carried out on the sequence diagram and the geographic diagram according to the previous interest point access history of the user, the soft attention mechanism is introduced, and a sequence embedded representation and a geographic embedded representation of the interest points are respectively generated:
after decoupling characterization is carried out on the two graphs according to the previous interest point access history of the user, a soft attention mechanism is introduced, and according to the current interest pointAggregate coding, coding according to a given geographical map +.>Will->Hidden representation +.>Generating a geo-embedded representation as a query vector>：
；
；
Wherein the method comprises the steps ofRepresenting sigmoid function->Representing attention vectors, queries and key matrices +.>Is a trainable parameter for +.>Will->Is +.>Generating a sequential embedded representation as a query vector>：
；
；
Step 4: the self-supervision decoupling process comprises the steps of firstly, acting on a graph propagation layer through two reading functions to obtain output, and then designing a projection head structure contrast loss for each decoupling representation:
since the sequence influence and the geographic influence have different influences on the interest points of the next access, the method is toAnd->The two representation forms are separated from each other, the click rate prediction is carried out by fully utilizing the information, the two embedded agents are obtained under the inspired by the prior contrast learning work, namely, the two embedded agents are acted on a graph propagation layer through two reading functions to obtain output, and the average value is selected to be pooled as the reading function so as to generate/>And->Is a proxy for the two agents:
；
；
respectively toIs a one-hop geographical neighbor->And->Averaging is performed due to ∈ ->And->Is read from two different graphs, so that each agent carries representative information of its respective graph characteristics, in particular +.>Carrying rich information about the geographical preferences of the user, whereas +.>Carrying information about the historical interests of the user;
before sending the embedding and its proxy to generate the self-supervising signals, a projection head is designed for each decoupling representation, the projection representation of which is calculated as follows:
；
；
wherein the method comprises the steps ofIs a linear transformation with trainable parameters, the projection head projects the embedding into another potential space, obtaining information from different aspects;
representing the projection poolAnd->Regarded as->And->I.e. both agents act as positive samples of the respective embeddings, while acting as negative samples of each other, according to the above assumption the following contrast losses are introduced:
；
wherein the method comprises the steps ofThe Bayesian personalized ranking loss is represented by the following formula:
；
wherein the method comprises the steps ofRepresenting the inner product of two given representations.
Step 5: constructing a click rate prediction model based on interest points, and performing model prediction and loss function construction:
model prediction: at the time of obtaining user access historyAnd->After embedding the representation of the decoupling representation, embedding the decoupling representation with the target +.>Is embedded->And geographical characterization->In connection, a 2-layer multi-layer perceptron is applied to predict +.>Click rate of->，/>The calculation formula of (2) is as follows:
；
s52, objective function and course learning: given a label y, carrying out supervised interest point click rate prediction by adopting binary cross entropy loss, wherein the formula is as follows:
；
after adding decoupling to the model, the overall model loss is obtained:
；
wherein the method comprises the steps ofThe model is trained by dynamically increasing a contrast loss weight preheating program:
wherein the method comprises the steps ofAnd->Is a super parameter, k represents the current course number.
Claims (5)
1. The interest point recommendation method based on the sequential and the understood coupling characterization is characterized by comprising the following steps of:
s1: based on the user and interest point interaction history record, respectively constructing a sequence chart and a geographic chart to recommend the interest points, specifically, using geographic information in the user and interest point interaction history record to represent the user set asThe set of points of interest is denoted +.>Where U represents each user, |U| represents the number of users, V represents each point of interest, |V| represents the number of points of interest; every interest point->By its coordinates (+)>) The tuples are geocoded, i.e.)>For the followingEvery user +.>The access sequence of the interest points is expressed asGiven user +.>And target interest point->The goal of the point of interest recommendation is to predict +.>Access->Is expressed as +.>Wherein->Is a learning function parameterized as +.>；
Constructing a sequence chart: given userIs>Then construct the directed sequence diagram +.>For each edge->Indicating that the user is +.>Middle visit->After which access is continued +.>Thus, the sequence diagram->Collecting sequence information from a history record of a user access interest point to perform decoupling characterization of sequence influence;
constructing a geographic diagram: given the point of interest location information, then construct an undirected geographic mapUndirected edgeRepresentation->And->The distance between them is at a specific distance threshold +.>In, edge weight matrix->Is interest point->And->Geographic distance between, and therefore, a geographic map/>Performing decoupling characterization of geographic influence from the point of interest location information;
s2: customizing the propagation process of the sequence diagram and the geographic diagram, constructing a diagram propagation layer to respectively carry out decoupling characterization on the sequence diagram and the geographic diagram, and outputting two groups of interest point representations with sequence and geographic information so as to better utilize the inherent characteristics of each diagram;
s3: after decoupling characterization is carried out on the sequence diagram and the geographic diagram according to the previous interest point access history of the user, the soft attention mechanism is introduced, and a sequence embedded representation and a geographic embedded representation of the interest points are respectively generated;
s4: the self-supervision decoupling process comprises the steps of firstly, acting on a graph propagation layer through two reading functions to obtain output, and then designing a projection head structure contrast loss for each decoupling characterization;
s5: and constructing a click rate prediction model based on the interest points, and performing model prediction and loss function construction.
2. The method for recommending points of interest based on sequential and well-understood coupling characterization according to claim 1, wherein the specific method in step S2 is as follows:
s21, a sequence diagram propagation layer: given target userAnd sequence diagram of construction->，/>From the connection matrix->Determining, applying a gated neural network +.>And decoupling characterization embedding +.>Is updated as:
1) The interest points represent:the method comprises the steps of carrying out a first treatment on the surface of the Wherein->The representation is decoupled from the token embedding,
2) The point of interest representation is then built as an input state:
；
wherein the method comprises the steps ofThe connection matrix, b is the parameter,
3) And then input it to the update gate for controlling the last memory information to continue to hold the amount of data to the current time:
；
wherein:is a parameter that can be trained and is,
4) After updating the gate and then determining that the network state at the previous moment is multiplied by the output of the reset gate through the reset gate, the network state is used as a parameter to calculate the candidate state at the current moment:
；
wherein the method comprises the steps ofIs a parameter that can be trained and is,
5) Candidate states are calculated after the update gate and the reset gate, and the formula is:
；
wherein:is a reset gate->The parameters may be a function of the training parameters,
6) Finally, the final state of the current time step is obtained:
；
wherein:is update gate->A current memory content;
designing a gating neural network to sequentially perform on all nodes of the sequence diagram, wherein hidden states are updated by the updating gatesAnd reset gate->Control, the sequential effect behind the user behavior is propagated by gating +.>Simulating to obtain the output of a hidden layer as the embedded representation of the sequential decoupling representation of the interest point:
；
s22, a geographical map propagation layer: 1) Message construction function: given a location-based graphConstructing a geographical map propagation layer using a messaging scheme of a map neural network, and then constructing a message, the +.>A pair of adjacent interest points on the layer->And->The message is expressed as:
；
wherein the method comprises the steps ofIs a message coding function, obtains interest point representation from the graph neural network of the previous layer>I.e. the input of the first layer is initialized with the embedded X of interest points:
；
message functionRepresenting the influence of the distance between the points of interest +.>Is realized by:
；
wherein the method comprises the steps ofIs a trainable weight matrix, as a lower layer representation a linear transformation taking into account distance effects; distance kernelAlong with->Exponentially decaying->And->The distance between them increases, using the graph Laplace norm of the graph roll-up neural network, where +.>And->Respectively indicate->Go up->And->Is the first hop neighbor of (a);
2) Message aggregation function: generating a node representation of each graph neural network layer by defining a message aggregation function:
；
wherein the method comprises the steps ofBy->Is derived from the representation of (a);
after propagation along the L-layer graph neural network, the hidden representation of the L-th layer is adopted as the geocoding of all interest points in V:
；
for the userSign-in history of->By aggregating users->Higher order geographic neighbors of points of interest in the point of interest access history to capture geographic preferences, user's geocode +.>Is->The geographical decoupling representation of all points of interest appearing in the list is embedded.
3. The method for recommending points of interest based on sequential and well-understood coupling characterization according to claim 1, wherein the specific method in step S3 is as follows:
s31, entering a sequence diagram and a geographic diagram according to the previous interest point access history of the userAfter the decoupling characterization, a soft attention mechanism is introduced according to the current interest pointAggregate coding, coding according to a given geographical map +.>Will->Hidden representation +.>Generating a geo-embedded representation as a query vector>：
；
；
Wherein the method comprises the steps ofRepresenting sigmoid function->Representing attention vectors, queries and key matrices +.>Is a trainable parameter for +.>Will->Is +.>Generating a sequential embedded representation as a query vector>：
；
。
4. The method for recommending points of interest based on sequential and well-understood coupling characterization according to claim 1, wherein the specific steps in step S4 are as follows:
s41 willAnd->The two representation forms are separated from each other, the order effect and the geographic effect are fully utilized to predict the click rate of interest points, the two embedded agents are obtained, the average pooling is selected as a reading function, and the average pooling is acted on a graph propagation layer to obtain output so as to generate ∈>And->Is->And->：
；
；
Wherein the method comprises the steps ofAnd->Representing a geographical map and a sequence map, < >>Carrying geographical preference information about the user's point of interest, < >>Carrying sequential preference information about user points of interest, < >>Representing a set of one-hop geographical neighbors,>an access order representation set representing points of interest thereof;
s42, before sending the embedding and its proxy to generate the self-supervising signals, designing a projection head for each decoupling token, the projection representation of which is calculated as follows:
；
；
wherein the method comprises the steps ofIs a linear transformation with trainable parameters, the projection head projects the embedding into another potential space, obtaining information from different aspects;
representing the projection poolAnd->Regarded as->And->I.e. the two agents act as positive samples of the respective embeddings, while acting as negative samples of each other, the design contrast loss function is:
；
wherein the method comprises the steps ofThe Bayesian personalized ranking loss is represented by the following formula:
；
wherein the method comprises the steps ofRepresenting the inner product of two given representations.
5. The method for recommending points of interest based on sequential and well-understood coupling characterization according to claim 1, wherein the specific steps in step S5 are as follows:
s51, model prediction: at the time of obtaining user access historyAnd->After embedding the representation of the decoupling representation, embedding the decoupling representation with the target +.>Is embedded->And geographical characterization->In connection, a 2-layer multi-layer perceptron is applied to predict +.>Click rate of->，/>The calculation formula of (2) is as follows:
；
s52, objective function and course learning: given a label y, carrying out supervised interest point click rate prediction by adopting binary cross entropy loss, wherein the formula is as follows:
；
after adding decoupling to the model, the overall model loss is obtained:
；
wherein the method comprises the steps ofIs a decoupled weight, training the model by dynamically increasing the contrast loss weight pre-heating program:
；
wherein the method comprises the steps ofAnd->Is a superparameter->Representing the current course number.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202410061576.2A CN117573986B (en) | 2024-01-16 | 2024-01-16 | Interest point recommendation method based on sequential and ground understanding coupling characterization |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202410061576.2A CN117573986B (en) | 2024-01-16 | 2024-01-16 | Interest point recommendation method based on sequential and ground understanding coupling characterization |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN117573986A true CN117573986A (en) | 2024-02-20 |
| CN117573986B CN117573986B (en) | 2024-03-26 |
Family
ID=89895922
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202410061576.2A Active CN117573986B (en) | 2024-01-16 | 2024-01-16 | Interest point recommendation method based on sequential and ground understanding coupling characterization |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN117573986B (en) |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20130024203A1 (en) * | 2011-07-20 | 2013-01-24 | Nternational Business Machines Corporation | Providing dynamic recommendations for points of interest utilizing automatically obtained collective telemetry to enhance user experience |
| CN111949865A (en) * | 2020-08-10 | 2020-11-17 | 杭州电子科技大学 | Interest point recommendation method based on graph neural network and user long-term and short-term preference |
| CN112784163A (en) * | 2021-01-29 | 2021-05-11 | 东北财经大学 | RS 24-based personalized POIs recommendation method |
| US20210248461A1 (en) * | 2020-02-11 | 2021-08-12 | Nec Laboratories America, Inc. | Graph enhanced attention network for explainable poi recommendation |
| CN115827974A (en) * | 2022-12-05 | 2023-03-21 | 北京理工大学 | Next interest point recommendation system based on spatio-temporal information representation |
| CN116805020A (en) * | 2022-12-14 | 2023-09-26 | 浙江工商大学 | Interest point recommendation method based on graphic neural network and context information perception |
| CN117194763A (en) * | 2023-02-27 | 2023-12-08 | 中国计量大学 | Method for recommending next POI based on user preference and space-time context information |
-
2024
- 2024-01-16 CN CN202410061576.2A patent/CN117573986B/en active Active
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20130024203A1 (en) * | 2011-07-20 | 2013-01-24 | Nternational Business Machines Corporation | Providing dynamic recommendations for points of interest utilizing automatically obtained collective telemetry to enhance user experience |
| US20210248461A1 (en) * | 2020-02-11 | 2021-08-12 | Nec Laboratories America, Inc. | Graph enhanced attention network for explainable poi recommendation |
| CN111949865A (en) * | 2020-08-10 | 2020-11-17 | 杭州电子科技大学 | Interest point recommendation method based on graph neural network and user long-term and short-term preference |
| CN112784163A (en) * | 2021-01-29 | 2021-05-11 | 东北财经大学 | RS 24-based personalized POIs recommendation method |
| CN115827974A (en) * | 2022-12-05 | 2023-03-21 | 北京理工大学 | Next interest point recommendation system based on spatio-temporal information representation |
| CN116805020A (en) * | 2022-12-14 | 2023-09-26 | 浙江工商大学 | Interest point recommendation method based on graphic neural network and context information perception |
| CN117194763A (en) * | 2023-02-27 | 2023-12-08 | 中国计量大学 | Method for recommending next POI based on user preference and space-time context information |
Non-Patent Citations (1)
| Title |
|---|
| 邵长城;陈平华;: "融合社交网络和图像内容的兴趣点推荐", 计算机应用, no. 05, 21 January 2019 (2019-01-21), pages 21 - 28 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN117573986B (en) | 2024-03-26 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Feng et al. | Poi2vec: Geographical latent representation for predicting future visitors | |
| Guo et al. | Combining geographical and social influences with deep learning for personalized point-of-interest recommendation | |
| Zhong et al. | Hybrid graph convolutional networks with multi-head attention for location recommendation | |
| Jiang et al. | Social recommendation with cross-domain transferable knowledge | |
| CN111949865A (en) | Interest point recommendation method based on graph neural network and user long-term and short-term preference | |
| Wang et al. | Geography-aware inductive matrix completion for personalized point-of-interest recommendation in smart cities | |
| CN111241419B (en) | Method for recommending next interest point based on user relation embedded model | |
| Wang et al. | Towards real-time demand-aware sequential POI recommendation | |
| CN107330798B (en) | Method for identifying user identity between social networks based on seed node propagation | |
| Cai et al. | FG-CF: Friends-aware graph collaborative filtering for POI recommendation | |
| Sun et al. | MFNP: A Meta-optimized Model for Few-shot Next POI Recommendation. | |
| Liu et al. | Mix geographical information into local collaborative ranking for POI recommendation | |
| CN109948066A (en) | A kind of point of interest recommended method based on Heterogeneous Information network | |
| Xiao et al. | MutualRec: Joint friend and item recommendations with mutualistic attentional graph neural networks | |
| CN115270007B (en) | POI recommendation method and system based on mixed graph neural network | |
| Liu et al. | GNNRec: gated graph neural network for session-based social recommendation model | |
| Sivaramakrishnan et al. | An effective user clustering-based collaborative filtering recommender system with grey wolf optimisation | |
| Lai et al. | Multi-view spatial-temporal enhanced hypergraph network for next poi recommendation | |
| Xu et al. | Deep convolutional recurrent model for region recommendation with spatial and temporal contexts | |
| CN113590976A (en) | Recommendation method of space self-adaptive graph convolution network | |
| Wu et al. | Graph-augmented social translation model for next-item recommendation | |
| CN117573986B (en) | Interest point recommendation method based on sequential and ground understanding coupling characterization | |
| Lu et al. | Social network alignment: a bi-layer graph attention neural networks based method | |
| Xi et al. | Exploiting bi-directional global transition patterns and personal preferences for missing POI category identification | |
| Li et al. | Multi-modal representation learning for successive poi recommendation |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |