Missing POI Identification Method
By constructing relationship graph grids and joint graph embeddings in location social network services and combining joint probability estimation with spatial proximity, the problems of missing POI identification and data sparseness are solved, and the accuracy and service quality of POI recommendations are improved.
Patent Information
- Application Number
- CN202210601769.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-05-30
- Publication Date
- 2024-11-19
- Estimated Expiration
- 2042-05-30
AI Technical Summary
The problem of missing POI in the location-based social network service has not been effectively solved, and the relationship between time periodicity, geographical proximity and user preferences is ignored, resulting in a relatively different deviation in POI recommendation and prediction. big.
By constructing a graph grid between user, time, POI and POI categories, joint graph embedding, capture sequential conversion patterns, time period impacts, and user preferences, and conduct joint probability estimation in combination with spatial proximity, identify users in the past Missing POI check-in information.
It improves the quality of POI recommendation services, alleviates the problem of data sparseness, and realizes accurate identification of missing POIs, which is suitable for recommendation systems, analysis of missing populations and disease tracing and other fields.
Smart Images

Figure CN114911996B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of geographic space information, and in particular relates to a method for identifying missing POIs. Background Art
[0002] POI is the abbreviation of "Point of Interest", which is usually used to represent a geographically located thing in the real world, such as a shopping mall or a restaurant, and is associated with the corresponding GPS coordinates. The popularity of location-based social networking services (LBSNs) has attracted more and more users to share their daily lives through location check-ins, which mainly consist of POI, check-in time, and some comments. With the continuous development of Internet technology and the increasing demand for personalization, the data from LBSN phones is effectively used for various applications (such as POI recommendation or prediction).
[0003] However, data quality issues (such as missing check-in POIs and data sparsity) always limit the effectiveness of the above research and applications. Specifically, it includes device failure or user cheating, which makes the spatial location (i.e., latitude and longitude) of a given check-in not have corresponding semantic POI check-in information, that is, the POI is missing. At present, POI-based research mainly focuses on recommending or predicting POIs that users may visit in the future, and assumes that all user check-ins specify clear POIs without considering the missing of large check-in POIs. However, according to reports, more than half of the user check-ins are missing in the LBSN platforms of Twitter and Foursquare. To this end, the method we proposed focuses on the identification of user missing POIs, that is, identifying where users have visited at a specific time and spatial location in the past. Its significance lies in: on the one hand, considering the candidate distribution of missing POIs, it helps to enrich user check-in data in order to better understand and model user mobility behavior and improve the service quality of POI recommendation; on the other hand, almost all POI recommendation or prediction tasks have data sparsity problems, and missing POI identification can alleviate the data sparsity problem; this task can also be used for social welfare, such as suspect tracking or missing people search analysis. In addition, by determining which places the patient has visited during the past period of loss, we can track people who may be in the same place with him / her to help avoid the risk of cross infection.
[0004] Among the existing studies, great progress has been made in POI recommendation and prediction, which are most relevant to our problem. Personalized Factorized Markov Chain (FPMC) is a widely used method for Bayesian Personalized Ranking (BPR) optimization, which embeds user preferences and personalized Markov chains for POI recommendation. Personalized Ranking Metric Embedding (PRME) jointly models sequential transitions, personalized preferences, and geographical influences to improve recommendation performance. Based on the extended RNN, a spatial-temporal recurrent neural network (STRNN) is proposed, which models the local spatiotemporal environment of each layer with time and distance transition matrices of time intervals and geographical distances, respectively. In addition, we also adopt the hyperbolic metric embedding (HME) method instead of the traditional Euclidean metric space to learn how to capture the underlying hierarchical structure and learn complex behavioral patterns.
[0005] Although these methods have achieved satisfactory results, they still have some shortcomings. First, for the temporal periodicity effect, geographic proximity effect, user preference and sequential transition pattern, existing methods rarely consider all these influencing factors at the same time, and ignore the relationship between these factors. Second, most of the current methods model relatively detailed POI-level user preferences and sequential transitions on relatively sparse data, which will lead to large deviations. Third, the above methods are not designed for missing POI identifiers, because such problems require the use of relevant contextual information before and after a given query time, rather than recommendation or prediction from a single perspective. Summary of the invention
[0006] The purpose of the present invention is to address the deficiencies of the prior art and provide a method for identifying missing POIs. Different from the existing prediction and recommendation of POIs at the next moment, this method aims to solve the new problem of "where have you been" identification by modeling the entire historical information and the previous and subsequent contextual information of the query time node.
[0007] In order to solve the above technical problems, the present invention adopts the following technical solutions:
[0008] A method for identifying missing POIs comprises the following steps:
[0009] Step 1. Input the user check-in data of the location-based social network service platform, clean the data, and unify the data format to ensure data consistency;
[0010] Step 2. Based on the data processed in step 1, a grid of relationship graphs between users, time and POI, POI category, and POI and POI category is constructed. Then, the above multiple relationship graphs are jointly embedded, so that the expression vectors of users, time, POI and POI category are embedded into a unified shared dimensional space, which jointly captures the sequential conversion pattern, time period influence and user preference.
[0011] Step 3. Considering the spatial proximity at the regional level, the urban area is divided into grids, and each POI is embedded in the corresponding network area, and then the distance in the geographic space is weighted;
[0012] Step 4. Combine the user preference obtained in step 2 and the distance weight obtained in step 3 to perform joint probability estimation, obtain the probability score of each candidate POI and rank it. Based on the ranking, obtain the POI check-ins that the user missed at some time in the past, and thus predict the location information where the user has been.
[0013] Furthermore, the data cleaning method in step 1 is to filter out POI locations and users with less than n check-in times, where n is set according to actual needs.
[0014] Furthermore, step 2 specifically includes the following sub-steps:
[0015] Step 2.1, based on the data processed in step 1, construct five relationship graph grids, namely user-POI, user-POI, time-POI, time-POI category, and POI category-POI category, where the weight in each graph represents the frequency of this check-in;
[0016] Step 2.2: Based on the five relationship graphs obtained in 2.1, the expression vectors of users, POIs, POI categories, and time related to user personalized preferences are embedded into a unified shared dimensional space using the joint graph embedding method to obtain the embedding vectors of users, POIs, POI categories, and time in the unified shared dimensional space: in, include and
[0017] Step 2.3: Based on the embedding vectors of user, POI, POI category, and time obtained in step 2.2, mine the static preference score S of the user at a certain time. p :
[0018]
[0019] In the formula, p represents the social place location (Point of Interest, POI), such as xx supermarket, xx noodle restaurant, etc.), c represents the semantic category of the social place location (such as shopping, dining), u represents the user, τ represents the time item of the query, t represents the time, for At the same time, in order to express the user's transfer preference between social location categories, we assign two roles to each semantic category c of social location: "out" and "in" correspond to and Therefore, use It represents the score of semantic category conversion of social locations at adjacent moments. The larger the value, the more τ The categories of social locations visited by the user at the moment c τ The more likely it is that at the next τ+1 time, the more likely it is to visit c τ+1 .
[0020] And based on the above embedding vector, the user's sequential transfer preference score S for the location is expressed q :
[0021]
[0022] In the formula, c τ 、c τ-1 and c τ+1 Respectively represent the social place location categories at the current moment, the previous moment, and the next moment, c τ Indicates that c τ-1 Indicates that c τ+1 express, represents the user personalized item, P(c τ |c τ-1 ,c τ+1 ) is the transition probability:
[0023]
[0024] in, represents the “out” component of the social location category feature at time -1, represents the “in” component of the social location category characteristics at time τ, represents the “in” component of the social location category feature at time +1, is the time interval between the time when the missing POI annotation occurs and the consecutive check-ins of the previous items. is the time interval between the moment when the missing POI annotation occurs and the consecutive check-in, η is the time interval threshold. When the consecutive check-in time is greater than η, it means that the user's check-in behavior is not affected by the sequence, but only by the long-term static preference;
[0025] Based on the above two factors, the user's probability score for the candidate POI is obtained by combining the user's static preference and sequence transfer preference:
[0026]
[0027] In the formula, θ represents the weight parameter, which ranges from [0,1]; when When it is greater than the time threshold η, it is considered that the user's check-in behavior is only affected by the long-term static preference; otherwise, it is affected by both factors.
[0028] Furthermore, step 3 specifically includes:
[0029] Divide the city into grid areas and obtain the center coordinates of each grid area. Then, according to the GPS coordinates of the user when the POI is missing, map it to the corresponding grid area, calculate the distance between the grid area and the grid center coordinates corresponding to other candidate POIs, and calculate the weight of the spatial distance:
[0030]
[0031] Where l represents the GPS coordinates when the POI is missing, p represents the GPS coordinates of the POI candidate, d(·) represents the GPS coordinate distance in L2 norm, a is a parameter, and X l represents the GPS coordinates of the center of the grid area mapped by l coordinates, and Xp represents the GPS coordinates of the center of the grid area mapped by the candidate POI.
[0032] Furthermore, in step 4, all influencing factors are integrated to perform joint probability estimation, and the probability score of each candidate POI is expressed as:
[0033] S total (p,c|u,l,t)=(1+w l,p )·S p+q (p,c|·);
[0034] Where p represents the social place POI, c represents the semantic category of the social place, u represents the user, l represents the GPS coordinates when the POI is missing, t represents the time when the POI is missing, and w l,p represents the spatial distance weight obtained in step 3, S p+q (p,c|·) represents the probability score of the user preference for the candidate POI obtained in step 2;
[0035] Finally, the missing POI check-ins are identified and predicted to obtain the final result:
[0036]
[0037] Among them, Accuary@k represents the probability of the standard answer appearing in the top-k candidates, and Top-k is the probability of the standard answer appearing in the top-k candidates. total The recommendation results are given in descending order of scores. hit@k means that if there is a correct result among the top-k candidate POIs with the highest scores, it is set to 1, otherwise it is 0; |Dtest| represents the number of samples in the test set.
[0038] Compared with the prior art, the beneficial effects of the present invention are as follows: the present invention proposes a category-aware network embedding model to solve the problem that about 30% of the check-in locations in the current location-based social network service (LBSN) software platform have missing social location locations and data quality problems (such as missing check-in POIs and sparse data), which also limits the effectiveness of existing location-based social network service research (such as POI recommendation or prediction). The method embeds five relationship graphs into a shared space to jointly capture time influence, user preferences and sequential transformation. For POI category sequential transformation modeling, the method innovatively assigns two roles to each POI category: the specific context identity of itself and another category. Finally, all the above-mentioned influencing factors are fused for joint probability estimation to obtain the probability score of each POI and rank it. According to the ranking, the POI check-in identification that the user missed at a certain time in the past is obtained to predict "where has he / she been?"; at the same time, the method can be naturally extended to the conventional next-moment POI recommendation and prediction tasks through deformation. BRIEF DESCRIPTION OF THE DRAWINGS
[0039] Figure 1 A framework diagram of a method for identifying missing POIs according to an embodiment of the present invention;
[0040] Figure 2 The figure is a flow chart of a missing POI identification method according to an embodiment of the present invention. DETAILED DESCRIPTION
[0041] The technical solutions in the embodiments of the present invention will be described clearly and completely below in combination with the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0042] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other.
[0043] The present invention will be further described below in conjunction with specific embodiments, but the present invention is not limited thereto.
[0044] like Figure 1 and Figure 2 As shown, the present invention provides a method for identifying missing POIs, comprising the following steps:
[0045] Step 1. Input the user check-in data of the location-based social network service platform (WeChat, Foursquare, Facebook, etc.), perform data cleaning, and unify the data format to ensure data consistency; in this step, the data cleaning method is to filter out POI locations and users with less than n check-in times, where n is set according to actual needs. In this embodiment, n is set to 10 times.
[0046] Step 2. Based on the data processed in step 1, a grid of relationship graphs between users, time and POI, POI category, and POI and POI category is constructed, and then the above multiple relationship graphs are jointly embedded, so as to embed the expression vectors of users, time, POI and POI category into a unified shared dimensional space, and jointly capture the sequential conversion pattern, time period influence and user preference; Step 2 specifically includes the following sub-steps:
[0047] Step 2.1: By dividing time, time can be divided into 42 time periods according to days (6 periods) and weeks (7 days). Then, based on the data processed in step 1, 5 relationship graph grids are constructed, namely user-POI: G up (U∪P,ε up ), User-POI: G uc (U∪C,ε uc ), Time—POI: G tp (T∪P,ε tp ), Time—POI Category: G tc (T∪C,ε tc ) and POI category—POI category: G cc (C∪C,ε cc ), where the weight in each graph represents the frequency with which this check-in occurs. up , G uc , G tp , G tc The construction of G is mainly used to capture the long-term and periodic preference patterns of users. cc Used to capture the user's behavior patterns during a sequence.
[0048] Step 2.2: Based on the five relationship graphs obtained in the previous step, perform joint graph embedding to embed the expression vectors of users, POIs, POI categories, and time related to the user's personalized preferences into a unified shared dimensional space, and obtain the embedding vectors of users, POIs, POI categories, and time in the unified shared dimensional space; specifically, this method extracts edges from each graph in turn, and then performs node embedding learning in turn, where the embedding method for each graph network is as follows:
[0049] (1) Given any bipartite graph network G AB =(VA ∪V B ,E), where V A and V B They represent two different types of nodes, and E represents the edges between the nodes;
[0050] (2) For graph G AB The edge e ij , its empirical distribution can be expressed as:
[0051]
[0052] Among them, w ij Represents edge e ij The weight of , deg(i) represents the degree of node i;
[0053] (3) At the same time, define the vertex v i Generate vertex v j The conditional probability of is defined as:
[0054]
[0055] in, and Represented as the embedding expression vector of node i and node j, v k Represents the negative sample obtained by negative sampling, Represents node v k Transpose of the embedding expression vector.
[0056] (4) In order to learn to ensure the conditional distribution P(v j |v i ) and experience distribution Close embedding, in Figure G AB Minimize the following objective function:
[0057]
[0058] In the formula, V represents the node vector set, λ i Represents node v i Importance parameter, d KL represents the KL divergence function, which is designed to measure the difference between two distributions, P θ (·) represents the node probability distribution under a given parameter θ;
[0059] After simplification, we get:
[0060]
[0061] (5) At the same time, in order to reduce the computational complexity, a negative sampling mechanism is introduced, and the loss function is finally obtained:
[0062]
[0063] Among them, σ(·) represents the Sigmoid activation function, Represents negative sample v k The sampling of v k Represents the negative sample obtained by negative sampling, v k ~q(·|i) represents the negative sample node v k The sampling of the given node v i The q(·|i) distribution of and Represented as the embedded expression vector of node i and node j, Represents a negative node v k Embed the transpose of the feature vector and take its negation;
[0064] Using the above method, each graph network is embedded into a graph to obtain the overall loss function:
[0065] O=O up +O uc +O tp +O tc +O cc ;
[0066] The model is trained according to the obtained overall loss function. After the above operations, the embedding vectors of users, POIs, POI categories, and time in a unified shared dimensional space are obtained: Among them, when the related variables are jointly embedded, in order to capture the sequential transfer pattern of users between visiting different social locations, compared with the traditional single identity, this method uses two independent feature vectors to represent the unified POI category, namely include and Indicates that two feature vectors are assigned to each POI category node and Respectively represent the exit and entry identities of this node; the algorithm pseudo code of this step is shown in Table 1:
[0067] Table 1 is the pseudo code of the joint graph embedding algorithm
[0068]
[0069]
[0070] Step 2.3: Based on the embedding vectors of user, POI, POI category, and time obtained in step 2.2, mine the static preference score S of the user at a certain time. p :
[0071]
[0072] In the formula, p represents the social place location (Point of Interest, POI), such as xx supermarket, xx noodle restaurant, etc.), c represents the semantic category of the social place location (such as shopping, dining), u represents the user, τ represents the time item of the query, t represents the time, for At the same time, in order to express the user's transfer preference between social location categories, we assign two roles to each semantic category c of social location: "out" and "in" correspond to and Therefore, use It represents the score of semantic category conversion of social locations at adjacent moments. The larger the value, the more τ The categories of social locations visited by the user at the moment c τ The more likely it is that at the next τ+1 time, the more likely it is to visit c τ+1 ;
[0073] And based on the above embedding vector, the user's sequential transfer preference score S for the location is expressed q :
[0074]
[0075] In the formula, c τ 、c τ-1 and c τ+1 They correspond to the social location categories at the current moment, the previous moment, and the next moment, respectively. represents the user's personalized preference item, P(c τ |c τ-1 ,c τ+1 ) is the transition probability:
[0076]
[0077] in, represents the “out” component of the social location category feature at time -1, represents the “in” component of the social location category characteristics at time τ, represents the “in” component of the social location category feature at time +1, is the time interval between the time when the missing POI annotation occurs and the consecutive check-ins of the previous items. is the time interval between the moment when the missing POI annotation occurs and the consecutive check-in, η is the time interval threshold. When the consecutive check-in time is greater than η, it means that the user's check-in behavior is not affected by the sequence, but only by the long-term static preference;
[0078] Based on the above two factors, the user's probability score for the candidate POI is obtained by combining the user's static preference and sequence transfer preference:
[0079]
[0080] In the formula, θ represents the weight parameter, which ranges from [0,1]; when When it is greater than the time threshold η, it is considered that the user's check-in behavior is only affected by the long-term static preference; otherwise, it is affected by both factors.
[0081] Step 3. Considering the spatial proximity at the regional level, the urban area is divided into linear grids, each POI is embedded in the corresponding network area, and then the distance in the geographic space is weighted;
[0082] Specifically, the city is divided into grid areas, and the center coordinates of each grid area are obtained. According to the GPS coordinates of the user when the POI is missing, it is mapped to the corresponding grid area after mapping, and the distance between the grid area and the grid center coordinates corresponding to other candidate POIs is calculated, and this is used as the weight based on spatial distance:
[0083]
[0084] Where l represents the GPS coordinates when the POI is missing, p represents the GPS coordinates of the POI candidate, d(·) represents the GPS coordinate distance in L2 norm, a is a parameter, and X l represents the GPS coordinates of the center of the grid area mapped by l coordinates, and Xp represents the GPS coordinates of the center of the grid area mapped by the candidate POI.
[0085] Step 4. Combine all the influencing factors obtained in Step 2 and Step 3 to perform joint probability estimation, obtain the probability score of each candidate POI and rank it. According to the ranking, the POI check-in that the user missed at a certain time in the past is obtained, thereby predicting the location information that the user has visited; specifically,
[0086] The above influencing factors are combined to obtain the joint probability score:
[0087] S total (p,c|u,l,t)=(1+w l,p )·S p+q (p,c|·);
[0088] Then test it:
[0089]
[0090] Among them, Accuary@k represents the probability of the standard answer appearing in the top-k candidates, and Top-k is the probability of the standard answer appearing in the top-k candidates.total The recommendation results are given in descending order of scores. hit@k means that if there is a correct result among the top-k candidate POIs with the highest scores, it is set to 1, otherwise it is 0; |Dtest| represents the number of samples in the test set.
[0091] The applicant runs on a computer with Intel(R) Core(TM) i7-7700K CPU@4.20GHz and 2080Ti GPU, uses the method of this embodiment, uses the public data sets NYC and TKY and the literature (Q.Liu, S.Wu, L.Wang, and T.Tan, "Predicting the next location: A recurrent model with spatial and temporal contexts," in Proceedings of the AAAI Conference on ArtificialIntelligence, vol.30, no.1, 2016.), (C.Yang, L.Bai, C.Zhang, Q.Yuan, and J.Han, "Bridging collaborative filtering and semi-supervised learning: a neural approach for poi recommendation," in Proceedings of the 23rd ACM SIGKDDInternational Conference on Knowledge Discovery and Data Mining, 2017, pp.1245–1254.), (D.Xi, F.Zhuang, Y.Liu, J.Gu, H.Xiong, and Q.He, "Modelling of bidirectionalspatio-temporal dependence and users'dynamic preferences for missing poicheck-in identification," in Proceedings of the AAAI Conference on ArtificialIntelligence, vol.33, no.01, 2019, pp.5458–5465.) compared the proposed method and found that the recognition accuracy was greatly improved, so it can be applied to recommendation systems, missing population analysis, disease tracing, personalized services and other fields.
[0092] The above are only preferred embodiments of the present invention, and are not intended to limit the implementation methods and protection scope of the present invention. Those skilled in the art should be aware that all solutions obtained by equivalent substitutions and obvious changes made using the contents of the specification of the present invention should be included in the protection scope of the present invention.
Claims
1. A method for identifying missing POIs, characterized in that: The steps include: Step 1. Input the user check-in data of the location-based social network service platform, clean the data, and unify the data format to ensure data consistency; Step 2. Based on the data processed in step 1, multiple relationship graphs are constructed based on users, time and POI, POI category, and POI and POI category, and then the above multiple relationship graphs are jointly embedded, so as to embed the expression vectors of users, time, POI and POI category into a unified shared dimensional space, and jointly capture the sequential conversion pattern, time period influence and user preference; Step 3. Considering the spatial proximity at the regional level, the urban area is divided into grids, and each POI is embedded in the corresponding network area, and then the distance in the geographic space is weighted; Step 4. Combine the user preference obtained in step 2 and the distance weight obtained in step 3 to perform joint probability estimation, obtain the probability score of each candidate POI and rank them. According to the ranking, the POI check-ins that the user missed at a certain time in the past are obtained, thereby predicting the location information that the user has visited; Wherein, step 2 specifically also includes the following sub-steps: Step 2.1, based on the data processed in step 1, construct five relationship graph grids, namely user-POI, user-POI, time-POI, time-POI category, and POI category-POI category, where the weight in each graph represents the frequency of this check-in; Step 2.2: Based on the five relationship graphs obtained in 2.1, the expression vectors of users, POIs, POI categories, and time related to user personalized preferences are embedded into a unified shared dimensional space using the joint graph embedding method to obtain the embedding vectors of users, POIs, POI categories, and time in the unified shared dimensional space: in, include and Step 2.3: Based on the embedding vectors of user, POI, POI category, and time obtained in step 2.2, mine the static preference score S of the user at a certain time. p : Where p represents the location of a social place, c represents the semantic category of the location of a social place (, u represents the user, τ represents the time term of the query, t represents the time, for At the same time, in order to express the user's transfer preference between social location categories, we assign two roles to each semantic category c of social location: out and in correspond to and Therefore, use It represents the score of semantic category conversion of social locations at adjacent moments. The larger the value, the more τ The categories of social locations visited by the user at the moment c τ The more likely it is that at the next τ+1 time, the more likely it is to visit c τ+1 ; And based on the above embedding vector, the user's sequential transfer preference score S for the location is expressed q : In the formula, c τ 、c τ-1 and c τ+1 Respectively represent the social place location categories at the current moment, the previous moment, and the next moment, represents the user's personalized preference item, P(c τ |c τ-1 ,c τ+1 ) is the transition probability: in, represents the outgoing component of the social location category feature at time τ-1, represents the input component of the social location category feature at time τ, represents the input component of the social location category feature at time τ+1, is the time interval between the time when the missing POI annotation occurs and the consecutive check-ins of the previous items. is the time interval between the moment when the missing POI annotation occurs and the consecutive check-in, η is the time interval threshold. When the consecutive check-in time is greater than η, it means that the user's check-in behavior is not affected by the sequence, but only by the long-term static preference; Based on the above two factors, the user's probability score for the candidate POI is obtained by combining the user's static preference and sequence transfer preference: In the formula, θ represents the weight parameter, which ranges from [0,1]; when When it is greater than the time threshold η, it is considered that the user's check-in behavior is only affected by the long-term static preference; otherwise, it is affected by both factors.
2. The missing POI identification method according to claim 1, characterized in that: The method of data cleaning in step 1 is to filter out POI locations and users with less than n check-in times, where n is set according to actual needs.
3. The missing POI identification method according to claim 1, characterized in that: Step 3 specifically includes: Divide the city into grid areas and obtain the center coordinates of each grid area. Then, according to the GPS coordinates of the user when the POI is missing, map it to the corresponding grid area, calculate the distance between the grid area and the grid center coordinates corresponding to other candidate POIs, and calculate the weight of the spatial distance: l represents the GPS coordinates when the POI is missing, p represents the GPS coordinates of the POI candidate POI, d(·) represents the GPS coordinate distance in L2 norm, a is a parameter, X l It represents the GPS coordinates of the center of the grid area where the coordinates are mapped, and Xp represents the GPS coordinates of the center of the grid area where the candidate POI is mapped.
4. The missing POI identification method according to claim 1, characterized in that: In step 4, all influencing factors are integrated to perform joint probability estimation, and the probability score of each candidate POI is expressed as: S total (p,c|u,l,t)=(1+w l,p )·S p+q (p,c|·); In the formula, p represents POI, c represents POI category, u represents user, l represents GPS coordinates when POI is missing, t represents the time when POI is missing, and w represents the time when POI is missing. l,p represents the spatial distance weight obtained in step 3, S p+ q(p, c|·) represents the probability score of the user preference for the candidate POI obtained in step 2; Finally, the missing POI check-ins are identified and predicted to obtain the final result: Among them, Accuary@k represents the probability of the standard answer appearing in the top-k candidates, and Top-k is the probability of the standard answer appearing in the top-k candidates. total The recommendation results are given in descending order of scores. hit@k means that if there is a correct result among the top-k candidate POIs with the highest scores, it is set to 1, otherwise it is 0; |D test | represents the number of samples in the test set.
