Interest point sorting method, device and equipment based on deep learning

Abstract

The invention discloses an interest point sorting method, device and equipment based on deep learning, and relates to the technical field of data analysis, the method comprises the steps that a multi-label sample set is constructed, the multi-label sample set comprises samples corresponding to a plurality of interest points POI, the samples carry user behavior labels, the user behavior labels comprise click labels and order issuing labels, and the click labels and the order issuing labels are selected from the click labels and the order issuing labels; the click tag indicates whether the user clicks the POI, and the order issuing tag indicates whether the user issues an order; positive samples and negative samples are determined in the multi-label sample set; inputting the multi-label sample set into a deep learning model, and training the deep learning model based on a multi-label weighted loss function to obtain an interest point sorting model, the deep learning model being used for scoring each POI and sorting according to the score of each POI to obtain an interest point sorting result; the multi-label loss function is used for increasing the score difference value of the POI corresponding to the positive sample and the POI corresponding to the negative sample. Therefore, the accuracy of model POI sorting is improved, and the user experience is improved.

Description

Technical Field

[0001] The present application relates to the field of data analysis technology, and in particular to a method, device and equipment for sorting points of interest based on deep learning. Background Art

[0002] In the freight search scenario, the deep ranking model of POIs aims to optimize the user's search experience by accurately evaluating the attractiveness of each POI to a specific user. The current mainstream technical solution is to construct a large number of POIs and user features including statistical features, category features, ID features, and text relevance features, and then encode these features into embedding vectors, which are then integrated into the model input through splicing operations to capture complex user-POI relationships.

[0003] Since the sorting model needs to score a large number of POIs in an online environment and quickly filter out a small number of results that best meet user needs, it is usually necessary to select a model with high computational efficiency and relatively simple structure, such as a deep neural network DNN model. The DNN model can efficiently score all candidate POIs and select the top N POIs to recommend to users based on the scores, where N is a set value. For example, in CN112364238A, the user's historical check-in data is obtained, and the deep learning model is trained based on the historical check-in data; the user's latest check-in data is input into the trained deep learning model, and the predicted user points of interest are output; then, the two features are calculated with the candidate set for Euclidean distance, sorted by score, and the top N POIs are output.

[0004] However, although this model can capture the general preferences of users, it does not further consider the impact of user behavior, resulting in recommendations based on ranking results that do not bring a good experience to users. Summary of the invention

[0005] In view of this, the main purpose of the present application is to provide a method, device and equipment for sorting points of interest based on deep learning, with the aim of improving the user experience by improving the accuracy of sorting points of interest.

[0006] The first aspect of the present application provides a method for sorting points of interest based on deep learning, characterized in that the method comprises:

[0007] Construct a multi-label sample set, which includes samples corresponding to multiple points of interest (POIs). The samples carry user behavior labels, which include click labels and order labels. The click label indicates whether the user clicks on the POI, and the order label indicates whether the user places an order.

[0008] In the multi-label sample set, positive samples and negative samples are determined based on the click labels;

[0009] The multi-label sample set is input into the deep learning model, and the deep learning model is trained based on the multi-label weighted loss function to obtain a POI ranking model, wherein the deep learning model is used to score each POI and sort them according to the score of each POI, and the multi-label loss function is used to increase the score difference between the POI corresponding to the positive sample and the POI corresponding to the negative sample.

[0010] In some implementations of the first aspect of the present application, determining positive samples and negative samples according to the click labels includes:

[0011] The sample whose click label indicates that the user clicks the POI is determined as a positive sample, and the sample whose click label indicates that the user does not click the POI is determined as a negative sample.

[0012] In some implementations of the first aspect of the present application, the user behavior tag also includes: an unfixed point tag, which is used to indicate whether the distance between the POI and the boarding point exceeds a threshold.

[0013] In some implementations of the first aspect of the present application, in a multi-label sample set, positive samples whose click labels indicate that a user clicks a POI, whose order labels indicate that a user places an order, and whose unfixed-point labels indicate that the distance between a POI and a boarding point exceeds a threshold are eliminated.

[0014] In some implementations of the first aspect of the present application, the sample includes: a user query field, and the method further includes:

[0015] During the training process of the deep learning model, samples are classified into the following three types of user intents based on the user query field: general intent, portal intent, and other intents, and the weights corresponding to each user intent are obtained based on the three types of user intent training.

[0016] In some implementations of the first aspect of the present application, samples are classified into the following three user intents based on the user query field: general intent, portal intent, and other intents, including:

[0017] Classify samples whose user query field is less than or equal to a preset length and / or whose user query field ends with a road name as general intents;

[0018] Classify the samples whose user query fields are longer than the preset length and contain numbers as portal intent;

[0019] The samples in which the user query field does not end with a road name or does not contain a number when the user query field is greater than a preset length are classified as other intents.

[0020] In some implementations of the first aspect of the present application, training a deep learning model based on a multi-label weighted loss function includes:

[0021] Determine the negative sample position penalty coefficient according to the position of the POI corresponding to the negative sample, wherein the closer the position of the POI corresponding to the negative sample is to the front, the greater the negative sample position penalty coefficient;

[0022] The multi-label weighted loss function is obtained by multiplying the BPR loss function by the negative sample position penalty coefficient. The BPR loss function is used to calculate the loss based on the score of the POI corresponding to the positive sample and the score of the POI corresponding to the negative sample;

[0023] The deep learning model is trained based on a multi-label weighted loss function.

[0024] In some implementations of the first aspect of the present application, when the order label of the positive sample indicates that the user has issued an order and the unfixed point label is used to indicate that the distance between the POI and the boarding point does not exceed a threshold, the BPR loss function is multiplied by the negative sample position penalty coefficient to obtain a multi-label weighted loss function, including:

[0025] The BPR loss function is multiplied by the negative sample position penalty coefficient, the order weight and the indefinite point weight to obtain the multi-label weighted loss function. The order weight and the indefinite point weight are the learnable parameters obtained by the deep learning model during the model training process based on the order label and the indefinite item label training.

[0026] A second aspect of the present application provides a device for sorting points of interest based on deep learning, the device comprising:

[0027] A sample construction module is used to construct a multi-label sample set. The multi-label sample set includes samples corresponding to multiple points of interest (POIs). The samples carry user behavior labels. The user behavior labels include: a click label and an order label. The click label is used to indicate whether the user clicks on the POI, and the order label is used to indicate whether the user places an order.

[0028] The sample classification module is used to determine positive samples and negative samples based on click labels in a multi-label sample set;

[0029] The model training module is used to input the multi-label sample set into the deep learning model, and train the deep learning model based on the multi-label weighted loss function to obtain a point of interest ranking model, wherein the deep learning model is used to score each POI and sort the POI according to the score of each POI, and the multi-label loss function is used to increase the score difference between the POI corresponding to the positive sample and the POI corresponding to the negative sample.

[0030] The third aspect of the present application provides a deep learning-based point of interest sorting device, characterized in that the device includes a memory and a processor, and the processor is used to execute the program stored in the memory and run any deep learning-based point of interest sorting method provided in the first aspect of the present application.

[0031] The technical solution provided by this application has the following beneficial effects:

[0032] In the technical solution provided in the present application, first, a multi-label sample set is constructed, each sample in the multi-label sample set corresponds to a point of interest POI, and each sample carries a user behavior label, which includes a click label and an order label, wherein the click label is used to indicate whether the user clicks on the POI, and the order label is used to indicate whether the user has issued an order after clicking on the POI. Then, each sample in the multi-label sample set is divided into positive samples and negative samples, so that the model can better predict the user's interest in different POIs. Finally, the multi-label samples are input into the deep learning model for training, and the multi-label weighted loss function is used as the model loss function. The design of the multi-label weighted loss function increases the score difference between the POIs corresponding to the positive samples and the negative samples, so that the model is more inclined to push the POI corresponding to the positive sample, that is, push the POI that the user is more likely to be interested in and more likely to issue an order. Under the guidance of the above loss function, the trained deep learning model is used to score the POI, and the scoring results are used to sort and display to the user. Due to the introduction of multi-label modeling of user behavior, the model can better predict which POIs are more likely to be clicked and ordered. Therefore, the final sorted POI list will be more in line with the user's real interests and needs, improving the user's overall usage experience and increasing the order rate while ensuring the click-through rate. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 A flowchart of a method for sorting points of interest based on deep learning provided in an embodiment of the present application;

[0034] Figure 2 A flowchart of another method for sorting points of interest based on deep learning provided in an embodiment of the present application;

[0035] Figure 3 A flowchart of another method for sorting points of interest based on deep learning provided in an embodiment of the present application;

[0036] Figure 4 A schematic diagram of the structure of a device for sorting points of interest based on deep learning provided in an embodiment of the present application;

[0037] Figure 5A schematic diagram of the structure of a point of interest sorting device based on deep learning provided in an embodiment of the present application. DETAILED DESCRIPTION

[0038] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0039] The embodiments of the present application provide a method, device and apparatus for sorting points of interest based on deep learning.

[0040] See also Figure 1 As shown, an embodiment of the present application provides a method for sorting points of interest based on deep learning, the method comprising:

[0041] S101: Construct a multi-label sample set, where the multi-label sample set includes samples corresponding to multiple points of interest (POIs), and the samples carry user behavior labels. The user behavior labels include: a click label and an order label. The click label is used to indicate whether the user clicks on the POI, and the order label is used to indicate whether the user places an order.

[0042] In the embodiment of the present application, a training set for model training is first constructed, that is, a multi-label sample set; each sample in the multi-label sample set corresponds to a point of interest POI, and each sample contains the interaction data between the user and the POI. The interaction data is a feature converted from various interactive behavior data between the user and the point of interest when using the application, and also contains user data and POI-related data. For example, when the user enters a certain place in the search box and a location recommendation list is displayed below the search box, the user data can include the information entered by the user in the search box and the user's own information, and the POI-related data includes the point of interest information of a POI in the recommendation list, such as the name, the corresponding longitude and latitude coordinates, etc.; and the interactive behavior data is whether the user clicks on the POI in the recommendation list and whether to place an order after clicking on the POI. It should be noted that the interactive data can be samples collected from POIs or constructed by technicians themselves, which does not affect the implementation of the embodiments of the present application.

[0043] In addition, the above interaction data will be converted into feature vectors and further encoded into embedding vectors, and then spliced ​​together as input for the deep learning model.

[0044] In addition, each sample also carries a user behavior tag for describing user behavior. Specifically, the user behavior tag includes a click tag and an order tag. The click tag indicates whether the user clicks on a POI. For example, when a user clicks to view a POI, the value of the click tag is set to 1; if the user does not click, the value of the click tag is 0. The order tag indicates whether the user places an order after clicking on a POI. For example, if the user places an order, the value of the order tag is 1; if the user does not place an order, the value of the order tag is 0.

[0045] It is understandable that when the program is running normally, since the user needs to click on the POI before further placing an order, there is no situation where the user does not click on the POI but places an order based on the POI. In other words, in the above example, there are three types of labels: 1. The click label is 1 and the order label is 0; 2. The click label is 1 and the order label is 1; 3. The click label is 1 and the order label is 0. However, the situation where the click label is 0 and the order label is 1 does not exist.

[0046] In some implementations of the embodiments of the present application, in order to reduce user complaints caused by inaccurate recommended locations and to further improve user experience, an unfixed-point tag is further introduced into the user behavior tag. That is, the user behavior tag includes, in addition to the click tag and the order placement tag, an unfixed-point tag indicating whether the distance between the POI and the boarding point exceeds a threshold.

[0047] The unfixed point tag indicates whether the distance between the boarding point specified by the user and the location corresponding to the POI exceeds the threshold. For example, if the user clicks on a POI and the boarding point determined by the user is no more than 1 km away from the POI, the value of the unfixed point tag is 1; if it is more than 1 km away, the value of the unfixed point tag is 0.

[0048] S102: In the multi-label sample set, positive samples and negative samples are determined according to the clicked labels.

[0049] In an embodiment of the present application, in order to enable the model to better learn the user's interest preferences in subsequent model training and improve the model prediction accuracy, each sample in the multi-label sample set is divided into two categories, one is the positive samples in which the user shows interest, and the other is the negative samples in which the user does not show interest.

[0050] In some implementations of the embodiments of the present application, whether the user is interested in a certain POI is measured by whether the user clicks on the POI. Therefore, positive samples and negative samples are determined according to the click label, specifically: samples whose click labels indicate that the user clicks on the POI are determined as positive samples, and samples whose click labels indicate that the user does not click on the POI are determined as negative samples. For example, when the value of the click label is 1, indicating that the user clicks on the POI, and 0, indicating that the user does click on the POI; when the value of the click label of a certain sample is 1, it will be regarded as a positive sample; when the value of the click label of a certain sample is 0, it will be regarded as a negative sample.

[0051] In some implementations of the present application, positive samples can be further divided. For example, samples that meet both "clicked" and "ordered" are identified as strong positive samples, while samples that only meet "clicked" are identified as weak positive samples. In addition, other methods can be used to divide positive and negative samples. For example, only samples that meet both "clicked" and "ordered" are identified as positive samples, while other samples are identified as negative samples.

[0052] In some implementations of the embodiments of the present application, in order to optimize the training effect of the deep learning model, in the multi-label sample set, positive samples whose click labels indicate that the user clicks on the POI, whose order labels indicate that the user places an order, and whose indeterminate point labels indicate that the distance between the POI and the boarding point exceeds a threshold are eliminated.

[0053] In this implementation, the embodiment of the present application removes positive samples in the following situations: the click label indicates that the user clicked on the POI, and the order label indicates that the user clicked on the POI and then issued the order, and the unfixed point label indicates that the distance between the POI and the actual boarding point exceeds a preset threshold. That is, the positive samples with a click label of 1, an order label of 1, and an unfixed point label of 0 are removed. Although such positive samples meet the conditions of "clicking on the POI" and "successfully issuing the order", the "unfixed point label" shows that the distance exceeds the set threshold. In this case, although the user clicks and issues the order, the location of the boarding point is not ideal, which reflects that there is a gap between the recommended POI and the actual needs of the user, which brings a bad experience to the user in actual use, and therefore is not used as a positive sample.

[0054] S103: Input the multi-label sample set into the deep learning model, and train the deep learning model based on the multi-label weighted loss function to obtain a point of interest ranking model, wherein the deep learning model is used to score each POI and sort the POI according to the score of each POI, and the multi-label loss function is used to increase the score difference between the POI corresponding to the positive sample and the POI corresponding to the negative sample.

[0055] In an embodiment of the present application, the features of each positive sample and its corresponding negative sample are spliced ​​together in the sample construction. For example, if there is 1 positive sample and 9 negative samples, then 9 groups of samples will be generated, which are then input into the deep learning model for model training. It should be noted that the deep learning model used in this application is a model specially designed for scoring and sorting POIs.

[0056] After the multi-label sample set is input into the deep learning model, the deep learning model will learn the user's preferences for different POIs based on the multi-label sample set, thereby achieving the effect of predicting user preferences. During the model training process, the multi-label weighted loss function is used to guide the model training. The multi-label weighted loss function is used to amplify the score difference between positive and negative samples, so that the model can give better scores to POIs that users are more likely to click and place orders. The multi-label weighted loss function will continuously adjust the parameters of the model during the training process to maximize the score difference between positive and negative samples.

[0057] In some implementations of the embodiments of the present application, the sample includes: a user query field, which is the field entered by the user in the search box. The embodiments of the present application further classify the user intent of the sample based on the user query field, and perform targeted optimization according to different intent classifications during the training process of the model.

[0058] at this time, Figure 1 The process shown also includes the following steps:

[0059] During the training process of the deep learning model, samples are classified into the following three types of user intents based on the user query field: general intent, portal intent, and other intents, and the weights corresponding to each user intent are obtained based on the three types of user intent training.

[0060] Among them, general intent refers to the content of the user's query is vague or broad, and cannot clearly specify a specific location; door-to-door intent refers to the content of the user's query is very specific, often including house numbers or detailed address information; and other intents are other intents except general intent and door-to-door intent.

[0061] During the training process, the deep learning model learns the corresponding weights for each type according to different user intent types. These weights are the parameters used by the model to learn user behaviors under different intents. By continuously optimizing these weights, the model can better distinguish samples with different intents and improve the accuracy of sorting. It is understandable that during the model training process, the weights are adjusted dynamically. The model will continuously update the weights through multiple iterative calculations so that samples of each intent type can be correctly classified and sorted.

[0062] After the model training is completed, the model will use corresponding weights to score and sort POIs for query scenarios with different user intentions. By classifying user queries by intent and learning and adjusting the weights corresponding to each intent during the training process, each intent corresponds to a scenario, so that the model can understand and predict user needs more accurately, thereby more effectively sorting POIs, improving the adaptability and accuracy of the model in different scenarios, and improving the accuracy and top click-through rate of the model in various scenarios, which helps to improve the success rate of site selection in business.

[0063] In some implementations of the embodiments of the present application, the deep learning model adopts an architecture of 4 fully connected dense layers, the bottom three dense layers share weights, and the last layer trains weights separately according to different scenarios, so as to further learn and optimize for different scenarios while extracting common features, thereby improving the prediction accuracy of the model in different scenarios.

[0064] Specifically, the first dense layer is used to extract primary features from the model input and compress information; the second layer is used to further refine and abstract features, and then process the primary features to obtain more expressive intermediate features; the third layer is used to perform high-level abstraction of intermediate features to improve feature expression capabilities; based on the shared weights of the first three layers to extract common features, the fourth layer trains weights separately for three usage scenarios: general intent, portal address intent, and other intent. The data characteristics and user needs in each scenario may be different, so independent weight learning is needed to adapt to these differences. The fourth layer trains weights separately according to different scenarios, allowing the model to make fine-grained adjustments and optimizations in different scenarios, thereby making more accurate predictions.

[0065] In some implementations of the embodiments of the present application, samples are classified into the following three user intents based on the user query field: general intent, portal intent, and other intents, specifically including the following steps:

[0066] S201: Classify samples whose user query field is less than or equal to a preset length and / or whose user query field ends with a road name as general intents.

[0067] In an embodiment of the present application, if the length of the query field entered by the user is less than or equal to a preset length, the query is considered to be a general intent. Because in this case, the user may enter a very short and broad keyword instead of the name of a specific place. Preferably, the preset length is 4 characters. In addition, if the user query field ends with a road name such as "road", "street", "avenue", etc., it means that the user is only looking for a POI near a certain road, and does not have a clear and specific destination, so it will also be classified as a general intent.

[0068] S202: Classify the samples whose user query fields are longer than a preset length and contain numbers as portal intent.

[0069] In the embodiment of the present application, if the length of the query field entered by the user is greater than the preset length, it means that the user's query is more specific. At the same time, if the query field also contains numbers, it means that the user may have entered information such as house numbers and building numbers, which means that the user has a clear destination requirement, so the user intent of this type of sample can be classified as door address intent.

[0070] S203: Classify samples in which the user query field does not end with a road name, or the user query field does not contain numbers when it is greater than a preset length, as other intents.

[0071] In an embodiment of the present application, if the query field does not meet the conditions of the first two intents, it is classified as other intents.

[0072] pass Figure 2 Through the steps shown, the system can effectively classify the user's query fields, thereby providing more targeted training data for the model, helping the deep learning model to more accurately predict the user's needs in different scenarios and improve the overall performance of the model.

[0073] In some implementations of the embodiments of the present application, the multi-label weighted loss function can give different degrees of penalties according to the location of negative samples, thereby optimizing the performance of the deep learning model in the task of sorting points of interest, and further improving the prediction accuracy of the model and user experience. The deep learning model is trained based on the multi-label weighted loss function, including:

[0074] S301: Determine a negative sample position penalty coefficient according to the position of the POI corresponding to the negative sample, wherein the closer the position of the POI corresponding to the negative sample is to the front, the greater the negative sample position penalty coefficient is.

[0075] In an embodiment of the present application, if a negative sample ranks high in the sorting, it means that the model has problems in distinguishing positive samples from negative samples. Therefore, the penalty coefficient corresponding to the negative sample position will be determined according to the sorting position of the point of interest POI corresponding to the negative sample. Among them, the higher the negative sample is, the greater the penalty coefficient is. For example, if the negative sample appears in the first place, its penalty coefficient will be greater than that of the negative sample appearing in the fifth place, so that the model can try to put the positive samples in front and the negative samples in the back, thereby improving the accuracy of the model sorting.

[0076] In some implementations of the present application, the penalty coefficient may be calculated as: 1+1 / (1+log(position of negative sample)). At this time, if the negative sample is at the first position, the penalty coefficient is 2; if the negative sample is at the second position, the penalty coefficient is 1.59.

[0077] S302: Multiplying the BPR loss function by the negative sample position penalty coefficient to obtain a multi-label weighted loss function, where the BPR loss function is used to calculate the loss based on the score of the POI corresponding to the positive sample and the score of the POI corresponding to the negative sample.

[0078] In an embodiment of the present application, in order to strengthen the penalty for the position of negative samples, the model is made to pay more attention to placing positive samples before negative samples, and at the same time, a greater penalty is given to negative samples ranked higher. The present application combines the Bayesian Personalized Ranking (BPR) loss function with the negative sample position penalty coefficient to obtain a multi-label weighted loss function. The BPR loss function is a commonly used loss function for ranking models, and its goal is to maximize the score difference between positive samples and negative samples.

[0079] For each set of sample data, the embodiment of the present application uses the same corresponding tensorflow code, and the code of the basic BPR loss function is:

[0080] ctr_loss=-tf.math.log(tf.nn.sigmoid(pos_ctr_prob-neg_ctr_prob))

[0081] Among them, ctr_loss represents the multi-label weighted loss function, tf.math.log represents the mathematical function used to calculate the logarithm, tf.nn.sigmoid represents the activation function, pos_ctr_prob represents the score of the POI corresponding to the positive sample, and neg_ctr_prob represents the score of the POI corresponding to the negative sample.

[0082] When "1+1 / (1+log(negative sample position))" is used to determine the negative sample position penalty coefficient, the multi-label weighted loss function is:

[0083] ctr_loss=-tf.math.log(tf.nn.sigmoid(pos_ctr_prob-neg_ctr_prob))*(1+1 / (1+log(x))

[0084] Among them, x represents the position of the POI corresponding to the negative sample.

[0085] In some implementations of the embodiments of the present application, when the order label of the positive sample indicates that the user has issued an order and the unfixed point label is used to indicate that the distance between the POI and the boarding point does not exceed the threshold, the BPR loss function is multiplied by the negative sample position penalty coefficient to obtain a multi-label weighted loss function, including:

[0086] The BPR loss function is multiplied by the negative sample position penalty coefficient, the order weight and the indefinite point weight to obtain the multi-label weighted loss function. The order weight and the indefinite point weight are the learnable parameters obtained by the deep learning model during the model training process based on the order label and the indefinite item label training.

[0087] In this implementation, in order to optimize the deep learning model's sorting effect on POIs and improve user experience, the multi-label weighted loss function is further weighted. Specifically, if the positive sample's order label indicates that the user has issued an order and the unfixed point label is used to indicate that the distance between the POI and the boarding point does not exceed the threshold, then the BRP loss function is multiplied by the negative sample position penalty coefficient and then multiplied by the product of the order weight and the unfixed point weight.

[0088] Among them, the order weight and the unfixed point weight are both learnable parameters. The order weight is used to amplify the influence of samples related to order issuance, and the unfixed point weight is used to amplify the influence of samples whose distance between the boarding point and the clicked POI does not exceed the threshold. The model will automatically adjust the parameters during the training process to make the samples corresponding to the order label and the unfixed point label more influential.

[0089] For example, when the values ​​of the order label and the unfixed point label are both 1, the multi-label weighted loss function can be expressed as:

[0090] ctr_loss = -tf.math.log(tf.nn.sigmoid(pos_ctr_prob-neg_ctr_prob))*(1+1 / 1+log(x))*order weight*indeterminate point weight.

[0091] Among them, the order weight and the indeterminate point weight are learnable parameters, and their initial values ​​are 2.

[0092] S303: Training the deep learning model based on the multi-label weighted loss function.

[0093] exist Figure 1In the process shown, first, a multi-label sample set is constructed, each sample in the multi-label sample set corresponds to a point of interest POI, and each sample carries a user behavior label, which includes a click label and an order label, wherein the click label is used to indicate whether the user clicks on the POI, and the order label is used to indicate whether the user has placed an order after clicking on the POI. Then, each sample in the multi-label sample set is divided into positive samples and negative samples, so that the model can better predict the user's interest in different POIs. Finally, the multi-label samples are input into the deep learning model for training, and the multi-label weighted loss function is used as the model loss function. The design of the multi-label weighted loss function increases the score difference between the POIs corresponding to the positive samples and the negative samples, so that the model is more inclined to push the POI corresponding to the positive sample, that is, push the POI that the user is more likely to be interested in and more likely to place an order. Under the guidance of the above loss function, the trained deep learning model is used to score the POI, and the scoring results are used to sort and display to the user. Due to the introduction of multi-label modeling of user behavior, the model can better predict which POIs are more likely to be clicked and ordered. Therefore, the final sorted POI list will be more in line with the user's real interests and needs, improving the user's overall usage experience and increasing the order rate while ensuring the click-through rate.

[0094] See also Figure 4 As shown, an embodiment of the present application provides a device for sorting points of interest based on deep learning, the device comprising:

[0095] The sample construction module 401 is used to construct a multi-label sample set, which includes samples corresponding to multiple points of interest (POIs). The samples carry user behavior tags, which include: a click tag and an order tag. The click tag is used to indicate whether the user clicks on the POI, and the order tag is used to indicate whether the user places an order.

[0096] A sample classification module 402 is used to determine positive samples and negative samples in a multi-label sample set according to click labels;

[0097] The model training module 403 is used to input the multi-label sample set into the deep learning model, and train the deep learning model based on the multi-label weighted loss function to obtain a point of interest ranking model, wherein the deep learning model is used to score each POI and sort the POI according to the score of each POI, and the multi-label loss function is used to increase the score difference between the POI corresponding to the positive sample and the POI corresponding to the negative sample.

[0098] In some implementations of the embodiments of the present application, determining positive samples and negative samples according to click labels includes:

[0099] The sample whose click label indicates that the user clicks the POI is determined as a positive sample, and the sample whose click label indicates that the user does not click the POI is determined as a negative sample.

[0100] In some implementations of the embodiments of the present application, the user behavior tag also includes: an unfixed point tag, which is used to indicate whether the distance between the POI and the boarding point exceeds a threshold.

[0101] In some implementations of the embodiments of the present application, in the multi-label sample set, positive samples whose click labels indicate that the user clicks on the POI, whose order labels indicate that the user places an order, and whose unfixed point labels indicate that the distance between the POI and the boarding point exceeds a threshold are eliminated.

[0102] In some implementations of the embodiments of the present application, the sample includes: a user query field, and the method further includes:

[0103] During the training process of the deep learning model, samples are classified into the following three types of user intents based on the user query field: general intent, portal intent, and other intents, and the weights corresponding to each user intent are obtained based on the three types of user intent training.

[0104] In some implementations of the embodiments of the present application, samples are classified into the following three user intents based on the user query field: general intent, portal intent, and other intents, including:

[0105] Classify samples whose user query field is less than or equal to a preset length and / or whose user query field ends with a road name as general intents;

[0106] Classify the samples whose user query fields are longer than the preset length and contain numbers as portal intent;

[0107] The samples in which the user query field does not end with a road name or does not contain a number when the user query field is greater than a preset length are classified as other intents.

[0108] In some implementations of the embodiments of the present application, training a deep learning model based on a multi-label weighted loss function includes:

[0109] Determine the negative sample position penalty coefficient according to the position of the POI corresponding to the negative sample, wherein the closer the position of the POI corresponding to the negative sample is to the front, the greater the negative sample position penalty coefficient;

[0110] The multi-label weighted loss function is obtained by multiplying the BPR loss function by the negative sample position penalty coefficient. The BPR loss function is used to calculate the loss based on the score of the POI corresponding to the positive sample and the score of the POI corresponding to the negative sample;

[0111] The deep learning model is trained based on a multi-label weighted loss function.

[0112] In some implementations of the embodiments of the present application, when the order label of the positive sample indicates that the user has issued an order and the unfixed point label is used to indicate that the distance between the POI and the boarding point does not exceed the threshold, the BPR loss function is multiplied by the negative sample position penalty coefficient to obtain a multi-label weighted loss function, including:

[0113] The BPR loss function is multiplied by the negative sample position penalty coefficient, the order weight and the indefinite point weight to obtain the multi-label weighted loss function. The order weight and the indefinite point weight are the learnable parameters obtained by the deep learning model during the model training process based on the order label and the indefinite item label training.

[0114] like Figure 5 As shown, the embodiment of the present application also provides a device for sorting points of interest based on deep learning, including: a memory 501, a processor 502;

[0115] Wherein, the memory 501 is used to store programs;

[0116] The processor 502 is used to execute the program in the memory to implement each step of the interest point sorting method based on deep learning provided in the embodiment of the present application.

[0117] Finally, it should be noted that, in the embodiments of the present application, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply that there is any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that the process, method, article or equipment including a series of elements includes not only those elements, but also includes other elements that are not clearly listed, or also includes elements inherent to such process, method, article or equipment. In the absence of more restrictions, the elements limited by the statement "comprise one..." do not exclude the presence of other identical elements in the process, method, article or equipment including the elements.

[0118] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for sorting points of interest based on deep learning, characterized in that: The method comprises: Construct a multi-label sample set, wherein the multi-label sample set includes samples corresponding to multiple points of interest (POIs), and the samples carry user behavior labels, wherein the user behavior labels include: a click label and an order label, wherein the click label is used to indicate whether the user clicks the POI, and the order label is used to indicate whether the user places an order; In the multi-label sample set, determining positive samples and negative samples according to the click labels; The multi-label sample set is input into a deep learning model, and the deep learning model is trained based on a multi-label weighted loss function to obtain a point of interest ranking model, wherein the deep learning model is used to score each of the POIs and sort them according to the score of each POI, and the multi-label loss function is used to increase the score difference between the POI corresponding to the positive sample and the POI corresponding to the negative sample.

2. The method according to claim 1, characterized in that The determining of positive samples and negative samples according to the click labels includes: The sample in which the click tag indicates that the user clicks the POI is determined as the positive sample, and the sample in which the click tag indicates that the user does not click the POI is determined as the negative sample.

3. The method according to claim 1, characterized in that The user behavior tag also includes: an unfixed point tag, and the unfixed point tag is used to indicate whether the distance between the POI and the boarding point exceeds a threshold.

4. The method according to claim 3, characterized in that In the multi-label sample set, the positive samples in which the click label indicates that the user clicks the POI, the order placement label indicates that the user places an order, and the indeterminate point label indicates that the distance between the POI and the boarding point exceeds a threshold are eliminated.

5. The method according to claim 1, characterized in that The sample includes: a user query field, and the method further includes: During the training process of the deep learning model, the samples are classified into the following three user intents based on the user query field: general intent, portal intent, and other intents, and the weights corresponding to each user intent are obtained based on the training of the three user intents.

6. The method according to claim 5, characterized in that The sample is classified into the following three user intents based on the user query field: general intent, portal intent, and other intents, including: Classify the samples in which the user query field is less than or equal to a preset length and / or the user query field ends with a road name as the general intent; Classify the sample whose user query field is greater than the preset length and contains numbers as the portal address intention; The samples in which the user query field does not end with a road name, or in which the user query field does not contain a number when it is greater than the preset length, are classified as the other intentions.

7. The method according to claim 1, characterized in that The deep learning model is trained based on a multi-label weighted loss function, including: Determine a negative sample position penalty coefficient according to the position of the POI corresponding to the negative sample, wherein the closer the position of the POI corresponding to the negative sample is to the front, the greater the negative sample position penalty coefficient is; The multi-label weighted loss function is obtained by multiplying the Bayesian personalized ranking BPR loss function by the negative sample position penalty coefficient, wherein the BPR loss function is used to calculate the loss based on the score of the POI corresponding to the positive sample and the score of the POI corresponding to the negative sample; The deep learning model is trained based on the multi-label weighted loss function.

8. The method according to claims 3 and 7, characterized in that When the order label of the positive sample indicates that the user has issued an order and the distance between the unfixed point label and the boarding point does not exceed a threshold, the multi-label weighted loss function is obtained by multiplying the BPR loss function by the negative sample position penalty coefficient, including: The BPR loss function is multiplied by the negative sample position penalty coefficient, the order weight and the indefinite point weight to obtain the multi-label weighted loss function. The order weight and the indefinite point weight are learnable parameters obtained by the deep learning model during the model training process based on the order label and the indefinite item label.

9. A device for sorting points of interest based on deep learning, characterized in that: The device comprises: A sample construction module is used to construct a multi-label sample set, wherein the multi-label sample set includes samples corresponding to multiple points of interest (POIs), and the samples carry user behavior tags, wherein the user behavior tags include: a click tag and an order tag, wherein the click tag is used to indicate whether the user clicks on the POI, and the order tag is used to indicate whether the user places an order; A sample classification module, used to determine positive samples and negative samples in the multi-label sample set according to the click label; A model training module is used to input the multi-label sample set into a deep learning model, and train the deep learning model based on a multi-label weighted loss function to obtain a point of interest ranking model, wherein the deep learning model is used to score each of the POIs and sort them according to the score of each POI, and the multi-label loss function is used to increase the score difference between the POI corresponding to the positive sample and the POI corresponding to the negative sample.

10. A device for sorting points of interest based on deep learning, characterized in that: The device includes a memory and a processor, and the processor is used to execute a program stored in the memory to run the deep learning-based point of interest ranking method as described in any one of claims 1 to 8.

Images