Article recommendation method and device and computer readable storage medium

Abstract

The embodiment of the invention provides an article recommendation method and device and a computer readable storage medium, and the method comprises the steps: obtaining discrete features and interest coefficients of an object; the discrete features comprise a plurality of interest vectors; performing intention classification processing on the plurality of interest vectors through a preset intention classification model to obtain a click intention interest vector, a novel click intention interest vector and a purchase intention vector; the preset intention classification model is used for performing intention disassembly on the interest vector to obtain a plurality of intention interest vectors; determining a multi-dimensional interest portrait based on the click intention interest vector, the novel click intention interest vector and the purchase intention vector; recalling a plurality of articles according to the multi-dimensional interest portraits and the interest coefficients; and screening the plurality of articles to determine to-be-recommended articles.

Description

Technical Field

[0001] The present invention relates to the field of computers, and in particular to an item recommendation method and device, and a computer-readable storage medium. Background Art

[0002] Although the existing multi-interest recall model can recall users' diverse interests (for example, users like clothes and mobile phones at the same time), in the e-commerce recommendation scenario, users' interests will change with their intentions. User interests will vary greatly under different intentions, so the corresponding recommendation strategy should also change under different intentions.

[0003] In the recall stage, existing recommendation systems build multiple models with different intent targets to predict user interests. However, existing models have the problem of cold start of products. Medium and long-tail products appear less frequently in user sequences, which makes it impossible to fully train product representations and leads to poor accuracy of item recommendations. Summary of the invention

[0004] Embodiments of the present invention provide an item recommendation method and device, and a computer-readable storage medium, which can improve the accuracy of item recommendation.

[0005] The technical solution of the present invention is achieved in this way:

[0006] An embodiment of the present invention provides an item recommendation method, the method comprising:

[0007] Acquire discrete features and interest coefficients of an object; the discrete features include a plurality of interest vectors;

[0008] The multiple interest vectors are subjected to intent classification processing by using a preset intent classification model to obtain a click intention interest vector, a novel click intention interest vector, and a purchase intention vector; the preset intent classification model is used to perform intent decomposition on the interest vector to obtain multiple intent interest vectors;

[0009] Determining a multi-dimensional interest portrait based on the click intention interest vector, the novel click intention interest vector, and the purchase intention vector;

[0010] According to the multi-dimensional interest portrait and the interest coefficient, multiple items are recalled; and the multiple items are screened to determine the items to be recommended.

[0011] In the above scheme, before performing intent classification processing on the multiple interest vectors by using a preset intent classification model to obtain a click intention interest vector, a novel click intention interest vector and a purchase intention vector, the method further includes:

[0012] Acquire different target training samples and multiple historical interest vectors of the object; the different target training samples include target click samples, target novel click samples and target purchase samples;

[0013] Based on the different target training samples and the multiple historical interest vectors, an initial intent classification model is trained to determine the preset intent classification model.

[0014] In the above scheme, the different target training samples include target click samples, target novel click samples and target purchase samples;

[0015] The training of the initial intent classification model based on the different target training samples and the multiple historical interest vectors to determine the preset intent classification model includes:

[0016] By using the initial intention classification model, the multiple historical interest vectors are subjected to intention classification processing to obtain a first click intention interest vector, a first novel click intention interest vector, and a first purchase intention vector related to the different target training samples;

[0017] Based on the first click intention interest vector, the first novel click intention interest vector, and the first purchase intention vector, respectively determine a second click intention interest vector, a second novel click intention interest vector, and a second purchase intention vector;

[0018] Determine a loss function value based on the target click sample, the target novel click sample, the target purchase sample, the second click intention interest vector, the second novel click intention interest vector, and the second purchase intention vector;

[0019] The initial intent classification model is optimized by using the loss function value until the loss function value reaches a preset loss function value, thereby determining the preset intent classification model.

[0020] In the above scheme, determining the loss function value based on the target click sample, the target novel click sample, the target purchase sample, the second click intention interest vector, the second novel click intention interest vector and the second purchase intention vector includes:

[0021] Calculating a first loss function value based on the target click sample and the second click intention interest vector;

[0022] Calculating a second loss function value based on the target novel click sample and the second novel click intention interest vector;

[0023] Calculating a third loss function value based on the target purchase sample and the second purchase intention vector;

[0024] The loss function value is obtained by performing a sum operation on the first loss function value, the second loss function value and the third loss function value.

[0025] In the above scheme, the intention classification process is performed on the multiple interest vectors by using a preset intention classification model to obtain a click intention interest vector, a novel click intention interest vector and a purchase intention vector, including:

[0026] By using the neural network in the preset intention classification model, each interest vector is decomposed into intentions to obtain a plurality of expert interest components corresponding to each interest vector;

[0027] The interest vectors of the multiple expert interest components are combined on different intentions through the gating network in the preset intention classification model to determine the click intention interest vector, the novel click intention interest vector and the purchase intention vector.

[0028] In the above solution, determining a multi-dimensional interest portrait based on the click intention interest vector, the novel click intention interest vector and the purchase intention vector includes:

[0029] Based on the click intention interest vector, the novel click intention interest vector and the purchase intention vector, respectively determine a plurality of category identifiers corresponding to the click intention interest vector, the novel click intention interest vector and the purchase intention vector;

[0030] Screening the multiple category identifiers to determine final category identifiers corresponding to the click intention interest vector, the novel click intention interest vector, and the purchase intention vector;

[0031] The final category identification is used as the multi-dimensional interest portrait.

[0032] In the above solution, the step of obtaining the discrete features and interest coefficients of the object includes:

[0033] Obtain the attribute characteristics of the object and the historical browsing data within a preset time period;

[0034] Determining a browsing identification relationship graph according to identification attribute information in the historical browsing data;

[0035] Determining an identification sequence according to the browsing identification relationship diagram;

[0036] Based on the identification sequence, characterization processing is performed to obtain multiple interest vectors;

[0037] The attribute features of the object and the multiple interest vectors are used as the discrete features.

[0038] In the above solution, the characterization process is performed based on the identification sequence to obtain multiple interest vectors, including:

[0039] Through the self-attention mechanism network, feature extraction is performed on the identification sequence to obtain high-order features;

[0040] The high-order features are processed by a dynamic routing algorithm to obtain the multiple interest vectors.

[0041] In the above scheme, the discrete features include attribute features;

[0042] The determining a multi-dimensional interest portrait based on the click intention interest vector, the novel click intention interest vector and the purchase intention vector includes:

[0043] determining the multi-dimensional interest portrait according to the click intention interest vector, the novel click intention interest vector, the purchase intention vector and the attribute feature; or,

[0044] The multi-dimensional interest portrait is determined according to the click intention interest vector, the novel click intention interest vector and the purchase intention vector.

[0045] In the above scheme, the method further includes:

[0046] The preset intention classification model is evaluated through the online historical browsing data of the object to obtain multiple evaluation results;

[0047] Based on the multiple evaluation results, calculating the average index of the preset intent classification model;

[0048] After completing an online item recommendation, the evaluation index of the current preset intention classification model is calculated. If the evaluation index is greater than or equal to the average index, the preset intention classification model is updated.

[0049] An embodiment of the present invention provides an item recommendation device, which includes an acquisition unit and a determination unit; wherein:

[0050] The acquisition unit is used to acquire discrete features and interest coefficients of the object; the discrete features include multiple interest vectors, and the multiple interest vectors are subjected to intent classification processing through a preset intent classification model to obtain a click intention interest vector, a novel click intention interest vector, and a purchase intention vector; the preset intent classification model is used to perform intent decomposition on the interest vector to obtain multiple intent interest vectors;

[0051] The determination unit is used to determine a multi-dimensional interest portrait based on the click intention interest vector, the novel click intention interest vector and the purchase intention vector; recall multiple items according to the multi-dimensional interest portrait and the interest coefficient; and screen the multiple items to determine the items to be recommended.

[0052] An embodiment of the present invention provides an item recommendation device, the item recommendation device comprising:

[0053] A memory for storing executable instructions;

[0054] A processor is used to execute the executable instructions stored in the memory. When the executable instructions are executed, the processor executes the item recommendation method.

[0055] An embodiment of the present invention provides a computer-readable storage medium, wherein the storage medium stores executable instructions. When the executable instructions are executed by one or more processors, the processors execute the item recommendation method.

[0056] The embodiment of the present invention provides an item recommendation method and device, and a computer-readable storage medium, wherein the method includes: obtaining discrete features and interest coefficients of an object; the discrete features include multiple interest vectors; performing intent classification processing on the multiple interest vectors through a preset intent classification model to obtain a click intention interest vector, a novel click intention interest vector, and a purchase intention vector; the preset intent classification model is used to perform intent decomposition on the interest vector to obtain multiple intent interest vectors; based on the click intention interest vector, the novel click intention interest vector, and the purchase intention vector, a multi-dimensional interest portrait is determined; based on the multi-dimensional interest portrait and the interest coefficient, multiple items are recalled; and the multiple items are screened to determine the items to be recommended. In the above scheme, the multiple interest vectors of the object are subjected to intent classification processing through a preset intent classification model to obtain a click intention interest vector, a novel click intention interest vector, and a purchase intention vector, and then the multi-dimensional interest portrait of the object is determined based on the click intention interest vector, the novel click intention interest vector, and the purchase intention vector. Since the multi-dimensional interest portrait is determined by multiple types of interest vectors, the multi-dimensional interest portrait is also multi-type, which can more comprehensively and accurately understand the needs of the object. Based on multi-dimensional interest portraits and interest coefficients, multiple items are recalled, and multiple items can accurately meet the browsing needs of the object. Then, by screening multiple items, the accuracy of recommended items is improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 An optional process diagram of an item recommendation method provided in an embodiment of the present invention Figure 1 ;

[0058] Figure 2 An optional process diagram of an item recommendation method provided in an embodiment of the present invention Figure 2 ;

[0059] Figure 3 An optional process diagram of an item recommendation method provided in an embodiment of the present invention Figure 3 ;

[0060] Figure 4a An optional method for determining an identification sequence of an item recommendation method according to an embodiment of the present invention is provided. Figure 1 ;

[0061] Figure 4b An optional method for determining an identification sequence of an item recommendation method according to an embodiment of the present invention is provided. Figure 2 ;

[0062] Figure 4c An optional method for determining an identification sequence of an item recommendation method according to an embodiment of the present invention is provided. Figure 3 ;

[0063] Figure 5 An optional framework for an item recommendation method is provided for an embodiment of the present invention Figure 1 ;

[0064] Figure 6 An optional framework for an item recommendation method is provided for an embodiment of the present invention Figure 2 ;

[0065] Figure 7 A schematic diagram of a device for recommending items according to an embodiment of the present invention is provided. Figure 1 ;

[0066] Figure 8 A schematic diagram of a device for recommending items according to an embodiment of the present invention is provided. Figure 2 . DETAILED DESCRIPTION

[0067] To make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the specific technical solution of the present invention will be further described in detail below in conjunction with the drawings in the embodiments of the present invention. The following embodiments are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0068] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art to which the present invention belongs. The terms used herein are only for the purpose of describing the embodiments of the present invention and are not intended to limit the present invention.

[0069] In the following description, reference is made to “some embodiments”, “this embodiment”, “embodiments of the present invention” and examples, etc., which describe a subset of all possible embodiments, but it can be understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0070] If similar descriptions of "first / second" appear in the invention document, the following description is added. In the following description, the terms "first\second\third" involved are merely used to distinguish similar objects and do not represent a specific ordering of the objects. It can be understood that "first\second\third" can be interchanged in a specific order or sequence where permitted, so that the embodiments of the invention described herein can be implemented in an order other than that illustrated or described herein.

[0071] The present invention proposes a novel method for constructing a multi-dimensional user interest portrait, and proposes a novel method for item recommendation (item recommendation) based on the multi-dimensional user interest portrait. The multi-dimensional user interest portrait can not only characterize the diverse interests of users, but also construct different interest portraits of users according to different intentions, and split the user interests into purchase interests, click interests, novel click interests, etc., thereby realizing a multi-intention, multi-category and multi-dimensional characterization of user interests. For the first time, the present invention constructs a multi-dimensional user interest portrait based on a user multi-interest network and a multi-task learning model (Multi-gate Mixture-of-Experts model, MMOE), realizing the construction of a multi-dimensional interest portrait with a single model, greatly saving resources; the model introduces more intent target knowledge, so that the single target effect is also improved. In addition, in order to solve the cold start problem, the GNN model is also introduced to initialize the item representation, which effectively solves the cold start problem of commodities, further improves the effect of user interest portraits, and thus improves the accuracy of item recommendations.

[0072] In the embodiment of the present invention, the item recommendation device may be a terminal. The terminal may be a device with a wireless communication function, and the terminal in the embodiment of the present invention may be a mobile phone, a tablet computer, a computer with a wireless transceiver function, a PDA, a desktop computer, etc.

[0073] Figure 1 This is an optional process diagram of an item recommendation method provided by an embodiment of the present invention. Figure 1 , will combine Figure 1 The steps shown are explained.

[0074] S101. Obtain discrete features and interest coefficients of an object; the discrete features include multiple interest vectors.

[0075] In an embodiment of the present invention, the item recommendation method is applicable to an e-commerce scenario, and uses the historical browsing data of an object to recommend items that meet the object's interest requirements.

[0076] In the embodiment of the present invention, the discrete features of the object include multiple interest vectors and attribute features, wherein each of the multiple interest vectors represents an interest in a certain aspect of the object, and the attribute features represent some other discrete attribute features of the object except the interest vector.

[0077] In the embodiment of the present invention, the interest vector represents the interest in a certain item, for example, interest in mobile phones, interest in clothes, interest in shoes, etc.

[0078] In the embodiment of the present invention, the interest coefficient includes a click interest coefficient, a novel click interest coefficient, and a purchase interest coefficient, and the sum of the three different interest coefficients is 1. The interest coefficient can be directly obtained by performing intent analysis on the historical browsing data of the object.

[0079] In an embodiment of the present invention, a terminal obtains attribute characteristics of an object and historical browsing data within a preset time period. A browsing identification relationship graph is determined based on identification attribute information in the historical browsing data. An identification sequence is determined based on the browsing identification relationship graph. Based on the identification sequence, a characterization process is performed to obtain multiple interest vectors. Finally, the multiple interest vectors of the object and the attribute characteristics of the object are used as discrete features of the object.

[0080] In the embodiment of the present invention, the object generally refers to a user, and the embodiment of the present invention does not limit this.

[0081] S102. Perform intent classification processing on multiple interest vectors through a preset intent classification model to obtain a click intention interest vector, a novel click intention interest vector, and a purchase intention vector; the preset intent classification model is used to perform intent decomposition on the interest vector to obtain multiple intent interest vectors.

[0082] In an embodiment of the present invention, a preset intent classification model is used to decompose the interest vector into multiple intent interest vectors; the preset intent classification model can be a multi-task learning model (Multi-gate Mixture-of-Expertsmodel, MMOE). Multi-task learning aims to build a single model to learn multiple goals and tasks at the same time. The multi-task model can improve the learning efficiency and quality of each task by learning the connections and differences between different tasks.

[0083] In the embodiment of the present invention, the framework of multi-task learning widely adopts the shared-bottom structure, and the bottom hidden layer is shared between different tasks. This structure can essentially reduce the risk of overfitting, but the effect may be affected by task differences and data distribution. There are also some other structures, such as the parameters of the two tasks are not shared, but the L2 norm is restricted by adding restrictions to the parameters of different tasks; there are also some that learn a set of hidden layers for each task and then learn the combination of all hidden layers. General multi-task learning model: the hidden layer close to the input layer is shared as a whole; MMOE: the shared bottom representation layer is divided into multiple expert layers (it can be regarded as a three-dimensional Tensor, that is, similar to a Shared Bottom but with a latitude of 3), and a Gate network is set up at the same time, so that different tasks can use the shared layer in a diversified manner; Gate network function: the gating networks of each task realize the selective use of experts through different final output weights. The gating networks of different tasks can learn different patterns of combining experts, so the model takes into account the capture of the relevance and difference of the tasks.

[0084] In the embodiment of the present invention, the terminal decomposes the intent of each interest vector through the neural network in the preset intent classification model to obtain multiple expert interest components corresponding to each interest vector. Then, the gated network in the preset intent classification model combines the interest vectors of multiple expert interest components on different intentions to determine the click intention interest vector, novel click intention interest vector and purchase intention vector.

[0085] Exemplarily, after obtaining the object interest vector i 1 ,i 2 ,…,i n After that, each interest representation i k (i k represents the kth interest representation) is input into the MMOE network to obtain the click intention interest representations respectively. Novel click intent interest representation and purchase intention representation Among them, click intention interest representation That is, click intention interest vector; novel click intention interest representation That is, the novel click intention interest vector; purchase intention representation That is the purchase intent vector.

[0086] S103: Determine a multi-dimensional interest portrait based on the click intention interest vector, the novel click intention interest vector, and the purchase intention vector.

[0087] In an embodiment of the present invention, a multi-dimensional interest portrait constructs an object interest in a hierarchical manner. First, the portrait splits the interest into three categories: click interest, novel click interest, and purchase interest according to the object intent. Then, each category of intent contains K combinations of category identifiers that the user is most interested in.

[0088] In an embodiment of the present invention, the terminal determines a multi-dimensional interest portrait based on a click intention interest vector, a novel click intention interest vector, a purchase intention vector and attribute characteristics; or, determines a multi-dimensional interest portrait based on a click intention interest vector, a novel click intention interest vector and a purchase intention vector.

[0089] In an embodiment of the present invention, the terminal determines multiple category identifiers corresponding to the click intention interest vector, the novel click intention interest vector, and the purchase intention vector, respectively, based on the click intention interest vector, the novel click intention interest vector, and the purchase intention vector. The multiple category identifiers are screened to determine the final category identifiers corresponding to the click intention interest vector, the novel click intention interest vector, and the purchase intention vector, and the final category identifier is used as a multi-dimensional interest portrait.

[0090] In an embodiment of the present invention, the terminal determines the multi-dimensional interest portrait according to the click intention interest vector, the novel click intention interest vector, the purchase intention vector and the attribute characteristics, including: splicing the attribute characteristics of the objects in the click intention interest vector, the novel click intention interest vector and the purchase intention vector to obtain the spliced ​​click intention interest vector, the spliced ​​novel click intention interest vector and the spliced ​​purchase intention vector. According to the spliced ​​click intention interest vector, the spliced ​​novel click intention interest vector and the spliced ​​purchase intention vector, multiple category identifiers corresponding to the spliced ​​click intention interest vector, the spliced ​​novel click intention interest vector and the spliced ​​purchase intention vector are determined, the multiple category identifiers are screened, the final category identifiers corresponding to the click intention interest vector, the novel click intention interest vector and the purchase intention vector are determined, and the final category identifiers are used as the multi-dimensional interest portrait.

[0091] S104. Recall multiple items based on the multi-dimensional interest portrait and interest coefficient; and screen the multiple items to determine the items to be recommended.

[0092] In an embodiment of the present invention, the terminal recalls multiple items based on multi-dimensional interest portraits and interest coefficients; sorts the multiple items through a refined ranking model to obtain a sorting result, and selects a preset number of items from the sorting result from high to low as items to be recommended.

[0093] In an embodiment of the present invention, after determining the multidimensional interest portrait based on the click intention interest vector, the novel click intention interest vector and the purchase intention vector, the terminal recalls items according to the identifiers and interest coefficients in the multidimensional interest portrait to obtain a preferred item pool corresponding to each of the multiple identifiers, thereby obtaining multiple items.

[0094] In the embodiment of the present invention, after recalling multiple items according to the specific intentions of different objects, the terminal evaluates the multiple items respectively through the refined ranking model to obtain the evaluation values ​​corresponding to the multiple items. By sorting the evaluation values ​​corresponding to the multiple items, a sorting result of the multiple items is obtained, and a preset number of items are selected from the sorting result from high to low as the items to be recommended.

[0095] For example, in the recall phase, for a large number of commodities (ie, items), according to different exclusive specific intentions, recall a|P click |+b|P novel_click |+c|P order | category identifiers and corresponding preferred commodity pools (i.e., multiple items), and send the relevant commodities (items) to the refined ranking stage, where a, b, c (i.e., interest coefficients) are the proportions corresponding to the three types of interests, click interest, novelty click interest, and purchase interest, respectively, and satisfy a+b+c=1. In the refined ranking stage, a refined ranking model such as the Deep Interest Network (DIN) model is used to sort the results of the recall stage and output them to the downstream, which determines the items to be recommended.

[0096] It is understandable that the terminal performs intent classification processing on multiple interest vectors of the object through a preset intent classification model, obtains a click intention interest vector, a novel click intention interest vector and a purchase intention vector, and then determines the multi-dimensional interest portrait of the object based on the click intention interest vector, the novel click intention interest vector and the purchase intention vector. Since the multi-dimensional interest portrait is determined by multiple types of interest vectors, the multi-dimensional interest portrait is also multi-type, which can understand the needs of the object more comprehensively and accurately. Based on the multi-dimensional interest portrait and interest coefficient, multiple items are recalled, and multiple items can accurately meet the browsing needs of the object. Then, by screening multiple items, the accuracy of item recommendation is improved.

[0097] In an embodiment of the present invention, Figure 2 An optional process diagram of an item recommendation method is provided for an embodiment of the present invention Figure 2 ,like Figure 2 As shown, before S102, S105 and S106 are also executed as follows:

[0098] S105: Obtain multiple historical interest vectors of different target training samples and objects.

[0099] In the embodiment of the present invention, different target training samples include target click samples, target novelty click samples and target purchase samples.

[0100] In the embodiment of the present invention, the items in different target training samples may be commodities, and the embodiment of the present invention does not limit this.

[0101] In the embodiment of the present invention, the terminal uses the user behavior sequence to click on the next product as the click interest label (i.e., the target click sample), the next novel click on the product as the novel click interest label (i.e., the target novel click sample), and the next order product as the purchase interest label (i.e., the target purchase sample). The target click sample, the target novel click sample, and the target purchase sample are respectively denoted as D click , D novel_click , D order , let label be m j .

[0102] In an embodiment of the present invention, the terminal obtains attribute features of the object and historical browsing data within a preset historical time period. According to the identification attribute information in the historical browsing data, a browsing identification relationship graph is determined. According to the browsing identification relationship graph, an identification sequence is determined. Based on the identification sequence, a characterization process is performed to obtain multiple historical interest vectors.

[0103] S106: Based on different target training samples and multiple historical interest vectors, the initial intent classification model is trained to determine a preset intent classification model.

[0104] In the embodiment of the present invention, the initial intent classification model is an untrained multi-task learning model. The preset intent classification model is obtained after training the initial intent classification model.

[0105] In an embodiment of the present invention, the terminal can perform intent classification processing on multiple historical interest vectors through an initial intent classification model to obtain a historical click intention interest vector, a historical novel click intention interest vector, and a historical purchase intention vector. The loss function value is determined by the target click sample, the target novel click sample, the target purchase sample, the historical click intention interest vector, the historical novel click intention interest vector, and the historical purchase intention vector. The initial intent classification model is optimized by the loss function value until the loss function value reaches the preset loss function value, and the preset intent classification model is determined.

[0106] In an embodiment of the present invention, the terminal can perform intent classification processing on multiple historical interest vectors through an initial intent classification model to obtain a first click intention interest vector, a first novel click intention interest vector, and a first purchase intention vector related to different target training samples. Based on the first click intention interest vector, the first novel click intention interest vector, and the first purchase intention vector, the second click intention interest vector (i.e., the historical click intention interest vector), the second novel click intention interest vector (i.e., the historical novel click intention interest vector), and the second purchase intention vector (i.e., the historical purchase intention vector) are determined respectively.

[0107] In an embodiment of the present invention, the terminal calculates a first loss function value based on the target click sample and the second click intention interest vector; calculates a second loss function value based on the target novel click sample and the second novel click intention interest vector; calculates a third loss function value based on the target purchase sample and the second purchase intention vector; and obtains the loss function value by performing an AND operation on the first loss function value, the second loss function value, and the third loss function value. The initial intent classification model is optimized by the loss function value until the loss function value reaches a preset loss function value, and the preset intent classification model is determined.

[0108] It is understandable that the terminal obtains multiple historical interest vectors of different target training samples and objects. Based on different target training samples and multiple historical interest vectors, the initial intent classification model is trained to determine the preset intent classification model, which can facilitate the subsequent use of the preset intent classification model for intent splitting and multi-dimensional interest profiling.

[0109] In an embodiment of the present invention, Figure 3 An optional process diagram of an item recommendation method is provided for an embodiment of the present invention Figure 3 ,like Figure 3 As shown, S106 can be implemented by S1061, S1062, S1063 and S1064 as follows:

[0110] S1061. Perform intent classification processing on multiple historical interest vectors through an initial intent classification model to obtain a first click intention interest vector, a first novel click intention interest vector, and a first purchase intention vector related to different target training samples.

[0111] In an embodiment of the present invention, the terminal performs intent classification processing on multiple historical interest vectors through an initial intent classification model to obtain a first click intention interest vector, a first novel click intention interest vector and a first purchase intention vector related to different target training samples.

[0112] In the embodiment of the present invention, the terminal decomposes the intent of each interest vector through the initial neural network in the initial intent classification model to obtain multiple expert interest components corresponding to each interest vector. Then, the initial gating network in the initial intent classification model combines the interest vectors of multiple expert interest components on different intentions to determine the first click intention interest vector, the first novel click intention interest vector and the first purchase intention vector.

[0113] S1062. Based on the first click intention interest vector, the first novel click intention interest vector, and the first purchase intention vector, determine a second click intention interest vector, a second novel click intention interest vector, and a second purchase intention vector, respectively.

[0114] In an embodiment of the present invention, the terminal determines the second click intention interest vector, the second novel click intention interest vector and the second purchase intention vector respectively by means of vector inner product based on the first click intention interest vector, the first novel click intention interest vector and the first purchase intention vector.

[0115] Exemplarily, after obtaining the user's multi-dimensional interest representation (i.e., the first click intention interest vector, the first novel click intention interest vector, and the first purchase intention vector), hard attention is used to obtain and label the sample m j The most relevant user interest representation (i.e., the second click intention interest vector, the second novel click intention interest vector, and the second purchase intention vector). In the embodiment of the present invention, the interest vectors of different intentions are respectively calculated based on the vector inner product method to be related to m j The three most similar interest representations (i.e., the second click intent interest vector, the second novel click intent interest vector, and the second purchase intent vector)

[0116]

[0117]

[0118]

[0119] Among them, m j represents a positive sample; Represents the second click intention interest vector; represents the second novel click intention interest vector; Represents the second purchase intention vector.

[0120] S1063. Determine the loss function value based on the target click sample, the target novel click sample, the target purchase sample, the second click intention interest vector, the second novel click intention interest vector, and the second purchase intention vector.

[0121] In an embodiment of the present invention, the terminal calculates a first loss function value based on the target click sample and the second click intention interest vector; calculates a second loss function value based on the target novel click sample and the second novel click intention interest vector; calculates a third loss function value based on the target purchase sample and the second purchase intention vector; and obtains the loss function value by performing a sum operation on the first loss function value, the second loss function value and the third loss function value.

[0122] Exemplarily, the loss function is shown in formula (4), as follows:

[0123]

[0124] Among them, L is the loss function value and mw represents all samples; Represents the first loss function value; Represents the second loss function value; Represents the value of the third loss function.

[0125] S1064. Optimize the initial intent classification model through the loss function value until the loss function value reaches the preset loss function value, and determine the preset intent classification model.

[0126] In an embodiment of the present invention, the terminal optimizes the initial intent classification model by comparing the loss function value with a preset loss function value until the loss function value reaches the preset loss function value, and determines the preset intent classification model.

[0127] It is understandable that the terminal can perform intent classification processing on multiple historical interest vectors through the initial intent classification model to obtain historical click intention interest vectors, historical novel click intention interest vectors, and historical purchase intention vectors. The loss function value is determined by the target click sample, the target novel click sample, the target purchase sample, the historical click intention interest vector, the historical novel click intention interest vector, and the historical purchase intention vector. The initial intent classification model is optimized by the loss function value until the loss function value reaches the preset loss function value, and the preset intent classification model is determined, which can facilitate the subsequent use of the preset intent classification model for intent splitting and multi-dimensional interest profiling.

[0128] In the embodiment of the present invention, S102 may be implemented by S1021 and S1022 as follows:

[0129] S1021. Decompose the intent of each interest vector through a neural network in a preset intent classification model to obtain a plurality of expert interest components corresponding to each interest vector.

[0130] In an embodiment of the present invention, after obtaining multiple interest vectors, the terminal may decompose the intent of each interest vector through a neural network in a preset intent classification model to obtain multiple expert interest components corresponding to each interest vector.

[0131] It should be noted that the neural network in the preset intent classification model is multiple neural networks.

[0132] Exemplarily, after obtaining the object interest vector i 1 ,i 2 , ..., i n After that, each interest representation i k (i k represents the kth interest representation) is input into the MMOE network, which first passes through the neural network f 1 (i k ), f 2 (i k ),…,f e (i k ) (also known as neural network) to learn user interests i k And i k It is broken down into e expert interest components, representing different potential intentions of the interest.

[0133] S1022. By using a gating network in a preset intent classification model, interest vectors of multiple expert interest components are combined on different intents to determine a click intention interest vector, a novel click intention interest vector, and a purchase intention vector.

[0134] In an embodiment of the present invention, the terminal can determine a click intention interest vector, a novel click intention interest vector and a purchase intention vector by combining interest vectors of multiple expert interest components on different intentions through a gating network in a preset intent classification model.

[0135] It should be noted that the gated network in the preset intent classification model is a plurality of gated networks.

[0136] Exemplarily, after obtaining multiple expert interest components, through the gating network g click (i k ), g novel_click (i k ), g ordler (i k )Learn kDifferent weight combinations are used for the interest representations of different experts on the three intention targets of click, novel click, and purchase, and finally the interest representations of different user intentions (i.e., click intention interest vector, novel click intention interest vector, and purchase intention vector) are obtained by multiplication. This process can be implemented by formula (5), formula (6), and formula (7), as follows:

[0137]

[0138]

[0139]

[0140] Among them, f j (i k ) represents a neural network; g click (i k ) j , g novel_click (i k ) j and g order (i k ) j Both represent gating networks; Represents the click intention interest vector; Represents novel click intention interest vector; Represents the purchase intention vector.

[0141] In this process, 3n interest representations are obtained. I click (Right now, ), represents all the click interest representations of the user (i.e., click intention interest vector). Similarly, Denoted as l novel_click (Right now, ), represents all the novel click interest representations of the user (i.e., novel click intention interest vector). I order (Right now, ), which represents all the purchase interest representations of the user (i.e., purchase intention vector).

[0142] It is understandable that the terminal decomposes the intent of each interest vector through the neural network in the preset intent classification model to obtain multiple expert interest components corresponding to each interest vector. Through the gated network in the preset intent classification model, the interest vectors of multiple expert interest components are combined on different intentions to determine the click intention interest vector, novel click intention interest vector and purchase intention vector, thus realizing the interest intention splitting and improving the classification of interest vectors, which is convenient for the subsequent characterization of multi-dimensional interest portraits, thereby improving the accuracy of item recommendations.

[0143] In the embodiment of the present invention, S103 may be implemented by S1031, S1032, and S1033 as follows:

[0144] S1031. Based on the click intention interest vector, the novel click intention interest vector, and the purchase intention vector, determine a plurality of category identifiers corresponding to the click intention interest vector, the novel click intention interest vector, and the purchase intention vector, respectively.

[0145] In an embodiment of the present invention, the terminal can calculate multiple category identifiers corresponding to the click intention interest vector, the novel click intention interest vector and the purchase intention vector respectively through cosine similarity based on the click intention interest vector, the novel click intention interest vector and the purchase intention vector.

[0146] In the embodiment of the present invention, after obtaining multiple category identifiers corresponding to the click intention interest vector, the novelty click intention interest vector, and the purchase intention vector, the terminal can obtain multiple category identifier candidate set results.

[0147] For example, for a user's multi-dimensional interest representation (i.e., a click intention interest vector, a novel click intention interest vector, and a purchase intention vector) I click ,I novel_click and I order , based on cosine similarity, calculate the TOP K category identifiers with the most similar interests, totaling K|I click |+K|I novel_click |+K|I order |Results of candidate sets of category identifiers.

[0148] S1032: Screen multiple category identifiers to determine the final category identifiers corresponding to the click intention interest vector, the novel click intention interest vector, and the purchase intention vector.

[0149] In an embodiment of the present invention, the terminal selects the final category identifier with the highest click intention interest vector similarity from multiple category identifiers, selects the final category identifier with the highest novel click intention interest vector similarity, and selects the final category identifier with the highest purchase intention vector similarity. The final category identifier includes one category identifier.

[0150] For example, in the category identifier candidate set results, each intent target selects the K category identifiers with the highest similarity, and the results are recorded as P click (i.e., the final category identifier with the highest click intention interest vector similarity), P novel_click (i.e., the final category identifier with the highest novel click intention interest vector similarity), P order (i.e., the final category identifier with the highest purchase intent vector similarity).

[0151] S1033. Use the final category identification as a multi-dimensional interest portrait.

[0152] In an embodiment of the present invention, after obtaining the final category identifier with the highest similarity to the click intention interest vector, the final category identifier with the highest similarity to the novel click intention interest vector, and the final category identifier with the highest similarity to the purchase intention vector, the terminal uses all the final category identifiers as multi-dimensional interest portraits.

[0153] It is understandable that the terminal determines multiple category identifiers corresponding to the click intention interest vector, the novel click intention interest vector, and the purchase intention vector, respectively, based on the click intention interest vector, the novel click intention interest vector, and the purchase intention vector. The multiple category identifiers are screened to determine the final category identifiers corresponding to the click intention interest vector, the novel click intention interest vector, and the purchase intention vector. The final category identifier is used as a multi-dimensional interest portrait to achieve multi-dimensional characterization of user interests such as multi-intentions and multiple categories, saving resources and improving efficiency.

[0154] In the embodiment of the present invention, obtaining discrete features of the object in S101 may be implemented through S1011, S1012, S1013, S1014 and S1015 as follows:

[0155] S1011. Acquire attribute characteristics of the object and historical browsing data within a preset time period.

[0156] In the embodiment of the present invention, the attribute characteristics of the object are some inherent characteristics of the object. The historical browsing data within a preset time period is generally the object's item browsing behavior in the last six months.

[0157] In the embodiment of the present invention, the terminal can obtain the attribute characteristics of the object and the historical browsing data within a preset time period.

[0158] S1012: Determine a browsing identification relationship graph according to identification attribute information in the historical browsing data.

[0159] In the embodiment of the present invention, after acquiring the historical browsing data, the terminal constructs a browsing identification relationship graph according to the identification attribute information in the historical browsing data.

[0160] In the embodiment of the present invention, the terminal constructs an edge between two vertices based on the identification attribute information in the historical browsing data using the items that are clicked together within one hour, and further constructs a browsing identification relationship graph.

[0161] S1013. Determine the identification sequence according to the browsing identification relationship diagram.

[0162] In the embodiment of the present invention, after obtaining the browsing identification relationship diagram, the terminal learns and obtains the representation of each identification to obtain an identification sequence.

[0163] In the embodiment of the present invention, the terminal learns and obtains the representation of each vertex (ie, identifier) ​​based on the GraphSAGE algorithm, denoted as m, thereby obtaining an identifier sequence.

[0164] For example, the GraphSAGE algorithm scheme is as follows Figure 4a , Figure 4b and Figure 4c As shown in the figure, first obtain the vertices corresponding to all browsed products and establish a sample neighborhood (i.e., Sample neighborhood). Specifically, the first neighborhood is established according to the step length k=1, and the second neighborhood is established according to the step length k=2, so as to obtain the sample neighborhood. Then, aggregate the feature information from neighbors (i.e., Aggregate feature information from neighbors), and build the connection relationship between vertices (i.e., the edge between vertices) according to the items clicked together. The features between the connection relationships are Aggregator represents information aggregation, aggregates vertices from the outside to the inside, and uses aggregated information to predict graph context and label (i.e., Predict graph context and label using aggregated information) to obtain the identification representation of the item. This solves the problem that the post-order algorithm cannot handle the problem that the number of times the mid- and long-tail items appear in the user sequence is small, resulting in poor representation learning effect.

[0165] S1014: Perform characterization processing based on the identification sequence to obtain multiple interest vectors.

[0166] In an embodiment of the present invention, a self-attention mechanism network is used to extract features from a logo sequence to obtain high-order features; a dynamic routing algorithm is used to process the high-order features to obtain multiple interest vectors.

[0167] For example, the terminal uses the Multi-Head Self-attention Layer in the Transformer model (such as Figure 5 As shown in formula (8), m 1 ,m 2 ,…,m x represents the category identification sequence of the items clicked by the subject in the past six months, l is a multi-head self-attention mechanism network, e 1 ,e 2 ,…,ex is a high-order representation of the network output. In the embodiment of the present invention, x is set to 600.

[0168] e 1 , e 2 , ..., e x = l(m 1 , m 2 , ..., m x ) (8)

[0169] Then, the Dynamic Routing algorithm ( Figure 6 The dynamic routing algorithm in the algorithm) obtains the object interest vector representation (i.e., multiple interest vectors), and the calculation is shown in formula (9), where h represents the dynamic routing algorithm, i 1 ,i 2 ,…,i n Indicates different interest vectors of objects, each interest vector represents a certain aspect of the user's interest, such as mobile phone interest, clothing interest, etc. n represents the number of user interest vectors, and n is set to 8 in the present invention.

[0170] i 1 ,i 2 , ..., i n =h(e 1 , e 2 , ..., e x ) (9)

[0171] It should be noted that if Figure 5 The Transformer model shown includes input, input embedding, positional encoding, multi-head attention mechanism, residual connection (Add&Norm) and feed forward; among them, the multi-head attention mechanism, residual connection and feed forward are all used for training.

[0172] S1015: Use attribute features of the object and multiple interest vectors as discrete features.

[0173] In the embodiment of the present invention, the terminal uses the attribute features of the object and multiple interest vectors as discrete features.

[0174] It is understandable that the terminal obtains the attribute characteristics of the object and the historical browsing data within a preset time period. According to the identification attribute information in the historical browsing data, a browsing identification relationship diagram is determined. According to the browsing identification relationship diagram, an identification sequence is determined. Based on the identification sequence, characterization processing is performed to obtain multiple interest vectors, providing data resources for subsequent intention splitting of multiple interest vectors.

[0175] In the embodiment of the present invention, after the item recommendation method completes the item recommendation, S201, S202 and S203 are further executed as follows:

[0176] S201. Evaluate a preset intent classification model through the online historical browsing data of the object to obtain multiple evaluation results.

[0177] In the embodiment of the present invention, the terminal uses multiple indicators to evaluate the preset intention classification model through the online historical browsing data of the object, and obtains multiple evaluation results.

[0178] Exemplarily, the multiple indicators may be AUC, HR@50, etc.

[0179] S202. Based on multiple evaluation results, calculate the average index of the preset intent classification model.

[0180] In an embodiment of the present invention, the terminal may calculate an average index of a preset intent classification model based on multiple evaluation results.

[0181] S203: After completing an online item recommendation, calculate the evaluation index of the current preset intention classification model. If the evaluation index is greater than or equal to the average index, update the preset intention classification model.

[0182] In an embodiment of the present invention, after completing an online item recommendation, the terminal calculates the evaluation index of the current preset intention classification model. If the evaluation index is greater than or equal to the average index, the preset intention classification model is updated; if the evaluation index is less than the average index, the preset intention classification model is not updated.

[0183] For example, in actual online use, when the index (i.e., evaluation index) of the model (i.e., preset intent classification model) is at an average level (i.e., average index) or above the average level, the model file can be updated to the online model library at a regular time. If the model evaluation index is lower than the average level of the model, it will not be updated. The training and evaluation cycle of the offline model is at the daily level, that is, the model effect is trained and evaluated every day.

[0184] It is understandable that the terminal evaluates the preset intent classification model through the online historical browsing data of the object and obtains multiple evaluation results. Based on the multiple evaluation results, the average index of the preset intent classification model is calculated. After completing an online item recommendation, the evaluation index of the current preset intent classification model is calculated. If the evaluation index is greater than or equal to the average index, the preset intent classification model is updated to improve the accuracy of the preset intent classification model.

[0185] In an embodiment of the present invention, Figure 6An optional framework of an item recommendation method provided by an embodiment of the present invention Figure 2 ,like Figure 6 As shown in the figure, the terminal obtains the long-term sequence (i.e., Long Term Sequence, which is historical browsing data), user features (i.e., User Features, which is attribute features) and target training samples (i.e., Target), obtains the initial representation of the item through the graph deep neural network algorithm (Embedding Layer), and then learns the high-order representation of the object behavior sequence based on the multi-head self-attention mechanism network (Multi-Head Self-attention Layer), and then obtains the object's diverse interest representation through the dynamic routing algorithm (Multi-Interest Extractor Layer). Finally, based on the multi-task learning algorithm (MMoE layer), the exclusive interest is split into click interest representation, novel click interest representation, and purchase interest representation. At the same time, the user features are processed through the connection layer (Concat Layer) to obtain the processed features, and then the user's multi-dimensional interest portrait (Concat Layer&RELU) is constructed through the click interest representation, novel click interest representation, purchase interest representation and processed features. The multi-dimensional interest portrait of the object is used as a separate recall source, and the items of interest to the user are recalled in different proportions according to different intentions, and provided to the sorting stage, and finally the items are recommended to the user. Among them, the label-aware attention mechanism (Label-aware Attention) and the sampled softmax loss (Sampled Softmax Loss) are used to train the multi-task learning algorithm. K and V are user embeddings, and Q is label embedding.

[0186] It is understandable that the user's multi-dimensional interest portrait can not only characterize the user's diverse interests, but also construct different interest portraits of users according to different intentions, and split the user's interests into purchase interests, click interests, novel click interests, etc., thereby realizing multi-intention, multi-category and other multi-dimensional characterizations of user interests; in terms of model details, the GNN model is introduced to initialize the item representation, which effectively solves the item cold start problem, and for the first time integrates the dynamic routing algorithm and the multi-task learning model (MMOE) to construct the user's multi-dimensional interest representation, realizing a single model to build a multi-dimensional interest portrait, saving resources and improving efficiency; the model introduces more intent knowledge, which also improves the single-target effect.

[0187] The present invention provides a structural diagram of an item recommendation device. Figure 1 ,like Figure 7As shown, the item recommendation device 7 includes: an acquisition unit 701 and a determination unit 702; wherein,

[0188] The acquisition unit 701 is used to acquire discrete features and interest coefficients of an object; the discrete features include multiple interest vectors, and the multiple interest vectors are subjected to intent classification processing through a preset intent classification model to obtain a click intention interest vector, a novel click intention interest vector, and a purchase intention vector; the preset intent classification model is used to perform intent decomposition on the interest vector to obtain multiple intent interest vectors;

[0189] The determination unit 702 is used to determine a multi-dimensional interest portrait based on the click intention interest vector, the novel click intention interest vector and the purchase intention vector; recall multiple items according to the multi-dimensional interest portrait and the interest coefficient; and screen the multiple items to determine the items to be recommended.

[0190] In the embodiment of the present invention, the acquisition unit 701 is further configured to perform intent classification processing on the multiple interest vectors through a preset intent classification model to obtain a plurality of historical interest vectors of different target training samples and objects before obtaining the click intention interest vector, the novel click intention interest vector and the purchase intention vector;

[0191] The determination unit 702 is further configured to train an initial intent classification model based on the different target training samples and the multiple historical interest vectors to determine the preset intent classification model.

[0192] In an embodiment of the present invention, the different target training samples include target click samples, target novel click samples and target purchase samples;

[0193] The acquisition unit 701 is further configured to perform intent classification processing on the multiple historical interest vectors through the initial intent classification model to obtain a first click intention interest vector, a first novel click intention interest vector, and a first purchase intention vector related to the different target training samples;

[0194] The determination unit 702 is also used to determine the second click intention interest vector, the second novel click intention interest vector and the second purchase intention vector respectively based on the first click intention interest vector, the first novel click intention interest vector and the first purchase intention vector; determine the loss function value based on the target click sample, the target novel click sample, the target purchase sample, the second click intention interest vector, the second novel click intention interest vector and the second purchase intention vector; optimize the initial intention classification model by the loss function value until the loss function value reaches a preset loss function value, and determine the preset intention classification model.

[0195] In the embodiment of the present invention, the item recommendation device 7 further includes a calculation unit 703; wherein,

[0196] The calculation unit 703 is used to calculate a first loss function value based on the target click sample and the second click intention interest vector; calculate a second loss function value based on the target novel click sample and the second novel click intention interest vector; calculate a third loss function value based on the target purchase sample and the second purchase intention vector;

[0197] The acquisition unit 701 is further configured to obtain the loss function value by performing a sum operation on the first loss function value, the second loss function value, and the third loss function value.

[0198] In the embodiment of the present invention, the acquisition unit 701 is further used to perform intention decomposition on each interest vector through the neural network in the preset intention classification model to obtain a plurality of expert interest components corresponding to each interest vector;

[0199] The determination unit 702 is further used to determine the click intention interest vector, the novel click intention interest vector and the purchase intention vector by combining the interest vectors of the multiple expert interest components on different intentions through the gating network in the preset intention classification model.

[0200] In an embodiment of the present invention, the determination unit 702 is further used to determine, based on the click intention interest vector, the novel click intention interest vector and the purchase intention vector, a plurality of category identifiers corresponding to each of the click intention interest vector, the novel click intention interest vector and the purchase intention vector; filter the plurality of category identifiers to determine the final category identifiers corresponding to each of the click intention interest vector, the novel click intention interest vector and the purchase intention vector; and use the final category identifier as the multi-dimensional interest portrait.

[0201] In the embodiment of the present invention, the acquisition unit 701 is further used to acquire the attribute characteristics of the object and the historical browsing data within a preset time period;

[0202] The determining unit 702 is further configured to determine a browsing identification relationship graph according to identification attribute information in the historical browsing data; and determine an identification sequence according to the browsing identification relationship graph;

[0203] The acquisition unit 701 is further configured to perform characterization processing based on the identification sequence to obtain multiple interest vectors;

[0204] The determining unit 702 is further configured to use the attribute features of the object and the multiple interest vectors as the discrete features.

[0205] In an embodiment of the present invention, the acquisition unit 701 is further used to extract features of the identification sequence through a self-attention mechanism network to obtain high-order features; and process the high-order features through a dynamic routing algorithm to obtain the multiple interest vectors.

[0206] In an embodiment of the present invention, the discrete features include attribute features;

[0207] The determination unit 702 is also used to determine the multi-dimensional interest portrait based on the click intention interest vector, the novel click intention interest vector, the purchase intention vector and the attribute characteristics; or, to determine the multi-dimensional interest portrait based on the click intention interest vector, the novel click intention interest vector and the purchase intention vector.

[0208] In the embodiment of the present invention, the item recommendation device 7 further includes an updating unit 704; wherein,

[0209] The acquisition unit 701 is further used to evaluate the preset intention classification model through the online historical browsing data of the object to obtain multiple evaluation results;

[0210] The calculation unit 703 is further configured to calculate an average index of the preset intent classification model based on the multiple evaluation results;

[0211] The updating unit 704 is used to calculate the evaluation index of the current preset intention classification model after completing an online item recommendation, and if the evaluation index is greater than or equal to the average index, update the preset intention classification model.

[0212] Based on the method of the above embodiment, a structural diagram of an item recommendation device provided by an embodiment of the present invention is as follows: Figure 2 ,like Figure 8 As shown, the item recommendation device 8 includes: a processor 801 and a memory 802; the memory 802 stores one or more programs executable by the processor 801, and when one or more programs are executed, the item recommendation method is executed.

[0213] The embodiment of the present invention also provides a computer-readable storage medium, wherein the computer storage medium stores executable instructions, and when the executable instructions are executed, the item recommendation method is executed.

[0214] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of hardware embodiments, software embodiments, or embodiments combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage and optical storage, etc.) containing computer-usable program code.

[0215] It should be noted that the description of the above storage medium and device embodiments is similar to the description of the above method embodiments, and has similar beneficial effects as the method embodiments. For technical details not disclosed in the storage medium, storage medium and device embodiments of the present invention, please refer to the description of the method embodiments of the present invention for understanding.

[0216] It should be understood that "one embodiment" or "an embodiment" or "some embodiments" mentioned throughout the specification means that specific features, structures or characteristics related to the embodiment are included in at least one embodiment of the present invention. Therefore, "in one embodiment" or "in an embodiment" or "in some embodiments" appearing throughout the specification do not necessarily refer to the same embodiment. In addition, these specific features, structures or characteristics can be combined in one or more embodiments in any suitable manner. It should be understood that in various embodiments of the present invention, the size of the serial number of the above-mentioned processes does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiment of the present invention. The serial numbers of the above-mentioned embodiments of the present invention are for description only and do not represent the advantages and disadvantages of the embodiments. The above description of the various embodiments tends to emphasize the differences between the various embodiments. The same or similar aspects can be referenced to each other. For the sake of brevity, this article will not go into details.

[0217] The term "and / or" in this article is only a description of the association relationship of associated objects, indicating that there may be three relationships. For example, object A and / or object B can represent three situations: object A exists alone, object A and object B exist at the same time, and object B exists alone.

[0218] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the existence of other identical elements in the process, method, article or device including the element.

[0219] In the several embodiments provided by the present invention, it should be understood that the disclosed devices and methods can be implemented in other ways. The embodiments described above are only schematic. For example, the division of the modules is only a logical function division. There may be other division methods in actual implementation, such as: multiple modules or components can be combined, or can be integrated into another system, or some features can be ignored, or not executed. In addition, the coupling, direct coupling, or communication connection between the components shown or discussed can be through some interfaces, and the indirect coupling or communication connection of devices or modules can be electrical, mechanical or other forms.

[0220] The modules described above as separate components may or may not be physically separated, and the components displayed as modules may or may not be physical modules; they may be located in one place or distributed on multiple network units; some or all of the modules may be selected according to actual needs to achieve the purpose of the present embodiment.

[0221] In addition, all functional modules in the embodiments of the present invention may be integrated into one processing unit, or each module may be a separate unit, or two or more modules may be integrated into one unit; the above-mentioned integrated modules may be implemented in the form of hardware or in the form of hardware plus software functional units.

[0222] A person skilled in the art can understand that all or part of the steps of implementing the above method embodiment can be completed by hardware related to program instructions, and the aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it executes the steps of the above method embodiment; and the aforementioned storage medium includes: mobile storage devices, read-only memories (ROM), magnetic disks or optical disks, etc., various media that can store program codes.

[0223] Alternatively, if the above-mentioned integrated unit of the present invention is implemented in the form of a software function module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the embodiment of the present invention can be essentially or partly reflected in the form of a software product that contributes to the relevant technology. The computer software product is stored in a storage medium and includes several instructions for enabling an electronic device to execute all or part of the methods described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as mobile storage devices, ROMs, magnetic disks or optical disks.

[0224] The methods disclosed in the several method embodiments provided by the present invention can be arbitrarily combined without conflict to obtain new method embodiments.

[0225] The features disclosed in several product embodiments provided by the present invention can be arbitrarily combined without conflict to obtain new product embodiments.

[0226] The features disclosed in several method or device embodiments provided by the present invention can be arbitrarily combined without conflict to obtain new method embodiments or device embodiments.

[0227] The above is only an embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. An item recommendation method, characterized in that: include: Acquire discrete features and interest coefficients of an object; the discrete features include a plurality of interest vectors; The multiple interest vectors are subjected to intent classification processing by using a preset intent classification model to obtain a click intention interest vector, a novel click intention interest vector, and a purchase intention vector; the preset intent classification model is used to perform intent decomposition on the interest vector to obtain multiple intent interest vectors; Determining a multi-dimensional interest portrait based on the click intention interest vector, the novel click intention interest vector, and the purchase intention vector; According to the multi-dimensional interest portrait and the interest coefficient, multiple items are recalled; and the multiple items are screened to determine the items to be recommended.

2. The method according to claim 1, characterized in that Before performing intent classification processing on the plurality of interest vectors by using a preset intent classification model to obtain a click intention interest vector, a novel click intention interest vector, and a purchase intention vector, the method further includes: Obtain multiple historical interest vectors for different target training samples and objects; Based on the different target training samples and the multiple historical interest vectors, an initial intent classification model is trained to determine the preset intent classification model.

3. The method according to claim 2, characterized in that The different target training samples include target click samples, target novel click samples and target purchase samples; The training of the initial intent classification model based on the different target training samples and the multiple historical interest vectors to determine the preset intent classification model includes: By using the initial intention classification model, the multiple historical interest vectors are subjected to intention classification processing to obtain a first click intention interest vector, a first novel click intention interest vector, and a first purchase intention vector related to the different target training samples; Based on the first click intention interest vector, the first novel click intention interest vector, and the first purchase intention vector, respectively determine a second click intention interest vector, a second novel click intention interest vector, and a second purchase intention vector; Determine a loss function value based on the target click sample, the target novel click sample, the target purchase sample, the second click intention interest vector, the second novel click intention interest vector, and the second purchase intention vector; The initial intent classification model is optimized by using the loss function value until the loss function value reaches a preset loss function value, thereby determining the preset intent classification model.

4. The method according to claim 3, characterized in that The determining of the loss function value based on the target click sample, the target novel click sample, the target purchase sample, the second click intention interest vector, the second novel click intention interest vector, and the second purchase intention vector includes: Calculating a first loss function value based on the target click sample and the second click intention interest vector; Calculating a second loss function value based on the target novel click sample and the second novel click intention interest vector; Calculating a third loss function value based on the target purchase sample and the second purchase intention vector; The loss function value is obtained by performing a sum operation on the first loss function value, the second loss function value and the third loss function value.

5. The method according to any one of claims 1 to 4, characterized in that: The method of performing intent classification processing on the plurality of interest vectors by using a preset intent classification model to obtain a click intent interest vector, a novel click intent interest vector and a purchase intent vector includes: By using the neural network in the preset intention classification model, each interest vector is decomposed into intentions to obtain a plurality of expert interest components corresponding to each interest vector; The interest vectors of the multiple expert interest components are combined on different intentions through the gating network in the preset intention classification model to determine the click intention interest vector, the novel click intention interest vector and the purchase intention vector.

6. The method according to any one of claims 1 to 4, characterized in that: The determining a multi-dimensional interest portrait based on the click intention interest vector, the novel click intention interest vector and the purchase intention vector includes: Based on the click intention interest vector, the novel click intention interest vector and the purchase intention vector, respectively determine a plurality of category identifiers corresponding to the click intention interest vector, the novel click intention interest vector and the purchase intention vector; Screening the multiple category identifiers to determine final category identifiers corresponding to the click intention interest vector, the novel click intention interest vector, and the purchase intention vector; The final category identification is used as the multi-dimensional interest portrait.

7. The method according to claim 1, characterized in that The step of obtaining the discrete features and interest coefficients of the object includes: Obtain the attribute characteristics of the object and the historical browsing data within a preset time period; Determining a browsing identification relationship graph according to identification attribute information in the historical browsing data; Determining an identification sequence according to the browsing identification relationship diagram; Based on the identification sequence, characterization processing is performed to obtain multiple interest vectors; The attribute features of the object and the multiple interest vectors are used as the discrete features.

8. The method according to claim 7, characterized in that The characterization process is performed based on the identification sequence to obtain multiple interest vectors, including: Through the self-attention mechanism network, feature extraction is performed on the identification sequence to obtain high-order features; The high-order features are processed by a dynamic routing algorithm to obtain the multiple interest vectors.

9. The method according to any one of claims 1 to 4, characterized in that: The discrete features include attribute features; The determining a multi-dimensional interest portrait based on the click intention interest vector, the novel click intention interest vector and the purchase intention vector includes: Determining the multi-dimensional interest portrait according to the click intention interest vector, the novel click intention interest vector, the purchase intention vector and the attribute characteristics; or, The multi-dimensional interest portrait is determined according to the click intention interest vector, the novel click intention interest vector and the purchase intention vector.

10. The method according to claim 1, characterized in that The method further comprises: The preset intention classification model is evaluated through the online historical browsing data of the object to obtain multiple evaluation results; Based on the multiple evaluation results, calculating the average index of the preset intent classification model; After completing an online item recommendation, the evaluation index of the current preset intention classification model is calculated. If the evaluation index is greater than or equal to the average index, the preset intention classification model is updated.

11. An item recommendation device, characterized in that: It includes an acquisition unit and a determination unit; wherein, The acquisition unit is used to acquire discrete features and interest coefficients of the object; the discrete features include multiple interest vectors, and the multiple interest vectors are subjected to intent classification processing through a preset intent classification model to obtain a click intention interest vector, a novel click intention interest vector, and a purchase intention vector; the preset intent classification model is used to perform intent decomposition on the interest vector to obtain multiple intent interest vectors; The determination unit is used to determine a multi-dimensional interest portrait based on the click intention interest vector, the novel click intention interest vector and the purchase intention vector; recall multiple items according to the multi-dimensional interest portrait and the interest coefficient; and screen the multiple items to determine the items to be recommended.

12. An item recommendation device, characterized in that: include: A memory for storing executable instructions; A processor is used to implement the item recommendation method described in any one of claims 1 to 9 when executing the executable instructions stored in the memory.

13. A computer-readable storage medium, characterized in that: The storage medium stores executable instructions, which, when executed, are used to cause the processor to execute the item recommendation method according to any one of claims 1 to 9.

Images