CN115495661A - Self-adaptive interest point recommendation method based on long-term and short-term preference of user - Google Patents

Self-adaptive interest point recommendation method based on long-term and short-term preference of user Download PDF

Info

Publication number
CN115495661A
CN115495661A CN202211232832.7A CN202211232832A CN115495661A CN 115495661 A CN115495661 A CN 115495661A CN 202211232832 A CN202211232832 A CN 202211232832A CN 115495661 A CN115495661 A CN 115495661A
Authority
CN
China
Prior art keywords
check
user
recent
term
interest
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.)
Pending
Application number
CN202211232832.7A
Other languages
Chinese (zh)
Inventor
司亚利
李峰
聂盼红
刘井莲
赵卫绩
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.)
Changshu Institute of Technology
Original Assignee
Changshu Institute of Technology
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 Changshu Institute of Technology filed Critical Changshu Institute of Technology
Priority to CN202211232832.7A priority Critical patent/CN115495661A/en
Publication of CN115495661A publication Critical patent/CN115495661A/en
Pending legal-status Critical Current

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
    • G06N3/049—Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
    • 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 Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Biomedical Technology (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Health & Medical Sciences (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Evolutionary Computation (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

The invention discloses a self-adaptive interest point recommendation method based on long-term and short-term preference of a user, which comprises the following steps: constructing a track sequence set for the user to sign in based on the LBNS historical sign-in data set, and dividing the track sequence into a historical sign-in track sequence and a recent sign-in track sequence by adopting a dynamic time window according to a time interval of sign-in records; learning the historical preference of the user by adopting an LSTM model fused with a time factor based on the historical check-in track sequence, learning the recent preference of the user by adopting an RNN model combined with a space-time factor based on the recent check-in track sequence, and obtaining the long-term and short-term interest preference of the user by combining the historical preference and the recent preference; and performing probability prediction on the candidate interest points by combining the long-term and short-term interest preferences, and recommending a plurality of interest points with the highest probability value ranking to the user. The invention utilizes the long and short term learning result to recommend the interest point and optimize the recommendation effect.

Description

Self-adaptive interest point recommendation method based on long-term and short-term preference of user
Technical Field
The invention relates to an interest point recommendation method, in particular to a self-adaptive interest point recommendation method based on long-term and short-term preference of a user.
Background
Point of interest recommendations based on location-based social networks (lbs) can provide mobile users with a variety of places, personalized, and never visited. Most of the existing interest point recommendation methods adopt a deep learning method to model the continuous check-in behavior of the user aiming at the continuous check-in track sequence of the user, but the following problems exist:
(1) The check-in track of the user cannot obviously and accurately reflect the long-term preference and the short-term preference of the user. In the existing method, all check-in records of a user are regarded as a whole to generate a continuous interest point check-in track sequence, and due to the irregularity, randomness and difference characteristics of check-in of the mobile user in a social network, the whole preference of the user can only be reflected, and the specific historical preference and the recent preference of the user are difficult to be clearly reflected. Although some methods divide the check-in track sequence of the user, the recent track only depends on one recently checked-in interest point, and the selection of the next interest point of the user in real life is often influenced by a plurality of recently continuous check-in interest points, so that the recent interest preference of the user is difficult to accurately reflect.
(2) The problem of short recent check-in trajectory sequences is not solved. In the mobile social network, a part of users with low check-in activity exist, the check-in number is small, and the time interval and the distance interval between the interest points of adjacent check-in are large, so that the inactive users only have short check-in track sequences and poor continuity, especially lack enough recent check-in information seriously, short-term interest mining is difficult to perform, the short-term interest preference of the users cannot be accurately obtained, and the users are difficult to have good recommendation effect.
(3) A dynamically adaptive user preference learning model is lacking. LBNS users have different sign-in behavior characteristics, however, the existing interest point recommendation method does not consider the interest preference diversity of the users, and adopts a single preference learning model to carry out uniform preference mining and recommendation on all the users, so that the learning model of the user preference is poor in flexibility, different preferences of the users cannot be better learned, and the recommendation performance is reduced.
Disclosure of Invention
Aiming at the defects of the prior art, the invention aims to provide a self-adaptive interest point recommendation method based on long-term and short-term preferences of a user, and solves the problems that the interest point recommendation in the prior art is difficult to reflect the long-term and short-term preferences of the user and has poor recommendation effect.
The technical scheme of the invention is as follows: a self-adaptive interest point recommendation method based on long-term and short-term preference of a user comprises the following steps:
step 1, constructing a track sequence set for a user to sign in based on an LBNS historical sign-in data set, and dividing the track sequence into a historical sign-in track sequence and a recent sign-in track sequence by adopting a dynamic time window according to a time interval of sign-in records;
step 2, learning the historical preference of the user by adopting an LSTM model fused with time factors based on the historical sign-in track sequence, learning the recent preference of the user by adopting an RNN model combined with space-time factors based on the recent sign-in track sequence, and obtaining the long-term and short-term interest preference of the user by combining the historical preference and the recent preference;
and step 3: and performing probability prediction on the candidate interest points by combining the long-term and short-term interest preferences, and recommending a plurality of interest points with the highest probability value ranking to the user.
Further, in step 1, the check-in records are sorted by check-in time according to the time interval of the check-in records by adopting dynamic time window division, the time interval of adjacent check-in records in the sorted check-in records is calculated, the maximum time interval is used as the dynamic time window, the check-in track sequence of the earliest first check-in to the dynamic time window is defined as the historical check-in track sequence, and the check-in track sequence after the dynamic time window is defined as the recent check-in track sequence.
Further, dividing the recent check-in records of the recent check-in track sequence of the users into active users and inactive users, calculating the similarity of the active users and the inactive users, and selecting the recent check-in track sequence of a plurality of active users with the highest similarity for the inactive users to be merged into the recent check-in track sequence of the inactive users.
Further, when dividing active users and inactive users, setting an active threshold value and an inactive threshold value according to an average recent check-in record number of recent check-in track sequences of all users, wherein the active threshold value is greater than the average recent check-in record number, the inactive threshold value is less than the average recent check-in record number, users whose recent check-in record number of the recent check-in track sequences reaches the active threshold value are active users, and users whose recent check-in record number of the recent check-in track sequences does not reach the inactive threshold value are inactive users.
Further, when the historical sign-in track sequence is used for learning the historical preference of the user by adopting the LSTM model fused with the time factor, the cell state c of the LSTM model fused with the time factor k Is updated as follows
Figure BDA0003882182380000021
Figure BDA0003882182380000022
i k ＝σ(W i [h k-1 ,x k ]+b i )
f k ＝σ(W f [h k-1 ,x k ]+b f )
T k ＝σ(W q q lk +W t s k +b t )
Figure BDA0003882182380000023
Is a matrix of weights that is a function of,
Figure BDA0003882182380000024
as a bias vector, x k ＝[s k ,q lk ]Is the input of the LSTM model of the fusion time factor,
Figure BDA0003882182380000025
an embedded representation of the location is represented by,
Figure BDA0003882182380000026
a feature vector representing a time interval of two adjacent check-in records in the historical check-in trace sequence,
the historical preference is p k ＝tanh(V l c k )，
Figure BDA0003882182380000027
The LSTM model for the fusion time factor requires learned parameters.
Further, when the RNN model combined with the space-time factor is adopted to learn the recent preference of the user based on the recent sign-in track sequence, the hidden state of the RNN model combined with the space-time factor is updated to
Figure BDA0003882182380000031
Figure BDA0003882182380000032
Is a point of interest l i An embedded representation of (a);
Figure BDA0003882182380000033
the feature vectors are the time intervals of two adjacent check-in records in the recent check-in track sequence;
Figure BDA0003882182380000034
feature vectors of distance intervals of two adjacent interest points in the recent check-in track sequence;
Figure BDA0003882182380000035
for inputting points of interest l i And the state information of the post model is updated, the state is represented by a d-dimensional feature vector and is used for recording feature information of a recent sign-in track sequence, sigma is a sigmod activation function,
Figure BDA0003882182380000036
the weight matrix, which is relevant, is the parameter to be learned by the model,
the recent preference is p i ＝tanh(V s h i )，
Figure BDA0003882182380000037
The weight matrix for correlation is the parameter that the model is to learn.
Further, the distance interval between two adjacent interest points in the recent check-in track sequence is d i,i-1
Figure BDA0003882182380000038
C＝sin L at i *sin L at i-1 +cos L at i *cos L at i-1 *cos(Lon i-1 -Lon i )
Where R represents the mean radius of the earth, pi is the circumference ratio, l i ＝{Lon i ,Lat i And l i-1 ＝{Lon i-1 ,Lat i-1 And represents the longitude and latitude of two adjacent interest points.
Further, the user's long-short term interest preference is p u ，p u ＝p i +p k ，p i As recent preferences of the user, p k Is the user's historical preferences.
Further, the recommendation probability S of the user u to the interest point l during the probability prediction in step 3 u,l
S u,l ＝p u T q l
p u For the long-short term interest preference of the user obtained in step 2,
Figure BDA0003882182380000039
belongs to L for interest point L can Is embedded in the representation, L can Is a candidate interest point set.
Compared with the prior art, the invention has the advantages that:
1. according to the invention, in the division of the user check-in track, a dynamic time window is adopted to carry out self-adaptive division on the user interest point check-in track sequence, and the generated historical check-in track sequence and the recent check-in track sequence can obviously and accurately reflect the long-term preference and the short-term preference of the user.
2. The user sign-in track self-adaptive filling processing method provided by the invention can fill the recent sign-in record of the active similar user into the recent sign-in track sequence of the inactive user, effectively utilizes the place association and the user association, and solves the problems of short recent sign-in track and cold start of the user.
3. The self-adaptive learning model for the long-term and short-term interest preference of the user, provided by the invention, adopts an LSTM model with fusion time factors for a historical sign-in track sequence, and accurately learns and acquires the long-term stable historical preference of the user; and (3) adopting an RNN model combined with space-time factors for the recent sign-in track sequence to accurately learn and acquire the recent preference of the user. The self-adaptive learning model can more accurately, fully and effectively learn the long-term and short-term interest preferences of the user, thereby being beneficial to making personalized interest point recommendation for the user.
Drawings
Fig. 1 is a schematic flow chart of a self-adaptive interest point recommendation method based on long-term and short-term preferences of a user.
FIG. 2 is a diagram illustrating a process of learning the user's historical preference by the LSTM model with time factors fused.
FIG. 3 is a diagram illustrating a process of learning recent preferences of a user by an RNN model in combination with spatiotemporal factors.
Detailed Description
The present invention is further illustrated by the following examples, which are not intended to limit the scope of the invention.
Please refer to fig. 1, the method for adaptively recommending a point of interest based on long-term and short-term preferences of a user according to the present invention includes the following steps:
step 1, self-adaptive generation of long-term and short-term sign-in tracks of users. Based on the LBNS historical sign-in data set, a track sequence set of user sign-in is constructed, a historical sign-in track sequence and a recent sign-in track sequence are obtained by adopting self-adaptive division of sign-in tracks, and beneficial expansion is carried out by adopting a self-adaptive filling processing method of sign-in tracks aiming at the problem that the recent sign-in track sequence of the user is short. And finally obtaining a historical sign-in track sequence and a recent sign-in track sequence of the user.
And 2, self-adaptive learning of long-term and short-term interest preference of the user. And aiming at the historical sign-in track sequence of the user, learning the historical preference of the user by adopting an LSTM model fused with time factors. Meanwhile, aiming at the recent sign-in track sequence of the user, the recent preference of the user is learned by adopting an RNN model combined with space-time factors. Finally, the long-term and short-term interest preference of the user is obtained.
And step 3, recommending personalized interest points. And calculating recommendation probability of the candidate places by combining long-term and short-term interest preference results of the users, and recommending top-n places to the users after descending order, so as to realize personalized interest point recommendation.
In practical application, the specific implementation process of the self-adaptive interest point recommendation method based on the long-term and short-term preference of the user is as follows:
to model the user's long-short term interest preferences, the entire sequence of check-in trajectories C of the user is required u According to a time window tw u The method is divided into two parts: historical sign-in trajectory sequence
Figure BDA0003882182380000041
And recent sign-in trajectory sequence
Figure BDA0003882182380000042
The sign-in track and the sign-in frequency of each user are different, so that the personalized characteristics of the users can be reflected by adopting the dynamic time window for self-adaptive division, and the specific strategy for self-adaptively dividing the sign-in track of the users in the step 1 is as follows:
first, a user check-in track sequence set is constructed. Let a user U be U, and the record of check-in is represented as
Figure BDA0003882182380000043
Figure BDA0003882182380000044
Respectively representing user u, the place id of check-in, longitude, latitude, check-in time and check-in date. Sorting in ascending order according to date and time to obtain all sign-in track sequence sets of the user u
Figure BDA0003882182380000045
Wherein every two adjacent records are adjacent check-in.
Then, the dynamic time window tw of the user is solved u . The time interval, i.e., the time difference, of all adjacent check-ins of user u is calculated, as shown in equation (1). Then the dynamic time window tw for user u u A time interval defined as the maximum, that is, the maximum time difference is taken as the time point for dividing the track sequence, as shown in formula (2).
Δt i ＝t i -t i-1 ＝(date i -date i-1 )×24+(time i -time i-1 ) (1)
tw u ＝max(Δt 2 ,Δt 3 ,…,Δt i ) (2)
And finally, dividing the user check-in track sequence. Set of all sign-in trajectory sequences for user u
Figure BDA0003882182380000051
Figure BDA0003882182380000052
According to a dynamic time window tw u Divide by the earliest first sign-in to tw u Sequence of check-in tracks at time points
Figure BDA0003882182380000053
Defining a historical sign-in track sequence to reflect the long-term stable historical preference of the user; tw u Sequence of check-in trajectories after a point in time
Figure BDA0003882182380000054
Defining a recent sign-in track sequence which reflects recent preference of a user; and satisfies the conditions
Figure BDA0003882182380000055
The short-term interest of the user plays an important role in point of interest recommendation, but some users have short recent check-in track sequences, so that the model is difficult to accurately acquire the recent preference of the users. Therefore, the invention provides a track sequence self-adaptive filling method, which fills the recent check-in records of active similar users into the recent check-in track sequence of the inactive users, so as to effectively utilize place association and user association and solve the problems of short recent check-in track and cold start of the users.
Active and inactive users: order to
Figure BDA0003882182380000056
And (3) representing the number of records in the recent check-in track sequence of the user u, and av is the average number of recent check-in records of all users, as shown in formula (3). Active user u ac Defining as the users with the recent sign-in record number larger than 2av, and correspondingly generating an active user set U ac (ii) a Active user u in Correspondingly generating an inactive user set U for users with the number of recent check-in records smaller than av/2 in 。
Figure BDA0003882182380000057
For inactive users u in Recent sign-in track
Figure BDA0003882182380000058
And (3) performing adaptive filling treatment, wherein the specific process is as follows:
first, an inactive user u is calculated in ∈U in And active user u ac ∈U ac Similarity SU of in,ac As shown in equation (4), the more common places the users have gone to, the higher the similarity. Wherein r is u,l Is a binary value indicating whether the user u checked in at the point of interest l, i.e. if the user u checked in at the point of interest l, then r u,l =1, otherwise r u,l ＝0。
SU in,ac ＝∑ l∈L r in,l ·r ac,l (4)
Secondly, sorting the similarity results in a descending order, and then arranging the similarity from high to low to obtain the user u in G = { u = similar users of (G) } ac1 ,u ac2 ,…,u aci }. Obtaining recent check-in sequences of the top 10 most similar users in the set G, combining the recent check-in sequences to form a large check-in sequence, and deleting the repeated place sequence to obtain the check-in sequence
Figure BDA0003882182380000059
Figure BDA00038821823800000510
Thirdly, the sequence is
Figure BDA00038821823800000511
Filling in to inactive users u in Recent check-in track sequence
Figure BDA00038821823800000512
In generating a new recent check-in sequence, i.e.
Figure BDA00038821823800000513
Finally, if there are more check-in records in the new sequence, it is beyond the scope of model processing, i.e.
Figure BDA00038821823800000514
Wherein delta max The maximum number of check-in records required for the check-in sequence, then the check-in sequence
Figure BDA00038821823800000515
The record with earlier check-in time is intercepted, and only delta is reserved max A latest check-in record, i.e.
Figure BDA0003882182380000061
Wherein
Figure BDA0003882182380000062
Represented as the earlier in time check-in sequence in set G.
In step 2 of the method, a long-term memory model LSTM integrating time factors is provided to obtain the historical preference of the user, a forgetting gate of the LSTM is utilized to filter some unimportant characteristic information in a user historical sign-in track sequence, a hidden cell unit is adopted to memorize the long-term stable interest characteristics of the user, and meanwhile, the influence of the time factors on the long-term interest of the user is considered. And the long-term interest of the user is less influenced by the geographic space, so the influence of the spatial factor is not considered.
The learning process of the user's historical preferences is shown in fig. 2.
Two types of information need to be input each time: point of interest information, temporal context information. Prior to modeling long-term interest of user u, the sequence of tracks is first checked in from their history
Figure BDA0003882182380000063
In the method, a sequence l of the check-in interest points is extracted 1 →l 2 →…→l k-2 →l k-1 →l k . For any two adjacent check-in records, calculating the time interval delta t of the adjacent check-in records k As shown in equation (1). By using
Figure BDA0003882182380000064
Represents a point of interest l k Is to be used to represent the embedded representation of,
Figure BDA0003882182380000065
representing a time interval Δ t k Is determined by the feature vector of (a),
Figure BDA0003882182380000066
is the input of the model.
The new candidate state after each input is only related to the current input and the state passed by the previous step, so the candidate state is updated by formula (5)
Figure BDA0003882182380000067
Figure BDA0003882182380000068
Candidate states
Figure BDA0003882182380000069
But is merely used to indicate the current interests of the user,
Figure BDA00038821823800000610
is a matrix of the weights that is,
Figure BDA00038821823800000611
as bias vector, cell state
Figure BDA00038821823800000612
The historical information of the access points of interest of the user is memorized, and the historical preference of the user is reflected. Considering the characteristic of long-term interest decay with time, the long-term interest accumulation speed of the user is controlled by adding a time gate in the LSTM, some early interests of the user are slowly decayed, the long-term stable interests of the user are reserved, and the state c k Is as shown in equation (6).
Figure BDA00038821823800000613
i k ＝σ(W i [h k-1 ,x k ]+b i ) (7)
f k ＝σ(W f [h k-1 ,x k ]+b f ) (8)
Wherein the content of the first and second substances,
Figure BDA00038821823800000614
respectively an input door and a forgetting door,
Figure BDA00038821823800000615
is a matrix of the weights that is,
Figure BDA00038821823800000616
is a bias vector. The two gates control c mainly according to the currently input interest point and the previous state k Is updated as shown in equations (7) and (8).
Figure BDA00038821823800000617
Expressed as a time gate, the attenuation of long-term interest is controlled according to the input interest points and the time context, not only the forgetting gate can filter the long-term interest, but also the time gate can filter some interests which are not updated in the early stage, so that the long-term stable interest of the user can be reserved, and the formula (9) is a specific implementation mode of the time gate.
Figure BDA0003882182380000071
In the formula (I), the compound is shown in the specification,
Figure BDA0003882182380000072
is a weight matrix, is a parameter required to be learned by the model,
Figure BDA0003882182380000073
is a bias vector.
When the user's entire history sign-in track
Figure BDA0003882182380000074
After learning is complete, the template may be acquiredCell status of type c k ，c k Features that represent the long-term interest of the user, unlike the output of conventional LSTM, only the long-term accumulated interest features of the user are required here, for which the historical preferences of the user are derived using equation (10). In the formula (I), the compound is shown in the specification,
Figure BDA0003882182380000075
representing a sequence of tracks by check-in
Figure BDA0003882182380000076
The learned historical preferences of the user,
Figure BDA0003882182380000077
parameters that need to be learned for the model.
p k ＝tanh(V l c k ) (10)
Since recent preference of the user is greatly influenced by time and place factors, the influence of time and distance context on interest preference of the user is fully considered in modeling, and if the time interval of check-in of two adjacent interest points in the check-in track sequence is shorter, the two interest points are continuous check-in and have higher relevance. Similarly, if the distance interval between two adjacent interest points is smaller, it also indicates that the two interest points have higher relevance and continuity, because the user is generally more likely to access nearby interest points, the activity of the user is often limited to a certain area, and the access behavior is more affected by regions.
Short-term interest modeling for user u utilizes a recent check-in trajectory sequence
Figure BDA0003882182380000078
Extracting interest points from each check-in record in sequence to obtain an interest point sequence l i-k →…→l i-2 →l i-1 →l i . For any two adjacent check-in records
Figure BDA0003882182380000079
Calculating its time interval Δ t i As shown in equation (1). For is toAny two adjacent interest points calculate the distance interval delta d of any two adjacent interest points by utilizing the longitude and latitude of the two adjacent interest points i ＝d i,i-1 As shown in equation (11). Wherein the places of two adjacent check-in and the respective longitude and latitude coordinates are l i ＝{Lon i ,Lat i And l i-1 ＝{Lon i-1 ,Lat i-1 R =6371Km means the average radius of the earth, pi =3.14 is the circumferential rate.
Figure BDA00038821823800000710
C＝sin L at i *sin L at i-1 +cos L at i *cos L at i-1 *cos(Lon i-1 -Lon i ) (12)
The invention fuses the time context and distance context information implicit in the sequence into a Recurrent Neural Network (RNN) model for modeling the recent check-in track sequence of the user and the learning process of the short-term interest preference of the user, as shown in figure 3.
There are three types of input information at the input layer: the information of the current interest point, the check-in time interval and the distance interval information of the previous interest point, and the update of the node state are not only related to the input of the current moment and the output state of the previous moment, but also related to the check-in time interval and the distance interval of the previous interest point, so that the influence of time and distance context on the interest preference of the user is reflected, and the specific hidden state update is shown as a formula (13).
Figure BDA00038821823800000713
p i ＝tanh(V s h i ) (14)
Wherein the content of the first and second substances,
Figure BDA00038821823800000711
is a point of interest l i Is a d-dimensional feature vector;
Figure BDA00038821823800000712
is the time interval Δ t i The feature vector of (2);
Figure BDA0003882182380000081
is a distance interval Δ d i The feature vector of (2);
Figure BDA0003882182380000082
for inputting points of interest l i The state information of the post model is updated, the state is represented by a d-dimensional feature vector and is used for recording the feature information of a recent sign-in track sequence;
Figure BDA0003882182380000083
for inputting points of interest l i The user interest preference output by the rear model is also a d-dimension feature vector. Sigma is the sigmod activation function,
Figure BDA0003882182380000084
the weight matrix is a parameter to be learned by the model, and the user interest preference characteristic p is finally output i I.e. the recent preferences of the user.
After learning the long-term and short-term interest preferences of the user, the long-term and short-term interest preference result of the user is obtained by using the formula (15)
Figure BDA0003882182380000085
Importantly, the long-term interest preference feature and the short-term interest preference feature are dynamic, and can be changed dynamically along with the increase of the user check-in track sequence, so that the interest features of the user can be fully reflected.
p u ＝p i +p k (15)
When calculating the probability of the candidate points of interest in step 3, only considering the points within 20km of the current position of the user as the candidate recommended point of interest set L in order to reduce the calculation overhead and complexity of the system can ，
Figure BDA0003882182380000086
Belongs to L for interest point L can Is expressed in terms of user long-and-short-term interest preferences p u Calculating to obtain the recommendation probability S of the user u to the interest point l u,l As shown in equation (16).
S u,l ＝p u T q l (16)
If the interest point l is consistent with the preference of the interest point of the user, the values of the corresponding dimensions of the two feature representation vectors are close, so that the interest scores obtained after the two vectors are operated are higher, and otherwise, the interest scores are lower. According to a recommended probability value S u,l Sorting in a descending order, and selecting the places with the highest top-n probability values to recommend to the user.

Claims (10)

1. A self-adaptive interest point recommendation method based on long-term and short-term preference of a user is characterized by comprising the following steps:
step 1, constructing a track sequence set for a user to check in based on an LBNS historical check-in data set, and dividing a historical check-in track sequence and a recent check-in track sequence by adopting a dynamic time window according to a time interval of check-in records;
step 2, learning the historical preference of the user by adopting an LSTM model fused with time factors based on the historical sign-in track sequence, learning the recent preference of the user by adopting an RNN model combined with space-time factors based on the recent sign-in track sequence, and obtaining the long-term and short-term interest preference of the user by combining the historical preference and the recent preference;
and step 3: and performing probability prediction on the candidate interest points by combining the long-term and short-term interest preferences, and recommending a plurality of interest points with the highest probability value ranking to the user.
2. The method as claimed in claim 1, wherein the step 1 employs dynamic time window division according to time intervals of check-in records, the check-in records are sorted by check-in time, time intervals of adjacent check-in records in the sorted check-in records are calculated, a maximum time interval is used as the dynamic time window, a check-in track sequence of the earliest first check-in to the dynamic time window is defined as the historical check-in track sequence, and a check-in track sequence after the dynamic time window is defined as the recent check-in track sequence.
3. The adaptive point-of-interest recommendation method based on long-term and short-term preferences of users according to claim 1, wherein the active users and the inactive users are divided according to the number of recent check-in records of the recent check-in track sequence of users, the similarity between the active users and the inactive users is calculated, and the recent check-in track sequence of a plurality of active users with the highest similarity is selected for the inactive users and merged into the recent check-in track sequence of the inactive users.
4. The adaptive point of interest recommendation method based on long-term and short-term preference of users according to claim 3, wherein when dividing active users and inactive users, setting an active threshold and an inactive threshold according to an average number of recent check-in records of recent check-in track sequences of all users, wherein the active threshold is greater than the average number of recent check-in records, the inactive threshold is less than the average number of recent check-in records, wherein users whose number of recent check-in records of check-in track sequences reaches the active threshold are active users, and users whose number of recent check-in records of recent check-in track sequences does not reach the inactive threshold are inactive users.
5. The adaptive point-of-interest recommendation method based on long-term and short-term preference of users according to claim 3, wherein when merging the recent check-in track sequences of several active users into the recent check-in track sequences of the inactive users, check-in records with the same check-in place in the recent check-in track sequences of the active users are eliminated, and when the merged recent check-in track sequences of the inactive users exceed the maximum value of the number of check-in records, the number of check-in records with earlier elimination time reaches the maximum value of the number of check-in records.
6. The method as claimed in claim 1, wherein when learning the historical preference of the user by using the time-factor-fused LSTM model based on the historical check-in trajectory sequence, the cell state c of the time-factor-fused LSTM model is used k Is updated as follows
Figure FDA0003882182370000021
Figure FDA0003882182370000022
i k ＝σ(W i [h k-1 ，x k ]+b i )
f k ＝σ(W f [h k-1 ，x k ]+b f )
Figure FDA0003882182370000023
Figure FDA0003882182370000024
Is a matrix of weights that is a function of,
Figure FDA0003882182370000025
in order to be a vector of the offset,
Figure FDA00038821823700000217
is the input of the LSTM model of the fusion time factor,
Figure FDA0003882182370000026
an embedded representation of the location is represented by,
Figure FDA0003882182370000027
a feature vector representing a time interval of two adjacent check-in records in the historical check-in trace sequence,
the historical preference is p k ＝tanh(V l c k )，
Figure FDA0003882182370000028
The LSTM model for the fusion time factor requires learned parameters.
7. The method of claim 1, wherein when learning the recent preference of the user based on the recent sign-in trajectory sequence using a spatio-temporal factor-integrated RNN model, the spatio-temporal factor-integrated RNN model is updated to a hidden state based on the recent sign-in trajectory sequence
Figure FDA0003882182370000029
Figure FDA00038821823700000210
Is a point of interest l i The embedded representation of (a);
Figure FDA00038821823700000211
the feature vectors are the time intervals of two adjacent check-in records in the recent check-in track sequence;
Figure FDA00038821823700000212
feature vectors of distance intervals of two adjacent interest points in the recent check-in track sequence;
Figure FDA00038821823700000213
for inputting points of interest l i And state information updated by the post model, wherein the state is represented by a d-dimensional feature vector and is used for recording a recent sign-in track sequenceSigma is a sigmod activation function,
Figure FDA00038821823700000214
the weight matrix, which is relevant, is the parameter to be learned by the model,
the recent preference is p i ＝tanh(V s h i )，
Figure FDA00038821823700000215
The weight matrix for correlation is the parameter that the model is to learn.
8. The adaptive point-of-interest recommendation method based on long-term and short-term preference of user according to claim 7, wherein the distance interval between two adjacent points of interest in the recent check-in trajectory sequence is d i，i-1
Figure FDA00038821823700000216
C＝sin L at i *sinL at i-1 +cos L at i *cos L at i-1 *cos(Lon i-1 -Lon i )
Where R represents the mean radius of the earth, pi is the circumference ratio, l i ＝{Lon i ，Lat i And l i-1 ＝{Lon i-1 ，Lat i-1 And represents the longitude and latitude of two adjacent interest points.
9. The adaptive point-of-interest recommendation method based on user long-short term preference according to claim 1, wherein the user long-short term interest preference is p u ，p u ＝p i +p k ，p i As recent preferences of the user, p k Is the user's historical preferences.
10. The adaptive point-of-interest recommendation method based on long-term and short-term preference of user as claimed in claim 1, wherein said step 3 is performed with probabilityRecommendation probability S of user u to interest point l in prediction u，l
S u，l ＝p u T q l
p u For the long-short term interest preference of the user obtained in step 2,
Figure FDA0003882182370000031
belongs to L for interest point L can Is embedded in the representation, L can Is a candidate interest point set.
CN202211232832.7A 2022-10-10 2022-10-10 Self-adaptive interest point recommendation method based on long-term and short-term preference of user Pending CN115495661A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202211232832.7A CN115495661A (en) 2022-10-10 2022-10-10 Self-adaptive interest point recommendation method based on long-term and short-term preference of user

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202211232832.7A CN115495661A (en) 2022-10-10 2022-10-10 Self-adaptive interest point recommendation method based on long-term and short-term preference of user

Publications (1)

Publication Number Publication Date
CN115495661A true CN115495661A (en) 2022-12-20

Family

ID=84474079

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202211232832.7A Pending CN115495661A (en) 2022-10-10 2022-10-10 Self-adaptive interest point recommendation method based on long-term and short-term preference of user

Country Status (1)

Country Link
CN (1) CN115495661A (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117633371A (en) * 2024-01-25 2024-03-01 云南大学 Recommendation method, device and readable storage medium based on multi-attention mechanism

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117633371A (en) * 2024-01-25 2024-03-01 云南大学 Recommendation method, device and readable storage medium based on multi-attention mechanism
CN117633371B (en) * 2024-01-25 2024-04-12 云南大学 Recommendation method, device and readable storage medium based on multi-attention mechanism

Similar Documents

Publication Publication Date Title
CN110399565B (en) Recursive neural network interest point recommendation method based on space-time period attention mechanism
CN113111581B (en) LSTM trajectory prediction method combining space-time factors and based on graph neural network
CN107247961B (en) Track prediction method applying fuzzy track sequence
CN107018493B (en) Geographic position prediction method based on continuous time sequence Markov model
CN108829766B (en) Interest point recommendation method, system, equipment and computer readable storage medium
CN106850289B (en) Service combination method combining Gaussian process and reinforcement learning
CN113139140B (en) Tourist attraction recommendation method based on space-time perception GRU and combined with user relationship preference
CN111695737B (en) LSTM neural network-based group target traveling trend prediction method
CN106568445A (en) Indoor track prediction method based on bidirectional circulation neural network
CN115495661A (en) Self-adaptive interest point recommendation method based on long-term and short-term preference of user
CN116108984A (en) Urban flow prediction method based on flow-POI causal relationship reasoning
CN110222273B (en) Business point promotion method and system in social network based on geographic community
CN114637911A (en) Next interest point recommendation method of attention fusion perception network
CN114461931A (en) User trajectory prediction method and system based on multi-relation fusion analysis
CN115952355A (en) Next interest point recommendation system based on space-time power law attention
CN115774819A (en) Interest point recommendation method and system based on hierarchical recurrent neural network
CN111988744B (en) Position prediction method based on user moving mode
Xu et al. PSARE: A RL-based online participant selection scheme incorporating area coverage ratio and degree in mobile crowdsensing
CN112529294A (en) Training method, medium and equipment for individual random trip destination prediction model
CN115762147B (en) Traffic flow prediction method based on self-adaptive graph meaning neural network
CN113032688B (en) Method for predicting access position of social network user at given future time
CN116257704A (en) Point-of-interest recommendation method based on user space-time behaviors and social information
CN115510317A (en) Cloud-edge collaborative interest point recommendation method and device
CN114861936A (en) Feature prototype-based federated incremental learning method
CN109902883B (en) Sign-in position prediction method based on personalized hierarchical kernel density estimation

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