CN113158038A - Interest point recommendation method and system based on STA-TCN neural network framework - Google Patents

Interest point recommendation method and system based on STA-TCN neural network framework Download PDF

Info

Publication number
CN113158038A
CN113158038A CN202110362907.2A CN202110362907A CN113158038A CN 113158038 A CN113158038 A CN 113158038A CN 202110362907 A CN202110362907 A CN 202110362907A CN 113158038 A CN113158038 A CN 113158038A
Authority
CN
China
Prior art keywords
interest
time
vector
sequence
point
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
Application number
CN202110362907.2A
Other languages
Chinese (zh)
Other versions
CN113158038B (en
Inventor
江浩
欧俊杰
王孝诚
金海明
刘艺娟
黄建强
王新兵
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Shanghai Jiaotong University
Original Assignee
Shanghai Jiaotong University
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Shanghai Jiaotong University filed Critical Shanghai Jiaotong University
Priority to CN202110362907.2A priority Critical patent/CN113158038B/en
Publication of CN113158038A publication Critical patent/CN113158038A/en
Application granted granted Critical
Publication of CN113158038B publication Critical patent/CN113158038B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

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
    • G06F16/9537—Spatial or temporal dependent retrieval, e.g. spatiotemporal queries
    • 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
    • G06N3/045—Combinations of networks
    • 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
    • G06N3/08—Learning methods

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Databases & Information Systems (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Health & Medical Sciences (AREA)
  • Computational Linguistics (AREA)
  • Evolutionary Computation (AREA)
  • Biophysics (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Biomedical Technology (AREA)
  • Artificial Intelligence (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

The invention provides an interest point recommendation method based on an STA-TCN neural network framework, which comprises the following steps: step S1: preprocessing data of a user check-in sequence, and filtering inactive users and inactive interest points; step S2: converting the preprocessed data into a high-dimensional embedded vector sequence; step S3: learning the sequential transition correlation of the embedded vector sequence by using a neural network, and outputting to obtain a result vector containing sequential transition correlation information; step S4: learning global space-time correlation by using a space-time self-attention mechanism on a result vector containing sequence transition correlation information, and outputting a final expression vector; step S5: and obtaining the point of interest recommendation result according to the final expression vector. The invention also provides an interest point recommendation system of the STA-TCN neural network framework, and the ST-Attention is used for integrating the time information and the spatial information into the Attention mechanism, so that the model learning of various correlations among the interest points is greatly promoted.

Description

Interest point recommendation method and system based on STA-TCN neural network framework
Technical Field
The invention relates to the field of recommendation systems, in particular to a point of interest recommendation method and system based on an STA-TCN neural network framework.
Background
Nowadays, with the rapid development of location-based social networking service (lbs n) platforms such as Foursquare and Yelp, more and more users want to share their Point of Interest (Interest) check-in records at different locations, such as restaurants, museums, etc., with friends. The large amount of user check-in data is helpful for learning user preference research on points of interest. How to accurately recommend points of interest to a user has high value for the point of interest owners to attract potential users, as well as for users to explore the surrounding environment and discover potential interesting points. Some terms specific to the term are explained as follows:
point of Interest (POI, Point-of-Interest): the location on the map can be an entertainment place, a dining place, a scenic spot and the like, and the position of the location is represented by a GPS.
Location-based Social Networks (LBSN): wherein the set of users is represented as
Figure BDA0003006298710000011
The set of points of interest is represented as
Figure BDA0003006298710000012
Wherein each point of interest
Figure BDA0003006298710000013
At a position gp(lon, lat), where lon and lat denote their longitude and latitude coordinates.
sign-In record (Check-In)): user check-in records are represented as triplets
Figure BDA0003006298710000014
This indicates that the user visited the location at the past timestamp t
Figure BDA0003006298710000015
Point of interest p.
Check-In History (Check-In History): given a data set, a user
Figure BDA0003006298710000016
Is defined as the set of all check-in records for that user, with each element representing a user
Figure BDA0003006298710000017
The ith check-in record in the dataset.
Point of interest Recommendation (POI Recommendation): for a given target user
Figure BDA0003006298710000018
The purpose of the interest point recommendation problem is to recommend target users
Figure BDA0003006298710000019
A list of the top M points of interest that are next but never visited is preferred.
In fact, the point of interest access behavior of the user shows strong sequential transitional relevance. That is, the interest point that the user accesses next is highly correlated with the interest point that the user has accessed in the past. For example, after eating dinner at a restaurant on a weekend, it is likely that some users will subsequently go to a movie theater, bar, or other entertainment venue. Naturally, capturing this sequential transition correlation between check-in records is critical to the point of interest recommendation system.
A Recurrent Neural Network (RNN) based model was first used to solve the point of interest recommendation problem, which can learn the sequential transitional relevance between user point of interest check-in records through the user's check-in records. To train such neural network-based models, the user's historical check-in sequence must be divided into a number of shorter subsequences, which are then entered into the models one by one, which is inevitably very time consuming.
After retrieval, patent document CN111241306A discloses a path planning method based on knowledge graph and pointer network, which includes: acquiring interest points in the tourist map as nodes to construct a knowledge map, wherein each node comprises four-dimensional information of an interest point, and aggregating the four-dimensional information of each node in the knowledge map by using a map neural network to generate an embedded matrix of the interest point; inputting the embedded matrix as a training sample into a pointer network, and training the pointer network to obtain a trained pointer network; and aiming at the interest points to be tested in the tourist map, obtaining an embedded matrix of the interest points as a test sample, inputting the test sample into the trained pointer network, sequentially selecting the interest points with the highest output probability as the next interest point of the current route, and finishing path planning. In the prior art, after information such as longitude, latitude, heat, playing time and the like of an interest point is needed, a knowledge graph is constructed and stored according to the information data, the knowledge graph needs to be updated continuously in the actual application process, and the process is complicated. Meanwhile, time information is not used, the use of the spatial information is only limited to the construction of a knowledge graph, and the spatial correlation of the interest points cannot be fully mined for the knowledge graph.
Patent document CN109885756A discloses a CNN and RNN-based serialization recommendation method, in which the algorithm utilizes the local feature learning capability of CNN to capture the correlation existing in recent historical behavior data, and utilizes the global and sequence learning capabilities of RNN to learn the long-term and short-term preferences of user historical behaviors, and finally utilizes a multi-layer perceptron to predict the behaviors that will be generated in the future by the learned feature expression and provide recommendations. The prior art does not use the time information and the space information in the data, and cannot fully mine the time and space correlation of the interest points.
In addition, in both the above two prior arts, RNN is used for training, and RNN can handle common sequence problems, but for long sequence problems, RNN is unable to do so, RNN may have gradient disappearance/explosion problems, and long sequence information cannot be retained, and training of the model is time consuming abnormally only due to the characteristic of serial training.
Therefore, it is highly desirable to develop a new neural network framework to better learn the local transitional correlation and the global spatial correlation of the check-in history recommended by the user interest points.
Disclosure of Invention
Aiming at the defects in the prior art, the invention aims to provide an interest point recommendation method and system based on an STA-TCN neural network framework.
The invention provides an interest point recommendation method based on an STA-TCN neural network framework, which comprises the following steps:
step S1: preprocessing data of a user check-in sequence, and filtering inactive users and inactive interest points;
step S2: converting the preprocessed data into a high-dimensional embedded vector sequence;
step S3: learning the sequential transition correlation of the embedded vector sequence by using a neural network, and outputting to obtain a result vector containing sequential transition correlation information;
step S4: learning global space-time correlation by using a space-time self-attention mechanism on a result vector containing sequence transition correlation information, and outputting a final expression vector;
step S5: and obtaining the point of interest recommendation result according to the final expression vector.
Preferably, both data sets are preprocessed in step S1 by deleting less than 10 inactive users of the check-in record and less than 10 undesired points of interest of the accessing user.
Preferably, step S2 includes the following:
step S2.1: embedding GPS coordinates through Tile Map system and converting the GPS coordinates into r-dimensional GPS coordinate vectors;
step S2.2: embedding a timestamp vector;
step S2.3: and embedding the interest point vector.
Preferably, the sequential transition correlations between user check-in sequences are learned using a time-series convolutional network in step S3.
Preferably, step S3 includes: embedding interest points with length L into sequence X1:L＝(X1,X2,…,XL) Input into a time-sequential convolutional network, wherein
Figure BDA0003006298710000032
The embedded vector representing the ith point of interest, for X, using the following formula1:LA cause-and-effect convolution is performed,
Figure BDA0003006298710000031
wherein, denotes a dilated causal convolution operation, f denotes a convolution filter with kernel size H, e is a dilation factor that controls the receive window size of the convolution kernel, xj-ehThe (e × h) th vector before position j is represented.
Preferably, step S3 further includes: the same kernel weight matrix is used for all input vectors, and the output characteristic X is obtained by utilizing an activation function1:LThe non-linearity of (2).
Y1:L＝ReLU(W*X1:L)
Where W represents a kernel weight matrix shared in a time-series convolutional network, ReLU is a non-linear activation function, and Y1:LRepresenting the output characteristics.
Preferably, the spatiotemporal self-attention mechanism in step S4 includes a mesh distance learning mechanism and a time-sensitive learning mechanism.
Preferably, the grid distance learning mechanism obtains the grid distance vectors for the two GPS locations by performing the operation of the following formula
Figure BDA0003006298710000041
Figure BDA0003006298710000042
Wherein
Figure BDA0003006298710000044
Representing a vector
Figure BDA0003006298710000045
The qth element in (1), Abs (·) represents an absolute value calculator.
Preferably, the time-sensitive learning mechanism embeds time into the vector sequence C1:L＝(c1,c2,…,cL) And a temporally embedded vector sequence T1:L＝(t1,t2,…,tL) As an input, wherein
Figure BDA0003006298710000046
An embedded vector representing the ith point of interest in the check-in,
Figure BDA0003006298710000047
representing the time embedding vector of the ith time stamp in the embedding sequence, and outputting a time correlation scoring matrix A by a time sensitivity learning mechanismt：
Figure BDA0003006298710000043
Wherein, WtRepresenting parameters in the T-SL mechanism.
The invention provides an interest point recommendation system based on an STA-TCN neural network framework, which comprises the following steps:
inputting an embedding layer: the input embedding layer takes a user sign-in sequence as input, consists of interest points, a GPS position and a time stamp, and respectively outputs an embedded vector sequence of the interest points, the GPS position and the time stamp;
time sequence convolution network: the time sequence convolution network takes the interest point embedding vector as input and outputs a result vector containing sequence transition correlation information;
a spatiotemporal attention module: taking the output of the time sequence convolution network, the timestamp of the check-in record and the GPS position embedding vector as input, and outputting the final expression vector of the learned global space and time correlation;
an output module: and the output module uses the selector to obtain the point of interest recommendation result according to the final representation vector.
Compared with the prior art, the invention has the following beneficial effects:
1. the invention fully utilizes the time and space information, does not need to construct a knowledge map, and uses ST-Attention to integrate the time information and the space information into an Attention mechanism, thereby greatly promoting various correlations among the model learning interest points.
2. The invention utilizes STA-TCN neural network to search the interest point, wherein TCN adds a residual error model to solve the problem of gradient disappearance/explosion, and can use causal convolution and expansion convolution to capture long sequence information, and simultaneously, the shallow model can be easily expanded into the deep model.
3. The TCN utilizes causal convolution to shield future information, so that the model can process the time sequence problem, and the expansion convolution enables the model to capture longer sequence length.
4. The Attention in the invention can acquire global and local relation, and can not receive the limitation of sequence length on the capture of long-term dependence like an RNN (radio network); compared with CNN and RNN, the parameters are less, and the model complexity is low; the result of each step is independent of the previous step and can be calculated in parallel.
Drawings
Other features, objects and advantages of the invention will become more apparent upon reading of the detailed description of non-limiting embodiments with reference to the following drawings:
FIG. 1 is an overall architecture diagram of a point of interest recommendation system based on a STA-TCN neural network framework according to the present invention;
FIG. 2 is a hierarchical first level Map based on Tile Map system in accordance with the present invention;
FIG. 3 is a hierarchical second level Map based on Tile Map system in accordance with the present invention;
FIG. 4 is a graph of the performance of the method at HR @10 (hit rate at a recommendation list length of 10) on a Gowalla dataset under different hyper-parameters;
FIG. 5 shows the performance of HR @10 (hit rate for a recommendation list of 10 length) on a Foursquare data set under different hyper-parameters;
FIG. 6 is a graph of the performance of the method on the NDCG @10 (normalized break cumulative gain) on the Gowalla dataset under different hyper-parameters;
FIG. 7 shows the performance of the method on the NDCG @10 (normalized break cumulative gain) on the Fourier data set under different hyper-parameters;
FIG. 8 is a graph of the time it takes for the currently popular POI recommendation method and method to achieve the best performance (hit rate at recommendation list length of 5) on the Gowalla dataset;
FIG. 9 is a graph of the time it takes for the currently popular POI recommendation method and method to achieve the best performance (hit rate at recommendation list length of 5) on the Foursquare dataset;
FIG. 10 is a graph of the time it takes for the currently popular POI recommendation method and method to achieve the best performance (normalized discounted cumulative gain at recommendation list length of 5) on the Gowalla dataset;
fig. 11 is a graph of the time it takes for the currently popular POI recommendation method and method to achieve the best performance (normalized discount cumulative gain at recommendation list length of 5) on the Foursquare data set.
Detailed Description
The present invention will be described in detail with reference to specific examples. The following examples will assist those skilled in the art in further understanding the invention, but are not intended to limit the invention in any way. It should be noted that it would be obvious to those skilled in the art that various changes and modifications can be made without departing from the spirit of the invention. All falling within the scope of the present invention.
As shown in fig. 1 to 11, the present invention provides a point of interest recommendation method based on an STA-TCN neural network framework, including the following steps:
step S1: preprocessing data of a user check-in sequence, and filtering inactive users and inactive interest points;
step S2: converting the preprocessed data into a high-dimensional embedded vector sequence;
step S3: learning the sequential transition correlation of the embedded vector sequence by using a neural network, and outputting to obtain a result vector containing sequential transition correlation information;
step S4: learning global space-time correlation by using a space-time self-attention mechanism on a result vector containing sequence transition correlation information, and outputting a final expression vector;
step S5: and obtaining the point of interest recommendation result according to the final expression vector.
Specifically, step S1 further includes the following specific steps:
step S1.1: experiments were conducted on two real-world LBSN datasets, Gowalla and Foursquare, which are widely used to evaluate point-of-interest recommendation methods. The Gowalla dataset contains check-in records worldwide from 2 months 2009 to 10 months 2010, and check-in samples of the Foursquare dataset all occur in New York City from 2 months 2010 to 1 month 2011. Both data sets are preprocessed by deleting less than 10 inactive users and less than 10 visiting users' undesirable points of interest. In order to ensure the diversity of user interests, users with less than 5 different points of interest accessing the records are excluded at the same time. Table 1 gives detailed statistics of the two data sets.
Table 1 data set statistics
Figure BDA0003006298710000061
Step S2 includes: inputting data into the embedding layer due to data centralizationIndexing at a large number of points of interest
Figure BDA0003006298710000062
GPS coordinates
Figure BDA0003006298710000063
And time stamp t, each of which would produce a sparse, high-dimensional vector if represented directly using One-hot encoding (One-hot), which would degrade model performance. Thus, to embed each point of interest, GPS coordinates and a timestamp into a low-dimensional representation vector, the present invention designs an input embedding layer by performing the following operations.
Step S2.1: GPS coordinates are embedded. As shown in FIGS. 2, 3, and 4-7, the entire region is first hierarchically divided into multiple levels of grids using a Tile Map system, and each grid, which may be represented by a vector, is represented using hash keys
Figure BDA0003006298710000071
Is shown in which
Figure BDA0003006298710000078
And numbers with bases 4 indicate where the grid is located in the ith level area. Then, with such Tile Map system, all the interest points can be converted into r-dimensional GPS coordinate vectors by their GPS information.
Step S2.2: time stamp embedding: considering that users usually have different point-of-interest access behaviors at different times and different dates in a day, the day is firstly divided into N time periods with equal intervals, and then each time period is used for judging whether the time period is equal to the other time periods
Figure BDA0003006298710000072
Whether on weekdays or weekends, each time period 12i is divided into
Figure BDA0003006298710000073
Or
Figure BDA0003006298710000074
Thus, the time can beDivided into 2N time segments, respectively denoted as
Figure BDA0003006298710000075
For a given timestamp t, embed it into a one-hot vector
Figure BDA0003006298710000076
The position with a value of 1 corresponds to the time period to which this t belongs.
Step S2.3: interest points are embedded, and it is apparent that the probability of a user accessing a point of interest varies at different times of the day. The present invention utilizes this time sensitivity to classify points of interest, specifically, for each point of interest in the check-in sequence, the check-in frequencies for all users in the above 2N time periods are calculated in the entire dataset, and these check-in frequencies are further combined into a time-sensitive embedded vector
Figure BDA0003006298710000077
Wherein
Figure BDA0003006298710000079
Representing the user access frequency of the point of interest under consideration during the ith time period. Considering that the interest point has a unique GPS coordinate and a time sensitivity attribute at the same time, c is spliced with the GPS embedded vector g to form a final representation vector of the interest point with x ═ c | | g.
By performing the above three types of embedding together, the input embedding layer will convert a user sign-in sequence of length L into four embedded vector sequences, including the interest point embedded vector sequence X1:LThe time-sensitive embedded vector sequence is denoted G1:LThe GPS embedded vector sequence is denoted as G1:LThe temporally embedded vector sequence is denoted T1:L。
Step S3 includes the following specific steps:
step S3.1, the invention learns the sequential transition correlation between user check-in records by using a time sequence convolution network. The recurrent neural network processes and transmits input information by using a serial structure, and stores information by using a large number of unit states and hidden states, and the time sequence convolution network performs convolution operation on the input information by using a parallel architecture, so that less memory requirement and higher training speed are realized. Compared with the original convolutional neural network, the time sequence convolutional network adopts a special convolution structure called causal convolution, and the sequence transition correlation learning effect of the input sequence is promoted to a great extent. On one hand, the causal convolution operation of the time sequence convolution network embeds the interest points of the input and the past time stamps, and performs convolution operation on the input vector, so that the time sequence of the user sign-in sequence is reserved; on the other hand, the convolution operation of the time-series convolution network has a swelling property, so that the receiving domain can be expanded by times to process an input sequence with an excessively long length.
More specifically, interest points of length L are embedded in sequence X1:L＝(X1,X2,…,XL) Input into a time-sequential convolutional network, wherein
Figure BDA0003006298710000085
The embedded vector representing the ith point of interest, for X, using the following formula1:LAnd carrying out causal convolution.
Figure BDA0003006298710000081
Where denotes a dilated causal convolution operation, f denotes a convolution filter with kernel size H, e is a dilation factor that controls the receive window size of the convolution kernel, xj-ehThe (e × h) th vector before position j is represented.
S3.2, the invention uses the same kernel weight matrix for all input vectors, and then obtains the output characteristic X by using an activation function1:LThe non-linearity of (2).
Y1:L＝ReLU(W*X1:L)
Where W represents a kernel weight matrix shared in a time-sequential convolutional network, ReLU is a non-linear activation function, and Y1:LRepresenting the output characteristics. To produce output sequences of equal length, zeros are padded to the input sequence X1:LTo the end of (c).
In addition, in order to better represent the preference degree of the user for the interest point, the invention also designs a gating injection mechanism which adds the original input vector passing through the gating unit and the output characteristic of the time sequence convolution network.
Y1:L＝Y1:L+X1:L⊙σ(Wg·X1:L+bg)
Final output of Y1:LConvolution characteristic and input information X integrating TCN output1:L，X1:LThe amount of information injected is controlled by a gating mechanism, where wgAnd bgIndicating a parameter, σ (-) indicates a sigmoid type function, a "-" indicates a multiplication based on a matrix element.
Step S4 includes the following specific steps:
while time-series convolutional networks may help capture sequential transition correlations between user check-in sequences, considering only such sequential transition correlations is not sufficient to accurately learn how much a user prefers points of interest. In order to learn the global spatiotemporal correlation, the invention provides a spatiotemporal self-attention (STATT) module based on a time sequence convolutional network, and the module enhances the self-attention network by two novel methods. A grid-distance learning (G-DL) mechanism and a time-sensitivity learning (T-SL) mechanism, respectively.
Step S4.1: first, a GPS embedding sequence G is obtained according to the step S21:L＝(g1,g2,…,gL),
Figure BDA0003006298710000087
Embedded vectors representing the ith GPS coordinate in the check-in sequence, for the sequence G, in order to learn the global spatial correlation1:LEach pair of embedded vectors in
Figure BDA0003006298710000086
And
Figure BDA0003006298710000082
the G-DL mechanism obtains the grid distance vector of two GPS positions by performing the operation of the following formula
Figure BDA0003006298710000083
Figure BDA0003006298710000084
Wherein
Figure BDA00030062987100000913
Representing a vector
Figure BDA00030062987100000914
The qth element in (1), Abs (·) represents an absolute value calculator. Vector quantity
Figure BDA0003006298710000091
Is represented by
Figure BDA00030062987100000915
And
Figure BDA0003006298710000092
the distance between the two grids, then the mechanism will be compared with G1:LWherein each pair of GPS embedded vectors corresponds
Figure BDA0003006298710000093
Combined into a vector matrix
Figure BDA0003006298710000094
And calculating a scoring matrix A of the spatial correlation by applying a feedforward neural networksWherein W issAnd bsRepresenting parameters in G-DL.
As＝Ws·Ms+bs
Step S4.2 point of interest access preferences of the user show a time sensitivity attribute. Inspired by this, the invention proposesThe T-SL mechanism learns the global temporal correlation of the input sequence. This mechanism embeds time into a vector sequence C1:L＝(c1,c2,…,cL) And a temporally embedded vector sequence T1:L＝(t1,t2,…,tL) As an input, wherein
Figure BDA00030062987100000916
An embedded vector representing the ith point of interest in the check-in,
Figure BDA00030062987100000917
representing the time-embedded vector of the ith time stamp in the check-in sequence, and outputting a time correlation scoring matrix A by T-SL through calculationt. Wherein WtRepresenting parameters in the T-SL mechanism.
Figure BDA0003006298710000095
Step S4.3, in order to better learn the global space-time correlation between the user check-ins, STATT combines the space-time correlation score matrix with the traditional self-attention mechanism and outputs the final expression vector sequence Z1:L。
Figure BDA0003006298710000096
Wherein
Figure BDA0003006298710000097
WQ、WK、WVRespectively representing a query matrix, a keyword matrix and a value matrix, wherein d is the dimension of a key vector; m is a mask matrix, filled with- ∞ in all upper triangular elements to satisfy the time relation constraint of the input sequence.
Finally, STATt will Z1:LLast vector z in (2)LAs a final representation vector of the user preferences.
Step S5 includes the following specific steps:
step S5.1 As shown in FIG. 1, the output module outputs zLThe selector is entered to produce the recommendation. Specifically, the output module searches for the nearest user's current check-in position
Figure BDA0003006298710000098
Generating interest point candidate set by each interest point
Figure BDA0003006298710000099
The selector will then
Figure BDA00030062987100000918
As input, and embedding vectors based on their points of interest
Figure BDA00030062987100000910
And a user preference vector zLInner product of each candidate interest point to calculate each candidate interest point
Figure BDA00030062987100000911
Is scored as
Figure BDA00030062987100000912
The final selector outputs a recommended point of interest list by selecting top-M candidate points of interest with the highest scores.
The invention uses a novel deep learning model-STA-TCN neural network framework, and respectively performs experiments on Gowalla and Foursquare data sets with the prior model, and the performance of two standard comparison models, namely an HR standard comparison model and an NDCG standard comparison model, is obtained through the experiments.
The HR standard comparison model is a representative hit rate model, and the hit is the user concerned when the user appears in the recommendation list;
the NDCG standard comparison model refers to a normalized depreciation cumulative gain, which is obtained by dividing the Depreciation Cumulative Gain (DCG) by the ideally maximum DCG value (IDCG).
In the training process, the maximum length of an input check-in sequence is set as 100, and the long sequence is divided into non-overlapping subsequences with the length of 100 from right to leftAnd (4) columns. All check-in records of the users are used as input for recommending the next interest point in the test process, and for each check-in sequence of the users, the latest interest point which does not appear in the past records is used as a target interest point to evaluate the performance of the model. Setting the number N of time segments of each day as 12, and setting the level r of grid division of the hierarchical diagram as 17; setting the number of convolution layers and the size of an inner core of the time sequence convolution network to be 2 and 6 respectively; furthermore, the dilation factors in the two convolutional layers of the time-sequential convolutional network are 1 and 2, respectively; candidate set size of points of interest
Figure BDA0003006298710000103
Set to 100.
Two common evaluation indices, Hit Rate (HR) and normalized cumulative yield (NDCG), were used in the experiments. HR @ K calculates the proportion of target interest points among the first K interest points appearing in the recommendation list, and NDCG @ K further calculates the ranking score of the target interest points among the first K interest points in the list.
TABLE 2 Performance of models on the Gowalla dataset
Figure BDA0003006298710000101
TABLE 3 Performance of the models on the Foursquare dataset
Figure BDA0003006298710000102
Figure BDA0003006298710000111
Tables 2 and 3 summarize the results of the STA-TCN and all baseline methods, with the optimal results highlighted in bold. It is clear that the STA-TCN proposed by the present invention shows the best performance on all metrics on each data set. STA-TCN improves the self-attention architecture through G-DL and T-DL mechanisms and enhances its ability to learn spatio-temporal correlations, with STA-TCN significantly improved 8.53%/11.81% and 6.23%/10.26% HR @5/10 on Gowalla and Foursquad datasets, respectively, compared to suboptimal GeoSAN.
And carrying out a comparison test, and evaluating from two angles of the influence of the hyper-parameters on the model and the model training efficiency. In particular, the impact of the hyper-parameter settings of the time series convolutional network on the model performance is analyzed. The influence of the super-parameter setting on the model performance is analyzed by performing experiments on different combinations of the number of convolution layers and the size of an inner core in the time-series convolution network. The core size was increased from 2 to 12 starting with 2, the number of layers was increased from 2 to 4 starting with 2, and the performance variation of the STA-TCN was evaluated on both data sets. As shown in fig. 4-7, the performance of the proposed STA-TCN is very strong and stable with only small performance fluctuations in the number of convolutional layers and the kernel size.
Further, the training efficiencies of the different models were compared. Under the same training environment, the experiment counted the time each method spent from the start of training to the convergence of the training loss and calculated their expressed recommended performance on both data sets. FIGS. 8-11 show the training time consumption and the corresponding HR @5 and NDCG @5 for each method, with those markers closer to the top left indicating that the method has higher training efficiency and better performance. It can be seen that due to the serial training structure of the recurrent neural network, training of most models based on the recurrent neural network is time-consuming, and the STA-TCN provided by the invention significantly improves the training speed and the performance of the recommendation system.
Those skilled in the art will appreciate that, in addition to implementing the system and its various devices, modules, units provided by the present invention as pure computer readable program code, the system and its various devices, modules, units provided by the present invention can be fully implemented by logically programming method steps in the form of logic gates, switches, application specific integrated circuits, programmable logic controllers, embedded microcontrollers and the like. Therefore, the system and various devices, modules and units thereof provided by the invention can be regarded as a hardware component, and the devices, modules and units included in the system for realizing various functions can also be regarded as structures in the hardware component; means, modules, units for performing the various functions may also be regarded as structures within both software modules and hardware components for performing the method.
The foregoing description of specific embodiments of the present invention has been presented. It is to be understood that the present invention is not limited to the specific embodiments described above, and that various changes or modifications may be made by one skilled in the art within the scope of the appended claims without departing from the spirit of the invention. The embodiments and features of the embodiments of the present application may be combined with each other arbitrarily without conflict.

Claims (10)

1. An interest point recommendation method based on an STA-TCN neural network framework is characterized by comprising the following steps:
step S1: preprocessing data of a user check-in sequence, and filtering inactive users and inactive interest points;
step S2: converting the preprocessed data into a high-dimensional embedded vector sequence;
step S3: learning the sequential transition correlation of the embedded vector sequence by using a neural network, and outputting to obtain a result vector containing sequential transition correlation information;
step S4: learning global space-time correlation by using a space-time self-attention mechanism on a result vector containing sequence transition correlation information, and outputting a final expression vector;
step S5: and obtaining the point of interest recommendation result according to the final expression vector.
2. The STA-TCN neural network framework-based point of interest recommendation method of claim 1, wherein said step S1 preprocesses both data sets by deleting less than 10 check-in recorded inactive users and less than 10 visiting users' unwanted points of interest.
3. The method for point of interest recommendation based on STA-TCN neural network framework according to claim 1, wherein said step S2 comprises the following steps:
step S2.1: embedding GPS coordinates through Tile Map system and converting the GPS coordinates into r-dimensional GPS coordinate vectors;
step S2.2: embedding a timestamp vector;
step S2.3: and embedding the interest point vector.
4. The method for point of interest recommendation based on STA-TCN neural network framework according to claim 1, characterized in that in step S3 sequential transition correlation between user check-in sequences is learned using time-series convolution network.
5. The method for recommending point of interest based on STA-TCN neural network framework according to claim 1, wherein said step S3 embeds point of interest with length L into sequence X1:L＝(X1,X2,…,XL) Input into a time-series convolutional network, where xiThe embedded vector representing the ith point of interest, for X, using the following formula1:LA cause-and-effect convolution is performed,
Figure FDA0003006298700000011
wherein, denotes a dilated causal convolution operation, f denotes a convolution filter with kernel size H, e is a dilation factor that controls the receive window size of the convolution kernel, xj-ehThe (e × h) th vector before position j is represented.
6. The method for recommending a point of interest based on STA-TCN neural network framework according to claim 1, wherein said step S3 further uses the same kernel weight matrix for all inputted vectors, and obtains the output feature X by using the activation function1:LThe non-linearity of (2).
Y1:L＝ReLU(W*X1:L)
Where W represents a kernel weight matrix shared in a time-series convolutional network, ReLU is a non-linear activation function, andY1:Lrepresenting the output characteristics.
7. The STA-TCN neural network framework-based point of interest recommendation method of claim 1, wherein the spatiotemporal self-attention mechanism in step S4 comprises a grid distance learning mechanism and a time-sensitive learning mechanism.
8. The STA-TCN neural network framework-based point of interest recommendation method of claim 7, wherein the grid distance learning mechanism obtains grid distance vectors for two GPS locations by performing the operation of the following formula
Figure FDA0003006298700000021
Figure FDA0003006298700000022
Wherein g isi(q) represents a vector giThe qth element in (1), Abs (·) represents an absolute value calculator.
9. The STA-TCN neural network framework-based point of interest recommendation method of claim 7, wherein the time-sensitive learning mechanism embeds time into vector sequence C1:L＝(c1,c2,…,cL) And a temporally embedded vector sequence T1:L＝(t1,t2,…,tL) As an input, where ciEmbedded vector, t, representing the ith point of interest in a check-iniRepresenting the time embedding vector of the ith time stamp in the embedding sequence, and outputting a time correlation scoring matrix A by a time sensitivity learning mechanismt：
Figure FDA0003006298700000023
Wherein, WtTo representParameters in the T-SL mechanism.
10. A point of interest recommendation system based on a STA-TCN neural network framework, comprising:
inputting an embedding layer: the input embedding layer takes a user sign-in sequence as input, consists of interest points, a GPS position and a time stamp, and respectively outputs an embedded vector sequence of the interest points, the GPS position and the time stamp;
time sequence convolution network: the time sequence convolution network takes the interest point embedding vector as input and outputs a result vector containing sequence transition correlation information;
a spatiotemporal attention module: taking the output of the time sequence convolution network, the timestamp of the check-in record and the GPS position embedding vector as input, and outputting the final expression vector of the learned global space and time correlation;
an output module: and the output module uses the selector to obtain the point of interest recommendation result according to the final representation vector.
CN202110362907.2A 2021-04-02 2021-04-02 Interest point recommendation method and system based on STA-TCN neural network framework Active CN113158038B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110362907.2A CN113158038B (en) 2021-04-02 2021-04-02 Interest point recommendation method and system based on STA-TCN neural network framework

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110362907.2A CN113158038B (en) 2021-04-02 2021-04-02 Interest point recommendation method and system based on STA-TCN neural network framework

Publications (2)

Publication Number Publication Date
CN113158038A true CN113158038A (en) 2021-07-23
CN113158038B CN113158038B (en) 2022-06-14

Family

ID=76888431

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110362907.2A Active CN113158038B (en) 2021-04-02 2021-04-02 Interest point recommendation method and system based on STA-TCN neural network framework

Country Status (1)

Country Link
CN (1) CN113158038B (en)

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113656709A (en) * 2021-08-24 2021-11-16 东北大学 Interpretable interest point recommendation method fusing knowledge graph and time sequence characteristics
CN114186139A (en) * 2021-12-30 2022-03-15 重庆理工大学 Graph neural network session recommendation method based on time enhancement
CN115225731A (en) * 2022-07-29 2022-10-21 中国人民解放军陆军工程大学 Online protocol identification method based on hybrid neural network
WO2023015658A1 (en) * 2021-08-13 2023-02-16 浙江大学 Point of interest recommendation method and system based on brain-inspired spatiotemporal perceptual representation

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110399565A (en) * 2019-07-29 2019-11-01 北京理工大学 Based on when null cycle attention mechanism recurrent neural network point of interest recommended method
CN111506835A (en) * 2020-04-17 2020-08-07 北京理工大学 Data feature extraction method fusing user time features and individual features
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
CN112069415A (en) * 2020-08-13 2020-12-11 中国海洋大学 Interest point recommendation method based on heterogeneous attribute network characterization learning
US20210064665A1 (en) * 2019-01-19 2021-03-04 Beijing Didi Infinity Technology And Development Co., Ltd. Systems and methods for online to offline services

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20210064665A1 (en) * 2019-01-19 2021-03-04 Beijing Didi Infinity Technology And Development Co., Ltd. Systems and methods for online to offline services
CN110399565A (en) * 2019-07-29 2019-11-01 北京理工大学 Based on when null cycle attention mechanism recurrent neural network point of interest recommended method
CN111506835A (en) * 2020-04-17 2020-08-07 北京理工大学 Data feature extraction method fusing user time features and individual features
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
CN112069415A (en) * 2020-08-13 2020-12-11 中国海洋大学 Interest point recommendation method based on heterogeneous attribute network characterization learning

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
LI QUAN等: "Next location recommendation based on spatiotemporal-aware GRU and attention", 《JOURNAL OF COMPUTER APPLICATIONS》 *
刘婕等: "基于自注意力机制的多模态语义轨迹预测", 《计算机工程与科学》 *

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2023015658A1 (en) * 2021-08-13 2023-02-16 浙江大学 Point of interest recommendation method and system based on brain-inspired spatiotemporal perceptual representation
CN113656709A (en) * 2021-08-24 2021-11-16 东北大学 Interpretable interest point recommendation method fusing knowledge graph and time sequence characteristics
CN113656709B (en) * 2021-08-24 2023-07-25 东北大学 Interpretable interest point recommendation method integrating knowledge graph and time sequence characteristics
CN114186139A (en) * 2021-12-30 2022-03-15 重庆理工大学 Graph neural network session recommendation method based on time enhancement
CN115225731A (en) * 2022-07-29 2022-10-21 中国人民解放军陆军工程大学 Online protocol identification method based on hybrid neural network
CN115225731B (en) * 2022-07-29 2024-03-05 中国人民解放军陆军工程大学 Online protocol identification method based on hybrid neural network

Also Published As

Publication number Publication date
CN113158038B (en) 2022-06-14

Similar Documents

Publication Publication Date Title
CN113158038B (en) Interest point recommendation method and system based on STA-TCN neural network framework
Yin et al. Joint modeling of user check-in behaviors for real-time point-of-interest recommendation
Chen et al. Curriculum meta-learning for next POI recommendation
Li et al. Next and next new POI recommendation via latent behavior pattern inference
Chen et al. CEM: A convolutional embedding model for predicting next locations
CN110119475A (en) A kind of POI recommended method and recommender system
Chen et al. Next POI recommendation based on location interest mining with recurrent neural networks
CN110570044A (en) next-place prediction method based on recurrent neural network and attention mechanism
CN111708876A (en) Method and device for generating information
CN111553279B (en) Method, device, equipment and storage medium for learning and identifying characterization of interest points
CN113139140A (en) Tourist attraction recommendation method based on space-time perception GRU and combined with user relation preference
CN115408618B (en) Point-of-interest recommendation method based on social relation fusion position dynamic popularity and geographic features
CN111783895B (en) Travel plan recommendation method, device, computer equipment and storage medium based on neural network
Wang et al. ST-SAGE: A spatial-temporal sparse additive generative model for spatial item recommendation
US20210263939A1 (en) Transition regularized matrix factorization for sequential recommendation
Chen et al. Multi-objective reinforcement learning approach for trip recommendation
Baral et al. CAPS: Context aware personalized POI sequence recommender system
CN116503588A (en) POI recommendation method, device and equipment based on multi-element relation space-time network
CN116975359A (en) Resource processing method, resource recommending method, device and computer equipment
Tomasi et al. Automatic Music Playlist Generation via Simulation-based Reinforcement Learning
CN115048530A (en) Graph convolution recommendation system fusing neighbor importance and feature learning
CN114065024A (en) POI recommendation method based on user personalized life mode
CN117633371B (en) Recommendation method, device and readable storage medium based on multi-attention mechanism
CN114417166B (en) Continuous interest point recommendation method based on behavior sequence and dynamic social influence
Zhang et al. Behavior Modeling Network with Inter-layer Attention for Human Mobility Prediction

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