Address recommendation method and device, network equipment, storage medium and program product

Abstract

The invention provides an address recommendation method and device, network equipment, a storage medium and a program product, and is applied to the technical field of artificial intelligence. The method comprises the steps of obtaining recommendation demand information of a first user, wherein the recommendation demand information comprises a demand address and / or a demand type; acquiring a historical address sequence corresponding to each second user according to the historical track data of the plurality of second users; the historical address sequence comprises a plurality of addresses; hiding at least part of the first address in the plurality of addresses to obtain a hidden point location; the type of the first address accords with a demand type; and calculating address recommendation loss according to the predicted address corresponding to the hidden point location, the second address and the recommendation demand information, and when the address recommendation loss does not meet a preset condition, adjusting the predicted address until the address recommendation loss meets the preset condition, and obtaining recommendation address information according to the predicted address. By adopting the method, the problem of insufficient address recommendation accuracy caused by incapability of identifying key information in the resource data in the prior art can be solved.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to an address recommendation method, apparatus, network device, storage medium, and program product. Background Technology

[0002] With the continuous development of artificial intelligence and natural language processing technologies, using multi-dimensional data such as user resources, network resources, and channel resources, and conducting comprehensive analysis through deep learning algorithms to perform address prediction and / or recommendation has become an important aspect of current artificial intelligence technology applications. For example, combining artificial intelligence and natural language processing can be used for intelligent site selection of business premises and recommendation of dining addresses.

[0003] Existing technologies for address prediction and / or recommendation typically involve inputting various resource data into a prediction algorithm model for training, obtaining prediction metrics, and then determining recommendation information based on these metrics. However, this approach fails to identify key information within the resource data, leading to insufficient accuracy in address recommendations. Summary of the Invention

[0004] The purpose of this application is to provide an address recommendation method, apparatus, network device, storage medium, and program product to solve the problem that existing artificial intelligence technologies, when performing address prediction and / or recommendation, cannot identify key information in resource data, resulting in insufficient accuracy in address recommendation.

[0005] One embodiment of the present invention provides an address recommendation method, comprising:

[0006] Obtain the recommendation request information of the first user, the recommendation request information including the request address and / or request type;

[0007] Based on the historical trajectory data of multiple second users, obtain the historical address sequence corresponding to each second user; the historical address sequence includes multiple addresses;

[0008] Randomly hide at least a portion of the first addresses from the plurality of addresses to obtain hidden locations; wherein the type of the first addresses conforms to the required type;

[0009] Based on the predicted address corresponding to the hidden location, the second address among the plurality of addresses, and the recommendation requirement information, the address recommendation loss is calculated. If the address recommendation loss does not meet the preset conditions, the predicted address is adjusted until the address recommendation loss meets the preset conditions. Based on the predicted address, the recommended address information is obtained. The second address is an address other than the first address among the plurality of addresses.

[0010] Optionally, the address recommendation method, wherein calculating the address recommendation loss based on the predicted address corresponding to the hidden location, the second address among the plurality of addresses, and the recommendation demand information includes:

[0011] Calculate the type labeling loss based on the second address and the demand type;

[0012] The first address prediction loss is calculated based on the similarity between the predicted address and the actual address corresponding to the predicted address.

[0013] The second address prediction loss is calculated based on the text similarity between the real address corresponding to the predicted address and the required address.

[0014] The address recommendation loss is calculated based on the sum of the type labeling loss, the first address prediction loss, and the second address prediction loss.

[0015] Optionally, the address recommendation method, wherein calculating the type labeling loss based on the second address and the demand type includes:

[0016] Calculate the first embedding vector corresponding to the second address;

[0017] The first embedding vector is transformed into a feature representation to obtain the corresponding first feature sequence;

[0018] Calculate the type labeling loss based on the first feature sequence and the demand type.

[0019] Optionally, in the address recommendation method, calculating the first embedding vector corresponding to the second address includes:

[0020] The first embedding vector is obtained by summing the following vectors:

[0021] The location embedding vector obtained from the second address;

[0022] The time embedding vector obtained based on the time the second user passes through the second address; and

[0023] The embedding vector is obtained based on the relative distance between the second address and the target address.

[0024] Optionally, the address recommendation method, wherein calculating the type labeling loss based on the first feature sequence and the demand type includes:

[0025] Based on the first feature sequence and the demand type, a conditional random field layer is used to calculate the type labeling loss.

[0026] Optionally, the address recommendation method, wherein calculating a first address prediction loss based on the similarity between the predicted address and the corresponding real address includes:

[0027] Calculate the second embedding vector of the predicted address;

[0028] Calculate the third embedding vector based on the actual address corresponding to the predicted address;

[0029] Calculate the similarity between the second embedding vector and the third embedding vector to obtain the first address prediction loss.

[0030] Optionally, in the address recommendation method, calculating the second embedding vector of the predicted address includes:

[0031] The first embedding vector is transformed into a feature representation to obtain the corresponding second feature sequence;

[0032] The second feature sequence is processed by a fully connected layer to obtain the second embedding vector.

[0033] Optionally, in the address recommendation method, calculating the third embedding vector based on the real address corresponding to the predicted address includes:

[0034] The third embedding vector is obtained by summing the following vectors:

[0035] Location embedding vector obtained based on the actual address;

[0036] The time embedding vector is obtained based on the time the second user passes through the real address;

[0037] The embedding vector is obtained based on the relative distance between the real address and the target address.

[0038] Optionally, the address recommendation method further includes:

[0039] If the address recommendation loss is less than or equal to a preset value, the address recommendation loss is determined to satisfy the preset condition.

[0040] One embodiment of this application also provides an address recommendation device, comprising:

[0041] The information acquisition module acquires the recommended demand information of the first user, the recommended demand information including demand address and / or demand type;

[0042] The address acquisition module is used to acquire the historical address sequence corresponding to each second user based on the historical trajectory data of multiple second users; the historical address sequence includes multiple addresses;

[0043] The address processing module is used to randomly hide at least a portion of the first address among the plurality of addresses to obtain the hidden location; wherein the type of the first address conforms to the required type.

[0044] The calculation module is used to calculate the address recommendation loss based on the predicted address corresponding to the hidden point, the second address among the plurality of addresses, and the recommendation requirement information, and adjust the predicted address when the address recommendation loss does not meet the preset conditions until the address recommendation loss meets the preset conditions, and obtain the recommended address information based on the predicted address; wherein, the second address is an address other than the first address among the plurality of addresses.

[0045] One embodiment of this application also provides a network device, comprising: a processor, a memory, and a program stored in the memory and executable on the processor, wherein the program, when executed by the processor, implements the address recommendation method as described in any of the preceding claims.

[0046] One embodiment of this application also provides a readable storage medium, comprising: a program stored on the readable storage medium, wherein the program, when executed by a processor, implements the steps of the address recommendation method as described in any of the preceding claims.

[0047] One embodiment of this application also provides a computer program product, which includes computer instructions that, when executed by a processor, implement the steps of the address recommendation method as described in any of the preceding claims.

[0048] At least one of the above-mentioned technical solutions of this application has the following beneficial effects:

[0049] Using the method described in this application embodiment, based on the first user's required address and required type, the first addresses corresponding to the required type in the historical trajectory data of multiple second users are hidden. Address prediction loss is calculated based on the predicted addresses corresponding to the hidden points, thereby strengthening the model's identification of the user's required address type and recommending address information, ensuring the accuracy and generalization ability of the obtained recommended address information. In addition, this method uses the historical trajectory data of multiple second users to analyze the addresses recommended for the first user, so that the determined recommended addresses are obtained through a deep understanding of the behavioral patterns and preferences of multiple second users, thereby achieving accurate address recommendations. Attached Figure Description

[0050] Figure 1 This is a flowchart illustrating the address recommendation method described in an embodiment of this application;

[0051] Figure 2 This is a flowchart illustrating a specific implementation of the method described in the embodiments of this application;

[0052] Figure 3 This is a schematic diagram of the information recommendation device described in the embodiments of this application. Detailed Implementation

[0053] To make the technical problems, technical solutions and advantages of this application clearer, a detailed description will be provided below in conjunction with the accompanying drawings and specific embodiments.

[0054] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0055] The terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such use of data can be interchanged where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein, and the objects distinguished by "first," "second," etc., are generally of the same class and the number of objects is not limited; for example, a first object can be one or more. Furthermore, in the specification and claims, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0056] It is worth noting that the technologies described in this application are not limited to Long Term Evolution (LTE) / LTE-Advanced (LTE-A) systems, but can also be used in other wireless communication systems, such as Code Division Multiple Access (CDMA), Time Division Multiple Access (TDMA), Frequency Division Multiple Access (FDMA), Orthogonal Frequency Division Multiple Access (OFDMA), Single-carrier Frequency-Division Multiple Access (SC-FDMA), and other systems. The terms "system" and "network" in this application are often used interchangeably, and the described technologies can be used not only in the systems and radio technologies mentioned above, but also in other systems and radio technologies. The following description describes New Radio (NR) systems for illustrative purposes, and NR terminology is used in most of the following description; however, these technologies can also be applied to applications beyond NR systems, such as 6th Generation (6G) communication systems.

[0057] To address the problem in existing artificial intelligence technologies that fail to identify key information in resource data during address prediction and / or recommendation, resulting in insufficient accuracy in address recommendations, this application provides an address recommendation method. Based on the user's desired address and type, a first address matching the desired type from the historical trajectory data of multiple users is hidden. Address prediction loss is calculated based on the predicted address corresponding to the hidden point, thereby enhancing the model's ability to identify and predict key information, thus ensuring the accuracy and generalization ability of the obtained recommended address information. Furthermore, this method utilizes the historical trajectory data of multiple second users to analyze and recommend addresses for the first user, ensuring that the determined recommended address is obtained through a deep understanding of the behavioral patterns and preferences of multiple second users, thus achieving accurate address recommendations.

[0058] like Figure 1 As shown in the embodiments of this application, the address recommendation method includes:

[0059] S110, Obtain the recommendation request information of the first user, the recommendation request information including the request address and / or request type;

[0060] S120, based on the historical trajectory data of multiple second users, obtain the historical address sequence corresponding to each second user; the historical address sequence includes multiple addresses;

[0061] S130, randomly hide at least a portion of the first address among the plurality of addresses to obtain hidden locations; wherein, the type of the first address conforms to the required type;

[0062] S140, calculate the address recommendation loss based on the predicted address corresponding to the hidden point, the second address among the plurality of addresses, and the recommendation requirement information, and adjust the predicted address when the address recommendation loss does not meet the preset conditions until the address recommendation loss meets the preset conditions, and obtain the recommended address information based on the predicted address; wherein, the second address is an address other than the first address among the plurality of addresses.

[0063] Using the method described in the embodiments of this application, based on the required address and required type of the first user, the first address corresponding to the required type in the historical trajectory data of multiple second users is hidden, and the address prediction loss is calculated based on the predicted address corresponding to the hidden point, so as to enhance the pre-model's identification of the user's required address type and make address information recommendations, and ensure the accuracy and generalization ability of the obtained recommended address information.

[0064] In this embodiment of the application, in step S110, the obtained recommended demand information includes demand address and demand type, and optionally, may also include the number of recommended demand addresses.

[0065] In some embodiments, optionally, after obtaining the recommendation request information input by the first user, the request address, request type, and number of recommendations in the recommendation request information can be extracted by a Large Language Model (LLM).

[0066] Optionally, the request type indicates the type of recommended address requested, which can be a Point of Interest (POI) address type, such as supermarket, shopping mall, milk tea shop, fast food restaurant, and convenience store.

[0067] For example, the first user's input recommendation request information could be "I want to open a convenience store in a certain city, please recommend 10 locations". By using a large language model to extract the [required address, request type, and number of recommendations] from the recommendation request information, we can obtain the triple [a certain city, convenience store, 10] in the recommendation request information.

[0068] Optionally, after obtaining the demand address and demand type (optionally, the number of recommendations) from the recommendation demand information, a text embedding vector (emb) of the recommendation demand information is generated. sen .

[0069] The text embedding vector is a numerical vector that can be recognized or processed by a computer program based on the recommendation requirement information. The method of converting the phrase, sentence or document into a text embedding vector is not the focus of this application and will not be described in detail here.

[0070] In step S120, the historical trajectory data of multiple second users is the address data reached by the multiple second users in the target area (such as a city) within a certain period of time. Based on the historical trajectory data, the user identifiers of multiple second users, the locations they reached, and the time of reaching the corresponding locations can be obtained.

[0071] Optionally, in the historical trajectory data, the address data corresponding to each second user can be represented as {user identifier, time, address}, which indicates that the second user corresponding to the user identifier has arrived at the location corresponding to the address at that time.

[0072] Based on the aforementioned historical trajectory data, address data corresponding to multiple second users can be obtained. This address data can then be grouped according to user identifiers to obtain multiple addresses corresponding to the same second user. These addresses can be sorted chronologically to form a historical address sequence, with different first users corresponding to different historical address sequences. For example, the historical address sequence of one of the second users can be represented as {address 1, address 2, ..., address n}.

[0073] In this embodiment, the recommended address for the first user is determined based on the analysis of historical trajectory data of multiple second users. This ensures that the determined recommended address is obtained through a deep understanding of the behavioral patterns and preferences of multiple second users, thereby achieving accurate address recommendation. Based on the obtained historical address sequence, in step S130, at least a portion of the first addresses that match the first user's needs are randomly hidden, resulting in hidden locations.

[0074] Optionally, the first address can be hidden by randomly selecting an address from multiple addresses that matches the first user's needs.

[0075] For example, if the first user's demand type is "convenience store", at least some of the addresses in the historical address sequence obtained in step S120 that are marked as "convenience store" are randomly hidden, resulting in the modified historical address sequence as: {address 1, address 2, hiding point, ..., address n}.

[0076] After obtaining the historical address sequence after the first hidden address is obtained, in step S140, the address recommendation loss is calculated based on the predicted address generated corresponding to the hidden point, the second address among the multiple addresses, and the recommendation requirement information. If the address recommendation loss does not meet the preset conditions, the predicted address is adjusted until the address recommendation loss meets the preset conditions. Based on the predicted address, the recommended address information is obtained.

[0077] In this embodiment of the application, the execution process of the above steps S110 to S140 can be implemented by a large language model (LLM). This execution process can be a pre-training process of obtaining the address recommendation function using an initial large language model, or it can be a process of implementing user address recommendation using the above model obtained through pre-training.

[0078] Using the method described in this embodiment, based on the first user's required address and required type, and collecting historical trajectory data from multiple second users, a large language model is used for analysis. By combining the large language model with the historical trajectory data of multiple second users, the determined recommended address is obtained through a deep understanding of the behavioral patterns and preferences of multiple second users, enabling more accurate recommendations of addresses corresponding to the required type.

[0079] In some embodiments, in step S140, the address recommendation loss is calculated based on the predicted address corresponding to the hidden location, the second address among the plurality of addresses, and the recommendation demand information, including:

[0080] Calculate the type labeling loss based on the second address and the demand type;

[0081] The first address prediction loss is calculated based on the similarity between the predicted address and the actual address corresponding to the predicted address.

[0082] The second address prediction loss is calculated based on the text similarity between the real address corresponding to the predicted address and the required address.

[0083] The address recommendation loss is calculated based on the sum of the type labeling loss, the first address prediction loss, and the second address prediction loss.

[0084] By employing this implementation method, the recommended address information is determined by calculating the sum of the type labeling loss, the first address prediction loss, and the second address prediction loss, and by comprehensively evaluating the multi-dimensional evaluation parameters. This comprehensive processing method achieves more refined address recommendations, thereby improving the accuracy of site selection.

[0085] In this embodiment of the application, optionally, the predicted address corresponding to the hidden point can be a randomly generated address in real time.

[0086] In some embodiments, in step S140, the type labeling loss is calculated based on the second address and the demand type, including:

[0087] Calculate the first embedding vector corresponding to the second address;

[0088] The first embedding vector is transformed into a feature representation to obtain the corresponding first feature sequence;

[0089] Calculate the type labeling loss based on the first feature sequence and the demand type.

[0090] In some embodiments, optionally, calculating the first embedding vector corresponding to the second address includes:

[0091] The first embedding vector is obtained by summing the following vectors:

[0092] The location embedding vector obtained from the second address;

[0093] The time embedding vector obtained based on the time the second user passes through the second address; and

[0094] The embedding vector is obtained based on the relative distance between the second address and the target address.

[0095] Specifically, based on the historical address sequence {address 1, address 2, hidden point, ..., address n}, for each address (second address) g excluding the hidden point... i (i from 1 to n), calculate the sum of the following vectors respectively to obtain the first embedding vector, as shown in the following formula (1):

[0096]

[0097] Among them, h i This is the first embedding vector corresponding to the i-th address; This is the location embedding vector corresponding to the i-th address; This is the time embedding vector corresponding to the i-th address; Let be the relative distance embedding vector corresponding to the i-th address.

[0098] Among them, the location embedding vector For the i-th address g i It is obtained after passing through a fully connected layer, that is, using the following formula (2):

[0099]

[0100] Where FC stands for fully connected layer (also known as dense layer or linear layer), For address g i One-hot encoding.

[0101] In this embodiment, the fully connected layer mentioned is a functional part of the model implementing the method described in the embodiments of the present invention, that is, a basic layer structure in a neural network, used to realize the nonlinear transformation between the features of the input information and the output information.

[0102] Temporal embedding vector For the second user via address g i Time t i The result is obtained after inputting into the fully connected layer, that is, using the following formula (3):

[0103]

[0104] in, For time t i The input feature vector can optionally include time information in seven dimensions: year, month, day, hour, minute, second, and whether it is a holiday.

[0105] Relative distance embedding vector To make address g i The result is obtained by inputting the target address (e.g., a city center) into the fully connected layer, i.e., by using the following formula (4):

[0106]

[0107] in For address g i Distance d relative to the target address i The input representation.

[0108] Using the above method, the first embedding vector h corresponding to each second address can be determined based on each second address in the historical address sequence {address 1, address 2, hidden point, ..., address n}. i According to the first embedding vector h corresponding to each second address i Form an embedding vector set H = {h1, h2, ..., h} n}

[0109] The first embedding vector h corresponding to each second address is calculated using the above method. i Next, the first embedding vector is transformed into a feature representation to obtain the corresponding first feature sequence. Specifically, the first embedding vector h corresponding to each second address is transformed into a feature representation. i The resulting set of embedding vectors H = {h1, h2, ..., h...} n After passing through the encoder, the output vector is obtained, which is the corresponding first feature sequence Out obtained using the following formula (5):

[0110] Out = Encoder(H)(5);

[0111] In this embodiment of the application, the encoder may optionally be a 5-layer Transformer encoder.

[0112] The encoder is a functional part of the model used to implement the method described in the embodiments of this application. It is used to capture the global dependencies between elements in the input sequence using a self-attention mechanism and a feedforward neural network (FFN) to generate high-level semantic codes for downstream tasks (such as classification, translation, and generation).

[0113] After obtaining the first feature sequence Out, the type labeling loss is calculated based on the first feature sequence Out and the user's demand type, denoted as L. CRF .

[0114] Specifically, the first feature sequence Out and the actual point of interest type i of location i are processed through a Conditional Random Field (CRF) layer to calculate the point of interest type labeling loss.

[0115] The Conditional Random Field (CRF) layer is one functional part of the model used to implement the method described in the embodiments of this application. Optionally, the type labeling loss is calculated based on the first feature sequence and the demand type, including:

[0116] Based on the first feature sequence and the demand type, a conditional random field layer is used to calculate the type labeling loss.

[0117] Combination Figure 2 As shown, taking a historical address sequence containing three addresses as an example, for address 1 (i.e., location 1) and address 2 (i.e., location 2) in the historical address sequence {address 1, address 2, hidden location}, the location embedding vectors are calculated respectively based on the fully connected layer. Temporal embedding vector and relative distance embedding vector Embed the address vector Temporal embedding vector and relative distance embedding vector Combined, they form the first embedding vector h corresponding to an address i. i The first embedding vector h corresponding to multiple addresses i After the first embedding vector is processed by the encoder, the corresponding first feature sequence is obtained. Based on the obtained first feature sequence and the above-mentioned address 1 (i.e. location 1), address 2 (i.e. location 2), and user demand type, the type labeling loss can be calculated, that is, the interest point type labeling loss is obtained.

[0118] In this embodiment of the application, optionally, in step S140, a first address prediction loss is calculated based on the similarity between the predicted address and the real address corresponding to the predicted address.

[0119] Using this implementation method, a predicted address is generated corresponding to the hidden location in the historical address sequence, and the similarity between the predicted address and the corresponding real address is calculated to obtain the first address prediction loss.

[0120] Optionally, a first address prediction loss is calculated based on the similarity between the predicted address and the corresponding real address, including:

[0121] Calculate the second embedding vector of the predicted address;

[0122] Calculate the third embedding vector based on the actual address corresponding to the predicted address;

[0123] Calculate the similarity between the second embedding vector and the third embedding vector to obtain the first address prediction loss.

[0124] Optionally, calculating the second embedding vector of the predicted address includes:

[0125] The predicted address is transformed using feature representation to obtain the corresponding second feature sequence;

[0126] The second feature sequence is processed by a fully connected layer to obtain the second embedding vector.

[0127] Optionally, the predicted address can be transformed using an encoder to obtain a corresponding second feature sequence.

[0128] Combination Figure 2 As shown, the hidden location g in the corresponding historical address sequence mask The predicted address corresponding to the hidden location (also known as the hidden location) is directly processed by the encoder to obtain the corresponding second feature sequence; the obtained second feature sequence is then processed by a fully connected layer to obtain the second embedding vector. That is, to obtain the embedding vector of the hidden point.

[0129] When calculating the third embedding vector based on the real address corresponding to the predicted address, the following steps are included:

[0130] The third embedding vector is obtained by summing the following vectors:

[0131] Location embedding vector obtained based on the actual address;

[0132] The time embedding vector is obtained based on the time the second user passes through the real address;

[0133] The embedding vector is obtained based on the relative distance between the real address and the target address.

[0134] That is, the third embedding vector of the real address corresponding to the predicted address is calculated using formulas (1) to (4).

[0135] like Figure 2 As shown in the example, the predicted address corresponds to the actual address g. true For address 3, calculate the actual address g according to formulas (1) to (4). true Location embedding vector Temporal embedding vector and relative distance embedding vector Calculate the sum of these three vectors to obtain the real address g. true The corresponding third embedding vector Optionally, the third embedding vector It is further processed through a fully connected layer.

[0136] After obtaining the above-mentioned hidden point embedding vector With real address g true The corresponding third embedding vector Then, similarity calculation is performed to obtain the first address prediction loss.

[0137] In the embodiments of this application, such as Figure 2 As shown, optionally, the second address prediction loss is calculated based on the text similarity between the real address corresponding to the predicted address and the requested address in the recommended request information, including:

[0138] Calculate the real address g true The corresponding third embedding vector Compared with the text embedding vector emb obtained based on recommendation requirements information sen The similarity between them is used to obtain the second address prediction loss, denoted as L. SEN Optionally, this second address prediction loss can also be called the text and location prediction loss.

[0139] The method described in this application embodiment obtains the aforementioned type labeling loss L. CRF First address prediction loss L MASK Second address L SEN After predicting the loss, the total loss L can be calculated. TOTAL :

[0140] L TOTAL =L CRF +L MASK +L SEN (6);

[0141] Among them, in the above total loss L TOTAL If the value is less than or equal to the preset value, the recommended address information is obtained based on the predicted address.

[0142] The above address recommendation process is a multi-loop execution process until the total loss obtained (i.e., the address recommendation loss) is less than or equal to the predicted value. In this way, recommended address information can be output based on the recommendation request information input by the first user.

[0143] It should be noted that, in the embodiments of this application, the training process of the model for address recommendation can also adopt the above method. By using the above implementation process, the model algorithm is repeatedly trained and continuously optimized until the total loss obtained is less than or equal to the preset value. Then, the training is stopped, the optimal parameters are output, and the trained model is obtained, that is, the pre-trained model that can be used for address recommendation using the above process is obtained.

[0144] The address recommendation method described in the embodiments of this application has the following beneficial effects:

[0145] By combining large language models (LLM) with user trajectory data, more accurate point-of-interest recommendations can be achieved, which not only improves the intelligence level of site selection, but also enhances the model's in-depth understanding of user behavior patterns.

[0146] By randomly hiding specific points of interest (such as convenience stores) and calculating the prediction loss, the model's ability to identify and predict key information can be enhanced, which helps to improve the accuracy and generalization ability of recommended site selection.

[0147] By introducing multidimensional embedding vectors and conditional random field (CRF) layers, the model's ability to comprehensively process location, time, and relative distance can be enhanced, thereby achieving more refined location recommendations and improving the accuracy of site selection.

[0148] One embodiment of this application also provides an address recommendation device, such as... Figure 3 As shown, it includes:

[0149] Information acquisition module 310 acquires the recommended demand information of the first user, the recommended demand information including demand address and / or demand type;

[0150] Address acquisition module 320 is used to acquire the historical address sequence corresponding to each second user based on the historical trajectory data of multiple second users; the historical address sequence includes multiple addresses;

[0151] Address processing module 330 is used to randomly hide at least a portion of the first address among the plurality of addresses to obtain hidden locations; wherein the type of the first address conforms to the required type.

[0152] The calculation module 340 is used to calculate the address recommendation loss based on the predicted address corresponding to the hidden point, the second address among the plurality of addresses, and the recommendation requirement information, and adjust the predicted address when the address recommendation loss does not meet the preset conditions until the address recommendation loss meets the preset conditions, and obtain the recommended address information based on the predicted address; wherein, the second address is an address other than the first address among the plurality of addresses.

[0153] Using the apparatus described in this application embodiment, based on the first user's required address and required type, the first address corresponding to the required type from the historical trajectory data of multiple second users is hidden. Address prediction loss is calculated based on the predicted address corresponding to the hidden point, thereby enhancing the model's ability to identify the user's required address type and recommend address information, ensuring the accuracy and generalization ability of the obtained recommended address information. In addition, using this apparatus, the addresses recommended for the first user are analyzed using the historical trajectory data of multiple second users, so that the determined recommended addresses are obtained through a deep understanding of the behavioral patterns and preferences of multiple second users, thereby achieving accurate address recommendations.

[0154] Optionally, in the address recommendation device, the calculation module 340 calculates the address recommendation loss based on the predicted address corresponding to the hidden location, the second address among the plurality of addresses, and the recommendation demand information, including:

[0155] Calculate the type labeling loss based on the second address and the demand type;

[0156] The first address prediction loss is calculated based on the similarity between the predicted address and the actual address corresponding to the predicted address.

[0157] The second address prediction loss is calculated based on the text similarity between the real address corresponding to the predicted address and the required address.

[0158] The address recommendation loss is calculated based on the sum of the type labeling loss, the first address prediction loss, and the second address prediction loss.

[0159] Optionally, in the address recommendation device, the calculation module 340 calculates the type labeling loss based on the second address and the demand type, including:

[0160] Calculate the first embedding vector corresponding to the second address;

[0161] The first embedding vector is transformed into a feature representation to obtain the corresponding first feature sequence;

[0162] Calculate the type labeling loss based on the first feature sequence and the demand type.

[0163] Optionally, in the address recommendation device, the calculation module 340 calculates the first embedding vector corresponding to the second address, including:

[0164] The first embedding vector is obtained by summing the following vectors:

[0165] The location embedding vector obtained from the second address;

[0166] The time embedding vector obtained based on the time the second user passes through the second address; and

[0167] The embedding vector is obtained based on the relative distance between the second address and the target address.

[0168] Optionally, in the address recommendation device, the calculation module 340 calculates the type labeling loss based on the first feature sequence and the demand type, including:

[0169] Based on the first feature sequence and the demand type, a conditional random field layer is used to calculate the type labeling loss.

[0170] Optionally, in the address recommendation device, the calculation module 340 calculates a first address prediction loss based on the similarity between the predicted address and the actual address corresponding to the predicted address, including:

[0171] Calculate the second embedding vector of the predicted address;

[0172] Calculate the third embedding vector based on the actual address corresponding to the predicted address;

[0173] Calculate the similarity between the second embedding vector and the third embedding vector to obtain the first address prediction loss.

[0174] Optionally, in the address recommendation apparatus, the calculation module 340 calculates the second embedding vector of the predicted address, including:

[0175] The predicted address is transformed using feature representation to obtain the corresponding second feature sequence;

[0176] The second feature sequence is processed by a fully connected layer to obtain the second embedding vector.

[0177] Optionally, in the address recommendation device, the calculation module 340 calculates a third embedding vector based on the real address corresponding to the predicted address, including:

[0178] The third embedding vector is obtained by summing the following vectors:

[0179] Location embedding vector obtained based on the actual address;

[0180] The time embedding vector is obtained based on the time the second user passes through the real address;

[0181] The embedding vector is obtained based on the relative distance between the real address and the target address.

[0182] Optionally, in the address recommendation device, the calculation module 340 is further configured to:

[0183] If the address recommendation loss is less than or equal to a preset value, the address recommendation loss is determined to satisfy the preset condition.

[0184] One embodiment of this application also provides a network device, which includes a processor, a memory, and a program stored in the memory and executable on the processor, wherein the program, when executed by the processor, implements the address recommendation method as described in any of the preceding claims.

[0185] The specific implementation of the address recommendation method by the program running on the processor of the network device can be found in the detailed description of the address recommendation method, and will not be repeated here.

[0186] In addition, specific embodiments of this application also provide a readable storage medium having a computer program stored thereon, wherein the program, when executed by a processor, implements the steps in the address recommendation method as described in any of the above.

[0187] Specifically, the readable storage medium is applied to the aforementioned network device. When applied to the network device, the execution steps in the corresponding address recommendation method are described in detail above and will not be repeated here.

[0188] Another embodiment of the present invention also provides a computer program product, comprising computer instructions that, when executed by a processor, implement the steps in the address recommendation method as described above.

[0189] Optionally, embodiments of this application may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0190] The computer program product described in this embodiment of the invention includes computer instructions that, when executed by a processor, implement the various processes of the method embodiments shown above and achieve the same technical effects. To avoid repetition, these will not be repeated here.

[0191] In the several embodiments provided in this application, it should be understood that the disclosed methods and apparatus can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0192] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can be physically included separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or in the form of hardware plus software functional units.

[0193] The integrated units implemented as software functional units described above can be stored in a computer-readable storage medium. These software functional units, stored in a storage medium, include several instructions that cause a computer device (which may be a personal computer, server, or network device, etc.) to execute some steps of the transmission and reception methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0194] The above describes the preferred embodiments of this application. It should be noted that those skilled in the art can make several improvements and modifications without departing from the principles described in this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. An address recommendation method, characterized in that, include: Obtain the recommendation request information of the first user, the recommendation request information including the request address and / or request type; Based on the historical trajectory data of multiple second users, obtain the historical address sequence corresponding to each second user; the historical address sequence includes multiple addresses; Randomly hide at least a portion of the first addresses from the plurality of addresses to obtain hidden locations; wherein the type of the first addresses conforms to the required type; Based on the predicted address corresponding to the hidden location, the second address among the plurality of addresses, and the recommendation requirement information, the address recommendation loss is calculated. If the address recommendation loss does not meet the preset conditions, the predicted address is adjusted until the address recommendation loss meets the preset conditions. Based on the predicted address, the recommended address information is obtained. The second address is an address other than the first address among the plurality of addresses.

2. The address recommendation method according to claim 1, characterized in that, Based on the predicted address corresponding to the hidden location, the second address among the multiple addresses, and the recommendation demand information, the address recommendation loss is calculated, including: Calculate the type labeling loss based on the second address and the demand type; The first address prediction loss is calculated based on the similarity between the predicted address and the actual address corresponding to the predicted address. The second address prediction loss is calculated based on the text similarity between the real address corresponding to the predicted address and the required address. The address recommendation loss is calculated based on the sum of the type labeling loss, the first address prediction loss, and the second address prediction loss.

3. The address recommendation method according to claim 2, characterized in that, Based on the second address and the demand type, calculate the type labeling loss, including: Calculate the first embedding vector corresponding to the second address; The first embedding vector is transformed into a feature representation to obtain the corresponding first feature sequence; Calculate the type labeling loss based on the first feature sequence and the demand type.

4. The address recommendation method according to claim 3, characterized in that, Calculating the first embedding vector corresponding to the second address includes: The first embedding vector is obtained by summing the following vectors: The location embedding vector obtained from the second address; The time embedding vector obtained based on the time the second user passes through the second address; and The embedding vector is obtained based on the relative distance between the second address and the target address.

5. The address recommendation method according to claim 3, characterized in that, Based on the first feature sequence and the demand type, calculate the type labeling loss, including: Based on the first feature sequence and the demand type, a conditional random field layer is used to calculate the type labeling loss.

6. The address recommendation method according to claim 2, characterized in that, The first address prediction loss is calculated based on the similarity between the predicted address and the corresponding real address, including: Calculate the second embedding vector of the predicted address; Calculate the third embedding vector based on the actual address corresponding to the predicted address; Calculate the similarity between the second embedding vector and the third embedding vector to obtain the first address prediction loss.

7. The address recommendation method according to claim 6, characterized in that, Calculating the second embedding vector of the predicted address includes: The predicted address is transformed using feature representation to obtain the corresponding second feature sequence; The second feature sequence is processed by a fully connected layer to obtain the second embedding vector.

8. The address recommendation method according to claim 6, characterized in that, Based on the actual address corresponding to the predicted address, calculate the third embedding vector, including: The third embedding vector is obtained by summing the following vectors: Location embedding vector obtained based on the actual address; The time embedding vector is obtained based on the time the second user passes through the real address; The embedding vector is obtained based on the relative distance between the real address and the target address.

9. The address recommendation method according to claim 1, characterized in that, The method further includes: If the address recommendation loss is less than or equal to a preset value, the address recommendation loss is determined to satisfy the preset condition.

10. An address recommendation device, characterized in that, include: The information acquisition module acquires the recommended demand information of the first user, the recommended demand information including demand address and / or demand type; The address acquisition module is used to acquire the historical address sequence corresponding to each second user based on the historical trajectory data of multiple second users; the historical address sequence includes multiple addresses; The address processing module is used to randomly hide at least a portion of the first address among the plurality of addresses to obtain the hidden location; wherein the type of the first address conforms to the required type. The calculation module is used to calculate the address recommendation loss based on the predicted address corresponding to the hidden point, the second address among the plurality of addresses, and the recommendation requirement information, and adjust the predicted address when the address recommendation loss does not meet the preset conditions until the address recommendation loss meets the preset conditions, and obtain the recommended address information based on the predicted address; wherein, the second address is an address other than the first address among the plurality of addresses.

11. A network device, characterized in that, include: A processor, a memory, and a program stored in the memory and executable on the processor, the program implementing the address recommendation method as described in any one of claims 1 to 9 when executed by the processor.

12. A readable storage medium, characterized in that, include: The readable storage medium stores a program that, when executed by a processor, implements the steps of the address recommendation method as described in any one of claims 1 to 9.

13. A computer program product, characterized in that, It includes computer instructions that, when executed by a processor, implement the steps of the address recommendation method as described in any one of claims 1 to 9.

Images