Interest point recommendation method, storage medium and equipment
By combining term embedding, position embedding and frequency embedding vectors, using bidirectional self-attention layer and gated vectors, dynamically selecting user intention representations, and generating point-of-interest recommendations through pooling operations, solving the diversified user intentions in the prior art. The problem of inaccurate recommendations has achieved personalized and accurate recommendations of interest.
Patent Information
- Application Number
- CN202411516176.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-29
- Publication Date
- 2025-01-07
AI Technical Summary
The existing point of interest recommendation model focuses too much on the last point of interest item in the preference score and cannot accurately capture user intentions, especially in complex time and space, resulting in inaccurate recommendation results; at the same time, a fixed number of user interests may be missed or contain unnecessary interest.
By obtaining user check-in record items from location-based social networks, generating item embedding vectors, position embedding vectors and frequency embedding vectors, combining bidirectional self-attention layer and gated vectors, dynamically selecting key user intention representations, and passing through the maximum pool The pooling and average pooling operations generate the final recommendation of points of interest.
It realizes more personalized and accurate point-of-interest recommendations, can capture complex user intentions, adapt to long sequences and multi-intention scenarios, and improves the accuracy and personalization of recommendations.
Smart Images
Figure CN119271896A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of social network data mining, and in particular to a method, storage medium, and device for recommending points of interest. Background Art
[0002] With the continuous development of information and communication technology, people in modern society rely more on sharing their daily lives through social media and location-based services. This trend has triggered in-depth research on POI recommendation. POI recommendation technology is a personalized recommendation task that analyzes the current location of the target user, his / her historical check-in records at different POIs, and his / her personal preferences for POI visits to discover potential patterns to predict the next set of POIs suitable for a specific user and recommend places that the user may be interested in. Most existing POI recommendation methods focus on designing complex neural network-based encoders to capture the relationship between check-in items, mine user-preferred POIs, and make recommendations.
[0003] At present, scholars at home and abroad have done a lot of valuable research work on POI recommendation. The technologies used in existing POI recommendation algorithms mainly use various neural network-based encoders such as recurrent neural networks, graph neural networks or Transformers to model sequences containing multiple items. Recurrent neural networks are widely used to capture patterns in time series, while graph neural networks can mine the similarities of POIs and generate recommendations by modeling complex relationship graphs between users and POIs. Some methods use multi-head self-attention and different learning frameworks, and use attention mechanisms and graph convolutional networks to independently learn the content of spatiotemporal background and learn users' preferences for POIs. Some methods use the dual attention architecture of personalized item frequency to achieve personalized POI recommendation for users by aggregating the updated representations of all relevant positions in the user trajectory and matching the most reasonable candidate positions in the weighted representations.
[0004] Although the above neural network-based encoders have achieved good results in capturing the preference relationships of various aspects of user interest points, there are two important problems in interest point recommendation:
[0005] 1. The existing POI recommendation model pays too much attention to the last POI item in the preference score. However, if the last item is not important or noisy, it cannot accurately capture the user intention, ignoring the diverse user intentions in complex time and space, resulting in inaccurate recommendation results, especially poor performance in long sequence check-in sequences.
[0006] 2. Some methods try to identify multiple user interests in the long-term check-in sequence history of users. They adopt a fixed number of user interests and extract the same number of interests for all users. Since the number of interests of different users is different, this method may miss some interests or include unnecessary interests. Summary of the invention
[0007] The purpose of the present invention is to solve the problem that the recommendation results are inaccurate due to the diversity of user intentions when recommending user intentions, and to propose a method for recommending points of interest, comprising the following steps:
[0008] S1. Obtain all check-in record items of users from the location-based social network to obtain a sequence of check-in record items of each user;
[0009] S2. Generate an item embedding vector for each check-in record item, construct a position embedding vector to represent the order of the check-in record items in the sequence, construct a frequency embedding vector to represent the frequency of visiting the point of interest in the sequence, and concatenate the item embedding vector, the position embedding vector, and the frequency embedding vector to obtain a combined embedding vector;
[0010] S3, input the combined embedding vector into the bidirectional self-attention layer to obtain the representation of the combined embedding vector;
[0011] S4. According to the representation of the item embedding vector and the combination embedding vector, a gating vector is used to obtain the user intent representation;
[0012] S5. Calculate the importance of user intent representations, remove user intent representations with importance lower than a set threshold, and obtain key user intent representations;
[0013] S6. Based on the item embedding vector, decode the key user intention representation to obtain the item distribution;
[0014] S7. Aggregate the item distribution, and generate a recommendation of the points of interest that the target user is most likely to visit next based on the aggregated item distribution to the target user.
[0015] Furthermore, each check-in record includes: a user, a point of interest, the latitude and longitude of the point of interest, and a check-in timestamp.
[0016] Furthermore, the combined embedding vector expression is:
[0017]
[0018] Among them, x i represents the i-th combined embedding vector representation, represents the i-th item embedding vector, a i represents the embedding vector of the ith position, represents the frequency embedding vector of the i-th item in the sequence, |s| represents the number of check-in record items in the check-in record item sequence, and m represents the maximum number of check-in record items in the check-in record item sequence.
[0019] Furthermore, the expression of user intention is:
[0020] o i =g⊙v i +(1-g)⊙c i , i∈(1,...,|s|,m},
[0021] g=σ(W g [v i ;c i ] T ),
[0022] Among them, i represents the intention of the i-th user, is the gate vector, v i represents the i-th item embedding vector, c i represents the representation of the i-th combined embedding vector, |s| represents the number of check-in items in the check-in item sequence, m represents the maximum number of check-in items in the check-in item sequence, σ(·) is the sigmoid function, is a learnable weight matrix, ⊙ represents the dot product of the vector, [v i ;c i ] means to set v i and c i Make stitching.
[0023] Furthermore, the importance of user intent representation is calculated according to the following formula:
[0024]
[0025] γ=α-entmax(w·[o1;…;O |s| ;o m ] T ),
[0026] Among them, α-entmax(z) represents a sparse transformation about z. If α≠1, then If α = 1, p j represents the probability distribution of the jth element; p represents the vector of the entire probability distribution; Δ l represents all possible probability distributions in l dimension; z represents the weight vector; is the importance weight vector of each item; |s| represents the number of check-in record items in the check-in record item sequence; is a learnable parameter;
[0027] The key user intent is expressed as:
[0028] {h1, ..., h k}={γ i o i |γ i >0,i∈{1,...,|s|,m}},
[0029] Among them, h1, ..., h k Represents a sequence composed of k key user intention representations.
[0030] Furthermore, the expression of project distribution is:
[0031]
[0032] in, represents the sequence of k item distributions, represents the normalized h1,...,h k , represents the normalized V, and V represents the item embedding vector.
[0033] Furthermore, the maximum pooling operation and the average pooling operation are used to aggregate the item distribution, which is expressed as:
[0034]
[0035] in, represents the aggregated item distribution and β represents a hyperparameter.
[0036] Furthermore, the loss function during the training of the entire method is expressed as:
[0037]
[0038] in, represents the loss function; y represents the one-hot embedding vector of the target item. If the jth item is the target item, y(j) = 1, otherwise, y(j) = 0; τ is a hyperparameter. Represents the distribution of items in the aggregation; represents the distribution of items in the jth aggregation; Represents the distribution of the i-th aggregated item.
[0039] The present invention further provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the above-mentioned method for recommending points of interest is implemented.
[0040] The present invention also proposes an electronic device, including a processor and a memory, wherein the processor and the memory are interconnected, wherein the memory is used to store a computer program, the computer program includes computer-readable instructions, and the processor is configured to call the computer-readable instructions to execute the above-mentioned point of interest recommendation method.
[0041] The beneficial effects brought by the technical solution provided by the present invention are:
[0042] The present invention proposes to integrate the item embedding vector, position embedding vector, and frequency embedding vector of the check-in record item into a high-dimensional feature: a combined embedding vector. The relationship between the interest point and the user behavior is further obtained by combining the embedding vector, so as to achieve more personalized and accurate recommendations. This high-dimensional feature representation provides rich contextual information for the model, which helps to improve the quality of recommendations. The method of the present invention introduces a self-attention mechanism to capture the complex relationships in the item sequence, especially the associations between the interest points. The method of the present invention makes full use of each item in the check-in sequence through multiple feature representations (item embedding, sequence order, frequency vector, etc.). The method of the present invention can dynamically select the key user intentions that are most relevant to the user's current behavior to ensure that the key information is retained. In addition, these multiple representations processed by the self-attention mechanism are effectively decoded and aggregated to generate the final recommendation that can reflect the user's multiple intentions. By combining maximum pooling and average pooling, the model ensures that both the significant features in the sequence are captured and the consistency and integrity of the sequence are guaranteed. The effect is significant when dealing with long sequence and multi-intention user interactions. BRIEF DESCRIPTION OF THE DRAWINGS
[0043] Figure 1 is a flow chart of a method for recommending points of interest according to an embodiment of the present invention;
[0044] Figure 2 It is a block diagram of an electronic device in an exemplary embodiment of the present invention. DETAILED DESCRIPTION
[0045] To make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be further described below with reference to the accompanying drawings.
[0046] The flowchart of the method for recommending points of interest according to an embodiment of the present invention is as follows: Figure 1 , specifically including the following steps:
[0047] S1. Obtain all the user's check-in records from the location-based social network and obtain each user's check-in record sequence {t1, t2, ..., t |s|}, where |s| represents the number of check-in items in the check-in item sequence. The number of check-in items in each check-in item sequence is different, and the maximum value of |s| is m. Each check-in record includes check-in record information such as user, point of interest, latitude and longitude of the point of interest, and check-in timestamp.
[0048] Before obtaining all the user's check-in record items, it is necessary to perform the following data preprocessing on all the original check-in record items obtained: delete invalid check-in records with missing check-in record information, discard only one historical check-in record and the check-in records of points of interest that appear less than five times in the entire check-in sequence. After the above steps, all the check-in record items of the user in the embodiment of the present invention are obtained.
[0049] S2. Generate an item embedding vector for each check-in record item, construct a position embedding vector to represent the order of the check-in record items in the sequence, construct a frequency embedding vector to represent the frequency of visiting the point of interest in the sequence, and concatenate the item embedding vector, position embedding vector, and frequency embedding vector to obtain a combined embedding vector.
[0050] The combined embedding vector expression is:
[0051]
[0052] Among them, x i represents the i-th combined embedding vector representation, represents the i-th item embedding vector, a i represents the embedding vector of the ith position, represents the frequency embedding vector of the i-th item in the sequence, f i represents the frequency of the i-th item in the sequence, |s| represents the number of check-in record items in the check-in record item sequence, and m represents the maximum number of check-in record items in the check-in record item sequence.
[0053] Sort the check-in record sequence in reverse order according to the frequency of visiting the point of interest, so the last item t |s| Always corresponds to the first position embedding a1. In addition, the learnable parameters of the position and frequency embedding vectors are randomly initialized.
[0054] S3. Input the combined embedding vector into the bidirectional self-attention layer to capture the complex relationship between the check-in sequence items and generate a contextual representation associated with the check-in sequence information for each item to obtain the representation of the combined embedding vector.
[0055] {c1,…,c|s | , c m}=Self-attention([x1,...,x|s | , x m ])
[0056] where {c1,...,c |s| , c m} represents the sequence of representations of the combined embedding vector, c1 represents the representation of x1, that is, the context representation of the first check-in record item associated with the check-in record item sequence information; c |s| Represents x |s| , that is, the context representation of the |s|th sign-in record item associated with the sign-in record item sequence information; c m Represents x m , that is, the context representation of the mth check-in record item associated with the check-in record item sequence information.
[0057] S4. To ensure that multiple context representations do not become similar and can better reflect different user intentions, we use the Highway network to enhance these representations by emphasizing the term embedding vectors. i and term embedding v i Combined to derive user intent representation This representation takes into account both the overall information of the check-in sequence and the specific information of each item. Based on the representation of the item embedding vector and the combined embedding vector, the gating vector is used to obtain the user intent representation.
[0058] The expression of user intention is:
[0059] o i =g⊙v i +(1-g)⊙c i , i∈{1,...,|s|,m},
[0060] g=σ(W g [v i ;c i ] T ),
[0061] Among them, i represents the intention of the i-th user, is the gate vector, v i represents the i-th item embedding vector, c i represents the representation of the i-th combined embedding vector, |s| represents the number of check-in items in the check-in item sequence, m represents the maximum number of check-in items in the check-in item sequence, σ(·) is the sigmoid function, is a learnable weight matrix, ⊙ represents the dot product of the vector, [v i ;c i ] means to set v i and c i Make stitching.
[0062] S5. Calculate the importance of the user intent representation, remove the user intent representation whose importance is lower than a set threshold, and obtain the key user intent representation.
[0063] A sparse transformation α-entmax is used, which assigns zero probability to unimportant representations:
[0064]
[0065] γ=α-entmax(w·[o1;…;o |s| ;o m ] T ),
[0066] Among them, α-entmax(z) represents a sparse transformation about z. If α≠1, then If α = 1, pj represents the probability distribution of the jth element; p represents the vector of the entire probability distribution; Δ l represents the l-dimensional probability simplex, represents all possible probability distributions in the i-dimensional space; z represents the weight vector; is the importance weight vector of each item; |s| represents the number of check-in record items in the check-in record item sequence; is a learnable parameter; α is empirically set to 1.5. We extract sequence representations by masking unnecessary user intents The key user intent is expressed as:
[0067] {h1, ..., h k}={γ i o i |γ i >0,i∈{1,...,|s|,m}},
[0068] Among them, {h1,...,h k} represents a sequence of k key user intent representations. This allows the model to dynamically select multiple intent representations based on a given check-in sequence.
[0069] S6. Based on the item embedding vector, the key user intent representation is decoded to obtain the item distribution.
[0070] The expression of the project distribution is:
[0071]
[0072] in, represents the sequence of k item distributions, represents the normalized {h1, ..., h k}, represents the normalized V, and V represents the item embedding vector.
[0073] S7. Aggregate the item distribution, and generate a recommendation of the point of interest that the target user is most likely to visit next based on the aggregated item distribution (which is a probability distribution) to the target user.
[0074] The maximum pooling operation and the average pooling operation function are used to aggregate the item distribution. The average pooling operation retains the main features of multiple intents, while the average pooling operation captures the consistent intent in the check-in sequence. It is expressed as:
[0075]
[0076] in, denotes the aggregated item distribution, β denotes a hyperparameter, and MaxPool(·) and AvgPool(·) represent the maximum pooling and average pooling functions, respectively, which extract the maximum value (or average) for each dimension from multiple vectors.
[0077] The loss function during the training of the entire method is expressed as:
[0078]
[0079] in, represents the loss function; y represents the one-hot embedding vector of the target item. If the jth item is the target item, y(j) = 1, otherwise, y(j) = 0; τ is a hyperparameter for distribution smoothness. Represents the distribution of items in the aggregation; represents the distribution of items in the jth aggregation; represents the item distribution of the ith aggregation.
[0080] In an exemplary embodiment, a computer-readable storage medium is included, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the above-mentioned method for recommending points of interest is implemented.
[0081] See also Figure 2 In an exemplary embodiment, an electronic device is also included, including at least one processor, at least one memory, and at least one communication bus.
[0082] The memory stores a computer program, which includes computer-readable instructions. The processor calls the computer-readable instructions stored in the memory through a communication bus to execute the above-mentioned method for recommending points of interest.
[0083] Through the implementation of this method, the model can successfully capture multiple potential intentions in user sessions, thereby significantly improving the accuracy and personalization of POI recommendations. Unlike traditional recommendation models that rely only on a single sequence representation, this method fully utilizes each item in the check-in sequence through multiple feature representations (item embedding, sequence order, frequency vector, etc.). Under the guidance of the intent selection layer, the model can dynamically select the representation that is most relevant to the user's current behavior to ensure that key information is retained. In addition, the model effectively decodes and aggregates these multiple representations processed by the self-attention mechanism to generate final recommendations that reflect the user's multiple intentions. By combining maximum pooling and average pooling, the model ensures that both the salient features in the sequence are captured and the consistency and integrity of the sequence are guaranteed. This series of designs enables the model to perform well in conversational recommendation tasks that deal with complex user behaviors. In particular, when dealing with multi-intent scenarios, the model can accurately perceive the changes of users in different situations and flexibly adjust the recommendation strategy, and finally provide personalized POI recommendations. Experiments have shown that the implemented model has obvious superiority in indicators such as recommendation accuracy and user satisfaction, especially when dealing with long sequences and multi-intent user interactions.
[0084] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A method for recommending points of interest, characterized in that: The following steps are involved: S1. Obtain all check-in record items of users from the location-based social network to obtain a sequence of check-in record items of each user; S2. Generate an item embedding vector for each check-in record item, construct a position embedding vector to represent the order of the check-in record items in the sequence, construct a frequency embedding vector to represent the frequency of visiting the point of interest in the sequence, and concatenate the item embedding vector, the position embedding vector, and the frequency embedding vector to obtain a combined embedding vector; S3, input the combined embedding vector into the bidirectional self-attention layer to obtain the representation of the combined embedding vector; S4. Based on the representation of the item embedding vector and the combined embedding vector, a gating vector is used to obtain the user intent representation. S5. Calculate the importance of user intent representations, remove user intent representations with importance lower than a set threshold, and obtain key user intent representations; S6. Based on the item embedding vector, the key user intention representation is decoded to obtain the item distribution; S7. Aggregate the item distribution, and generate a recommendation of the points of interest that the target user is most likely to visit next based on the aggregated item distribution to the target user.
2. A method for recommending points of interest according to claim 1, characterized in that: Each check-in record includes: user, point of interest, latitude and longitude of the point of interest, and check-in timestamp.
3. A method for recommending points of interest according to claim 1, characterized in that: The combined embedding vector expression is: Among them, x i represents the i-th combined embedding vector representation, represents the i-th item embedding vector, a i represents the embedding vector of the ith position, represents the frequency embedding vector of the i-th item in the sequence, |s| represents the number of check-in record items in the check-in record item sequence, and m represents the maximum number of check-in record items in the check-in record item sequence.
4. A method for recommending points of interest according to claim 1, characterized in that: The expression of user intention is: o i =g⊙v i +(1-g)⊙c i ,i∈{1,...,|s|,m}, g=σ(W g [v i ;c i ] T ), Among them, i represents the intention of the i-th user, is the gating vector, vi represents the i-th item embedding vector, c i represents the representation of the i-th combined embedding vector, |s| represents the number of check-in items in the check-in item sequence, m represents the maximum number of check-in items in the check-in item sequence, σ(·) is the sigmoid function, is a learnable weight matrix, ⊙ represents the dot product of the vector, [v i ;c i ] means to set v i and c i Make stitching.
5. A method for recommending points of interest according to claim 4, characterized in that: The importance of user intent representation is calculated according to the following formula: γ=α-entmax(w·[o1;…;o |s| ;o m ] T ), Among them, α-entmax(z) represents a sparse transformation about z. If α≠1, then If α = 1, p j represents the probability distribution of the jth element; p represents the vector of the entire probability distribution; Δ 1 represents all possible probability distributions in l dimension; z represents the weight vector; is the importance weight vector of each item; |s| represents the number of check-in record items in the check-in record item sequence; is a learnable parameter; The key user intent is expressed as: {h1,...,h k }={γ i the i |c i >0,i∈{1,...,|s|,m}}, Among them, {h1,...,h k } represents a sequence composed of k key user intention representations.
6. A method for recommending points of interest according to claim 5, characterized in that: The expression of the project distribution is: in, represents the sequence of k item distributions, represents the normalized h1,...,h k , represents the normalized V, and V represents the item embedding vector.
7. A method for recommending points of interest according to claim 6, characterized in that: The maximum pooling operation and the average pooling operation are used to aggregate the item distribution, which is expressed as: in, represents the aggregated item distribution and β represents a hyperparameter.
8. A method for recommending points of interest according to claim 1, characterized in that: The loss function during the training of the entire method is expressed as: in, represents the loss function; y represents the one-hot embedding vector of the target item, if the jth item is the target item, then y(j) = 1, otherwise, y(j) = 0; τ is a hyperparameter for distribution smoothness; Represents the distribution of items in the aggregation; represents the distribution of items in the jth aggregation; Represents the distribution of the i-th aggregated item.
9. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 8 is implemented.
10. An electronic device, characterized in that: The method comprises a processor and a memory, wherein the processor and the memory are interconnected, wherein the memory is used to store a computer program, the computer program comprises computer-readable instructions, and the processor is configured to call the computer-readable instructions to execute the method according to any one of claims 1 to 8.