Method, device, equipment and storage medium for extracting feature of item identification

Abstract

The embodiment of the present invention discloses a method, device, equipment and storage medium for extracting features of item identification. The method includes: obtaining the historical access data of each user for each item, segmenting each historical access data, and obtaining the user access sequence of each user, wherein the user access sequence is a sequence used to represent the order of the user's access to the item identification of each item; constructing a graph network of each item identification according to each user access sequence, sampling each item identification in the graph network, and generating a first training sample according to the sampling result; inputting the first training sample into an untrained graph neural network, and extracting the identification features of each item identification according to the training results of the graph neural network. The technical solution of the embodiment of the present invention can deeply explore the correlation between the item identifications and compress them into understandable identification features of limited dimensions, thereby achieving the effect of reducing the dimension and efficiently utilizing the item identification.

Description

Technical Field

[0001] The embodiments of the present invention relate to the field of computer application technology, and in particular to a method, device, equipment and storage medium for extracting feature of an object identification. Background Art

[0002] For items on e-commerce platforms, usually, each item has its own unique item identification (ID) that can represent the entire item. The information contained in the item identification is not only the ID itself, but may also be various information of the corresponding item within the e-commerce platform. Therefore, the item identification is an important reference factor in many application scenarios.

[0003] In order to make better use of item identification, feature extraction can be performed on item identification to obtain identification features. However, the main feature extraction scheme currently used for item identification is hard coding technology, such as one-hot coding, hash coding, and so on.

[0004] In the process of implementing the present invention, the inventors discovered that the prior art has the following technical problems: the existing feature extraction scheme has the problems of dimensionality explosion and insufficient information mining. Summary of the invention

[0005] The embodiments of the present invention provide a method, device, equipment and storage medium for extracting features of an object identification, which solve the problems of dimensionality explosion and insufficient information mining in existing feature extraction solutions.

[0006] In a first aspect, an embodiment of the present invention provides a method for extracting features of an item identification, which may include:

[0007] Obtaining the historical access data of each user for each item, segmenting each historical access data, and obtaining the user access sequence of each user, wherein the user access sequence is a sequence used to represent the order in which the user accesses the item identifier of each item;

[0008] Constructing a graph network of each item identifier according to each user access sequence, sampling each item identifier in the graph network, and generating a first training sample according to the sampling result;

[0009] The first training sample is input into an untrained graph neural network, and the identification features of each object identification are extracted according to the training results of the graph neural network.

[0010] In a second aspect, an embodiment of the present invention further provides a feature extraction device for an item identification, which may include:

[0011] A user access sequence obtaining module is used to obtain the historical access data of each user to each item, segment each historical access data, and obtain the user access sequence of each user. The user access sequence is a sequence used to represent the order in which the user accesses the item identifier of each item;

[0012] A first training sample generating module is used to construct a graph network of each item identifier according to each user access sequence, sample each item identifier in the graph network, and generate a first training sample according to the sampling result;

[0013] The identification feature extraction module is used to input the first training sample into an untrained graph neural network, and extract the identification features of each item identification according to the training results of the graph neural network.

[0014] In a third aspect, an embodiment of the present invention further provides a feature extraction device for an item identification, which may include:

[0015] one or more processors;

[0016] A memory for storing one or more programs;

[0017] When one or more programs are executed by one or more processors, the one or more processors implement the feature extraction method for object identification provided by any embodiment of the present invention.

[0018] In a fourth aspect, an embodiment of the present invention further provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the feature extraction method for object identification provided by any embodiment of the present invention.

[0019] The technical solution of the embodiment of the present invention obtains the user access sequence of each user for representing the order of the user's access to the item identification of each item by segmenting the historical access data of each user to each item, that is, the user access sequence contains the association relationship between the item identifications; then the item identifications are associated by constructing a graph network of the item identifications according to the access sequence of each user, so that after sampling the item identifications in the graph network, a first training sample can be generated, and each group of first training samples can include item identifications with an association relationship; the first training sample is input into an untrained graph neural network, and the item identifications with an association relationship are deeply mined based on the graph neural network, so that the identification features of each item identification can be extracted respectively according to the training results of the graph neural network. The above technical solution deeply mines the association relationship between the item identifications based on the graph neural network, and compresses it into an understandable identification feature of a limited dimension, thereby achieving the effect of reducing the dimension and efficiently utilizing the item identification. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 is a flow chart of a method for extracting features of an object identification in Embodiment 1 of the present invention;

[0021] Figure 2a is a schematic diagram of a user access sequence in a feature extraction method for item identification in the first embodiment of the present invention;

[0022] Figure 2b is a schematic diagram of a graph network in a feature extraction method for object identification in Embodiment 1 of the present invention;

[0023] Figure 3 is a flow chart of a method for extracting features of an object identification in Embodiment 2 of the present invention;

[0024] Figure 4 is a flow chart of a method for extracting features of an object identification in Embodiment 3 of the present invention;

[0025] Figure 5 is a schematic diagram of a graph neural network in a feature extraction method for object identification in Embodiment 3 of the present invention;

[0026] Figure 6 is a flow chart of a feature extraction method for an object identification in a fourth embodiment of the present invention;

[0027] Figure 7 is a schematic diagram of a network structure of an item sorting model in a feature extraction method for item identification in a fourth embodiment of the present invention;

[0028] Figure 8 It is a structural block diagram of a feature extraction device for an object identification in Embodiment 5 of the present invention;

[0029] Fig. 9 It is a structural schematic diagram of a feature extraction device for object identification in Embodiment 6 of the present invention. DETAILED DESCRIPTION

[0030] The present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It is to be understood that the specific embodiments described herein are only used to explain the present invention, rather than to limit the present invention. It is also necessary to explain that, for ease of description, only parts related to the present invention, rather than all structures, are shown in the accompanying drawings.

[0031] Before introducing the embodiment of the present invention, the application scenario of the embodiment of the present invention is first exemplified: With the rapid development of e-commerce and the widespread popularity of mobile Internet, online shopping has become a common shopping choice for consumers. In order to encourage consumers to buy more goods, the e-commerce platform will push the most suitable goods to consumers. On this basis, since the product identification contains various information about the goods within the e-commerce platform, it is crucial to better construct the identification features of the product identification for the e-commerce platform to accurately push out the goods suitable for consumers.

[0032] Embodiment 1

[0033] Figure 1 It is a flow chart of a feature extraction method for an item identification provided in the first embodiment of the present invention. This embodiment can be applicable to the case of extracting features from item identification, and is particularly applicable to the case of extracting features from item identification based on a graph neural network. The method can be executed by a feature extraction device for item identification provided in an embodiment of the present invention, which can be implemented by software and / or hardware, and which can be integrated in a feature extraction device for item identification, which can be various user terminals or servers.

[0034] See also Figure 1 The method of the embodiment of the present invention specifically includes the following steps:

[0035] S110, obtaining historical access data of each user for each item, segmenting each historical access data, and obtaining a user access sequence of each user, wherein the user access sequence is a sequence for indicating the order in which the user accesses the item identifier of each item.

[0036] Among them, the items are items that can be obtained by users on the e-commerce platform. In practical applications, for example, they can be commodities that can be purchased by consumers on shopping websites. Each item has its own item ID, which is the unique identification information of the item, which can be randomly set or based on certain policy settings. It should be noted that even if the item IDs are randomly set, there are actually some hidden connections between the item IDs, which can be mined by the feature extraction method described in the embodiment of the present invention. In other words, the feature extraction method described in the embodiment of the present invention can be used to mine deep connections that cannot be directly obtained from the surface and are not limited to the item ID itself from other dimensions such as user behavior dimensions.

[0037] To illustrate the above point, see, for example, Figure 2a , which is a schematic diagram of items visited by the same user and / or different users on an e-commerce platform. Figure 2aIt can be seen that user 1 (i.e., U1) visited item D, item A, and item B in sequence, that is, user 1 visited the item identifiers corresponding to item D, item A, and item B in sequence. The situations of user 2 and user 3 are similar, and will not be repeated here. It can be seen that there must be a certain association relationship between these items. For example, after user 1 searched for "XX brand mobile phone", he visited item D, item A, and item B, and did not visit item C, item E, and item F. In fact, item D, item A, and item B are all items under the XX brand. Naturally, these items are similar in category, which shows that there is an association relationship between these items, that is, there is an association relationship between the item identifiers corresponding to these items.

[0038] The historical access data may be data on the access operations performed by each user on each item within a preset time period before the current time point, and the operations may be click operations, browse operations, add-to-cart operations, purchase operations, comment operations, etc., which are not specifically limited here. For example, the user who accessed what item in the above example is a historical access data.

[0039] Since each historical access data is all the historical access data of each user, in order to mine the association between each item or the item identifier corresponding to each item, all the historical access data of each user can be segmented by user, so that when at least one historical access data is sequenced based on the access order, the user access sequence of each user can be obtained, that is, the user access sequence of a certain user can be a sequence of the order in which the user accesses the item identifiers of each item. For example, see Figure 2a , where the dotted lines represent different user access sequences. For example, the item ID of item D, the item ID of item A and the item ID of item B constitute the user access sequence of user 1, and the item ID of item B and the item ID of item E constitute the user access sequence of user 2; the item ID of item D, the item ID of item E and the item ID of item F constitute another user access sequence of user 2; the user access sequence of user 3 is similar and will not be repeated here.

[0040] In actual applications, there are many optional ways to implement the segmentation of each historical access data. For example, in different application scenarios, each historical access data is segmented using corresponding preset segmentation factors. The preset segmentation factors can be access words, access time, access location, access device, access page, etc. For example, the historical access data of the same user at a preset time interval and located on the same access page are segmented to form the user access sequence of the user. For example, the historical access data of the same user on the same access device and based on the same access word are segmented to form the user access sequence of the user, and so on. No specific limitation is made here. It should be noted that the above-mentioned consideration of access time ensures that the user access sequence is the historical access data of the user occurring within a period of time, thereby ensuring that the historical access data in the same user access sequence will not have a large deviation. The above-mentioned historical access data under the same access page can be the data when a user accesses an item after entering an advertising page or a live broadcast page. From the above, it can be seen that the user access sequence can be a sequence of item identifiers that have the same identity in at least one dimension.

[0041] S120: construct a graph network of each item identifier according to each user access sequence, sample each item identifier in the graph network, and generate a first training sample according to the sampling result.

[0042] Among them, since the user access sequence is the sequence of the user's access order to the item identifier of each item, a graph network can be constructed based on all the user access sequences of each user. The graph network can be considered as a network structure formed based on the user's access behavior to items, which uses items or corresponding item identifiers as network nodes and the access sequence between items or the access sequence between corresponding item identifiers and item identifiers as edges. For example, based on Figure 2a The user access sequence shown can be constructed Figure 2b For the graph network shown in the figure, if we start traversing from network node D, its next network node may be A or E. Similarly, the next network node of network node E may be C or F.

[0043] Furthermore, sampling is performed on each item identifier in the graph network, such as sampling from one item identifier to another item identifier, and then sampling to another item identifier, thereby obtaining a sampling sample containing at least one item identifier. For the sake of clarity, the item identifier in the sampling sample can be referred to as a sample identifier, that is, there is a sampling order between each sample identifier in the sampling sample. For example, Figure 2bAs shown, a sampling sample can be DABEF. In practical applications, for example, the construction process of the above-mentioned graph network and the sampling process of item identification can be implemented by some technologies under the graph neural network, such as DeepWalk, line, node2vec, SDNE, etc. The above-mentioned sampling samples can also be called sampling results. According to the sampling results, a first training sample can be generated, and its generation process can be related to the actual application scenario. For example, in order to find associated item identifications, an item identification and an item identification that is adjacent to the item identification in the sampling results can be used as the first training sample, wherein the item identification can be used as the actual input and the item identification that is adjacent to the item identification can be used as the expected output.

[0044] S130: Input the first training sample into an untrained graph neural network, and extract identification features of each object identification according to the training results of the graph neural network.

[0045] Among them, the Graph Neural Networks (GNN) can be a network structure based on skip-gram or CBOW under Word2Vec, or FastText, transformed, etc. Since the first training sample is a sample composed of the identification of each item, after multiple groups of first training samples are input into the untrained graph neural network, the identification features of each item identification can be extracted respectively according to the training results of the graph neural network. It should be noted that as time goes by, the historical access data is constantly changing. In order to ensure the real-time accuracy of the identification features, the above steps can be repeated based on the preset time intervals, thereby improving the accuracy of the identification features.

[0046] It should be noted that the above technical solution is an extraction solution for identification features based on graph neural networks. Since graph neural networks are a machine learning model, they learn the identification features of object identification based on the machine learning model, which can not only solve the problem of dimensionality explosion, but also further explore the intrinsic connections between object identifications. The reason for this explanation is that, on the one hand, the object identification is compressed and represented as identification features of finite dimensions through neural network technology, thereby solving the problem of dimensionality explosion and improving the operating efficiency of the corresponding feature extraction equipment; on the other hand, the correlation between object identifications is deeply explored based on graph neural networks, and they are further processed into understandable identification features, thereby realizing efficient utilization of object identifications.

[0047] The technical solution of the embodiment of the present invention obtains the user access sequence of each user for representing the order of the user's access to the item identification of each item by segmenting the historical access data of each user to each item, that is, the user access sequence contains the association relationship between the item identifications; then the item identifications are associated by constructing a graph network of the item identifications according to the access sequence of each user, so that after sampling the item identifications in the graph network, a first training sample can be generated, and each group of first training samples can include item identifications with an association relationship; the first training sample is input into an untrained graph neural network, and the item identifications with an association relationship are deeply mined based on the graph neural network, so that the identification features of each item identification can be extracted respectively according to the training results of the graph neural network. The above technical solution deeply mines the association relationship between the item identifications based on the graph neural network, and compresses it into an understandable identification feature of a limited dimension, thereby achieving the effect of reducing the dimension and efficiently utilizing the item identification.

[0048] Embodiment 2

[0049] Figure 3 It is a flow chart of a feature extraction method for an item identification provided in the second embodiment of the present invention. This embodiment is optimized based on the above-mentioned technical solutions. In this embodiment, optionally, sampling each item identification in the graph network may include: for the current identification in each item identification in the graph network, respectively determining the probability of the current identification to the candidate identification in each item identification except the current identification; sampling the next identification of the current identification from each candidate identification according to each probability, and assigning the next identification to the sampling sample where the current identification is located; updating the next identification to the current identification, if the number of sample identifications in the sampling sample is less than the preset number threshold, repeating the step of respectively determining the probability of the current identification to the candidate identification in each item identification except the current identification; taking the sampled sample as the sampling result. Among them, the explanations of the terms that are the same or corresponding to the above-mentioned embodiments are not repeated here.

[0050] See also Figure 3 The method of this embodiment may specifically include the following steps:

[0051] S210. Obtain the historical access data of each user for each item, segment each historical access data, obtain the user access sequence of each user, and construct a graph network of each item identification according to each user access sequence, wherein the user access sequence is a sequence used to represent the order in which the user accesses the item identification of each item.

[0052] S220. For the current identifier among the item identifiers in the graph network, determine the probability of the current identifier to the candidate identifiers among the item identifiers other than the current identifier; sample the next identifier of the current identifier from each candidate identifier according to each probability, and assign the next identifier to the sampling sample where the current identifier is located.

[0053] Among them, when sampling each item identification in the graph network, the sample can be taken from the current identification to the next identification of the current identification. During the sampling process, the probability of the current identification to each candidate identification among the item identifications except the current identification can be determined respectively, and then the next identification can be sampled from each candidate identification according to each probability. For example, the candidate identification corresponding to the largest probability is taken as the next identification, and then the next identification can be attributed to the sampling sample where the current identification is located. At this time, the sampling sample can include at least two item identifications obtained by sampling in sequence.

[0054] On this basis, optionally, for each candidate identification in each of the item identifications except the current identification, the probability P(v j |v i ):

[0055]

[0056] Among them, v i is the current identifier, v j is a candidate identifier, M ij Yes i to v j The edge weights of Yes i The set of all outgoing edges of ∈ is contained in, that is, Indicates that there is a i to v j The edge of e ij Yes i to v j , ε is the set of all edges in the graph network, is not included in, that is Indicates that there is no i to v j On this basis, the above weights can be determined in many ways, such as from v i to v j There are O records of historical access data, and P records of all historical access data. Then M ij =O / P, the weight at this time is equivalent to the frequency of occurrence.

[0057] S230, updating the next identifier to the current identifier. If the number of sample identifiers in the sampling sample is less than the preset number threshold, repeating the steps of determining the probability of the current identifier to the candidate identifiers of each item identifier except the current identifier; and taking the sampling sample as the sampling result.

[0058] Among them, the next identifier is updated to the current identifier. In order to control the number of sample identifiers in each sampling sample, if the number of sample identifiers in the sampling sample is less than the preset number threshold, the above steps can be repeated to continue sampling, otherwise the sampling can be stopped and the sampling sample at this time is used as the sampling result. Of course, when the number of sample identifiers does not exceed the preset number threshold, but the next identifier of the current identifier no longer exists, the sampling can also be stopped, and the sampling sample at this time is used as the sampling result. Exemplarily, if the preset number threshold is 30, the sampling sample obtained thereby can be (item identifier 1->item identifier 2->...->item identifier 30), that is, the sampling sample can be 30 item identifiers obtained by sampling in sequence.

[0059] S240: Input the first training sample into an untrained graph neural network, and extract identification features of each object identification according to the training results of the graph neural network.

[0060] The technical solution of the embodiment of the present invention traverses each item identification in the graph network, takes any item identification as the starting point, and samples the next item identification according to the probability of the item identification to the remaining item identifications, and so on. In this way, a sampling sample including multiple item identifications with associated relationships is obtained, which is an important guarantee for the subsequent effective generation of the first training sample.

[0061] Embodiment 3

[0062] Figure 4 It is a flow chart of a feature extraction method for an object identification provided in the third embodiment of the present invention. This embodiment is optimized based on the above-mentioned technical solutions. In this embodiment, optionally, each group of first training samples includes a sample identification in the sampling result, and a context identification within a preset sampling window of the sample identification; the first training samples are input into an untrained graph neural network, and the identification features of each object identification are respectively extracted according to the training results of the graph neural network, including: inputting multiple groups of first training samples into an untrained graph neural network, and obtaining an input weight matrix in the graph neural network according to the training results of the graph neural network; for each object identification, extracting a set of weights corresponding to the object identification from the input weight matrix, and using the weights as the identification features of the object identification. Among them, the explanations of the same or corresponding terms in the above-mentioned embodiments are not repeated here.

[0063] See also Figure 4The method of this embodiment may specifically include the following steps:

[0064] S310, obtaining the historical access data of each user for each item, segmenting each historical access data, and obtaining the user access sequence of each user, wherein the user access sequence is a sequence used to represent the order in which the user accesses the item identifier of each item.

[0065] S320, constructing a graph network of item identifiers according to the access sequences of each user, sampling each item identifier in the graph network, and generating a first training sample according to the sampling result, wherein each group of first training samples includes a sample identifier in the sampling result and a context identifier within a preset sampling window of the sample identifier.

[0066] The sample identifier is an item identifier in the sampling result, and the context identifier is an item identifier within a preset sampling window of the sample identifier. For example, assuming that a sampling result is item identifier 1->item identifier 2->...->item identifier 30, the sample identifier is item identifier 3 and the preset sampling window is 2, then the context identifier can be item identifier 1, item identifier 2, item identifier 4 and item identifier 5. In practical applications, the sample identifier can be used as the actual input and the context identifier can be used as the expected output, so that the sample identifier and the context identifier in the first training sample are item identifiers with an associated relationship.

[0067] S330. Input multiple groups of first training samples into an untrained graph neural network, and obtain an input weight matrix in the graph neural network based on the training results of the graph neural network.

[0068] In which, multiple groups of first training samples are input into an untrained graph neural network, and the graph neural network can predict the neighboring identifiers adjacent to the sample identifiers based on the sample identifiers. The neighboring identifiers can be item identifiers that are directly adjacent to the sample identifiers, or item identifiers that are indirectly adjacent to the sample identifiers. Then, the graph neural network compares the neighboring identifiers with the context identifiers, and adjusts various hyperparameters in the graph neural network based on the comparison results, so that the input weight matrix of the graph neural network can be obtained after the network converges. The input weight matrix can be a matrix composed of multiple rows and columns of weights. In practical applications, optionally, the graph neural network can be a network structure based on the skip-gram structure under Word2Vec, such as Figure 5 As shown, the Input Matrix is ​​the input weight matrix.

[0069] S340: For each item identification, extract a set of weights corresponding to the item identification from the input weight matrix, and use the weights as identification features of the item identification.

[0070] Among them, for each item identification, a set of weights corresponding to the item identification can be extracted from the input weight matrix. The set of weights can be a row weight or a column weight in the input weight matrix, and then the set of weights is used as the identification feature of the item identification. Exemplarily, if the item identification is N, then the Nth row weight in the input weight matrix can be used as the identification feature of the item identification. In this way, no matter how many dimensions the item identification itself has, the dimension of the identification feature obtained can be the number of columns of the input weight matrix, and the number of columns is limited, thereby solving the problem of dimensionality explosion.

[0071] On this basis, optionally, since each item identifier may not be a continuous code, in order to make full use of each weight in the input weight matrix, before obtaining the historical access data of each user to each item, each item identifier can be mapped to obtain a mapping result with a continuous code, and then the corresponding item identifier is updated according to the mapping result, thereby obtaining an item identifier with a continuous code. For example, assuming that the item identifiers are 005, 008 and 0015, they are mapped to obtain 001, 002 and 003 with continuous codes, thereby the first row of weights in the input weight matrix can be used as the identification feature of 001 (i.e. 005), the second row of weights can be used as the identification feature of 002 (i.e. 008), and the third row of weights can be used as the identification feature of 003 (i.e. 015).

[0072] The technical solution of an embodiment of the present invention is to train an untrained graph neural network based on a first training sample including a sample identifier in a sampling result and a context identifier within a preset sampling window of the sample identifier. For each object identifier, a set of weights corresponding to the input weight matrix of the object identifier in the graph neural network after training can be used as the identification feature of the object identifier. Since the dimension of a set of weights is limited, the problem of dimensionality explosion is solved.

[0073] Embodiment 4

[0074] Figure 6 It is a flow chart of a feature extraction method for item identification provided in the fourth embodiment of the present invention. This embodiment is optimized based on the above-mentioned technical solutions. In this embodiment, optionally, after respectively extracting the identification features of each item identification according to the training results of the graph neural network, it may also include: obtaining the target features of the target user among each user, and the recall features of the recalled items associated with the target user among each item; inputting the target features and the recall features into the trained item sorting model, and displaying each recalled item according to the output results of the item sorting model. Among them, the explanations of the terms that are the same or corresponding to the above-mentioned embodiments are not repeated here.

[0075] See also Figure 6 The method of this embodiment may specifically include the following steps:

[0076] S410, obtaining the historical access data of each user for each item, segmenting each historical access data, and obtaining the user access sequence of each user, wherein the user access sequence is a sequence for indicating the order in which the user accesses the item identifier of each item.

[0077] S420: construct a graph network of each item identifier according to each user access sequence, sample each item identifier in the graph network, and generate a first training sample according to the sampling result.

[0078] S430: Input the first training sample into an untrained graph neural network, and extract identification features of each object identification according to the training results of the graph neural network.

[0079] S440, obtaining target features of a target user among each user, and recall features of recalled items associated with the target user among each item; inputting the target features and recall features into a trained item sorting model, and displaying each recalled item according to an output result of the item sorting model.

[0080] Among them, the target user may be a user who has performed certain access operations on the e-commerce platform after the current time point mentioned above. When the target user performs certain access operations, the e-commerce platform may recall some recalled items associated with the access operations from various items for the access operations. For example, after the target user searches based on a certain search term on the e-commerce platform, the e-commerce platform may recall some recalled items associated with the search term from various items. It should be noted that different target users still have different interests in the same recalled item. In order to enable the target user to see the recalled items of interest as soon as possible, the target feature of the target user and the recall features of the recalled items associated with the target user in each item may be input into the trained item sorting model. The target feature may be the user feature of the target user, and the recall feature may be the identification feature of the recalled item. The item sorting model may determine the probability of the target user performing access operations on each recalled item based on the target feature and each recall feature. Of course, the specific type of the access operation is related to the second training sample in the model training process, such as click, browse, add to cart, purchase, comment, etc. Thus, the target user's interest in each recalled item can be determined based on the output of the item ranking model, and then the recalled items are displayed based on the interest, for example, the recalled items that the user is more interested in are displayed in front. Exemplarily, the network structure of the item ranking model can be as follows: Figure 7As shown in the figure, the bottom-level input features only show sparse features (i.e., identification features), and other numerical features such as target features are not specifically marked at the bottom level. Each recall feature is processed by the embedding layer to obtain a dense embedding feature, which is then passed through a multi-layer neural network for gradient transfer, and the entire item sorting model is optimized through the back-propagation algorithm.

[0081] On this basis, optionally, the above-mentioned item sorting model can be pre-trained through the following steps: obtaining the historical features of historical users in each user, the sample features of sample items in each item, and the access operations of historical users on sample items; taking the historical features, sample features and access operations as a group of second training samples, and training the original sorting model based on multiple groups of second training samples to obtain an item sorting model. Wherein. The historical user is a user who has performed certain access operations on the e-commerce platform before the current time point, the historical feature is the user feature of the historical user, the sample feature is the identification feature of the item identification of the sample item, and the access operation is a certain operation when the historical user accesses the sample item. For example, assuming that the application scenario is click, the access operation can be click or no click. The historical features and sample features can be actual inputs and the access operation can be the expected output. Therefore, the original sorting model can be trained based on multiple groups of second training samples to obtain an item sorting model that can output the probability of the target user accessing each recalled item.

[0082] It should be noted that the above technical solution can be applied to a variety of application scenarios, such as displaying various items to make up a purchase in order to facilitate users to use coupons, and displaying recalled items in scenarios such as advertising, recommendation, and search, etc., etc., which are not specifically limited here.

[0083] The technical solution of the embodiment of the present invention improves the sorting effect of the item sorting model by inputting the learned identification features into the item sorting model. This helps users quickly find items that suit their interests, thereby optimizing the user's item acquisition experience.

[0084] Embodiment 5

[0085] Figure 8 This is a structural block diagram of a feature extraction device for item identification provided in Embodiment 5 of the present invention. The device is used to execute the feature extraction method for item identification provided in any of the above embodiments. The device and the feature extraction method for item identification in the above embodiments belong to the same inventive concept. For details not described in detail in the embodiment of the feature extraction device for item identification, please refer to the embodiment of the feature extraction method for item identification. Figure 8The device may specifically include: a user access sequence obtaining module 510, a first training sample generating module 520 and an identification feature extracting module 530. Among them,

[0086] The user access sequence obtaining module 510 is used to obtain the historical access data of each user to each item, segment each historical access data, and obtain the user access sequence of each user, wherein the user access sequence is a sequence used to represent the order in which the user accesses the item identifier of each item;

[0087] A first training sample generating module 520 is used to construct a graph network of each item identifier according to each user access sequence, sample each item identifier in the graph network, and generate a first training sample according to the sampling result;

[0088] The identification feature extraction module 530 is used to input the first training sample into an untrained graph neural network, and extract the identification features of each object identification according to the training results of the graph neural network.

[0089] Optionally, the user access sequence obtaining module 510 may specifically include:

[0090] The data segmentation unit is used to segment each historical access data based on preset segmentation factors, wherein the preset segmentation factors include at least one of the following: access word, access time, access location, access device and access page.

[0091] Optionally, the first training sample generating module 520 may specifically include:

[0092] A probability determination unit, for determining, for each item identifier in the graph network, the probability of the current identifier to each item identifier other than the current identifier;

[0093] A next identifier attribution unit is used to sample the next identifier of the current identifier from each candidate identifier according to each probability, and attribute the next identifier to the sampling sample where the current identifier is located;

[0094] A repeating execution unit, configured to update the next identifier to the current identifier, and if the number of sample identifiers in the sampled sample is less than a preset number threshold, repeatedly executing the step of respectively determining the probability of the current identifier to the candidate identifiers of each item identifier other than the current identifier;

[0095] The sampling result obtaining unit is used to take the sampling sample as the sampling result.

[0096] On this basis, optionally, the probability determination unit may specifically include:

[0097] The probability determination subunit is used to determine the probability P(v j |v i ):

[0098]

[0099] Among them, v i is the current identifier, v j is a candidate identifier, M ij Yes i to v j The edge weights of Yes i The set of all outgoing edges, e ij Yes i to v j , and ε is the set of all edges in the graph network.

[0100] Optionally, each group of first training samples includes a sample identifier in the sampling result and a context identifier within a preset sampling window of the sample identifier, and the identifier feature extraction module 530 may include:

[0101] An input weight matrix obtaining unit is used to input multiple groups of first training samples into an untrained graph neural network, and obtain an input weight matrix in the graph neural network according to the training result of the graph neural network;

[0102] The identification feature extraction unit is used to extract a set of weights corresponding to the item identification from the input weight matrix for each item identification, and use the weights as identification features of the item identification.

[0103] Optionally, based on the above device, the device may further include:

[0104] The item identification mapping module is used to map each item identification to obtain a mapping result with continuous coding, and to update the corresponding item identification according to the mapping result.

[0105] Optionally, the feature extraction device for the item identification may further include:

[0106] A recall feature acquisition module, used to acquire target features of a target user among all users, and recall features of recalled items associated with the target user among all items;

[0107] The recalled item display module is used to input the target features and the recalled features into the trained item sorting model, and display each recalled item according to the output results of the item sorting model.

[0108] On this basis, optionally, the item sorting model can be pre-trained through the following modules:

[0109] An access operation acquisition module, used to acquire historical features of historical users among all users, sample features of sample items among all items, and access operations of historical users on sample items;

[0110] The module for obtaining the item sorting model is used to use the historical features, sample features and access operations as a group of second training samples, train the original sorting model based on multiple groups of second training samples, and obtain the item sorting model.

[0111] The feature extraction device for item identification provided in the fifth embodiment of the present invention obtains the user access sequence of each user for representing the order of the user's access to the item identification of each item by means of segmenting the historical access data of each user to each item through the user access sequence obtaining module, that is, the user access sequence contains the association relationship between each item identification; then the first training sample generation module associates each item identification by means of constructing a graph network of each item identification according to each user access sequence, so after sampling each item identification in the graph network, a first training sample can be generated, and each group of first training samples can include item identifications with an association relationship; the identification feature extraction module inputs the first training sample into an untrained graph neural network, and deeply mines each item identification with an association relationship based on the graph neural network, thereby extracting the identification features of each item identification respectively according to the training results of the graph neural network. The above device deeply mines the association relationship between each item identification based on the graph neural network, and compresses it into an understandable identification feature of a limited dimension, thereby achieving the effect of reducing the dimension and efficiently utilizing the item identification.

[0112] The feature extraction device for object identification provided in the embodiment of the present invention can execute the feature extraction method for object identification provided in any embodiment of the present invention, and has the functional modules and beneficial effects corresponding to the execution method.

[0113] It is worth noting that in the embodiment of the feature extraction device for object identification mentioned above, the various units and modules included are only divided according to functional logic, but are not limited to the above division, as long as the corresponding functions can be achieved; in addition, the specific names of the functional units are only for the convenience of distinguishing each other, and are not used to limit the scope of protection of the present invention.

[0114] Embodiment 6

[0115] Fig. 9 A structural diagram of a feature extraction device for an item identification provided in Embodiment 6 of the present invention is shown in FIG. Fig. 9As shown, the device includes a memory 610, a processor 620, an input device 630, and an output device 640. The number of processors 620 in the device can be one or more. Fig. 9 A processor 620 is used as an example; the memory 610, the processor 620, the input device 630 and the output device 640 in the device can be connected via a bus or other means. Fig. 9 The connection via bus 650 is taken as an example.

[0116] The memory 610, as a computer-readable storage medium, can be used to store software programs, computer executable programs and modules, such as program instructions / modules corresponding to the feature extraction method for item identification in the embodiment of the present invention (for example, the user access sequence acquisition module 510, the first training sample generation module 520 and the identification feature extraction module 530 in the feature extraction device for item identification). The processor 620 executes various functional applications and data processing of the device by running the software programs, instructions and modules stored in the memory 610, that is, realizes the feature extraction method for item identification described above.

[0117] The memory 610 may mainly include a program storage area and a data storage area, wherein the program storage area may store an operating system and at least one application required for a function; the data storage area may store data created according to the use of the device, etc. In addition, the memory 610 may include a high-speed random access memory, and may also include a non-volatile memory, such as at least one disk storage device, a flash memory device, or other non-volatile solid-state storage device. In some instances, the memory 610 may further include a memory remotely arranged relative to the processor 620, and these remote memories may be connected to the device via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0118] The input device 630 may be used to receive input digital or character information and generate key signal input related to user settings and function control of the device. The output device 640 may include a display device such as a display screen.

[0119] Embodiment 7

[0120] Embodiment 7 of the present invention provides a storage medium including computer executable instructions, wherein the computer executable instructions are used to execute a method for extracting features of an object identification when executed by a computer processor, the method comprising:

[0121] Obtaining the historical access data of each user to each item, segmenting each historical access data, and obtaining the user access sequence of each user, wherein the user access sequence is a sequence used to represent the order in which the user accesses the item identifier of each item;

[0122] Constructing a graph network of each item identifier according to each user access sequence, sampling each item identifier in the graph network, and generating a first training sample according to the sampling result;

[0123] The first training sample is input into an untrained graph neural network, and the identification features of each object identification are extracted according to the training results of the graph neural network.

[0124] Of course, the computer executable instructions of a storage medium containing computer executable instructions provided by an embodiment of the present invention are not limited to the method operations described above, and can also execute related operations in the feature extraction method for object identification provided by any embodiment of the present invention.

[0125] Through the above description of the implementation methods, the technicians in the relevant field can clearly understand that the present invention can be implemented by means of software and necessary general hardware, and of course it can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory (FLASH), hard disk or optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, server, or network device, etc.) to execute the methods described in each embodiment of the present invention.

[0126] Note that the above are only preferred embodiments of the present invention and the technical principles used. Those skilled in the art will understand that the present invention is not limited to the specific embodiments described herein, and that various obvious changes, readjustments and substitutions can be made by those skilled in the art without departing from the scope of protection of the present invention. Therefore, although the present invention has been described in more detail through the above embodiments, the present invention is not limited to the above embodiments, and may include more other equivalent embodiments without departing from the concept of the present invention, and the scope of the present invention is determined by the scope of the appended claims.

Claims

1. A feature extraction method for object identification, characterized in that: include: Acquire historical access data of each user to each item, segment each of the historical access data, and obtain a user access sequence of each of the users, wherein the user access sequence is a sequence used to represent the order in which the user accesses the item identifiers of the items; Constructing a graph network of each item identifier according to each user access sequence, sampling each item identifier in the graph network, and generating a first training sample according to the sampling result; wherein each group of the first training samples includes a sample identifier in the sampling result and a context identifier within a preset sampling window of the sample identifier; The first training sample is input into an untrained graph neural network, and the identification features of each of the object identifications are extracted respectively according to the training results of the graph neural network.

2. The method according to claim 1, characterized in that The segmenting of each of the historical access data includes: Each of the historical access data is segmented based on preset segmentation factors, wherein the preset segmentation factors include at least one of the following: access words, access time, access location, access device and access page.

3. The method according to claim 1, characterized in that The sampling of each item identifier in the graph network includes: For a current identifier among each item identifier in the graph network, respectively determine the probability of the current identifier to a candidate identifier among each item identifier except the current identifier; Sampling the next identifier of the current identifier from each of the candidate identifiers according to each of the probabilities, and assigning the next identifier to the sampling sample where the current identifier is located; The next identifier is updated to the current identifier, and if the number of sample identifiers in the sampled sample is less than a preset number threshold, the step of respectively determining the probability of the current identifier to the candidate identifiers of each of the item identifiers other than the current identifier is repeated; The sampling sample is used as the sampling result.

4. The method according to claim 3, characterized in that: The respectively determining the probability of the current identification to the candidate identifications other than the current identification in the item identifications includes: For each candidate identifier among the item identifiers except the current identifier, the probability P(v j |v i ): Among them, v i is the current identifier, v j is the candidate identifier, M ij Yes i to v j The weight of the edge, Yes i The set of all outgoing edges, e ij Yes i to v j , ε is the set of all edges in the graph network.

5. The method according to claim 1, characterized in that The step of inputting the first training sample into an untrained graph neural network and extracting identification features of each of the item identifications according to the training results of the graph neural network includes: Inputting multiple groups of the first training samples into an untrained graph neural network, and obtaining an input weight matrix in the graph neural network according to the training results of the graph neural network; For each of the item identifiers, a set of weights corresponding to the item identifier is extracted from the input weight matrix, and the weights are used as identification features of the item identifier.

6. The method according to claim 5, characterized in that Before obtaining the historical access data of each user to each item, the method further includes: Each of the item identifiers is mapped to obtain a mapping result with a continuous code, and the corresponding item identifiers are updated respectively according to the mapping result.

7. The method according to claim 1, characterized in that After extracting the identification features of each of the item identifications according to the training results of the graph neural network, the method further includes: Acquire target features of a target user among the users, and recall features of recalled items associated with the target user among the items; The target feature and the recall feature are input into a trained item sorting model, and each of the recalled items is displayed according to an output result of the item sorting model.

8. The method according to claim 7, characterized in that The item sorting model is pre-trained through the following steps: Acquire historical features of historical users among the users, sample features of sample items among the items, and access operations of the historical users on the sample items; The historical features, the sample features, and the access operations are used as a group of second training samples, and the original sorting model is trained based on multiple groups of the second training samples to obtain the item sorting model.

9. A feature extraction device for object identification, characterized in that: include: A user access sequence obtaining module, used to obtain the historical access data of each user to each item, segment each of the historical access data, and obtain the user access sequence of each user, wherein the user access sequence is a sequence used to represent the order in which the user accesses the item identifiers of each item; A first training sample generation module is used to construct a graph network of each item identifier according to each user access sequence, sample each item identifier in the graph network, and generate a first training sample according to the sampling result; wherein each group of the first training samples includes a sample identifier in the sampling result and a context identifier within a preset sampling window of the sample identifier; The identification feature extraction module is used to input the first training sample into an untrained graph neural network, and extract the identification features of each of the object identifications according to the training results of the graph neural network.

10. A feature extraction device for object identification, characterized in that: include: one or more processors; A memory for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the feature extraction method for object identification as described in any one of claims 1-8.

11. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method for extracting features of an object identification as described in any one of claims 1 to 8 is implemented.

Images