Method and device for training point-of-interest classification model and point-of-interest classification method

Abstract

The invention provides an interest point classification model training method and device and an interest point classification method, and relates to the field of computers. And enhancing at least one piece of interest point training data in the interest point description data and the interest point category label description data by utilizing a self-supervision auxiliary task, and then training a classification model based on the enhanced interest point training data by utilizing an interest point classification task to obtain an interest point classification model. Therefore, the quality of training data actually used in training is improved, and the accuracy of the interest point classification model obtained through training is improved.

Description

Technical Field

[0001] The present disclosure relates to the field of computers, and in particular to a method and device for training a point of interest classification model and a method for classifying a point of interest. Background Art

[0002] In geographic information systems, Point of Interest (POI) refers to all geographical objects that can be abstracted as points, especially some geographical entities that are closely related to people's lives, such as schools, banks, restaurants, gas stations, hospitals, supermarkets, etc. Each point of interest contains four aspects of information, such as name, category, longitude, and latitude. Points of interest are classified into primary, secondary, and even more categories. Point of interest classification refers to classifying points of interest into corresponding categories based on their characteristics.

[0003] In some related technologies, a large amount of interest point training data with interest point category labels is obtained, and based on the classification task, the interest point training data is used to train the classification model to obtain the interest point classification model.

[0004] However, the sources of POI training data are wide-ranging and their quality varies, which will affect the accuracy of the trained POI classification model. Summary of the invention

[0005] In the disclosed embodiment, the self-supervised auxiliary task is used to enhance at least one of the interest point description data and interest point category label description data, and then the interest point classification task is used to train the classification model based on the enhanced interest point training data to obtain the interest point classification model. Thus, the quality of the training data actually used in the training is improved, and the accuracy of the interest point classification model obtained by training is further improved.

[0006] Some embodiments of the present disclosure propose a training method for a point of interest classification model, comprising: utilizing a self-supervised auxiliary task to enhance at least one item of point of interest training data among the point of interest description data and the point of interest category label description data; utilizing the point of interest classification task to train a classification model based on the enhanced point of interest training data to obtain a point of interest classification model.

[0007] In some embodiments, the classification model includes a first encoder for processing interest point description data, and the enhancing the interest point description data includes: forming a first positive sample pair according to two vector representations respectively output by the first encoder based on the same interest point description input; forming a first negative sample pair according to two vector representations respectively output by the first encoder based on different interest point descriptions input; and obtaining enhanced interest point training data according to the first positive sample pair and the first negative sample pair.

[0008] In some embodiments, the self-supervised auxiliary task includes a first self-supervised auxiliary task, and training the classification model includes: constructing a loss function of the first self-supervised auxiliary task based on the similarity between two samples in the first positive sample pair and the similarity between two samples in the first negative sample pair.

[0009] In some embodiments, the first encoder comprises a random dropout layer.

[0010] In some embodiments, the enhancing of the interest point category label description data includes: taking the interest point description as the first tuple, taking the interest point category label description to which the interest point description belongs as the second tuple, taking the interest point category label description not belonging to the interest point description as the third tuple, forming a second positive sample pair based on the first tuple and the second tuple, and forming a second negative sample pair based on the first tuple and each third tuple; and obtaining enhanced interest point training data based on the second positive sample pair and the second negative sample pair.

[0011] In some embodiments, the classification model includes a first encoder for processing interest point description data, a second encoder for processing interest point category label description data, and a multilayer perceptron cascaded with the first encoder and the second encoder, the self-supervised auxiliary task includes a second self-supervised auxiliary task, and training the classification model includes: calculating the similarity between the feature representation of the first tuple output by the multilayer perceptron and the feature representation of the second tuple output by the multilayer perceptron as the first similarity; calculating the similarity between the feature representation of the first tuple output by the multilayer perceptron and the feature representation of the third tuple output by the multilayer perceptron as the second similarity; and constructing a loss function for the second self-supervised auxiliary task based on the first similarity and the second similarity.

[0012] In some embodiments, the first encoder and the second encoder share parameters during training.

[0013] In some embodiments, the third tuple is determined according to other POI category label descriptions in the same level POI category category except the second tuple.

[0014] In some embodiments, the classification model also includes a classification head cascaded with a multi-layer perceptron, and the training of the classification model includes: constructing a loss function for the point of interest classification task based on the point of interest category and the point of interest category label description output by the classification head; using a total loss function formed by the loss function of the self-supervised auxiliary task and the loss function of the point of interest classification task, iteratively updating the parameters of the classification model to train the classification model.

[0015] In some embodiments, obtaining the interest point classification model includes: determining a model consisting of a first encoder, a multilayer perceptron, and a classification head in a trained classification model as the interest point classification model.

[0016] Some embodiments of the present disclosure propose a method for classifying points of interest, comprising: inputting a description of a point of interest into a point of interest classification model for processing, wherein the point of interest classification model is obtained according to a training method for the point of interest classification model; and obtaining a point of interest category corresponding to the point of interest description output by the point of interest classification model.

[0017] Some embodiments of the present disclosure provide a training device for a point of interest classification model, comprising: a memory; and a processor coupled to the memory, wherein the processor is configured to execute a training method for the point of interest classification model based on instructions stored in the memory.

[0018] Some embodiments of the present disclosure propose a training device for a point of interest classification model, comprising: an enhancement module, configured to enhance at least one item of interest point training data among interest point description data and interest point category label description data using a self-supervised auxiliary task; and a training module, configured to train a classification model based on the enhanced interest point training data using the interest point classification task to obtain an interest point classification model.

[0019] Some embodiments of the present disclosure provide a computer-readable storage medium having computer instructions stored thereon, which, when executed by a processor, implement the steps of a method for training an interest point classification model.

[0020] Some embodiments of the present disclosure provide a computer program product, including computer instructions, characterized in that when the computer instructions are executed by a processor, the steps of a method for training a point of interest classification model are implemented. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] The following is a brief introduction to the drawings required for use in the embodiments or related technical descriptions. The present disclosure can be more clearly understood according to the following detailed description with reference to the drawings.

[0022] Obviously, the drawings described below are only some embodiments of the present disclosure, and a person skilled in the art can obtain other drawings based on these drawings without creative work.

[0023] Figure 1 A schematic diagram showing a trained classification model and its training according to some embodiments of the present disclosure.

[0024] Figure 2 A flowchart illustrating a method for training a point of interest classification model according to some embodiments of the present disclosure is shown.

[0025] Figure 3 A schematic diagram showing some embodiments of the present disclosure of enhancing interest point description data using a first self-supervised auxiliary task.

[0026] Figure 4 A schematic diagram showing some embodiments of the present disclosure of enhancing interest point category label description data using a second self-supervised auxiliary task.

[0027] Figure 5 A schematic diagram showing a method for classifying points of interest according to some embodiments of the present disclosure.

[0028] Figure 6 A schematic diagram showing the structure of a training device for a point of interest classification model according to some embodiments of the present disclosure.

[0029] Figure 7 A schematic diagram showing the structure of a training device for a point of interest classification model according to some embodiments of the present disclosure. DETAILED DESCRIPTION

[0030] It should be noted that the relative arrangement of components and steps, the numerical expressions and numerical values ​​set forth in these embodiments do not limit the scope of the present disclosure unless specifically stated otherwise.

[0031] Those skilled in the art can understand that the terms "first" and "second" in the embodiments of the present disclosure are only used to distinguish different steps, devices or modules, etc., and neither represent any specific technical meaning nor indicate the necessary logical order between them.

[0032] It should also be understood that in the embodiments of the present disclosure, “plurality” may refer to two or more than two, and “at least one” may refer to one, two, or more than two.

[0033] It should also be understood that any component, data or structure mentioned in the embodiments of the present disclosure can generally be understood as one or more, unless explicitly limited or otherwise indicated in the context.

[0034] In addition, the term "and / or" in the present disclosure is only a description of the association relationship of associated objects, indicating that three relationships may exist. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. In addition, the character " / " in the present disclosure generally indicates that the associated objects before and after are in an "or" relationship.

[0035] It should also be understood that the description of the various embodiments in the present disclosure focuses on the differences between the various embodiments, and the same or similar aspects thereof can be referenced to each other, and for the sake of brevity, they will not be described one by one.

[0036] At the same time, it should be understood that for the convenience of description, the sizes of the various parts shown in the drawings are not drawn according to the actual proportional relationship.

[0037] The following description of at least one exemplary embodiment is merely illustrative in nature and is in no way intended to limit the present disclosure, its application, or uses.

[0038] Technologies, methods, and equipment known to ordinary technicians in the relevant art may not be discussed in detail, but where appropriate, the technologies, methods, and equipment should be considered as part of the specification.

[0039] It should be noted that like reference numerals and letters refer to similar items in the following figures, and therefore, once an item is defined in one figure, it need not be further discussed in subsequent figures.

[0040] In addition, in order to avoid obscuring the present disclosure due to unnecessary details, only the processing steps and / or device structures closely related to at least the scheme according to the present disclosure are shown in the drawings, and other details that are not closely related to the present disclosure are omitted. It should also be noted that similar reference numerals and letters in the drawings indicate similar items, and therefore once an item is defined in one drawing, it does not need to be discussed again for subsequent drawings.

[0041] In order to solve the problem that the sources of interest point training data are wide and the quality is uneven, which in turn affects the accuracy of the interest point classification model obtained through training, the embodiment of the present disclosure uses a self-supervised auxiliary task to enhance the interest point training data, and then uses the interest point classification task to train the classification model based on the enhanced interest point training data, and finally obtains the interest point classification model. Thus, the quality of the training data actually used in the training is improved, and the accuracy of the interest point classification model obtained through training is improved.

[0042] Figure 1 Schematic diagram showing the trained classification model and its training in some embodiments of the present disclosure. Figure 1As shown, the trained classification model includes an encoder, a multilayer perceptron (MLP) and a classification head. The encoder is cascaded with the multilayer perceptron, and the multilayer perceptron is cascaded with the classification head. The encoder and the multilayer perceptron are used to extract features, and the classification head uses the extracted features for classification. The main function of the classification head is to convert the high-dimensional feature vector into a low-dimensional category probability distribution, which can be achieved, for example, by a series of fully connected layers, which map the feature vector to an output space equal to the number of categories. The encoder includes a first encoder for processing interest point description data, and may also include a second encoder for processing interest point category label description data. The first encoder and the second encoder share parameters during the training process. The encoder can use a pre-trained language model, such as the BERT (Bidirectional Encoder Representations from Transformers, Bidirectional Encoder Representations based on Transformer) model. The first encoder may include a random dropout layer. The random dropout layer randomly discards some neurons (and their connections) from the neural network during training, thereby achieving regularization of the neural network, reducing its structural risk, preventing overfitting, and improving the generalization ability of the model. The multi-layer perceptron may be implemented, for example, using a two-layer MLP, but is not limited to this example.

[0043] Figure 2 A flowchart of a method for training a point of interest classification model according to some embodiments of the present disclosure is shown. Figure 2 As shown, the training method of the interest point classification model includes the following steps.

[0044] In step 210, at least one item of interest point training data among the interest point description data and the interest point category label description data is enhanced by using a self-supervised auxiliary task.

[0045] The self-supervised auxiliary task is an auxiliary task in self-supervised learning, which aims to enhance the ability to learn feature representations from unlabeled data. The first self-supervised auxiliary task and / or the second self-supervised auxiliary task are used to enhance the interest point training data. The interest point training data includes interest point description data and interest point category label description data. The interest point description data is enhanced by the first self-supervised auxiliary task. The interest point category label description data is enhanced by the second self-supervised auxiliary task.

[0046] The point of interest description data is the description data of the point of interest (such as a geographic object), which can be text data. The point of interest category label description data is the description data of the label of the category to which the point of interest belongs, which can be text data. For example, the point of interest description data is "a certain brand of electric vehicles (a certain branch)", and the corresponding point of interest category label description data is "shopping service / specialty store", where "shopping service" is a first-level category and "specialty store" is a second-level category.

[0047] In step 220, the classification model is trained based on the enhanced interest point training data using the interest point classification task to obtain an interest point classification model.

[0048] The trained classification model may be, for example, a general neural network model capable of implementing a classification function. The classification model is trained using the interest point training data to obtain an interest point classification model capable of classifying the interest point description data.

[0049] In the disclosed embodiment, the self-supervised auxiliary task is used to enhance at least one of the interest point description data and interest point category label description data, and then the interest point classification task is used to train the classification model based on the enhanced interest point training data to obtain the interest point classification model. Thus, the quality of the training data actually used in the training is improved, and the accuracy of the interest point classification model obtained by training is further improved.

[0050] Combine the following Figure 3 Description,The first self-supervised auxiliary task is used to enhance the interest point description data.

[0051] The enhancement of the interest point description data includes: forming a positive sample pair (first positive sample pair) according to the two vector representations respectively output by the first encoder based on the same interest point description input, that is, the two vector representations respectively output by the first encoder based on the same interest point description input as the positive sample pair; forming a negative sample pair (first negative sample pair) according to the two vector representations respectively output by the first encoder based on different interest point descriptions input, that is, the two vector representations respectively output by the first encoder based on different interest point descriptions input as the negative sample pair; according to the first positive sample pair and the first negative sample pair, the enhanced interest point training data is obtained. Correspondingly, training the classification model includes: constructing a loss function of the first self-supervised auxiliary task based on the similarity of the two samples in the first positive sample pair and the similarity of the two samples in the first negative sample pair.

[0052] For example, the interest point description data includes three different interest point descriptions, namely, interest point description a, interest point description b, and interest point description c. The first encoder forms a positive sample pair (first positive sample pair) based on the two vector representations respectively output by the interest point description a. The first encoder forms a positive sample pair (first positive sample pair) based on the two vector representations respectively output by the interest point description b. The first encoder forms a positive sample pair (first positive sample pair) based on the two vector representations respectively output by the interest point description c. The first encoder forms a negative sample pair (first negative sample pair) based on the vector representation output by the interest point description a and the vector representation output by the interest point description b. The first encoder forms a negative sample pair (first negative sample pair) based on the vector representation output by the interest point description a and the vector representation output by the interest point description c. The first encoder forms a negative sample pair (first negative sample pair) based on the vector representation output by the interest point description b and the vector representation output by the interest point description c. The loss function of the first self-supervised auxiliary task is, for example, Among them, τ is the temperature hyperparameter, z i and z′ i are two vector representations output by the first encoder based on the interest point description i, z′ j is a vector representation of the output of the first encoder based on the interest point description j, B is the batch size, batch in deep learning refers to a group of samples used to train the model in one iteration, N is the number of training samples, and the general calculation function of the similarity between information items h1 and h2 is

[0053] Considering that the POI description data is usually short, if the text enhancement method such as replacement or random addition / deletion is directly used, a large amount of noise will be introduced. This embodiment uses the first self-supervised auxiliary task to enhance the POI description data. Even if the original POI description data is noisy and inconsistent, high-quality training data can be obtained to enhance the robustness of the model.

[0054] Combine the following Figure 4 Description,The second self-supervised auxiliary task is used to enhance the interest point category label description data.

[0055] The enhancement of the interest point category label description data includes: taking the interest point description as the first tuple, taking the interest point category label description to which the interest point description belongs as the second tuple, and taking the interest point category label description not belonging to the interest point description as the third tuple. For example, other interest point category label descriptions except the second tuple in the same level interest point category are determined as the third tuple, forming a positive sample pair (second positive sample pair) according to the first tuple and the second tuple, that is, the first tuple and the second tuple are used as the positive sample pair, forming a negative sample pair (second negative sample pair) according to the first tuple and each third tuple, that is, the first tuple and each third tuple are used as the negative sample pair, and there may be multiple second negative sample pairs; according to the second positive sample pair and the second negative sample pair, the enhanced interest point training data is obtained. Accordingly, training the classification model includes: calculating the similarity between the feature representation of the first tuple output by the multilayer perceptron and the feature representation of the second tuple output by the multilayer perceptron as the first similarity; calculating the similarity between the feature representation of the first tuple output by the multilayer perceptron and the feature representation of the third tuple output by the multilayer perceptron as the second similarity; and constructing a loss function for the second self-supervised auxiliary task based on the first similarity and the second similarity. The feature representation of the first tuple output by the multilayer perceptron is the output of the first tuple passing through the first encoder and the multilayer perceptron in sequence. The feature representation of the second tuple output by the multilayer perceptron is the output of the second tuple passing through the second encoder and the multilayer perceptron in sequence. The feature representation of the third tuple output by the multilayer perceptron is the output of the third tuple passing through the second encoder and the multilayer perceptron in sequence.

[0056] For example, a certain point of interest description is "a certain brand of electric car (a certain branch)", and the point of interest category label description to which it belongs is "shopping service / specialty store", where the first-level category "shopping service" includes, for example, "specialty store", "convenience store", "convenience store" and other second-level categories, then the non-belonging point of interest category label description can be "shopping service / convenience store" or "shopping service / convenience store", then the triplet is, for example, ("a certain brand of electric car (a certain branch)", "shopping service / specialty store", "shopping service / convenience store" or "shopping service / convenience store"), where ("a certain brand of electric car (a certain branch)", "shopping service / specialty store") constitutes a second positive sample pair, ("a certain brand of electric car (a certain branch)", "shopping service / convenience store") constitutes a second negative sample pair, and ("a certain brand of electric car (a certain branch)", "shopping service / convenience store") constitutes a second negative sample pair. The loss function of the second self-supervised auxiliary task is, for example, Among them, h i is the feature representation of the first tuple of the multilayer perceptron output, is the feature representation of the second tuple of the multilayer perceptron output, It is the feature representation of the third tuple output by the multilayer perceptron. The meanings of other symbols are as mentioned above and will not be repeated here.

[0057] In the classification of interest points, the number of interest points of different categories may be seriously unbalanced, making it difficult for the classification model to identify the minority class. To address this problem, this embodiment uses the contrastive learning task of the second self-supervised auxiliary task to perform contrastive learning on the interest point category label description data (supervision information), enhance the supervision information, and improve the model classification effect when the interest point categories are unbalanced.

[0058] In step 220, the classification model is trained based on the enhanced POI training data to obtain the POI classification model, including: constructing a loss function for the POI classification task according to the POI category and the POI category label description output by the classification head; iteratively updating the parameters of the classification model by using the total loss function formed by the loss function of the self-supervised auxiliary task (the loss function of the first self-supervised auxiliary task and / or the loss function of the second self-supervised auxiliary task) and the loss function of the POI classification task to train the classification model; and determining the model composed of the first encoder, the multilayer perceptron and the classification head in the trained classification model as the POI classification model.

[0059] For example, the loss function for the interest point classification task is Among them, y i and are the interest point category label description corresponding to the interest point description i and the interest point category output by the classification head, and the meanings of other symbols are as mentioned above and will not be repeated here. The total loss function is, for example, l=l1+αl2+βl3 or l=γl1+αl2+βl3, where α, β, γ are weight parameters that can be set.

[0060] Figure 5 A schematic diagram showing a method for classifying points of interest in some embodiments of the present disclosure is shown. Figure 5 As shown, the interest point classification method of this embodiment includes the following steps.

[0061] In step 510, the description of the interest point to be classified is input into the interest point classification model for processing, wherein the interest point classification model is obtained according to the aforementioned training method.

[0062] The description of the point of interest to be classified is input from the first encoder of the point of interest classification model, and is processed by the first encoder, the multilayer perceptron and the classification head in the point of interest classification model, and the prediction result of the point of interest category is output from the classification head of the point of interest classification model.

[0063] In step 550, the interest point category corresponding to the interest point description output by the interest point classification model is obtained.

[0064] Based on the high-quality POI classification model obtained through the aforementioned enhanced training, the POI category corresponding to the POI description can be determined more accurately.

[0065] Point of interest categories are very important in many applications. For example, in community management scenarios, they help community managers understand the distribution of functional areas in the management area and facilitate relevant decision-making; in map service scenarios, they classify geographic objects for user query and navigation; in business analysis scenarios, companies can analyze the distribution of points of interest of different categories for market analysis and site selection decisions.

[0066] Figure 6 A schematic diagram showing the structure of a training device for a point of interest classification model according to some embodiments of the present disclosure is shown. Figure 6 As shown, the training device 600 for the interest point classification model of this embodiment includes: a memory 610 and a processor 620 coupled to the memory 610, and the processor 620 is configured to execute the training method for the interest point classification model in any of the aforementioned embodiments based on the instructions stored in the memory 610.

[0067] The training device 600 for the interest point classification model may further include an input / output interface 630 , a network interface 640 , a storage interface 650 , etc. These interfaces 630 , 640 , 650 , the memory 610 , and the processor 620 may be connected, for example, via a bus 660 .

[0068] The memory 610 may include, for example, a system memory, a fixed non-volatile storage medium, etc. The system memory may store, for example, an operating system, an application program, a boot loader, and other programs.

[0069] The processor 620 may be implemented by a general purpose processor, a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic devices, discrete gates or transistors and other discrete hardware components.

[0070] Among them, the input and output interface 630 provides a connection interface for input and output devices such as a display, a mouse, a keyboard, and a touch screen. The network interface 640 provides a connection interface for various networked devices. The storage interface 650 provides a connection interface for external storage devices such as SD cards and USB flash drives. The bus 660 can use any bus structure among a variety of bus structures. For example, the bus structure includes but is not limited to the Industry Standard Architecture (ISA) bus, the Micro Channel Architecture (MCA) bus, and the Peripheral Component Interconnect (PCI) bus.

[0071] Figure 7 A schematic diagram showing the structure of a training device for a point of interest classification model according to some embodiments of the present disclosure is shown. Figure 7 As shown, the training device 700 for the point of interest classification model of this embodiment includes: an enhancement module 710, which is configured to enhance at least one item of interest point training data among the interest point description data and the interest point category label description data by using a self-supervised auxiliary task; and a training module 720, which is configured to train the classification model based on the enhanced interest point training data by using the interest point classification task to obtain the interest point classification model.

[0072] The enhancement module 710 is configured to form a first positive sample pair according to two vector representations respectively output by the first encoder based on the same interest point description input; form a first negative sample pair according to two vector representations respectively output by the first encoder based on different interest point descriptions input; and obtain enhanced interest point training data according to the first positive sample pair and the first negative sample pair. The training module 720 is configured to construct a loss function of the first self-supervised auxiliary task based on the similarity of the two samples in the first positive sample pair and the similarity of the two samples in the first negative sample pair.

[0073] The enhancement module 710 is configured to take the description of the point of interest as the first tuple, the description of the interest point category label to which the description of the point of interest belongs as the second tuple, and the description of the interest point category label not belonging to the description of the point of interest as the third tuple, form a second positive sample pair according to the first tuple and the second tuple, and form a second negative sample pair according to the first tuple and each third tuple; and obtain enhanced interest point training data according to the second positive sample pair and the second negative sample pair. The training module 720 is configured to calculate the similarity between the feature representation of the first tuple output by the multilayer perceptron and the feature representation of the second tuple output by the multilayer perceptron as the first similarity; calculate the similarity between the feature representation of the first tuple output by the multilayer perceptron and the feature representation of the third tuple output by the multilayer perceptron as the second similarity; and construct a loss function for the second self-supervised auxiliary task based on the first similarity and the second similarity.

[0074] The training module 720 is configured to construct a loss function for the point of interest classification task based on the point of interest category and the point of interest category label description output by the classification head; iteratively update the parameters of the classification model using the total loss function formed by the loss function of the self-supervised auxiliary task and the loss function of the point of interest classification task to train the classification model; and determine the model composed of the first encoder, the multilayer perceptron and the classification head in the trained classification model as the point of interest classification model.

[0075] It should be understood by those skilled in the art that the embodiments of the present disclosure may be provided as methods, systems, or computer program products. Therefore, the present disclosure may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present disclosure may take the form of a computer program product implemented on one or more (non-transient) computer-readable storage media (including but not limited to disk storage, CD-ROM, optical storage, cloud storage, etc.) containing computer program code. A computer program product should be understood as a software product that implements its solution mainly through a computer program.

[0076] The present disclosure is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present disclosure. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0077] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0078] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process in the computer or other programmable device. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0079] The above description is only a preferred embodiment of the present disclosure and is not intended to limit the present disclosure. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present disclosure should be included in the protection scope of the present disclosure.

Claims

1. A method for training a point of interest classification model, comprising: At least one of the interest point description data and the interest point category label description data is enhanced by using the self-supervised auxiliary task; Using the interest point classification task, the classification model is trained based on the enhanced interest point training data to obtain the interest point classification model.

2. The training method according to claim 1, wherein: The classification model includes a first encoder for processing the point of interest description data, wherein enhancing the point of interest description data includes: Form a first positive sample pair according to two vector representations respectively output by the first encoder based on the same interest point description input; Form a first negative sample pair according to two vector representations respectively output by the first encoder based on different interest point descriptions of the input; Enhanced interest point training data is obtained according to the first positive sample pair and the first negative sample pair.

3. The training method according to claim 2, wherein: The self-supervised auxiliary task includes a first self-supervised auxiliary task, and training the classification model includes: Based on the similarity between the two samples in the first positive sample pair and the similarity between the two samples in the first negative sample pair, a loss function for the first self-supervised auxiliary task is constructed.

4. The training method according to claim 2, wherein: The first encoder includes a random dropout layer.

5. The training method according to any one of claims 1 to 4, wherein: The step of enhancing the interest point category label description data includes: Taking the description of the point of interest as the first tuple, taking the description of the category label of the point of interest to which the description of the point of interest belongs as the second tuple, taking the description of the category label of the point of interest not belonging to the description of the point of interest as the third tuple, forming a second positive sample pair according to the first tuple and the second tuple, and forming a second negative sample pair according to the first tuple and each third tuple; Enhanced interest point training data is obtained according to the second positive sample pair and the second negative sample pair.

6. The training method according to claim 5, wherein: The classification model includes a first encoder for processing interest point description data, a second encoder for processing interest point category label description data, and a multilayer perceptron cascaded with the first encoder and the second encoder, the self-supervised auxiliary task includes a second self-supervised auxiliary task, and training the classification model includes: Calculate the similarity between the feature representation of the first tuple output by the multilayer perceptron and the feature representation of the second tuple output by the multilayer perceptron as a first similarity; Calculate the similarity between the feature representation of the first tuple output by the multilayer perceptron and the feature representation of the third tuple output by the multilayer perceptron as the second similarity; Based on the first similarity and the second similarity, a loss function for the second self-supervised auxiliary task is constructed.

7. The training method according to claim 6, wherein: The first encoder and the second encoder share parameters during training.

8. The training method according to claim 5, wherein: The third tuple is determined according to other interest point category label descriptions in the same level interest point category except the second tuple.

9. The training method according to claim 6, wherein: The classification model also includes a classification head cascaded with a multi-layer perceptron, and the training of the classification model includes: According to the POI categories and POI category label descriptions output by the classification head, a loss function for the POI classification task is constructed; The total loss function formed by the loss function of the self-supervised auxiliary task and the loss function of the interest point classification task is used to iteratively update the parameters of the classification model to train the classification model.

10. The training method according to claim 9, wherein: The interest point classification model includes: A model consisting of the first encoder, the multilayer perceptron and the classification head in the trained classification model is determined as the interest point classification model.

11. A method for classifying points of interest, comprising: Inputting the interest point description into an interest point classification model for processing, wherein the interest point classification model is obtained according to the training method according to any one of claims 1 to 10; Obtain the point of interest category corresponding to the point of interest description output by the point of interest classification model.

12. A training device for a point of interest classification model, comprising: Memory; And a processor coupled to the memory, characterized in that the processor is configured to execute the training method of the interest point classification model according to any one of claims 1-11 based on instructions stored in the memory.

13. A training device for a point of interest classification model, comprising: An enhancement module is configured to enhance at least one item of interest point training data among the interest point description data and the interest point category label description data by using a self-supervised auxiliary task; The training module is configured to train the classification model based on the enhanced interest point training data using the interest point classification task to obtain the interest point classification model.

14. A computer-readable storage medium having computer instructions stored thereon, characterized in that: When the computer instructions are executed by a processor, the steps of the method for training the interest point classification model described in any one of claims 1-11 are implemented.

15. A computer program product comprising computer instructions, characterized in that When the computer instructions are executed by a processor, the steps of the method for training the interest point classification model described in any one of claims 1-11 are implemented.

Images