A location recommendation system and method

Abstract

The application provides a place recommendation system, comprising an encoding module, a sequence period feature extraction module, a geographical feature extraction module and a query mapping module. The encoding module is used for encoding original data to form a user sequence. The sequence period feature extraction module comprises a time interval feature extraction module for extracting interval rules in the user sequence, a time span feature extraction module for extracting time span rules in the user sequence, and an aggregation module for aggregating the interval rules and the time span rules. The geographical feature extraction module is used for learning place features at different geographical scales. The query mapping module is used for place recommendation according to the sequence period features and the place features. The system captures the period features of multiple time intervals and multiple time spans in the user sequence, extends to a physical geographical map, adds geographical feature information to the place recommendation, and effectively improves the reliability of the recommendation result.

Description

Technical Field

[0001] This invention relates to the field of recommendation system technology, and in particular to a location recommendation system and method. Background Technology

[0002] Location recommendation aims to suggest locations of interest to users and is widely used in numerous location-based social networks, such as Dianping and Gewara. On these platforms, users communicate with other users in the community by sharing their browsing history, and this browsing information also provides data for location recommendations. As one of the core services of social networks, location recommendation typically utilizes location and time information from users' browsing history, combined with historical records, to make recommendations. Most location recommendation methods abstract users' browsing history into a time-sorted sequence, and these methods can be broadly classified into three categories:

[0003] (1) Recommendation models based on recurrent networks. Models based on recurrent networks make recommendations by learning the continuous correlations between user access sequences. Markov chains and Long Short-Term Memory (LSTM) networks belong to this category. Their characteristic is that the output of the previous step is used as the input for the current step. However, this type of method is limited by the model's memory capacity and easily loses long sequence features, making it difficult to handle long sequence data. For the sparse and excessively long sequences commonly found in recommendations, this type of model has limited effectiveness and is difficult to train.

[0004] (2) Attention-based recommendation models. Attention-based models have been widely used in the recommendation field in recent years because they have strong generalization ability and can capture global features in user sequences. However, these methods usually have more training parameters, and they do not have the ability to learn complex sequence features.

[0005] (3) Recommendation models based on convolutional networks. Models based on convolutional networks typically use traditional convolutions on user sequences to extract local features. These methods often use 3×3 or 5×5 convolutional kernels, which, limited by the finite kernel size, often cannot fully capture the feature information in the sequence. As an improvement, some methods use dilated convolutions to expand the receptive field and capture a larger range of sequence features. However, these methods only utilize the single feature of dilated convolutions in expanding the receptive field, neglecting the ability of dilated convolutions to capture the periodic features of the sequence. Summary of the Invention

[0006] To address some or all of the problems in existing technologies, this invention provides a location recommendation system that focuses on leveraging temporal features to enhance recommendation results. It performs multi-scale time-dimensional recommendations by mining user patterns within these temporal features. The system includes:

[0007] The encoding module is used to encode the raw data to form a user sequence;

[0008] A sequence periodic feature extraction module, used to extract the periodic features of the user sequence, including:

[0009] The time interval feature extraction module is used to extract the interval pattern in the user sequence;

[0010] The time span feature extraction module is used to extract the time span pattern in the user sequence; and

[0011] The aggregation module aggregates the interval patterns and time span patterns to obtain the sequence periodic characteristics;

[0012] The geographic feature extraction module is used to learn location features at different geographic scales; and

[0013] The query mapping module is used to recommend locations based on the sequence periodicity features and location features.

[0014] Furthermore, the raw data includes the user's access location, access time, access coordinates, and location text.

[0015] Furthermore, the user sequences include time series with equal time intervals and geographic maps with equal physical spatial distances.

[0016] Furthermore, the time interval feature extraction module includes a multi-level binary search tree and a dilated convolutional network, wherein the features of the rightmost and leftmost leaf nodes searched in each layer of the binary search tree are combined with the dilated convolutional network.

[0017] Furthermore, the time span feature extraction module includes a multi-level line segment search tree and a temporal convolutional network, wherein each segment of the line segment search tree is independent of each other and there are no overlapping regions.

[0018] Furthermore, the aggregation module employs a nonlinear weighted aggregation method.

[0019] Furthermore, the geographic feature extraction module includes a multi-level line segment search tree and a temporal convolutional network, wherein adjacent left and right regions in each layer of the line segment search tree are combined through the temporal convolutional network.

[0020] Based on the system described above, another aspect of the present invention provides a location recommendation method, comprising:

[0021] Obtain user data;

[0022] Convert the user data into a user sequence;

[0023] Extract the sequence periodicity features and geographical features of the user sequence; and

[0024] Location recommendations are made based on the sequence periodicity and geographical features.

[0025] Further, converting the user data into a user sequence includes:

[0026] The user data is sorted according to the time of access records;

[0027] Divide the sorted user data into finite-length sequence groups; and

[0028] The sequence group is divided into time series with equal time intervals and geographical maps with equal physical spatial distances.

[0029] This invention provides a location recommendation system and method. To more accurately capture user preferences and provide users with recommendation results that conform to their historical access patterns at the current time, a hierarchical architecture model based on temporal convolutional networks is adopted. This model uses multi-periodic convolution for location recommendation, thereby acquiring sequence periodic features across multiple time intervals and spans. It can learn user access preferences under complex temporal patterns, resulting in higher accuracy of recommendation results in actual experiments. The system and method design two new types of tree-structured temporal convolutional modules to extract periodic features across multiple intervals and spans, and possess a larger receptive field, making it applicable to long-sequence data. Furthermore, the system and method extend the new tree-structured temporal convolution from sequences to physical geographic maps. Compared to other methods that only consider single distances and calculate absolute distances between locations, this grid segmentation method, combined with the relative physical distances of locations in multi-scale geographic space, makes distance factor calculation more efficient and considers more comprehensive factors. It also alleviates the problem of sparse distances between locations and can estimate physical distances between locations at multiple geographic scales. Attached Figure Description

[0030] To further illustrate the above and other advantages and features of the various embodiments of the present invention, a more specific description of the various embodiments of the present invention will be presented with reference to the accompanying drawings. It is to be understood that these drawings depict only typical embodiments of the invention and are therefore not intended to limit its scope. In the drawings, identical or corresponding parts will be indicated by identical or similar reference numerals for clarity.

[0031] Figure 1This diagram illustrates the structure of a location recommendation system according to an embodiment of the present invention.

[0032] Figure 2 A schematic flowchart illustrating a location recommendation method according to an embodiment of the present invention is shown.

[0033] Figure 3 This diagram illustrates a comparison of the effects of a location recommendation method according to an embodiment of the present invention with the Xi'an-Yo technology. Detailed Implementation

[0034] In the following description, the invention is described with reference to various embodiments. However, those skilled in the art will recognize that the embodiments may be practiced without one or more specific details or in conjunction with other alternatives and / or additional methods or components. In other instances, well-known structures or operations are not shown or described in detail so as not to obscure the inventive points of the invention. Similarly, for illustrative purposes, specific numbers and configurations are set forth to provide a comprehensive understanding of embodiments of the invention. However, the invention is not limited to these specific details.

[0035] In this specification, references to "an embodiment" or "this embodiment" mean that a particular feature, structure, or characteristic described in connection with that embodiment is included in at least one embodiment of the invention. The phrase "in one embodiment" appearing throughout this specification does not necessarily refer to the same embodiment in all instances.

[0036] It should be noted that the embodiments of the present invention describe the method steps in a specific order; however, this is only for illustrating the specific embodiment and not for limiting the order of the steps. On the contrary, in different embodiments of the present invention, the order of the steps can be adjusted according to actual needs.

[0037] Traditional location recommendation systems lack the ability to model and leverage periodic visit patterns in user behavior. Existing methods typically combine geographical factors, such as distance between locations, and semantic information, such as location features, with temporal features for location recommendations. They only consider the impact of very basic and simple time factors on recommendation results, such as assuming that visits more recent than the current time are more likely to be relevant. However, they struggle to learn complex temporal features such as visit cycle patterns and habits. In other words, they often overlook the periodic patterns in user visit records, such as daily visits to restaurants or weekly visits to supermarkets. The inventors, through research, discovered that mining the periodic patterns in user visit records can enhance location recommendations, enabling the model to fully utilize multiple sequence features in the time dimension and increasing the reliability of recommendation results.

[0038] Based on this, this patent proposes a location recommendation system and method that focuses on learning the periodic features of multiple time intervals and time spans in user access sequences, such as user access characteristics every four hours and user access characteristics during morning, noon, and evening time periods. This model can capture the periodic features of multiple time intervals and time spans in user sequences, flexibly learn various periodic features, and extend them to physical geographic maps, while adding geographic feature information to location recommendations. The system and method utilize temporal features to enhance recommendation results, performing multi-scale time-dimensional recommendations by mining user patterns existing in temporal features, thereby effectively improving the reliability of recommendation results.

[0039] The present invention will be further described below with reference to the accompanying drawings of the embodiments.

[0040] Figure 1 This diagram illustrates the structure of a location recommendation system according to an embodiment of the present invention. Figure 1 As shown, a location recommendation system is implemented in software and runs on a CPU or GPU server. It includes an encoding module 101, a sequence periodic feature extraction module 102, a geographic feature extraction module 103, and a query mapping module 104. The encoding module 101 encodes the raw data to form a user sequence. The sequence periodic feature extraction module 102 extracts periodic features based on the user sequence. The geographic feature extraction module 103 extracts geographic features based on the user sequence. The query mapping module 104 makes location recommendations based on the periodic and geographic features.

[0041] In one embodiment of the present invention, the original data mainly refers to the captured user access records, which include at least information such as user access location, access time, access coordinates, and location text. In one embodiment of the present invention, the encoding module 101 is used to extract user access sequences based on the user access records, then sort them according to the order of access time, and divide them into sequence groups of finite length. Finally, the sequence groups are preprocessed into time series and geographic maps with equal time intervals and equal physical spatial distances to obtain the user sequence.

[0042] like Figure 1 As shown, the sequence periodicity feature extraction module 102 includes a time interval feature extraction module 121, a time span feature extraction module 122, and an aggregation module 123. The time interval feature extraction module 121 is used to extract the interval pattern in the user sequence, the time span feature extraction module 122 is used to extract the time span pattern in the user sequence, and the aggregation module 123 is used to aggregate the interval pattern and the time span pattern to obtain the sequence periodicity feature.

[0043] In one embodiment of the present invention, the time interval feature extraction module 121 employs a multi-time interval convolutional model, which aims to capture the interval patterns in user sequences, such as the pattern of user visits every hour or every four hours. The multi-time interval convolutional model combines a multi-level binary search tree network structure with traditional temporal convolution, using a multi-level tree architecture to capture multi-time interval features in the sequence. In one embodiment of the present invention, the multi-time interval convolutional model combines the features of the rightmost and leftmost leaf nodes of each layer of the binary search tree with dilated convolutions in the temporal convolutional network, thereby achieving the effect of obtaining periodic features of different time intervals in each layer of the network. Specifically, in one embodiment of the present invention, the multi-time interval convolutional model includes a tree-like dilated convolutional network with multiple unit intervals, which has a binary tree structure and contains multiple levels of networks. The input data, as the root node, is located at the bottom layer of the network, and dilated convolutions at specified intervals are performed on the network to extract the left and right node information. The left and right node information is the output of this layer of the network and is further fed into the upper layer network as input data. The unit intervals of the dilated convolutions in each layer are inconsistent. By combining multi-level networks, the model can obtain the temporal features of user sequences at multiple intervals.

[0044] In one embodiment of the present invention, the time span feature extraction module 122 employs a multi-time span convolutional model, which aims to capture the time span patterns in user sequences, such as user access patterns in two-hour increments or in six-hour increments. The multi-time span convolutional model combines a multi-level segment search tree network structure with traditional temporal convolution, using a multi-level tree architecture to capture multi-time span features in the sequence. Specifically, in one embodiment of the present invention, the time span feature extraction module includes a multi-unit span tree-structured dilated convolutional network, based on a binary tree structure and composed of multiple layers. Multiple units of the network are considered as a single entity, and temporal features of the user sequence at different time spans are obtained through unions of different unit lengths. In each layer of the network, a specific dilated convolutional kernel is set, such as size and stride, to obtain node information at specified intervals, and the model output at multiple unit spans is obtained through layer combination. In one embodiment of the present invention, unlike the design of overlapping segments in traditional temporal networks, the multi-time-span convolutional model strictly defines the range of each segment on the segment tree, making each segment independent and non-overlapping. This allows the multi-time-span convolutional model to learn the sequence periodic features on each complete span segment, while also distinguishing it from other temporal network models. Specifically, this is achieved by coordinating the effective range of the upper and lower convolutional kernels, ensuring that the data's effective region is independent during data transmission. In one embodiment of the present invention, when setting the range of the convolutional kernel, considering the exponential growth characteristic of the segment search tree, the convolutional kernel can use a 2 / 4 / 8 growth pattern. Independent regions, on the one hand, prevent older sequences from acquiring more recent information during the transmission of sequence data from the lower layer to the top layer of the network, thus ensuring that temporal information is not leaked during network transmission. On the other hand, non-overlapping regions ensure that the temporal features learned by the network are not interfered with, which is beneficial for learning strongly correlated patterns. In one embodiment of the present invention, the aggregation module 123 uses a non-linear weight aggregation method to aggregate the interval pattern and time span pattern to obtain the sequence periodic features.

[0045] The geographic feature extraction module 103 aims to learn location features at different geographic scales, thereby further enriching location recommendations. In one embodiment of the invention, the geographic feature extraction module 103 divides the same physical geographic map into multi-level sub-maps to represent features at different scales, such as geographic features per unit square kilometer or per 500 square meters. The geographic feature extraction module 103 is a geographic-level extension of the multi-temporal-span convolutional model. Its biggest difference from the multi-temporal-span convolutional model is that it extends the 1D temporal-domain convolution originally applied to sequences to a 2D geographic map. The geographic feature extraction module 103 uses a multi-level architecture to capture multi-scale geographic features on the physical map, combining adjacent left and right regions in each layer of the two-dimensional map through temporal-domain convolution to obtain periodic features at different scales in each layer of the network.

[0046] Before using the location recommendation system described above for location recommendation, the system first needs to be trained. For system training, the selection and preprocessing of the dataset are crucial. In one embodiment of the present invention, real location information is selected as the dataset, which includes information such as user access time, access coordinates, and location text. User access sequences are extracted from this dataset and further preprocessed to provide reliable data input for model training. Specifically, considering the input requirements of the system, user access records are sorted chronologically and divided into finite-length sequence groups. According to the model characteristics, the sequences are preprocessed into time series with equal time intervals and geographical maps with equal physical spatial distances. Simultaneously, the processed data is divided into a 90% training dataset and a 10% test dataset, and the training and test sets are further divided into a 10% validation dataset. The system can be trained using the training dataset, optimized using the test dataset, and finally validated using the validation dataset.

[0047] Based on the location recommendation system described above Figure 2 This diagram illustrates a flow chart of a location recommendation method according to an embodiment of the present invention. Figure 2 As shown, a location recommendation method includes:

[0048] First, in step 201, user data is obtained. This involves acquiring the user's access records shared on the platform to form raw data.

[0049] Next, in step 202, data transformation. The original data is transformed into user sequences. In one embodiment of the present invention, the transformation of the user sequences includes: extracting user access sequences from the original data, sorting the user access sequences chronologically and dividing them into sequence groups of finite length, and finally preprocessing the sequence groups into time series with equal time intervals and geographic maps with equal physical spatial distances.

[0050] Next, in step 203, periodic features and geographical features are extracted. The time interval feature extraction module 121 extracts the interval patterns in the user sequence, and the time span feature extraction module 122 extracts the time span patterns in the user sequence. Then, the aggregation module 123 aggregates the interval patterns and time span patterns to obtain the sequence periodic features; and the geographical feature extraction module 103 extracts location features at different geographical scales; and

[0051] Finally, in step 204, location recommendation is performed. Locations are recommended based on the sequence periodicity and geographical features.

[0052] To verify the effectiveness of the system and method, experiments were conducted on a real dataset and compared with other existing models. The results are as follows: Figure 3 As shown, on different datasets, the location recommendation system and method provided by this invention achieves a recommendation accuracy that is superior to other related technologies and methods currently available.

[0053] Although various embodiments of the invention have been described above, it should be understood that they are presented by way of example only and not as limitations. It will be apparent to those skilled in the art that various combinations, modifications, and alterations can be made without departing from the spirit and scope of the invention. Therefore, the breadth and scope of the invention disclosed herein should not be limited by the exemplary embodiments disclosed above, but should be defined solely by the appended claims and their equivalents.

Claims

1. A place recommendation system, characterized by, The method comprises the following steps: encoding the original data to form a user sequence, wherein the original data comprises user access locations, access times, access coordinates and location texts, and the user sequence comprises time sequences with equal time intervals and geographical maps with equal physical space distances; extracting periodic features of the user sequence, comprising: extracting interval rules in the user sequence, wherein the interval feature extraction module is a binary tree structure and comprises a multi-level network, the root node of the multi-level network is input data, each layer comprises a specified interval convolution, and the left and right node information of any layer convolution is configured as the input data of the lower layer convolution; extracting time span rules in the user sequence, wherein the time span feature extraction module is a binary tree structure and comprises a multi-level network, each network layer comprises a convolution, the size and / or step length of the convolution are determined according to the required time span, and the slice ranges of each network layer are independent of each other without overlapping areas; and aggregating the interval rules and the time span rules to obtain sequence periodic features; learning location features of the user sequence at different geographical scales; and performing location recommendation according to the sequence periodic features and the location features.

2. The place recommendation system of claim 1, wherein, The unit intervals of each layer convolution are not the same.

3. The place recommendation system of claim 1, wherein, The aggregation module adopts a nonlinear weight aggregation method.

4. The place recommendation system of claim 1, wherein, The geographical feature extraction module comprises a multi-level line segment search tree and a time domain convolution network, wherein adjacent left and right regions in each layer of the line segment search tree are combined through the time domain convolution network.

5. A place recommendation method characterized by comprising: The method comprises the following steps: obtaining user data, wherein the user data comprises user access locations, access times, access coordinates and location texts; sorting the user data according to the time sequence of access records, dividing the sorted user data into sequence groups with limited lengths, dividing the sequence groups into time sequences with equal time intervals and geographical maps with equal physical space distances, and converting the user data into a user sequence; extracting sequence periodic features and geographical features of the user sequence, wherein extracting the sequence periodic features of the user sequence comprises: combining the features of the rightmost and leftmost leaf nodes searched by each layer of the binary search tree with a convolution network to obtain periodic features of different time intervals at each network layer, combining a multi-level line segment search tree on a time domain convolution network to capture multi-time span features on the user sequence; and performing location recommendation according to the sequence periodic features and the location features. ​

Images