CN115687801A - Position recommendation method based on position timeliness characteristics and time perception dynamic similarity - Google Patents
Position recommendation method based on position timeliness characteristics and time perception dynamic similarity Download PDFInfo
- Publication number
- CN115687801A CN115687801A CN202211180601.6A CN202211180601A CN115687801A CN 115687801 A CN115687801 A CN 115687801A CN 202211180601 A CN202211180601 A CN 202211180601A CN 115687801 A CN115687801 A CN 115687801A
- Authority
- CN
- China
- Prior art keywords
- time
- user
- recommendation
- time slot
- location
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 238000000034 method Methods 0.000 title claims abstract description 138
- 230000008447 perception Effects 0.000 title claims abstract description 32
- 230000032683 aging Effects 0.000 claims abstract description 53
- 239000011159 matrix material Substances 0.000 claims abstract description 48
- 238000001914 filtration Methods 0.000 claims abstract description 32
- 238000011156 evaluation Methods 0.000 claims abstract description 30
- 238000005065 mining Methods 0.000 claims abstract description 6
- 230000007246 mechanism Effects 0.000 claims abstract description 5
- 238000005516 engineering process Methods 0.000 claims description 20
- 230000006870 function Effects 0.000 claims description 8
- 238000012935 Averaging Methods 0.000 claims description 7
- 238000006243 chemical reaction Methods 0.000 claims description 7
- 235000006679 Mentha X verticillata Nutrition 0.000 claims description 6
- 235000002899 Mentha suaveolens Nutrition 0.000 claims description 6
- 235000001636 Mentha x rotundifolia Nutrition 0.000 claims description 6
- 230000009471 action Effects 0.000 claims description 5
- 230000003679 aging effect Effects 0.000 claims description 3
- 238000009499 grossing Methods 0.000 claims description 3
- 230000002123 temporal effect Effects 0.000 claims description 2
- 238000003491 array Methods 0.000 claims 1
- 238000010606 normalization Methods 0.000 claims 1
- 230000006399 behavior Effects 0.000 description 11
- 230000008569 process Effects 0.000 description 7
- 238000010586 diagram Methods 0.000 description 5
- 230000008901 benefit Effects 0.000 description 4
- 238000013461 design Methods 0.000 description 3
- 238000011161 development Methods 0.000 description 3
- 230000018109 developmental process Effects 0.000 description 3
- 238000011160 research Methods 0.000 description 3
- 238000004364 calculation method Methods 0.000 description 2
- 238000000354 decomposition reaction Methods 0.000 description 2
- 230000007547 defect Effects 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 238000012545 processing Methods 0.000 description 2
- 235000010627 Phaseolus vulgaris Nutrition 0.000 description 1
- 244000046052 Phaseolus vulgaris Species 0.000 description 1
- 206010044565 Tremor Diseases 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 238000013473 artificial intelligence Methods 0.000 description 1
- 235000021152 breakfast Nutrition 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 238000004891 communication Methods 0.000 description 1
- 230000001747 exhibiting effect Effects 0.000 description 1
- 230000004927 fusion Effects 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 230000003993 interaction Effects 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000001737 promoting effect Effects 0.000 description 1
- 238000012163 sequencing technique Methods 0.000 description 1
- 238000012360 testing method Methods 0.000 description 1
- 230000036962 time dependent Effects 0.000 description 1
- 230000007704 transition Effects 0.000 description 1
Images
Landscapes
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The invention discloses a position recommendation method based on position timeliness characteristics and time perception dynamic similarity, which comprises the following steps: generating a user-time-position three-dimensional scoring matrix according to the original check-in data set; and extracting a user-position two-dimensional scoring matrix of each time slot, and calculating the dynamic similarity of the positions in different time slots. Predicting a score of the unaccessed address using an improved project-based collaborative filtering method; calculating the aging characteristic values of the positions in different time slots based on time perception; realizing personalized probability density modeling by using a kernel density estimation method, and mining geographical influence; constructing a scoring prediction mechanism fusing user history preference, geographic distance influence, position aging characteristics and position dynamic similarity, and recommending the position with higher final prediction score to the user; defining a timeliness evaluation system of the recommendation system, and comparing the prediction accuracy and recommendation timeliness of different recommendation systems. The invention has strong transportability and wide industrial application prospect.
Description
Technical Field
The invention relates to a position recommendation method based on position timeliness characteristics and time perception dynamic similarity, and belongs to the technical field of artificial intelligence and machine learning.
Background
In recent years, with the rapid popularization of mobile intelligent terminals and the rapid development of wireless communication technologies, location-based Social Networks (lbs ns) are becoming popular worldwide, such as the famobook, twitter, yelp and domestic bean, surf microblog, popular comment, etc., which are well-known foreign Social network platforms. As an internet bridge connecting the physical world with the virtual network, the location social network does not simply add location information rigidly to the traditional social network, but rather forms a more complex social network through the reconstruction of the traditional social network platform and data structures. In the position social network, as a consumer of information, a user can learn related knowledge through information shared by others, and finds out interested merchants and services at any time and place, thereby greatly facilitating daily life of people. As a producer of information, a user can also actively share own consumption experience through the sign-in behavior of the intelligent terminal at any time and place.
The advantages and convenience of the location social network attract a large number of users and merchants to perform information interaction in the location social network platform, and a large amount of information such as locations, pictures, audio, videos and comments is accumulated. The mass data provides new opportunities for the industry and academia to research the behavior preference of the user, but also increases the difficulty of the user to accurately find the interested goods or services. In order to solve the problem of Information Overload (Information Overload) in the big data era, a recommendation System (recommendation System) has become an indispensable technical service means in a location social network as an important Information retrieval tool. Through analysis of large-scale spatiotemporal data in the position social network, the recommendation system can mine behavior patterns, rules and preferences of the user, and recommend corresponding commodities, services and the like to the user according to the existing item information.
In the Location social network, location Recommendation (Location Recommendation) is a research hotspot and emerging application in the field of Recommendation systems. The main task of location recommendation is to fully mine the user's check-in preferences at different locations, recommending to the user where they may be interested and would like to check-in the future. The position recommendation system can help the user to explore a new interesting place in a city on one hand, and can also help a merchant accurately place advertisements to a target customer on the other hand, and an unprecedented business opportunity is provided for the merchant. At present, domestic popular social network platforms (such as American groups, hungry, tremble, microblog, weChat friend circles and the like) provide position recommendation services.
As an important component of a recommendation system, no matter development process or key technology, the position recommendation system is in a way similar to a traditional recommendation system, and some position recommendation systems consider positions as common items similar to movies, music and the like and generate recommendation results by utilizing a traditional recommendation method. The main technologies of the conventional recommendation system include two aspects, namely a content-based method and a collaborative filtering method, wherein the collaborative filtering method is divided into a memory-based collaborative filtering algorithm and a model-based collaborative filtering algorithm. Content-based Recommendation algorithms (Content-based Recommendation) utilize description information of commodities and individuals to match description attributes of the commodities with personal information of users, interest descriptions and the like, have a particularly obvious advantage in solving the cold start problem, but have great limitations due to difficulty in extracting useful information and features when recommending rich media (such as videos, pictures and music). The memory-based collaborative filtering algorithm mainly includes a user-based collaborative filtering algorithm (UBCF) and an item-based collaborative filtering algorithm (IBCF). The UBCF method recommends the commodities visited by the user similar to the recommended user to the user by calculating the similarity between the user and the user. The IBCF method recommends to a user similar items of an item that the user has visited by calculating the similarity between items. The UBCF and IBCF methods are similar, but compared with a large number of commodities, the number of commodities visited by a user is very small, which brings a certain difficulty to calculate the similarity of the users, and from the commodity perspective, since there are many records of visits of each commodity by the user, it is easier to find similar commodities. Therefore, the IBCF method tends to be superior to the UBCF method in terms of the recommended effect. The model-based collaborative filtering algorithm is a generic term of a class of algorithms, and a typical representative of the algorithm is a matrix decomposition algorithm, for example, a low-dimensional orthogonal matrix decomposed by using a Singular Value Decomposition (SVD) technology reduces noise on the basis of an original matrix, and can effectively reveal potential association of users and commodities.
Unlike recommendations for traditional items (e.g., movies, music, jokes, etc.), the subject of the location recommendation is an address with a geographic factor. User access to a down-line address is highly susceptible to complex contextual environments, exhibiting social, geographic, and temporal characteristics. For example, from the perspective of social relationships, a user's likeability of a location is often influenced by comments from friends when selecting the location. Users are more willing to trust the sharing of friends than strangers; from a geographical point of view, most check-ins occur in certain restricted areas, such as the user's address or areas around the office; from a time profile, check-in activities also exhibit some specific time patterns, such as a user checking in at a location near an office during the day and checking in at a bar, movie theater, or gym at night. These unique features make location recommendation different from traditional recommendation systems, and therefore, how to further introduce relationship context, location context and time context into traditional recommendation algorithms has become an urgent need for various social application platforms to recommend location lists of interest to users in real time.
The multi-source heterogeneous information such as rich check-in records, social relations, space-time data and the like in the position social network plays an important auxiliary role in user behavior modeling, and at present, some recommendation systems integrate different types of contexts into the position recommendation problem, but still have some defects and shortcomings, and the following points are summarized:
(1) Most position recommendation systems mainly analyze and mine context information such as user preference, social relationship, geographic influence and the like, a global recommendation list is provided for a user, and real-time recommendation cannot be realized according to the current time. However, the user's access needs to the location at different time periods are not the same, e.g., 8 am user may go to a breakfast shop, 16 pm user may go to a coffee shop near the office, and 20 pm user may be more likely to access movie theaters, bars, etc. locations. Therefore, the current requirements and preferences of the user are analyzed according to the time information, and the real-time position recommendation result is provided for the user, so that the recommendation accuracy and the user satisfaction can be effectively improved.
(2) Dynamic similarity of locations over different time periods is ignored. In the existing research, when the position similarity is mined, the time dimension dynamic characteristics of the position similarity are not considered, and a global similarity matrix is shared in different time periods. How to mine fine-grained characteristics of the position similarity and enable the position similarity to generate different position similarity matrixes according to time variables is a technical problem which needs to be solved urgently.
(3) The aging characteristics of the location are ignored. The aging characteristics represent the novelty and popularity of the location at the current time, and existing studies have ignored mining of location aging characteristics. In fact, locations with a high frequency of recent access are more popular with users than locations with a high frequency of early access, and there is a necessary correlation between the aging characteristics of the locations and the probability of access. Therefore, a clear definition and a calculation method must be provided for the aging characteristics of the location, and the aging characteristic result is applied to the recommendation process to improve the aging performance of the recommendation system.
(4) A time efficiency evaluation system for a location recommendation system is lacking. In a location social network, the access habits of users are particularly affected by location aging characteristics, for example, a mall will always attract more customers when it is just starting a business. Therefore, whether the position more conforming to the recent interest and preference of the user can be timely recommended to the user should be an important evaluation index in the evaluation of the position recommendation system. However, the existing recommendation technology mainly studies on the accuracy of recommendation results, and ignores the timeliness evaluation in the position recommendation scene.
The above-mentioned disadvantages of the existing location recommendation technology bring great disadvantages in the design, development, deployment and operation of different location social network platforms, and especially cause the service quality of the recommendation system to be reduced on the network platform of massive project information, thereby affecting the sales performance of the e-commerce system.
Disclosure of Invention
The invention aims to provide a position recommendation method based on position timeliness characteristics and time perception dynamic similarity aiming at the defects of the prior art, the method takes a real-time position recommendation system with high accuracy and strong timeliness as a target, considers the dynamic rule of the position similarity along with the time dimension change, innovatively excavates the position dynamic similarity based on time perception, and effectively improves the prediction accuracy of the position recommendation system. Meanwhile, the invention provides a method for defining and calculating the position timeliness characteristics, and the position timeliness characteristic value is fused in the recommendation process, so that the recommendation timeliness of the position recommendation system is improved. In addition, the system theory is taken as a theoretical basis, the evaluation system is taken as a necessary component of the recommendation system, the prediction accuracy of the recommendation system is measured, and meanwhile, the timeliness of the recommendation system is also investigated. The invention innovatively provides a timeliness evaluation system of the position recommendation system, and provides important technical support for quantifying the novelty of the output result of the position recommendation system.
The technical scheme adopted by the invention for solving the technical problems is as follows: dividing a day into 8 time slots, dividing a user-time-position three-dimensional scoring matrix according to the time slots, extracting a user-position two-dimensional scoring matrix corresponding to each time slot, and calculating the dynamic similarity of positions in different time slots by using Jaccard coefficients aiming at each scoring matrix; the traditional project-based collaborative filtering method is improved, and the score of the user on the inaccessible address is calculated by utilizing the dynamic similarity of the position; defining a position aging characteristic calculation method based on time perception, and excavating aging characteristic values of each position in each time slot one by one; fitting the personalized distribution of the sign-in of the user by using a kernel density estimation method, quantifying the influence of the geographic distance on the sign-in of the user, constructing a personalized probability density model, and calculating the sign-in probability of the user to the inaccessible address; and comprehensively considering the historical preference, the geographic distance influence, the position aging characteristic and the dynamic similarity of the positions of the users, calculating the final prediction scores under the fusion of the user context, the position context and the time context, sequencing the final prediction scores of all the inaccessible addresses, and recommending a plurality of positions with the top rank to the users.
The specific process of the method comprises the following steps:
step 1: collecting and sorting original sign-in data sets, converting specific time in the sign-in records into different time slots, counting sign-in times of users at each position in different time slots, and converting the sign-in records into a user-time-position three-dimensional scoring matrix according to a counting result.
Step 2: and segmenting the user-time-position three-dimensional scoring matrix according to the time slots, extracting the user-position two-dimensional scoring matrix in each time slot, and calculating the dynamic similarity of the positions in different time slots by using the Jaccard coefficient aiming at each scoring matrix. The traditional project-based collaborative filtering method is improved, and the dynamic similarity of the positions and the scores already completed by the user are utilized to predict the scores of the positions to the inaccessible addresses.
And 3, step 3: and respectively recording the specific time of each user in each time slot for accessing the position for the last time aiming at each position, and calculating the aging characteristic value of the position in each time slot for each user one by one. And averaging the position aging characteristic values of all users who visit the address in different time slots, and calculating the aging characteristic value of the position based on time perception.
And 4, step 4: and calculating the geographic distance between each position in the check-in data set according to the latitude and longitude information, realizing personalized probability density modeling by using a kernel density estimation method, and individually mining the influence of geographic features.
And 5: the influence of the user context, the position context and the time context on the check-in behavior of the user is comprehensively considered, a scoring prediction mechanism fusing the historical preference, the geographic distance influence, the position aging characteristic and the position dynamic similarity of the user is constructed, and a plurality of positions with high final prediction scoring are recommended to the user.
And 6: and (5) providing an timeliness evaluation index and defining a timeliness evaluation system of the recommendation system. And comparing the prediction accuracy and the recommendation timeliness of the recommendation system and other classical recommendation systems provided by the invention by using the accuracy evaluation index and the timeliness evaluation index respectively, and evaluating the accuracy and the effectiveness of the provided technology.
Has the advantages that:
1. the position recommendation method based on the position timeliness characteristics and the time perception similarity further considers the influence of time context on the sign-in behavior of the user on the basis of comprehensively considering context information such as user preference, geographic influence and the like, and expands the recommendation result from a user-position binary relation to a user-time-position three-dimensional model. The technical scheme provided by the invention analyzes the current requirements and preferences of the user according to the time information, and provides a real-time position recommendation result for the user dynamically, so that the user stickiness of the position social network platform is enhanced, a merchant can be helped to push advertisements for the user in real time, and greater commercial benefits are brought to the merchant.
2. The method and the device consider the dynamic rule that the position similarity changes along with the time dimension, innovatively excavate the position similarity based on time perception, realize self-adaptive recommendation on the basis of project-based collaborative filtering, and effectively improve the prediction accuracy of the position recommendation system. The technical scheme can not only be oriented to a project-based collaborative filtering system, but also be applied to a user-based collaborative filtering method, and has certain portability.
3. The invention innovatively excavates the aging characteristics of the position, and quantifies the degree of the attenuation of the attraction of the position to the user along with the time, so that the recommendation result can better accord with the aging preference of the user. The proposed recommendation method can provide more addresses which accord with recent interest and preference of the user, greatly improves the use satisfaction of the user on the social network platform, and has very important significance for practical application.
4. The method defines a timeliness evaluation system of the position recommendation system, provides technical indexes for detecting whether the recommendation result meets the recent interest preference of the user, and fills the blank of the timeliness evaluation field of the position recommendation system. In practical application, the timeliness evaluation result is fed back to the recommendation system, so that the innovation promoting capability of the recommendation system is favorably improved, and the robustness of the recommendation system is enhanced.
5. The method and the device can be applied to a position recommendation system, can be applied to the field of personalized recommendation of other traditional projects, and have strong portability and wide industrial application prospect.
Drawings
Fig. 1 is a flowchart of a location recommendation method based on location aging characteristics and time perception similarity according to the present invention.
Fig. 2 is a flowchart of specific steps of a location recommendation method based on location aging characteristics and time perception similarity according to the present invention.
Fig. 3 is a diagram illustrating the statistical results of the number of location visits and the number of users of the embodiment of the present invention.
Fig. 4 is a box diagram of Precision after the position recommendation method based on the position aging characteristics and the time perception similarity in the embodiment of the present invention is run for 100 times.
Fig. 5 is a box diagram of Recall rate Recall after 100 runs of the location recommendation method based on location aging characteristics and time perception similarity in the embodiment of the present invention.
Fig. 6 is a box diagram of a recommendation accuracy index F1 after the position recommendation method based on the position aging characteristics and the time perception similarity operates 100 times in the embodiment of the present invention.
FIG. 7 is a histogram of the Precision comparison of the recommendation method and the classical project-based collaborative filtering (IBCF), user-based collaborative filtering (UBCF), kernel density estimation-based access probability prediction method (KDE) in an embodiment of the present invention.
FIG. 8 is a histogram comparing Recall rates Recall of the recommendation method and classical project-based collaborative filtering (IBCF), user-based collaborative filtering (UBCF), and kernel density estimation-based access probability prediction (KDE) in an embodiment of the present invention.
FIG. 9 is a histogram comparing F1 values of the comprehensive accuracy index of the recommendation method and the classical project-based collaborative filtering (IBCF), user-based collaborative filtering (UBCF), and kernel density estimation-based access probability prediction (KDE) methods in the embodiment of the present invention.
FIG. 10 is a histogram of the time-dependent index value comparison of the recommendation method and a classical project-based collaborative filtering (IBCF), user-based collaborative filtering (UBCF), kernel density estimation-based access probability prediction method (KDE) in an embodiment of the present invention.
Detailed Description
The invention is described in further detail below with reference to the drawings.
As shown in fig. 1 and fig. 2, the specific process of the design and implementation of the present invention includes the following steps:
step 1: collecting and sorting original sign-in data sets, converting specific time in the sign-in records into different time slots, counting sign-in times of users at each position in different time slots, and converting the sign-in records into a user-time-position three-dimensional scoring matrix according to a counting result. The operation steps are as follows:
step 1-1: and collecting an original check-in data set C, rounding the check-in Time in each check-in record, recording the rounded check-in Time as Time, and setting the set of Time values as Time = {0,1,2,3, \ 8230;, 23}. The time of day is divided into 8 discrete time slots T, the set of time slots denoted T = {0,1,2, \8230;, 7}. The corresponding conversion relation between the rounded check-in time and the time slot t is as follows:
step 1-2: sorting the check-in data set converted into the time slot to obtain n check-in records, and recording the n check-in records as C = { C = 1 ,c 2 ,…,c n }. Wherein each check-in record contains the ID of the user who has checked-in behavior, the check-in time slot t and the ID, longitude and latitude information of the visited position, which are marked as c i =<userID,t,locationID,longitude,latitude>,i∈[1,n]. All user sets in the check-in dataset are U, all location sets are L, and the number of users and locations are denoted NU and NL, respectively.
Step 1-3: the score of user u for location l at time slot t is defined as: if the user u visits the position l in the sign-in time period corresponding to the time slot t, scoring r u,t,l =1; otherwise, r u,t,l =0。
Summarizing all scores to form a user-time-position three-dimensional score matrix R = { R = { (R) } u,t,l },u∈U,t∈[0,7]L ∈ L, the scoring matrix R has a total of NU × 8 rows and a total of NL columns, where NU and NL are the total number of users and locations, respectively.
Step 2: and segmenting the user-time-position three-dimensional scoring matrix according to time slots, extracting a user-position two-dimensional scoring matrix in each time slot, and calculating the dynamic similarity of the positions in different time slots by using the Jaccard coefficient aiming at each scoring matrix. The traditional project-based collaborative filtering method is improved, and the dynamic similarity of the positions and the scores already completed by the user are utilized to predict the scores of the positions to the inaccessible addresses. The specific operation steps are as follows:
step 2-1: according to the value of the time slot t, the three-dimensional scoring matrix of the user-time-position is divided into eight two-dimensional scoring matrices of the user-position from t =0 to t =7, and the scoring matrix is recorded as R t ={r u,l },t∈[0,7],u∈U,l∈L。
Step 2-2: for each two-dimensional scoring matrix R t Selecting more suitable Jaccard coefficient to calculate position similarity, and calculating separatelyPosition l calculated at each time slot t i And l j Dynamic similarity based on time perception between the two:
wherein i ∈ [1, NL ]],j∈[1,NL],t∈[0,7],U i,t Indicating that location l was visited at time slot t i User set of U j,t Indicating that location l was visited at time slot t j Is selected.
Step 2-3: selecting a target user u in a location social network a As a recommended service object, the current recommended time is set r Conversion to corresponding time slot t r 。
Step 2-4: the traditional project-based collaborative filtering method is improved, and the dynamic similarity of the positions and the target user u are utilized a The already completed scoring predicts its scoring of the unaccessed address/:
wherein u is a Is the target object of the current service of the recommendation system, t r Is the corresponding time slot when the recommendation system provides the recommendation service, L is a position that the target user has not visited yet, L represents the set of all positions, sim (L, L', t) r ) Is indicated in time slot t r The dynamic similarity between the time positions l and l',representing target user u a At time slot t r The score of position l'.
And step 3: and respectively recording the specific time of each user in each time slot for accessing the position for the last time aiming at each position, and calculating the aging characteristic value of the position in each time slot for each user one by one. And averaging the position aging characteristic values of all users who visit the address in different time slots, and calculating the aging characteristic value of the position based on time perception. The method comprises the following implementation steps:
step 3-1: and recording the oldest time and the newest time of the sign-in action in the sign-in data set, and respectively recording the time as minT and maxT.
Step 3-2: for each location, all users that have visited a location/are recorded as a set U _l . If the user U belongs to U _l Then record the time that user u last accessed location l in time slot t as recentT (u, l, t). For user u, the aging characteristic value timelness (u, l, t) of location l at time slot t is calculated as follows:
wherein minT and maxT are the oldest time and the newest time of the sign-in behavior in the sign-in data set respectively, and the entry T (u, l, t) is the time when the user u accesses the position l for the latest time in the time slot t.
Step 3-3: for U visiting location l _l Averaging the timeless (u, l, t) values of all users in the set, and calculating the time perception-based aging characteristic value of the position l in a time slot t:
wherein U is _l To access all users in location l, timelness (u, l, t) represents the age eigenvalue of location l for user u at time slot t.
And 4, step 4: and calculating the geographic distance between each position in the check-in data set according to the latitude and longitude information, realizing personalized probability density modeling by using a kernel density estimation method, and individually mining the influence of geographic features. The method comprises the following implementation steps:
step 4-1: and acquiring all addresses and longitude and latitude information corresponding to the addresses in the check-in data set C, and calculating the geographic distance between the positions according to the longitude and latitude of each address. Set position l i Respectively of long and short lengths of long and short i And lat i Is recorded as i =<lng i ,lat i >Position l j Respectively of long and short lengths of long and short j And lat j Is recorded as l j =<lng j ,lat j >Then position l i And l j The geographical distance between is:
wherein R is the earth radius, R =6371km.
After the geographic distance between every two addresses is calculated, a distance matrix Dis = { Dis = is formed ij Where is ij Indicates the position l i And l j 1 ≦ i ≦ NL,1 ≦ j ≦ NL, the matrix having NL rows and NL columns, NL being the total number of addresses in the check-in data set.
Step 4-2: target user u a The set of locations visited is recorded as L _a Finding L from the distance matrix Dis _a The geographic distance d between each pair of positions in the set forms a set of distance samples X _a The distance distribution is estimated by a probability density function f over the distance d:
where σ is a smoothing coefficient, also called bandwidth, and K (.) is a gaussian kernel:
step 4-3: given target user u a Visited location set L _a Calculating target user u a Access candidate address l i The probability of (c) is:
where the f-function is a probability density function as shown in equation 7.
And 5: the influence of the user context, the position context and the time context on the check-in behavior of the user is comprehensively considered, a scoring prediction mechanism fusing the historical preference, the geographic distance influence, the position aging characteristic and the position dynamic similarity of the user is constructed, and a plurality of positions with high final prediction scoring are recommended to the user. The method comprises the following implementation steps:
step 5-1: to calculate a target user u a At the current recommended time slot t r For the final prediction score of the inaccessible address l, firstly, the min-max standardization processing is carried out on each pre-score generated in the step 2, the step 3 and the step 4:
wherein,utilizes an improved project-based collaborative filtering method to process the current recommendation time slot t r Calculating target user u a Scoring of the unaccessed addresses, l, as shown in steps 2-4; timeline (, l) r t) is the current recommended time slot t of the candidate address l r Time-based perceived age eigenvalues, as shown in steps 3-3, pr (L | L) _a ) Is based on the target user u a Visited location set L _a And predicting the access probability of the obtained candidate address L by using a kernel density estimation method, wherein L is a set of all positions as shown in a step 4-3.
Step 5-2: comprehensive consideration bitInfluence of the time perception-based dynamic similarity and geographic distance of the position and the aging characteristic on the position on the check-in habit of the user is calculated, and a target user u is calculated a At t r The final prediction score for candidate address/is:
step 5-3: for target user u a All the addresses which are not visited are sorted according to the final prediction score (formula 13), and N positions with the top rank are recommended to the target user u a 。
Step 6: and (5) providing an timeliness evaluation index and defining a timeliness evaluation system of the recommendation system. And comparing the prediction accuracy and the recommendation timeliness of the recommendation system and other classical recommendation systems provided by the invention by using the accuracy evaluation index and the timeliness evaluation index respectively, and evaluating the accuracy and the effectiveness of the provided technology. The method comprises the following implementation steps:
step 6-1: randomly selecting NU multiplied by 10% of users as a target user set TestU, wherein NU represents the total number of users in the check-in data set. For each target user u in the TestU set a Respectively operating recommendation algorithms to generate current recommendation time t r Recommendation list of (u) TopNList a ,t r )。
Step 6-2: calculating recommendation method in time slot t r Is a target user u a Providing a time efficiency index value when recommending service, and defining the value as a target user u a The average of the aging characteristic values of all the positions in the recommendation list of (1):
wherein u is a Is the target user, t r Is the time slot corresponding to the current recommended time, topNList (u) a ,t r ) Is that the recommendation method is in time slot t r Is a target user u a Provided recommendation list, timeline (u) a ,l,t r ) Is to forUser u a In other words, position l is at time slot t r The aging characteristic value (as shown in equation 4).
Step 6-3: calculating recommendation method in time slot t r Aging property of (2):
where TestU is the target user set, timeliness (u, t) r ) Is in time slot t r And (4) providing a timeliness index value of the recommendation method when the target user u is provided with the recommendation service (shown in the formula 14).
Step 6-4: defining the final Timeliness of the recommendation method as the average value of Timeliness indexes of each time slot:
where T is a time slot set, T = {0,1,2,3,4,5,6,7}, and Timeliness (T) is the Timeliness of the recommended method at time slot T (equation 15).
Step 6-5: calculating recommendation method in time slot t r Precision and recall in hours:
where TestU is the set of all target users, TP (u, t) r )、FP(u,t r ) And FN (u, t) r ) The position numbers of positive case halving, negative case misclassification and positive case misclassification in the recommendation list are respectively.
Step 6-6: calculating the final accuracy and recall rate of the recommendation method, wherein the value of the final accuracy and recall rate is the average value of the corresponding evaluation indexes in each time slot:
where T is a time slot set, T = {0,1,2,3,4,5,6,7}, precision (T) and recall (T) are the precision and recall, respectively, of the recommended method within the time slot T.
Step 6-7: calculating the comprehensive accuracy F1 value of the recommendation system:
wherein precision and recall are the overall accuracy and recall of the recommended method running once, respectively.
And 6-8: and (4) repeatedly executing the step 6-1 to the step 6-7 for Ntimes, wherein the values of the final Timeliness index value Timeliness, the prediction accuracy Precision, the Recall rate Recall and the comprehensive accuracy index F1 of the recommendation method are the average values of the index results corresponding to the Ntimes.
Step 6-9: and comparing and analyzing the results of all indexes: if Precision of the position recommendation method based on the position timeliness characteristics and the time perception dynamic similarity is larger than Precision values of other recommendation algorithms, the prediction accuracy of the recommendation technology provided by the invention is higher; if the Recall rate Recall of the algorithm provided by the invention is greater than the Recall values of other recommended algorithms, the technical search capability provided by the invention is stronger; if the comprehensive precision index F1 value of the algorithm provided by the invention is larger than the F1 values of other recommended algorithms, the comprehensive capability of the technology provided by the invention in the aspect of prediction accuracy is stronger; if the Timeliness of the recommendation method provided by the invention is larger than the Timeliness values of other recommendation algorithms, the recommendation technology provided by the invention can better mine the recent preference of the user and has stronger Timeliness.
The following takes a specific location-based social network brightkit as an example to describe in detail how the location recommendation method based on location aging characteristics and time-aware dynamic similarity in the present invention operates.
The brightkit dataset is the social relationship and check-in information collected by the united states stanford university SNAP laboratory for 58228 users on the brightkit website during 2008 to 201010 months. The number of positions in the Brightkite data set is 693362, the check-in records of users are 4747281, and 214078 social relationships are formed among the users. The brightkit dataset is one of the most commonly used test datasets by researchers of the location recommendation system.
The invention selects check-in data of five hot areas in Los Angeles, san Francisco, new York, maricopa and King county in a BrightKite data set as an example for instantiation description.
Step 1: collecting and sorting original check-in data sets, converting specific time in the check-in records into different time slots, counting check-in times of users at various positions in different time slots, and converting the check-in records into user-time-position three-dimensional scoring matrixes according to counting results. The operation steps are as follows:
step 1-1: and collecting an original check-in data set C, rounding the check-in Time in each check-in record, recording the rounded check-in Time as Time, and setting the set of Time values as Time = {0,1,2,3, \ 8230;, 23}. For example, if the check-in time is 00; if the check-in time is 23.
The time of day is divided into 8 discrete time slots T, the set of time slots denoted T = {0,1,2, \8230;, 7}. The corresponding conversion relation between the rounded check-in time and the time slot t is as follows:
specifically, the conversion manner of the rounded check-in time and the time slot t is shown in table 1:
TABLE 1 transition relationships corresponding to check-in times and time slots
Step 1-2: sorting the check-in data set after being converted into the time slot to obtain 56861 check-in records, and recording the records as C = { C = 1 ,c 2 ,…,c 56861 }. Wherein each check-in record contains the ID of the user who has checked-in behavior, the check-in time slot t and the ID, longitude and latitude information of the visited position, which are marked as c i =<userID,t,locationID,longitude,latitude>,i∈[1,56861]. All user sets in the check-in dataset are U, all location sets are L, the number of users NU =1963, the number of locations NL =2574. Fig. 3 is a schematic diagram showing statistical results of the number of location visits and the number of users in the present embodiment.
Step 1-3: the score of user u for location l at time slot t is defined as: if user u has accessed location l within three check-in time periods (Table 1) corresponding to time slot t, score r is given u,t,l =1; otherwise, r u,t,l =0。
Summarizing all scores to form a user-time-position three-dimensional score matrix R = { R = u,t,l },u∈U,t∈[0,7]L ∈ L, scoring matrix R has 1963X 8 rows and 2574 columns.
Step 2: and segmenting the user-time-position three-dimensional scoring matrix according to the time slots, extracting the user-position two-dimensional scoring matrix in each time slot, and calculating the dynamic similarity of the positions in different time slots by using the Jaccard coefficient aiming at each scoring matrix. The traditional project-based collaborative filtering method is improved, and the dynamic similarity of the positions and the scores already completed by the user are utilized to predict the scores of the positions to the inaccessible addresses. The specific operation steps are as follows:
step 2-1: according to the value of the time slot t, the three-dimensional scoring matrix of the user-time-position is divided into eight two-dimensional scoring matrices of the user-position from t =0 to t =7, and the scoring matrix is recorded as R t ={r u,l },t∈[0,7],u∈U,l∈L。
Step 2-2:for each two-dimensional scoring matrix R t Selecting more suitable Jaccard coefficient to calculate position similarity, and calculating position l in each time slot t i And l j Dynamic similarity based on time perception between the two:
wherein i ∈ [1,2574 ]],j∈[1,2574],t∈[0,7],U i,t Indicating that location l was visited at time slot t i User set of (2), U j,t Indicating that location l was visited at time slot t j Is selected.
Step 2-3: target user u in selected location social network a As a recommended service object, the current recommended time is set r Conversion to corresponding time slots t r . For example, if the current recommended time is 22 r =7。
Step 2-4: the traditional project-based collaborative filtering method is improved, and the dynamic similarity of the positions and the target user u are utilized a The scoring that has been done to predict its scoring of the inaccessible address/is:
wherein u is a Is the target object of the current service of the recommendation system, t r Is the corresponding time slot when the recommendation system provides the recommendation service, L is a position that the target user has not visited yet, L represents the set of all positions, sim (L, L', t) r ) Is indicated in time slot t r The dynamic similarity between the time positions l and l',representing a target user u a At time slot t r Scoring position l'.
And step 3: and respectively recording the specific time of each user in each time slot for accessing the position for the last time aiming at each position, and calculating the aging characteristic value of the position in each time slot for each user one by one. And averaging the position aging characteristic values of all users who access the address in different time slots, and calculating the aging characteristic value of the position based on time perception. The method comprises the following implementation steps:
step 3-1: and recording the oldest time and the newest time of the sign-in action in the sign-in data set, and respectively recording the time as minT and maxT. In this embodiment, the oldest time of the check-in action is 58, which is recorded as minT =20080404003758; the latest time of check-in occurred was 2010, 10 months, 18 days, 14, 04, noted maxT =20101018143104.
Step 3-2: for each location, all users that have visited a location/are recorded as a set U _l . If the user U belongs to U _l Then record the time that user u last accessed location l in time slot t as recentT (u, l, t). For user u, the aging characteristic value timeliness (u, l, t) of location l at time slot t is calculated as follows:
step 3-3: for U having visited location l _l Averaging the timeless (u, l, t) values of all users in the set, and calculating the time perception-based aging characteristic value of the position l in a time slot t:
wherein U is _l To access all users in location/, timeservice (u, l, t) represents the age characteristic value of location/for user u at time slot t.
And 4, step 4: and calculating the geographic distance between each position in the check-in data set according to the latitude and longitude information, realizing personalized probability density modeling by using a kernel density estimation method, and performing personalized mining on geographic characteristic influences. The realization steps are as follows:
step 4-1: number of check-inAnd acquiring all the addresses and the corresponding longitude and latitude information thereof from the data set C, and calculating the geographic distance between the positions according to the longitude and latitude of each address. Setting a position l i Respectively have a longitude and latitude of lng i And lat i Is recorded as i =<lng i ,lat i >Position l j Respectively of long and short lengths of long and short j And lat j Is recorded as l j =<lng j ,lat j >Then position l i And l j The geographical distance between is:
after the geographic distance between every two addresses is calculated, a distance matrix Dis = { Dis } is formed ij Where is ij Indicates the position l i And l j 1 ≦ i ≦ 2574,1 ≦ j ≦ 2574, and the matrix has 2574 rows and 2574 columns.
Step 4-2: target user u a The set of locations visited is recorded as L _a Finding L from the distance matrix Dis _a The geographic distance d between each pair of positions in the set forms a set of distance samples X _a The distance distribution is estimated by a probability density function f over the distance d:
where σ is a smoothing coefficient, also called bandwidth, and K (.) is a gaussian kernel:
step 4-3: given target user u a Visited location set L _a Calculating a target user u a Access candidate address l i The probability of (c) is:
where the f function is a probability density function as shown in equation 28.
And 5: the influence of the user context, the position context and the time context on the sign-in behavior of the user is comprehensively considered, a scoring prediction mechanism fusing the historical preference, the geographic distance influence, the position timeliness characteristic and the position dynamic similarity of the user is constructed, and a plurality of positions with high final prediction scores are recommended to the user. The realization steps are as follows:
step 5-1: to calculate a target user u a At the current recommended time slot t r For the final prediction score of the inaccessible address l, firstly, the min-max standardization processing is carried out on each pre-score generated in the steps 2,3 and 4:
wherein,utilizes an improved project-based collaborative filtering method to currently recommend a time slot t r Calculating target user u a Scoring of the unaccessed addresses l, as in steps 2-4); tlmiet (, r ) Is that the candidate address l is in the current recommended time slot t r Time-based perception of aging characteristic, pr (L | L), as shown in step 3-3 _a ) Is based on the target user u a Visited location set L _a Predicting the access probability of the obtained candidate address l by using a kernel density estimation method, as shown in step 4-3And L is the set of all locations.
Step 5-2: comprehensively considering the time efficiency characteristics of the position, the influence of the dynamic similarity of the position based on time perception and the geographic distance on the sign-in habit of the user, and calculating the target user u a At t r The final prediction score for candidate address/is:
step 5-3: for target user u a All the addresses which are not visited are sorted according to the final prediction score, and N positions with the top rank are recommended to the target user u a . N is a multiple of 10 and is respectively 10, 20, 30, 40 and 50.
Step 6: and (5) providing an timeliness evaluation index and defining a timeliness evaluation system of the recommendation system. And comparing the prediction accuracy and the recommendation timeliness of the recommendation system and other classical recommendation systems provided by the invention by using the accuracy evaluation index and the timeliness evaluation index respectively, and evaluating the accuracy and the effectiveness of the provided technology. The method comprises the following implementation steps:
step 6-1: 196 users are randomly selected as a target user set TestU, and each target user u in the TestU set a Respectively operating recommendation algorithms to generate current recommendation time t r Recommendation list of (u) TopNList a ,t r )。
Step 6-2: calculating recommendation method in time slot t r Is a target user u a Providing a time efficiency index value when recommending service, and defining the value as a target user u a The average of the aging characteristic values of all the positions in the recommendation list of (1):
wherein u is a Is the target user, t r Is the time slot corresponding to the current recommended time, topNList (u) a ,t r ) Is that the recommendation method is in time slot t r Is a target user u a Provided recommendation list, timeline (u) a ,l,t r ) Is for user u a In other words, position i is in time slot t r The aging characteristic value (equation 25).
And 6-3: calculating recommendation method in time slot t r Aging property:
where TestU is the set of target users, timeliness (u, t) r ) Is in time slot t r And (3) providing a timeliness index value of the recommendation method when the recommendation service is provided for the target user u, as shown in formula 35.
And 6-4: defining the final Timeliness of the recommendation method as the average value of Timeliness indexes of each time slot:
where T is the time slot set, T = {0,1,2,3,4,5,6,7}, and Timeliness (T) is the Timeliness of the recommended method at time slot T, as shown in equation 36.
Step 6-5: calculating recommendation method in time slot t r Precision and recall in hours:
where TestU is the set of all target users, TP (u, t) r )、FP(u,t r ) And FN (u, t) r ) The number of positions in the recommendation list for positive case halves, negative case misclassifications, and positive case misclassifications, respectively.
Step 6-6: calculating the final accuracy and recall rate of the recommendation method, wherein the value of the final accuracy and recall rate is the average value of the corresponding evaluation indexes in each time slot:
where T is a time slot set, T = {0,1,2,3,4,5,6,7}, precision (T) and recall (T) are the precision and recall, respectively, of the recommended method within the time slot T.
Step 6-7: calculating the comprehensive accuracy F1 value of the recommendation system:
wherein precision and recall are the overall accuracy and recall of the recommended method running once, respectively.
And 6-8: the steps 6-1 to 6-7 are repeatedly executed for 100 times, and the box charts of the Precision, the Recall Precision and the comprehensive Precision index F1 after the recommendation method provided by the invention is operated for 100 times are respectively shown in fig. 4,5 and 6.
The values of the final Timeliness index value Timeliness, the prediction accuracy Precision, the Recall rate Recall and the comprehensive accuracy index F1 of the recommendation method are the average values of the corresponding index results of 100 times. When N is 10, 20, 30, 40, and 50, the results of Precision, recall, integrated Precision index F1, and Timeliness of each recommended method are shown in tables 2,3,4, and 5, respectively, where each row has a bold-format value representing the maximum value of the row index:
TABLE 2 Precision index values for different recommendation methods
TABLE 3 Recall ratio Recall index values for different recommendation methods
TABLE 4 recommendation precision F1 index values of different recommendation methods
TABLE 5 Timeliness index values for different recommended methods
In the case of the present embodiment, histograms of comparison between the recommendation method and a classic project-based collaborative filtering method IBCF, user-based collaborative filtering (UBCF), precision and Recall of a kernel density estimation-based access probability prediction method KDE, and Timeliness are respectively shown in FIG. 7, FIG. 8, FIG. 9 and FIG. 10.
Step 6-9: and comparing and analyzing the results of all indexes: the Precision of the position recommendation method based on the position timeliness characteristics and the time perception dynamic similarity is greater than Precision values of other recommendation algorithms, and therefore the prediction accuracy of the recommendation technology is higher; the Recall rate Recall of the algorithm provided by the invention is greater than the Recall values of other recommended algorithms, which shows that the technical search capability provided by the invention is stronger; the comprehensive precision index F1 value of the algorithm provided by the invention is larger than the F1 values of other recommended algorithms, which shows that the comprehensive capability of the technology provided by the invention in the aspect of prediction accuracy is stronger; the Timeliness of the recommendation method provided by the invention is larger than the Timeliness values of other recommendation algorithms, so that the recommendation technology provided by the invention can better mine the recent preference of the user and has stronger Timeliness.
The method is different from the existing position recommendation algorithm, aims to construct a real-time, high-accuracy and strong-timeliness position recommendation system, emphasizes the consideration of the differences of position similarity in different time periods and the inherent timeliness characteristics of positions, innovatively excavates the position dynamic similarity based on time perception, and effectively improves the prediction accuracy of the position recommendation system; the method for defining and calculating the position timeliness characteristics is provided, and the recommendation timeliness of the position recommendation system is greatly improved. In addition, the invention designs a set of timeliness evaluation system aiming at the position recommendation system while measuring the prediction accuracy of the recommendation system, and provides important technical support for quantifying the novelty of the recommendation result. The technology provided by the invention has wide application prospect and is expected to be widely applied to the social network market based on the position.
The above-described process flow is only a preferred embodiment of the present invention, but does not represent all the details of the present invention. Any modification, equivalent replacement, and improvement made by those skilled in the art within the technical scope of the present disclosure within the spirit and principle of the present invention should be included in the protection scope of the present invention.
Claims (7)
1. A position recommendation method based on position aging characteristics and time perception dynamic similarity is characterized by comprising the following steps:
step 1: collecting and sorting an original check-in data set, converting specific time in the check-in records into different time slots, counting check-in times of users at each position in different time slots, and converting the check-in records into a user-time-position three-dimensional scoring matrix according to a counting result;
step 2: the user-time-position three-dimensional scoring matrix is segmented according to time slots, a user-position two-dimensional scoring matrix in each time slot is extracted, dynamic similarity of positions in different time slots is calculated by using a Jaccard coefficient aiming at each scoring matrix, and scoring of an inaccessible address is predicted by using the dynamic similarity of the positions and scoring completed by a user;
and step 3: aiming at each position, respectively recording the specific time of each user in each time slot for accessing the position most recently, calculating the time efficiency characteristic value of the position in each time slot for each user one by one, averaging the position time efficiency characteristic values of all users accessing the address in different time slots, and calculating the time-sensing-based time efficiency characteristic value of the position;
and 4, step 4: calculating the geographic distance between each position in the check-in data set according to the longitude and latitude information, realizing personalized probability density modeling by using a kernel density estimation method, and personalized mining geographic characteristic influence;
and 5: a scoring prediction mechanism comprehensively considering the user context, the position context and the time context to sign in the behavior of the user recommends a plurality of positions with higher final prediction scores to the user;
step 6: and providing an timeliness evaluation index, defining a timeliness evaluation system of the recommendation system, comparing the prediction accuracy and the recommendation timeliness of the recommendation system and other classical recommendation systems provided by the invention by using the accuracy evaluation index and the timeliness evaluation index respectively, and evaluating the accuracy and the effectiveness of the provided technology.
2. The location recommendation method based on location aging characteristics and time-aware dynamic similarity according to claim 1, wherein step 1 of the method comprises:
step 1-1: collecting an original check-in data set C, rounding the check-in Time in each check-in record, recording the rounded check-in Time as Time, setting the set of Time values as Time = {0,1,2,3, \ 8230;, 23}, dividing the Time of one day into 8 discrete Time slots T, and expressing the Time slot set as T = {0,1,2, \ 8230, 7}, wherein the corresponding conversion relation between the rounded check-in Time and the Time slots T is as follows:
step 1-2: sorting the check-in data set converted into the time slot to obtain n check-in records, and recording the n check-in records as C = { C = 1 ,c 2 ,…,c n And each check-in record comprises the ID of the user with the check-in behavior, the check-in time slot t and the ID, longitude and latitude information of the visited position, and is marked as c i =<userID,t,locationID,longitude,latitude>,i∈[1,n]All user sets in the check-in data set are U, all position sets are L, and the number of the users and the number of the positions are respectively recorded as NU and NL;
step 1-3: the score of user u for location l at time slot t is defined as: if the user u visits the position l in the sign-in time period corresponding to the time slot t, scoring r u,t,l =1; otherwise, r u,t,l =0;
Summarizing all scores to form a user-time-position three-dimensional score matrix R = { R = { (R) } u,t,l },u∈U,t∈[0,7]And L ∈ L, the scoring matrix R has NU × 8 rows and NL columns, where NU and NL are the total number of users and locations, respectively.
3. The location recommendation method based on location aging characteristics and time-aware dynamic similarity according to claim 1, wherein step 2 of the method comprises:
step 2-1: according to the value of the time slot t, the three-dimensional scoring matrix of the user-time-position is divided into eight two-dimensional scoring matrices of the user-position from t =0 to t =7, and the scoring matrix is recorded as R t ={r u,l },t∈[0,7],u∈U,l∈L;
Step 2-2: for each two-dimensional scoring matrix R t Selecting more suitable Jaccard coefficient to calculate position similarity, and calculating position l in each time slot t i And l j Dynamic similarity based on time perception between the two:
wherein i ∈ [1, NL ]],j∈[1,NL],t∈[0,7],U i,t Indicating that location l was visited at time slot t i User set of (2), U j,t Indicating that location l was visited at time slot t j A set of users of (a);
step 2-3: target user u in selected location social network a As a recommended service object, the current recommended time is taken r Conversion to corresponding time slots t r ;
Step 2-4: the traditional project-based collaborative filtering method is improved, and the dynamic similarity of the position and the target user u are utilized a The scoring that has been done to predict its scoring of the inaccessible address/is:
wherein u is a Is a target object of the current service of the recommendation system, t r Is the corresponding time slot when the recommendation system provides the recommendation service, L is a position that the target user has not visited yet, L represents the set of all positions, sim (L, L', t) r ) Is indicated in time slot t r The dynamic similarity between the time positions l and l',representing target user u a At time slot t r Scoring position l'.
4. The location recommendation method based on location aging characteristics and time-aware dynamic similarity according to claim 1, wherein step 3 of the method comprises:
step 3-1: recording the oldest time and the newest time of the sign-in action in the sign-in data set, and respectively recording the time as minT and maxT;
step 3-2: for each location, all users that have visited a location/are recorded as a set U _l If the user U belongs to U _l If the time when the user u last visits the position l in the time slot t is recorded as a recording t (u, l, t), for the user u, the aging characteristic value timeiiness (u, l, t) of the position l in the time slot t is calculated as follows:
wherein minT and maxT are respectively the oldest time and the newest time of the sign-in action in the sign-in data set, and the recentT (u, l, t) is the time when the user u accesses the position l at the latest time in the time slot t;
step 3-3: for U visiting location l _l Averaging the timeiness (u, l, t) values of all users in the set, and calculating the time perception-based aging characteristic value of the position l in a time slot t:
wherein U is _l To access all users in location l, timelness (u, l, t) represents the age eigenvalue of location l for user u at time slot t.
5. The location recommendation method based on location aging characteristics and time-aware dynamic similarity according to claim 1, wherein step 4 of the method comprises:
step 4-1: acquiring all addresses and longitude and latitude information corresponding to the addresses in the sign-in data set C, calculating the geographic distance between the positions according to the longitude and latitude of each address, and setting a position l i Respectively have a longitude and latitude of lng i And lat i Is recorded as i =<lng i ,lat i >Position l j Respectively have a longitude and latitude of lng j And lat j Is recorded as l j =<lng j ,lat j >Then position l i And l j The geographical distance between is:
wherein R is the earth radius, R =6371km;
calculate every two addressesAfter the geographical distance between them, a distance matrix Dis = { Dis = is formed ij Where is ij Indicates the position l i And l j The geographical distance between the first and second arrays is 1-NL, and NL, the matrix has NL rows and NL columns, and NL is the total number of addresses in the check-in data set;
step 4-2: target user u a The set of locations visited is recorded as L _a Finding L from the distance matrix Dis _a The geographic distance d between each pair of positions in the set forms a set of distance samples X _a The distance distribution is estimated by a probability density function f over the distance d:
where σ is a smoothing coefficient, also called bandwidth, and K (·) is a gaussian kernel function:
step 4-3: given target user u a Visited location set L _a Calculating a target user u a Access candidate address l i The probability of (c) is:
where the f-function is the probability density function (equation 7).
6. The location recommendation method based on location aging characteristics and time-aware dynamic similarity according to claim 1, wherein the step 5 of the method comprises:
step 5-1: to calculate a target user u a At the current recommended time slot t r The final prediction score of the non-accessed address l is firstly compared with the final prediction scores generated in the second step, the third step and the fourth stepEach pre-score was subjected to min-max normalization:
wherein,utilizes an improved project-based collaborative filtering method to currently recommend a time slot t r Calculating target user u a Scoring of the unaccessed address/; timeiiness (l, t) r ) Is that the candidate address l is in the current recommended time slot t r Temporal perceptual age-related feature value, pr (L | L) over time _a ) Is based on the target user u a Visited location set L _a Predicting the access probability of the obtained candidate address L by using a kernel density estimation method, wherein L is a set of all positions;
step 5-2: comprehensively considering the time efficiency characteristics of the position, the influence of the dynamic similarity of the position based on time perception and the geographic distance on the sign-in habit of the user, and calculating the target user u a At t r The final prediction score for candidate address/is:
step 5-3: for target user u a All the addresses which are not visited are sorted according to the final prediction scores, and N positions which are ranked at the top are recommended to the target user u a 。
7. The location recommendation method based on location aging characteristics and time-aware dynamic similarity according to claim 1, wherein step 6 of the method comprises:
step 6-1: randomly selecting NU multiplied by 10% of users as a target user set TestU, wherein NU represents the total number of users in the check-in data set and is each target user u in the TestU set a Respectively operating recommendation algorithms to generate current recommendation time t r Recommendation list of (u) TopNList a ,t r );
Step 6-2: calculating recommendation method in time slot t r Is a target user u a Providing a time efficiency index value when recommending service, and defining the value as a target user u a The average of the aging characteristic values of all the positions in the recommendation list of (1):
wherein u is a Is the target user, t r Is the time slot corresponding to the current recommended time, topNList (u) a ,t r ) Is that the recommendation method is in time slot t r Is a target user u a The provided recommendation list, timelness (u) a ,l,t r ) Is for user u a In other words, position i is in time slot t r The aging characteristic value (as shown in equation 4);
step 6-3: calculating recommendation method in time slot t r Aging property:
where TestU is the target user set, timeliness (u, t) r ) Is in time slot t r Providing a target user u with a timeliness index value of a recommendation method when providing a recommendation service (as shown in formula 14);
step 6-4: defining the final Timeliness of the recommendation method as the average value of Timeliness indexes of each time slot as follows:
wherein T is a time slot set, T = {0,1,2,3,4,5,6,7}, timelining (T) is the Timeliness of the recommended method at time slot T (as shown in equation 15);
and 6-5: calculating recommendation method in time slot t r Precision and recall in hours:
where TestU is the set of all target users, TP (u, t) r )、FP(u,t r ) And FN (u, t) r ) The position numbers of positive case halving, negative case misclassification and positive case misclassification in the recommendation list are respectively;
step 6-6: calculating the final accuracy and recall rate of the recommendation method, wherein the value of the final accuracy and recall rate is the average value of the corresponding evaluation indexes in each time slot:
wherein T is a time slot set, T = {0,1,2,3,4,5,6,7}, precision (T) and recall (T) are the precision rate and recall rate of the recommended method in the time slot T, respectively;
step 6-7: calculating the comprehensive accuracy F1 value of the recommendation system:
wherein precision and call are respectively the total accuracy and recall rate of the recommended method running once;
and 6-8: repeatedly executing the step 6-1 to the step 6-7 for Ntimes, wherein the values of the final Timeliness index value Timeliness, the prediction accuracy Precision, the Recall rate Recall and the comprehensive accuracy index F1 of the recommendation method are the average values of the index results corresponding to the Ntimes;
and 6-9: and comparing and analyzing the results of all indexes: if Precision of the method is greater than Precision values of other recommendation algorithms, the prediction accuracy of the recommendation technology of the method is higher; if the Recall rate Recall of the method is greater than the Recall values of other recommended algorithms, the technical searching capability of the method is stronger; if the comprehensive precision index F1 value of the method is larger than the F1 values of other recommended algorithms, the comprehensive capability of the technology of the method in the aspect of prediction accuracy is stronger; if the Timeliness of the method is larger than the Timeliness values of other recommendation algorithms, the recommendation technology provided by the method can better mine recent preferences of the user and has stronger Timeliness.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202211180601.6A CN115687801B (en) | 2022-09-27 | 2022-09-27 | Position recommendation method based on position aging characteristics and time perception dynamic similarity |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202211180601.6A CN115687801B (en) | 2022-09-27 | 2022-09-27 | Position recommendation method based on position aging characteristics and time perception dynamic similarity |
Publications (2)
Publication Number | Publication Date |
---|---|
CN115687801A true CN115687801A (en) | 2023-02-03 |
CN115687801B CN115687801B (en) | 2024-01-19 |
Family
ID=85063463
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202211180601.6A Active CN115687801B (en) | 2022-09-27 | 2022-09-27 | Position recommendation method based on position aging characteristics and time perception dynamic similarity |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN115687801B (en) |
Citations (16)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
KR20100048218A (en) * | 2008-10-30 | 2010-05-11 | 에스케이커뮤니케이션즈 주식회사 | Collaborative filtering system and method and computer readable medium processing the method |
US20100153292A1 (en) * | 2008-12-11 | 2010-06-17 | Microsoft Corporation | Making Friend and Location Recommendations Based on Location Similarities |
KR101259604B1 (en) * | 2011-11-15 | 2013-05-14 | 경기대학교 산학협력단 | Contents recommendation system and method based on social network |
CN106776930A (en) * | 2016-12-01 | 2017-05-31 | 合肥工业大学 | A kind of location recommendation method for incorporating time and geographical location information |
CN107341261A (en) * | 2017-07-13 | 2017-11-10 | 南京邮电大学 | A kind of point of interest of facing position social networks recommends method |
CN107515949A (en) * | 2017-09-14 | 2017-12-26 | 云南大学 | User's space-time method for measuring similarity in interest point prediction and recommendation |
CN108074016A (en) * | 2017-12-25 | 2018-05-25 | 苏州大学 | Customer relationship intensity prediction method, device and equipment based on position social networks |
CN108875007A (en) * | 2018-06-15 | 2018-11-23 | 腾讯科技(深圳)有限公司 | The determination method and apparatus of point of interest, storage medium, electronic device |
CN110334293A (en) * | 2019-07-12 | 2019-10-15 | 吉林大学 | A kind of facing position social networks has Time Perception position recommended method based on fuzzy clustering |
KR20190124524A (en) * | 2018-04-26 | 2019-11-05 | 충북대학교 산학협력단 | Method and system for recommending point of interest |
CN111475744A (en) * | 2020-04-03 | 2020-07-31 | 南京理工大学紫金学院 | Personalized position recommendation method based on ensemble learning |
CN112487307A (en) * | 2020-12-10 | 2021-03-12 | 武汉大学 | Time-aware POI recommendation method |
CN112784160A (en) * | 2021-01-29 | 2021-05-11 | 电子科技大学 | Interest point recommendation method based on tensor decomposition and considering context semantic information |
CN114036376A (en) * | 2021-10-26 | 2022-02-11 | 南京理工大学紫金学院 | Time-aware self-adaptive interest point recommendation method based on K-means clustering |
CN114528480A (en) * | 2022-01-21 | 2022-05-24 | 朱俊 | Time-sensing self-adaptive interest point recommendation method based on K-means clustering |
CN114861081A (en) * | 2022-05-18 | 2022-08-05 | 南京审计大学 | Collaborative filtering recommendation method and system based on time-efficient neighbor credible selection |
-
2022
- 2022-09-27 CN CN202211180601.6A patent/CN115687801B/en active Active
Patent Citations (16)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
KR20100048218A (en) * | 2008-10-30 | 2010-05-11 | 에스케이커뮤니케이션즈 주식회사 | Collaborative filtering system and method and computer readable medium processing the method |
US20100153292A1 (en) * | 2008-12-11 | 2010-06-17 | Microsoft Corporation | Making Friend and Location Recommendations Based on Location Similarities |
KR101259604B1 (en) * | 2011-11-15 | 2013-05-14 | 경기대학교 산학협력단 | Contents recommendation system and method based on social network |
CN106776930A (en) * | 2016-12-01 | 2017-05-31 | 合肥工业大学 | A kind of location recommendation method for incorporating time and geographical location information |
CN107341261A (en) * | 2017-07-13 | 2017-11-10 | 南京邮电大学 | A kind of point of interest of facing position social networks recommends method |
CN107515949A (en) * | 2017-09-14 | 2017-12-26 | 云南大学 | User's space-time method for measuring similarity in interest point prediction and recommendation |
CN108074016A (en) * | 2017-12-25 | 2018-05-25 | 苏州大学 | Customer relationship intensity prediction method, device and equipment based on position social networks |
KR20190124524A (en) * | 2018-04-26 | 2019-11-05 | 충북대학교 산학협력단 | Method and system for recommending point of interest |
CN108875007A (en) * | 2018-06-15 | 2018-11-23 | 腾讯科技(深圳)有限公司 | The determination method and apparatus of point of interest, storage medium, electronic device |
CN110334293A (en) * | 2019-07-12 | 2019-10-15 | 吉林大学 | A kind of facing position social networks has Time Perception position recommended method based on fuzzy clustering |
CN111475744A (en) * | 2020-04-03 | 2020-07-31 | 南京理工大学紫金学院 | Personalized position recommendation method based on ensemble learning |
CN112487307A (en) * | 2020-12-10 | 2021-03-12 | 武汉大学 | Time-aware POI recommendation method |
CN112784160A (en) * | 2021-01-29 | 2021-05-11 | 电子科技大学 | Interest point recommendation method based on tensor decomposition and considering context semantic information |
CN114036376A (en) * | 2021-10-26 | 2022-02-11 | 南京理工大学紫金学院 | Time-aware self-adaptive interest point recommendation method based on K-means clustering |
CN114528480A (en) * | 2022-01-21 | 2022-05-24 | 朱俊 | Time-sensing self-adaptive interest point recommendation method based on K-means clustering |
CN114861081A (en) * | 2022-05-18 | 2022-08-05 | 南京审计大学 | Collaborative filtering recommendation method and system based on time-efficient neighbor credible selection |
Non-Patent Citations (1)
Title |
---|
阚绪驰: ""基于位置社交网络的兴趣点推荐算法研究"", pages 138 - 1562 * |
Also Published As
Publication number | Publication date |
---|---|
CN115687801B (en) | 2024-01-19 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Huang et al. | Activity patterns, socioeconomic status and urban spatial structure: what can social media data tell us? | |
Kefalas et al. | A time-aware spatio-textual recommender system | |
Su et al. | A method for discovering clusters of e-commerce interest patterns using click-stream data | |
Li et al. | Mining user similarity based on location history | |
Xing et al. | Points-of-interest recommendation based on convolution matrix factorization | |
Bernardes et al. | A social formalism and survey for recommender systems | |
CN106960044B (en) | Time perception personalized POI recommendation method based on tensor decomposition and weighted HITS | |
CN111475744B (en) | Personalized position recommendation method based on ensemble learning | |
CN115408618B (en) | Point-of-interest recommendation method based on social relation fusion position dynamic popularity and geographic features | |
Xia et al. | Vrer: context-based venue recommendation using embedded space ranking SVM in location-based social network | |
CN114036376A (en) | Time-aware self-adaptive interest point recommendation method based on K-means clustering | |
Li et al. | Social recommendation based on trust and influence in SNS environments | |
Kang et al. | A personalized point-of-interest recommendation system for O2O commerce | |
CN114528480A (en) | Time-sensing self-adaptive interest point recommendation method based on K-means clustering | |
Aggarwal et al. | Context-sensitive recommender systems | |
Hong et al. | Multi-criteria tensor model consolidating spatial and temporal information for tourism recommendation | |
Wang et al. | A contextual collaborative approach for app usage forecasting | |
Shi et al. | Mining significant crisp-fuzzy spatial association rules | |
Lou | E‐Commerce Recommendation Technology Based on Collaborative Filtering Algorithm and Mobile Cloud Computing | |
CN118071400A (en) | Application method and system based on graph computing technology in information consumption field | |
Mao et al. | Utilizing multi-source data in popularity prediction for shop-type recommendation | |
Sharma et al. | A location based novel recommender framework of user interest through data categorization | |
Meng et al. | A survey of context-aware recommender systems: from an evaluation perspective | |
CN107784095A (en) | A kind of education resource auto recommending method based on mobile learning | |
Margaris et al. | Enhancing rating prediction quality through improving the accuracy of detection of shifts in rating practices |
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 |