Feature Recognition and Data Recommendation Method, Device and Storage Medium for Target Object

Abstract

Feature recognition method and data recommendation method, device, electronic device and storage medium for target objects, which relate to the Internet and artificial intelligence. The feature recognition method for target objects includes: obtaining a behavior sequence that includes numerical expressions of multiple operation behaviors of a target object and is used to characterize the recommended object involved in the operation behavior, calculating the correlation between the operation behavior and the target recommended object through the correlation between each operation behavior and the category to which the target recommended object belongs and the correlation between the numerical expression of the operation behavior and the numerical expression of the target recommended object, and selecting, based on this, the operation behaviors that meet the preset conditions for the correlation with the target recommended object from the behavior sequence; and on this basis, calculating the degree of interest of the target object in each recommended object and obtaining the features of the target object. It realizes automatically screening out behaviors with a high degree of relevance to the target recommended object for feature recognition and data recommendation. It can reduce the computational amount of feature recognition and data recommendation and improve the effect.

Description

Technical Field

[0001] This application relates to the fields of Internet and artificial intelligence technologies. Specifically, this application relates to a method for identifying features of a target object, a data recommendation method, an apparatus, an electronic device, a computer-readable storage medium, and a computer program product. Background Art

[0002] In Internet applications, in application scenarios where a platform actively displays products, videos, pictures, etc. to a target object, in order to better convert the displayed content into effective operations of the target object, such as being clicked or favorited by the target object, personalized display needs to be performed according to the characteristics and interests of the target object. The related technologies involved in this process are collectively referred to as data recommendation algorithms. One way to understand the characteristics and interests of the target object is to identify the characteristics and interests of the target object based on the historical behavior of the target object.

[0003] In related technologies, an artificial neural network model is used to finally obtain a digital expression of the characteristics and interests of the target object based on the records of the existing behaviors of the target object, and to predict the single item that the target object is most likely to click or favorite and recommend it to the target object according to the relevance between the digital expression of the characteristics and interests of the target object and the digital expression of the target product, video, picture, etc.

[0004] Among them, there are mainly two ways to use the records of the existing behaviors of the target object: one is to use the specified number of behaviors within the recent time range of the target object as training data and perform behavior prediction. In this way, the training data can only reflect the recent characteristics of the target object, but the historical information is lost; the other way is to perform model training and interest prediction based on all the historical behaviors of the target object over a long period of time. In this way, the all historical behaviors usually include extremely long behavior sequences, which bring great difficulties to data processing, and there is a lot of noise information in them, which is also not conducive to accurately predicting the characteristics and interests of the target object. Summary of the Invention

[0005] The purpose of the embodiments of this application aims to solve at least one aspect of the above technical problems to a certain extent, and provides a technical solution for automatically screening behavior sequences and performing feature recognition and data recommendation of target objects based on the screened behavior sequences.

[0006] According to the first aspect of the embodiments of this application, a method for identifying features of a target object is provided. The method includes:

[0007] Obtain the numerical representation of the behavior sequence of the target object, where the numerical representation of the behavior sequence includes the numerical representations of multiple operation behaviors of the target object, and the numerical representation of the operation behavior is used to characterize the recommended object involved in the operation behavior;

[0008] For multiple target recommended objects in the object data set composed of recommended objects, respectively obtain the relative interest degree of the target object in each target recommended object according to the numerical representation of the behavior sequence;

[0009] Obtain the characteristics of the target object according to the relative interest degree of the target object in each target recommended object;

[0010] Among them, obtaining the relative interest degree of the target object in each target recommended object specifically includes:

[0011] According to the numerical representation of the behavior sequence, calculate the correlation between each operation behavior in the behavior sequence and the target recommended object, specifically including: calculating the correlation between the category to which the operation behavior belongs and the category to which the target recommended object belongs as the first component, and calculating the correlation between the numerical representation of the operation behavior and the numerical representation of the target recommended object as the second component, and obtaining the correlation between each operation behavior in the behavior sequence and the target recommended object according to the first component and the second component;

[0012] According to the correlation between each operation behavior in the behavior sequence and the target recommended object, screen out the operation behaviors in the behavior sequence whose correlation with the target recommended object meets the preset conditions as the reference behavior sequence;

[0013] According to the reference behavior sequence, obtain the relative interest degree of the target object in the target recommended object.

[0014] According to the second aspect of the embodiments of the present application, a data recommendation method is provided, and the method includes:

[0015] Obtain a recommended object data set including multiple recommended objects;

[0016] Obtain the historical behavior sequence of the target object;

[0017] According to the target object feature recognition method of the first aspect of the present application, obtain the characteristics of the target object according to the historical behavior sequence of the target object, where the characteristics of the target object are used to describe the relative interest degree of the target object in the recommended objects in the recommended object data set;

[0018] Select one or more recommended objects from the recommended object data set according to the characteristics of the target object and perform data recommendation to the target object.

[0019] According to the third aspect of the embodiments of the present application, a feature recognition device for a target object is provided. The device includes:

[0020] A behavior sequence acquisition module, configured to acquire a numerical expression of the behavior sequence of the target object. The numerical expression of the behavior sequence includes numerical expressions of multiple operation behaviors of the target object, and the numerical expression of the operation behavior is used to characterize the recommended object involved in the operation behavior.

[0021] An interest recognition module, configured to acquire the relative interest degree of the target object in the target recommended object. The interest recognition module specifically includes:

[0022] A relevance calculation unit, configured to calculate the relevance between each operation behavior in the behavior sequence and the target recommended object according to the numerical expression of the behavior sequence. Specifically, it includes: calculating the relevance between the category to which the operation behavior belongs and the category to which the target recommended object belongs as the first component, and calculating the relevance between the numerical expression of the operation behavior and the numerical expression of the target recommended object as the second component, and obtaining the relevance between each operation behavior in the behavior sequence and the target recommended object according to the first component and the second component;

[0023] An operation behavior screening unit, configured to screen out operation behaviors that satisfy a preset condition with the target recommended object from the behavior sequence according to the relevance between each operation behavior in the behavior sequence and the target recommended object, as a reference behavior sequence;

[0024] An interest estimation unit, configured to acquire the relative interest degree of the target object in the target recommended object according to the reference behavior sequence; and

[0025] A feature recognition module, configured to acquire the features of the target object according to the relative interest degree of the target object in each target recommended object.

[0026] According to the fourth aspect of the embodiments of the present application, a data recommendation device is provided. The device includes:

[0027] A historical behavior sequence acquisition module, configured to acquire the historical behavior sequence of the target object;

[0028] A recommended object set acquisition module, configured to acquire a recommended object data set including multiple recommended objects;

[0029] A feature recognition module, configured to acquire the features of the target object according to the feature recognition method of the target object in the first aspect of the embodiments of the present application, where the features of the target object are used to describe the relative interest degree of the target object in the recommended objects in the recommended object data set; and

[0030] The data recommendation module selects one or more recommended objects from the recommended object data set according to the characteristics of the target object and recommends the data to the target object.

[0031] According to the fifth aspect of the embodiments of the present application, an electronic device is provided. The electronic device includes: a memory, a processor, and a computer program stored on the memory. When the computer program is executed by the processor, it implements the steps of the automatic screening method of the behavior sequence according to the embodiments of the first aspect of the present application, or when the computer program is executed by the processor, it implements the steps of the data recommendation method according to the embodiments of the second aspect of the present application.

[0032] According to the sixth aspect of the embodiments of the present application, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by the processor, it implements the steps of the method for identifying the characteristics of the target object according to the embodiments of the first aspect of the present application, or when the computer program is executed by the processor, it implements the steps of the data recommendation method according to the embodiments of the second aspect of the present application.

[0033] According to the seventh aspect of the embodiments of the present application, a computer program product includes a computer program. When the computer program is executed by the processor, it implements the steps of the method for identifying the characteristics of the target object according to the embodiments of the first aspect of the present application, or when the computer program is executed by the processor, it implements the steps of the data recommendation method according to the embodiments of the second aspect of the present application.

[0034] The beneficial effects brought by the technical solutions provided by the embodiments of the present application include:

[0035] For the series of technical solutions for identifying the characteristics of the target object of the present application, by comprehensively using a variety of relevance evaluation methods, the relevance between each operation behavior in the historical behavior sequence of the target object and the target recommended object is calculated. The behavior most relevant to the target recommended object can be screened out from the ultra-long behavior sequence as a reference basis for interest recognition. Compared with directly processing the ultra-long historical behavior sequence, the computational amount of subsequent processing can be greatly reduced. At the same time, compared with directly selecting several most recent behaviors from the historical behavior sequence, the selected operation behaviors according to the embodiments of the present application preserve long-term historical information, are more complete in information, and can better represent the true intention of the target object's interest in the target commodity, obtaining better feature recognition results.

[0036] In the technical solutions of the data recommendation series of the present application, by first screening the historical behavior sequence of the target object, the behaviors most relevant to the target recommendation object are selected to form a reference behavior sequence, which is used as the reference basis for interest recognition and data recommendation. Compared with directly using the extremely long historical behavior sequence as the reference behavior sequence, the amount of computation for subsequent processing can be greatly reduced. At the same time, compared with directly selecting several most recent behaviors from the historical behavior sequence as the reference behavior sequence, the operation behaviors selected according to the solution of the embodiment of the present application preserve long-term historical information, have more complete information, and can better represent the true intention of the target object's interest in the target commodity.

[0037] Based on the reference behavior sequence screened by the automatic screening series of technical solutions of the behavior sequence of the present application, better data processing effects can be obtained whether for model training, interest recognition or data recommendation, including but not limited to: less computation amount, faster computation speed, faster model training, better model training results, recommendation results more in line with the interest characteristics of the target object, etc. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings required for the description of the embodiments of the present application.

[0039] Figure 1 It is a schematic structural diagram of a recommendation algorithm for a large-scale object in the related art;

[0040] Figure 2 It is a schematic flowchart of a method for feature recognition of a target object provided by an embodiment of the present application;

[0041] Figure 3 It is a schematic flowchart of a data recommendation method provided by an embodiment of the present application;

[0042] Figure 4 It is a schematic structural diagram of a device for feature recognition of a target object provided by an embodiment of the present application;

[0043] Figure 5 It is a schematic structural diagram of a data recommendation device provided by an embodiment of the present application;

[0044] Figure 6 It is a schematic structural diagram of an electronic device provided by an embodiment of the present application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0045] The following describes the embodiments of the present application with reference to the drawings in the present application. It should be understood that the embodiments described below in conjunction with the drawings are exemplary descriptions for explaining the technical solutions of the embodiments of the present application, and do not constitute limitations on the technical solutions of the embodiments of the present application.

[0046] Those skilled in the art can understand that, unless specifically stated otherwise, the singular forms "a", "an", "the", and "said" used herein may also include plural forms. It should be further understood that the terms "comprise" and "include" used in the embodiments of the present application mean that the corresponding features can be implemented as the presented features, information, data, steps, operations, elements, and / or components, but do not exclude the implementation of other features, information, data, steps, operations, elements, components, and / or their combinations supported by the technical field of the present invention. It should be understood that when we say an element is "connected" or "coupled" to another element, the one element can be directly connected or coupled to the other element, or it can mean that the one element and the other element establish a connection relationship through an intermediate element. In addition, the "connection" or "coupling" used herein may include wireless connection or wireless coupling. The term "and / or" used herein indicates at least one of the items defined by the term, for example, "A and / or B" can be implemented as "A", or implemented as "B", or implemented as "A and B".

[0047] To make the objectives, technical solutions, and advantages of the present application clearer, the following will further describe the embodiments of the present application in detail with reference to the accompanying drawings.

[0048] To more clearly illustrate the technical solutions of the present application, the relevant technical background will be introduced first.

[0049] Artificial Intelligence (AI) is a theory, method, technology, and application system that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use knowledge to obtain the best results. In other words, artificial intelligence is a comprehensive technology in computer science that attempts to understand the essence of intelligence and produce a new intelligent machine that can react in a way similar to human intelligence. Artificial intelligence also studies the design principles and implementation methods of various intelligent machines to enable the machines to have the functions of perception, reasoning, and decision-making.

[0050] Artificial intelligence technology is an interdisciplinary subject involving a wide range of fields, including both hardware-level technologies and software-level technologies. The basic technologies of artificial intelligence generally include technologies such as sensors, dedicated artificial intelligence chips, cloud computing, distributed storage, big data processing technology, operation / interaction systems, and mechatronics. The software technologies of artificial intelligence mainly include several major directions such as computer vision technology, speech processing technology, natural language processing technology, and machine learning / deep learning, autonomous driving, and intelligent transportation.

[0051] Machine learning (ML) is a multidisciplinary field that encompasses probability theory, statistics, approximation theory, convex analysis, and algorithmic complexity theory. It specifically studies how computers can simulate or implement human learning behaviors to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Machine learning is at the core of artificial intelligence and the fundamental way to make computers intelligent. Its applications span all areas of AI. Machine learning and deep learning typically include techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and learning through demonstration.

[0052] Selecting appropriate recommendation objects from large-scale data sets and recommending them to target users is one of the important application aspects of artificial intelligence and machine learning technologies, and is also a key mechanism for utilizing massive data in the contemporary Internet field. The following is a brief introduction to the recommendation methods in related technologies. Figure 1 , Figure 1 This is a schematic diagram of the architecture of a large-scale object recommendation algorithm in related technology. This framework, first proposed by YouTube, includes two main operations: recall and ranking. It has since been widely adopted across the industry, and most current large-scale object recommendation algorithms are based on similar ideas.

[0053] In various application scenarios, such as e-commerce shopping platforms, video platforms, news platforms, and advertising push systems, the platform resources (i.e., the objects that can be recommended to the target users) are enormous, for example, in the millions, tens of millions, or even more. One approach to selecting these objects is to first select a group of objects that are relevant to the target user and their current context based on the target user's historical behavior and current operational context. This process is called recall. The number of recalled objects is significantly reduced compared to the dataset of all objects, for example, from millions or tens of millions to hundreds or thousands. These recalled objects are then sorted according to a specific ranking strategy, and the top-ranked objects are displayed to the target user based on the ranking results. To avoid data traps or to include objects that the target user has not previously expressed interest in, candidate recommended objects are also included in the ranking phase. Throughout the recommendation process, the target user's historical behavior plays a crucial role in both the recall and ranking stages.

[0054] Currently, in the recall stage, a commonly used method in the industry for processing historical behavior sequences is to perform weighted average fusion on the numerical representations of the N objects of the target object's most recent behaviors to generate a vector representation of the target object's behavior sequence, thereby characterizing the features of the target object, and performing recall based on the fused feature vector. However, this method cannot handle the situation where the sequence length is too long in real time. If the behavior cycle of the target object is very long, the sequence length may be several thousand or even tens of thousands. Then, the existing computing resources cannot support the use of this method, and the online recommendation system has very high requirements for the calculation latency of each request. If only the most recent N behaviors are used for feature representation, the long-term behavior history of the target object is ignored, and a lot of information about the target object is lost.

[0055] Moreover, the inventors found that in the direct weighted average based on the target object's behavior sequence in the related art, since the interest and preference of the target object represented by each behavior for the object of this behavior - the recommended object are completely unknown, the selection of weights is more based on factors such as time distance. Therefore, this way of describing the features of the target object cannot accurately reflect the interest of the target object when performing recall.

[0056] Based on the above problems of the prior art, the present application proposes a method for identifying the features of a target object and a data recommendation method, which are used to screen out the behaviors of the target object that are most relevant to the target recommended object, and thereby identify the interest of the target object in the target recommended object and the features of the target object, and finally perform more effective data recommendation.

[0057] The technical solutions of the embodiments of the present application and the technical effects produced by the technical solutions of the present application will be described below by describing several exemplary embodiments. It should be noted that the following embodiments can be referred to, learned from, or combined with each other. For the same terms, similar features, and similar implementation steps in different embodiments, they will not be repeatedly described.

[0058] It should be noted that in the optional embodiments of the present application, for relevant data such as object information (such as the historical behavior sequence of the object, non-behavior features), when the above embodiments of the present application are applied to specific products or technologies, object permission or consent needs to be obtained, and the collection, use, and processing of relevant data need to comply with the relevant laws, regulations, and standards of relevant countries and regions. That is to say, if the embodiments of the present application involve data related to the object, these data need to be obtained with the authorization and consent of the object and comply with the relevant laws, regulations, and standards of the country and region.

[0059] See Figure 2 , Figure 2It is a schematic flowchart of a method for automatically screening behavior sequences provided by an embodiment of the present application. Among them, the method 100 for automatically screening behavior sequences includes the following steps S110 to S130.

[0060] In step S110, obtain the numerical representation of the behavior sequence of the target object. Among them, the numerical representation of the behavior sequence includes the numerical representations of multiple operation behaviors of the target object, and the numerical representation of the operation behavior is used to characterize the recommended object involved in the operation behavior.

[0061] Specifically, when obtaining the behavior sequence of the target object, what is most concerned about is which recommended objects are involved in the operation behaviors in the behavior sequence and what operations are performed on the target recommended object. Therefore, the numerical representation of the operation behavior often includes information about the target recommended object. At the same time, the numerical representation of the operation behavior can also include the type of operation performed on the recommended object, such as browsing, collecting, purchasing, blocking, etc. In some scenarios, such as news recommendation, the operations corresponding to the target object are relatively single, and may only be clicking or not clicking. At this time, the numerical representation of the recommended object can also be directly equivalent to the numerical representation of the operation behavior.

[0062] At the data processing level, specifically, the object space including each recommended object can be vectorized and described to obtain the vector representation of each recommended object in the object space; according to the vector representations of the recommended objects involved in each operation behavior in the behavior sequence, determine the numerical representation of the operation behavior; and obtain the numerical representation of the behavior sequence according to the numerical representations of each operation behavior in the behavior sequence.

[0063] Among them, when the number of recommended objects is large, the dimension of the vector space is high. Simply converting the recommended objects directly into vector representations, such as the item2vec method in the related art, the resulting dimension is too high, the calculation amount is large, and it may bring difficulties to subsequent calculations due to problems such as singular value decomposition of sparse non-full rank matrices. Therefore, the vectors of the recommended objects can be densely represented (embedding), and related calculations are performed according to the vectors after embedding. That is, the vector representations of the recommended objects involved in the operation behavior in the object space are densely vector represented, and the numerical representation of the operation behavior is determined according to the results of the dense vector representation.

[0064] For example, the relevant information of the recommended object can include predefined partitions such as the ID, category, and attributes of the object; and the relevant information of the operation behavior includes the relevant information of the recommended object involved, and optionally, can include the type of operation, etc.

[0065] For example, for a product object, its category may include the type of the product, and its attributes may include the brand, material, etc. of the product. For objects such as videos and news, the category may include the field to which they belong, such as technology, finance, real estate, etc., and the attributes may include the duration of the video, the release time of the news, etc.

[0066] In this application, aspects such as categories and attributes that can be pre-classified based on prior knowledge or can be clustered based on other automated methods (such as artificial neural networks) are collectively referred to as categories. Categories are used to represent aspects in recommended objects that can be determined according to prior or previous knowledge or calculations. Category information can be reflected in the numerical expression of the recommended object, or it may not be reflected in the numerical expression of the recommended object, but instead, a separate database is set up for storage.

[0067] Category information can include one or more aspects, and finally presented in the form of a vector or matrix. It can also be vectorized and expressed and stored after embedding according to needs.

[0068] In step S120, for multiple target recommended objects in the object data set composed of recommended objects, according to the numerical expression of the behavior sequence, the relative degree of interest of the target object in each target recommended object is obtained respectively.

[0069] Specifically, obtaining the relative degree of interest of the target object in each target recommended object specifically includes steps S121 to S123.

[0070] In step S121, according to the numerical expression of the behavior sequence, the relevance between each operation behavior in the behavior sequence and the target recommended object is calculated. Specifically, it includes: calculating the relevance between the category to which the operation behavior belongs and the category to which the target recommended object belongs as the first component, and calculating the relevance between the numerical expression of the operation behavior and the numerical expression of the target recommended object as the second component.

[0071] Whether it is the vectorized representation or the embedding vector of the result after embedding the vectorized representation, the relationship between recommended objects in the real world and the relationship between the numerical expressions of recommended objects are not completely corresponding. The corresponding relationship between the two is hidden in complex mathematical transformations, and it is difficult to summarize an explicitly expressible physical meaning. Moreover, existing computing technologies have no good way to model the numerical expressions of recommended objects.

[0072] The inventors noticed that, within the technical framework of the recommended object vector representation and embedding in the prior art, when characterizing the relevance between an operation behavior and a recommended object solely based on the correlation between the embedding vector of the operation behavior and the embedding vector of the recommended object, in many cases, there will be a deviation between the calculation result and the real world, which will seriously affect the accuracy of the recommendation algorithm. Furthermore, based on the above observations, a method for calculating the relevance between an operation behavior and a recommended object from multiple perspectives is proposed. By weighting the calculation results from different perspectives, compared with calculating from a single perspective, the deviation from the real situation can be probabilistically reduced, and unreasonable numerical results can be excluded.

[0073] Specifically, in the step of calculating the relevance between each operation behavior in the behavior sequence and the target recommended object according to the behavior sequence, calculating the relevance between the category to which the operation behavior belongs and the category to which the target recommended object belongs as the first component may include: obtaining the category to which the recommended object involved in the operation behavior belongs; determining the category to which the operation behavior belongs according to the category to which the recommended object involved in the operation behavior belongs; and comparing the category to which the operation behavior belongs with the category to which the target recommended object belongs, and calculating the relevance between the category to which the operation behavior belongs and the category to which the target recommended object belongs as the first component according to the comparison result.

[0074] Among them, for the category to which the operation behavior belongs and the category to which the target recommended object belongs, a relatively important classification label can be selected for comparison, or multiple classification labels can be selected for comparison, or the label information of various classification attributes, etc. can be sum-coded and mixed to form a unified category code, and then compared, etc. In the embodiments described in combination with mathematical expressions hereinafter, the category selection in the above various cases will be uniformly represented by the symbol C.

[0075] Calculating the relevance between the numerical expression of the operation behavior and the numerical expression of the target recommended object as the second component may include: measuring the relevance between the two according to various distances used to represent the relevance between the numerical expression of the operation behavior and the numerical expression of the target recommended object.

[0076] For example, the weighted inner product between the numerical expression of the operation behavior and the numerical expression of the target recommended object can be used to calculate the relevance between the numerical expression of the operation behavior and the numerical expression of the target recommended object as the second component.

[0077] In some embodiments, the numbers of the recommended objects are numbered continuously according to certain rules, or are numbered in segments according to rules. Numerically, this is reflected in that the correlation (such as similarity) between the vector representations corresponding to the objects is positively correlated with the correlation (similarity) between the recommended objects. That is, the higher the correlation (similarity) between the vector representations corresponding to Object 1 and Object 2, the higher the correlation (similarity) between Object 1 and Object 2. For the preprocessed vector representations, one of the ways to characterize the correlation between vectors is the magnitude of the inner product.

[0078] Of course, those skilled in the art, under the teaching of the present invention, can also use various distance functions commonly used in the art to represent the correlation between two vectors, such as Minkowski distance (such as Euclidean distance, Manhattan distance, Chebyshev distance), standardized Euclidean distance, Mahalanobis distance, cosine of the included angle, Hamming distance, Jaccard distance (or Jaccard similarity coefficient), correlation distance (or correlation coefficient), etc., to measure the correlation between the numerical expression of the operation behavior and the numerical expression of the target recommended object.

[0079] According to the specific application scenario, one or more of the above various distances representing correlation can be selected to calculate the second component.

[0080] Calculating the correlation between the numerical expression of the operation behavior and the numerical expression of the target recommended object as the second component may include using the hash distance. It should be noted that not all hash function distances are used to represent the correlation of vectors. In the present application, a hash function used to represent correlation can be used, such as the Locality-Sensitive Hashing (LSH) distance, to measure the correlation. The hash distance can be used in combination with or independently of the various distances for representing correlation introduced above.

[0081] Perform vector mapping on the numerical expression of the operation behavior according to a preset hash function matrix to obtain the hash vector of the operation behavior; perform vector mapping on the numerical expression of the target recommended object according to the preset hash function matrix to obtain the hash vector of the target recommended object; and obtain the hash distance between the numerical expression of the operation behavior and the numerical expression of the target recommended object according to the hash vector of the operation behavior and the hash vector of the target recommended object.

[0082] After the numerical expressions of the operation behavior and the recommended object are subjected to hash operations, by utilizing the advantages of the hash vector in aspects such as data accuracy identification, data storage size, performance, and retrieval speed, the correlation analysis of the two from another angle can be obtained with a relatively small amount of calculation and high precision.

[0083] It should be noted that since the hash function is essentially probabilistic, multiple input items can share the same hash value, that is, multiple recommended objects may correspond to the same hash vector, which is equivalent to a "classification" based on certain principles to a certain extent. This classification principle is specifically reflected in the expression of the hash function. That is to say, through the parameter selection of the hash function, certain characteristics related to the known real-world classification can be given to the hash distance, thereby introducing more reliable references for the judgment of relevance.

[0084] For example, by setting the form of the hash function matrix, the parameters of the hash function can be determined through data fitting based on the known label data. Another example is that a neural network can be used to implement the hash function, and the neural network can be trained according to the known label data to obtain a hash function neural network representing a certain label principle.

[0085] In step S122, according to the relevance between each operation behavior in the behavior sequence and the target recommended object, the operation behaviors that meet the preset conditions for the relevance with the target recommended object are screened out from the behavior sequence as the reference behavior sequence.

[0086] There are various options for how to use the first component and the second component to measure the relevance between the operation behavior and the target recommended object. For example, according to the judgment requirements and the actual situation of the recommended object, different weights can be assigned to each component and the multiple distances (if any) in the second component, and comprehensive consideration can be made according to different weights.

[0087]

[0088] The following uses a specific embodiment of parametric representation to illustrate the behavior sequence screening method of the present application. In this embodiment, the second component uses a combination of inner product distance and hash distance to measure the relevance.

[0089] First, feature construction is performed. The ID and category information of the recommended object are converted into vectors and densely vectorized (embedding) to obtain the numerical expression of each recommended object. Among them, for multiple aspects in the relevant information of the recommended object, vectorization and embedding can be respectively performed in their respective value spaces, and then the embedded vectors are concatenated to obtain the numerical expression of the recommended object.

[0090] For example, the numerical expression of the recommended object a is denoted as e a ​Let B represent the behavior sequence of the target object. B = [b1,..b i ..b T , where b i represents the recommended object involved in the i-th behavior of the target object u, and the numerical expression of the recommended object involved is used as the numerical expression of the corresponding operation behavior.

[0091] To simplify the description and highlight the main idea of this application, in this embodiment, each operation in the behavior sequence only represents the recommended object involved in the behavior. Those skilled in the art can, according to the needs of the specific application scenario, set the form including the operation category for the operation behavior vector, which will also fall within the protection scope of the claims of this application.

[0092] In the calculation of the first component, obtain the category C of the target recommended object a a and the category C of the recommended object i involved in the operation behavior i , and calculate the first component r according to formula (1) i 1 .

[0093] r i 1 = sign(C i = C a ) (Formula 1)

[0094] where sign(x) is the sign function, and the value is 1 when the target item category is the same as the behavior category, otherwise the value is 0.

[0095]

[0096] In the calculation of the second component, using the numerical expression of the recommended object involved in the operation behavior as the numerical expression of the corresponding operation behavior, each item in B can be represented by the numerical expression of the recommended object, that is, convert B = [b1,..b i ..b T to the vector E = [e1,..e i ..e T .

[0097] In the second component representing the correlation between the operation behavior b i and the target recommended object a, the inner product distance r i 2 is calculated as shown in formula 2:

[0098] r i 2 = (W b * e i ) * (W a * e a )T (Formula 2)

[0099] Wherein, W b and W a are predefined parameter matrices.

[0100] In the second component characterizing the relevance between the operation behavior b i and the target recommended object a, in the calculation of the hash distance r i 3 For the transformed behavior sequence E = [e1,..e i ..e T , assuming that the dimension of each vector is d, a hash function matrix H is set, the size of H is d*m, and according to the hash function matrix H, the e i vector is mapped to a new vector p i . In some embodiments, the functional form of the hash matrix can be selected such that the value of p i is 0 or 1.

[0101] The specific calculation method is as follows:

[0102] for(k≥0and≤m-1):

[0103]

[0104] Where e i [j] represents the j-th dimension of the vector e i , H[j][k] represents the hash function at the corresponding position (the j-th row, the k-th column) in the hash function matrix, and sign is the sign function. For example, the finally obtained new vector p i can be in the following form:

[0105] p i = [0, 1, 1, 0, 0,...]

[0106] For the vector e a corresponding to the target recommended object a, it can be transformed into a new vector p a through a hash operation, specifically as follows:

[0107] for(i≥0and≤m-1):

[0108]

[0109] Calculate the hash distance r i characterizing the relevance between the operation behavior b a and the target recommended object e i 3 , and the calculation formula is as follows:

[0110] r i 3 = cosin(p i *p a )

[0111] Then, the weighted sum of the two distances of the first component and the second component is calculated to obtain the importance score S of each operation behavior with respect to the target recommended object i , and the importance score S i is used to characterize the relevance between the operation behavior and the recommended object. For the target recommended object a, the importance score S i of the operation behavior b i can be expressed as:

[0112] S i = W1 * r i 1 + W2 * r i 2 + W3 * r i 3

[0113] where W1, W2, and W3 are predefined weight coefficients.

[0114] According to the importance score, the operation behaviors that meet the preset conditions in terms of relevance to the target recommended object are screened out from the behavior sequence. It can be to select the specified number of behaviors with the highest scores according to the importance score, or to select the behaviors greater than the predefined threshold, etc. Those skilled in the art can select specific parameters according to the specific application requirements under the guidance of this application.

[0115] Finally, according to the importance score S i , s operation behaviors with a relatively high degree of relevance to the target recommended object are obtained from the behavior sequence E of the target object u, constituting the reference behavior sequence Q, Q = [e i ... e s .[[]END]

[0116] The obtained reference behavior sequence Q can be used for subsequent operations such as interest recognition, feature extraction, and data recommendation of the target object.

[0117] In step S123, according to the reference behavior sequence, the relative degree of interest of the target object in the target recommended object is obtained.

[0118] This step can be specifically implemented as follows: First, obtain the numerical representation of the reference behavior sequence; input the numerical representation of the reference behavior sequence and the numerical representation of the target recommended object into the pre-trained self-attention neural network, and obtain the interest recognition result of the target object output by the self-attention neural network; and according to the interest recognition result, obtain the interest score of the target object for the target recommended object to represent the relative interest degree of the target object in the target recommended object.

[0119] Among them, the self-attention neural network is trained to calculate the interest recognition result of the target object according to the input numerical representation of the reference behavior sequence and the numerical representation of the target recommended object, based on the multi-head self-attention mechanism.

[0120] Estimating the interest score of the target object for the target recommended object according to the reference behavior sequence of the target object can realize the interest recognition of the target object through the pre-trained neural network, and then further predict the interest score according to the result of the interest recognition. Various neural network structures in the related art can be used to perform interest recognition according to the reference behavior sequence, such as single-head attention, multi-head self-attention, Max Pool (maximum pooling), Avg Pool (average pooling), CNN (convolutional neural network), etc. Those skilled in the art can make a choice under the guidance of the spirit of this application according to specific scenario requirements without exceeding the protection scope of this application.

[0121] In some embodiments, estimating the interest score of the target object for the target recommended object according to the reference behavior sequence may include: obtaining the numerical representation of the reference behavior sequence; and inputting the numerical representation of the reference behavior sequence and the numerical representation of the target recommended object into the pre-trained self-attention neural network, and obtaining the interest recognition result of the target object output by the self-attention neural network; according to the interest recognition result, obtaining the interest score of the target object for the target recommended object; among them, the self-attention neural network is trained to calculate the interest recognition result of the target object according to the input numerical representation of the reference behavior sequence and the numerical representation of the target recommended object, based on the multi-head self-attention mechanism.

[0122] The algorithm of the self-attention mechanism is suitable for large-scale parallel computing. For the case where there are many recommended objects in the recommended object space, it can quickly identify the interests of the target object from the behavior sequence of the target object, and is especially suitable for the data processing requirements of large-scale recommendation systems.

[0123] Based on the interest recognition result, obtaining the interest score of the target object for the target recommended object may include: inputting the interest recognition result of the target object, the numerical expression of the non-behavioral features of the target object, and the numerical expression of the features of the target recommended object into the trained interest score prediction neural network, and obtaining the interest score of the target object for the target recommended object output by the interest score prediction neural network; and performing data recommendation according to the interest score of the target object for the target recommended object. The non-behavioral features may include other features except the behavioral features of the target object.

[0124] Among them, the model structure of the interest score prediction neural network can be designed as a multi-layer perceptron MLP to obtain a faster operation speed. The model parameters are obtained through supervised training according to the historical behavior sequence or reference behavior sequence of the target object, and the interest score prediction neural network is trained to output a relatively high interest score when the historical behavior sequence or reference behavior sequence of the target object includes an operation behavior related to the target recommended object; when the historical behavior sequence or reference behavior sequence of the target object does not include an operation behavior related to the target recommended object, a relatively low interest score is output.

[0125] The interest score of the target object for the recommended objects in the recommended object dataset can be used to generate an interest description of the target object for the recommended object, and for predicting the operation probability for a specific recommended object, etc. Specifically, in the entire recommendation algorithm, those skilled in the art can use the interest score data of the target object for the target recommended object in the recall step of the recommended object and the sorting step of the recalled objects according to the usage requirements.

[0126] The following uses specific embodiments represented by parameters to illustrate the data recommendation method of the present application, especially the calculation method of the interest score and the model training process.

[0127] Let Q = [e i ...e s represent the reference behavior sequence used to characterize the feature embedding of the target object, and e k represents the embedding of the k-th feature of the target object. According to the reference behavior sequence Q, the interest of the target object can be identified in the following way.

[0128] Denote the interest score of the target object u for the target recommended object v as y uv , then the parameter matrices K and V corresponding to the multi-head attention mechanism can be set to the same value as Q, that is:

[0129] Q = K = V = [e i ...e s

[0130] The self-attention score SA(Q, K, V) calculated according to the multi-head self-attention mechanism is as follows:

[0131]

[0132] where d is the corresponding vector distance, and the multi-head self-attention score can be expressed as:

[0133] E = MultiHead(Q, K, V) = concat(head1,..head n ) * W O

[0134] where W O is a model parameter, MultiHead, and concat are the relevant standard functional forms for self-attention calculation. Those skilled in the art can refer to the algorithms of relevant self-attention mechanisms for specific implementation, which will not be elaborated here. The value E output by the MultiHead function is then output by the MLP layer as Mapping the output of the MLP layer through an activation function to obtain the output. For example, the sigmoid function can be selected. Then, the attention recognition result r output by the multi-head self-attention network u can be obtained through the following formula:

[0135]

[0136]

[0137] The attention recognition result r u , can be used to represent the interest recognition result of the target object. The interest recognition result can be further used to input an interest score prediction neural network to obtain the interest score. For example, the interest score prediction neural network can be implemented by a multi-layer perceptron MLP. For example, in some embodiments, a three-layer neural network can be used. The interest score y of the target object u for the target recommended object v uv can be expressed as:

[0138] y uv = f θ (r u , u, v)

[0139] where f θ is the prediction function, θ is the model parameter, u can include non-behavioral features such as the basic attribute features of the target object. For example, u can include features for identifying the target object, and v includes the features of the target recommended object.

[0140] The loss function L for model training can be designed as follows:

[0141] L(y uv , f θ (r u , u, v)) = y uv * log(f θ (r u , u, v)) + (1 - y uv ) * log(1 - f θ (r u , u, v))

[0142] In the model training stage, the value of y uv can be designed according to whether an operation behavior has occurred on the target recommendation object v by the target object u. For example, if an operation behavior has occurred, the interest score can take a relatively large value. For example, y uv = 1; otherwise, the interest score takes a relatively small value, for example, y uv = 0. It should be noted that the values 1 and 0 are just one way of taking values, and obviously it is not limited to this. The training dataset can be from the filtered behavior sequence Q or from the original behavior sequence data B, and those skilled in the art can make a choice according to different training requirements.

[0143] According to the loss function L(y uv , f θ (r u , u, v)), the model parameters are updated using stochastic gradient descent until convergence. Finally, the parameters of the interest score prediction model are obtained.

[0144] In step S130, the features of the target object are obtained according to the relative interest degree of the target object in each target recommendation object.

[0145] In the related art, the vector embedding of the behavior sequence of the target object is weighted and averaged as the embedding vector of the features of the target object. In the embodiments of the present application, due to the estimation of the interest scores of the target object for each recommendation object, the interest scores of the target object for each recommendation object can be comprehensively considered. For example, in the recall stage of the recommendation algorithm, the weighted average of the numerical expressions of each recommendation object, such as the embedding vector, can be used to describe the comprehensive features of the target object. And further, the embedding vector of the comprehensive features of the target object is used for the operations in the recall stage.

[0146] The feature recognition series of technical solutions for the target object of the present application calculates the relevance between each operation behavior in the historical behavior sequence of the target object and the target recommended object by comprehensively using various relevance evaluation methods. It can screen out the behavior most relevant to the target recommended object from the ultra-long behavior sequence as the reference basis for interest recognition. Compared with directly processing the ultra-long historical behavior sequence, it can greatly reduce the computational workload of subsequent processing. At the same time, compared with directly selecting several most recent behaviors from the historical behavior sequence, the operation behaviors selected according to the solution of the embodiment of the present application preserve long-term historical information, have more complete information, and can better represent the true intention of the target object's interest in the target commodity, thus obtaining better feature recognition results.

[0147] The following combines specific implementation manners to illustrate the subsequent relevant recommendation algorithms for the recognized target object features. Refer to Figure 3 , Figure 3 which is a schematic flowchart of a data recommendation method provided by an embodiment of the present application. Among them, the data recommendation method 200 may include the following steps S210 to S240.

[0148] In step S210, a recommended object data set including multiple recommended objects is obtained. According to different usage scenarios, the recommended objects may be various commodities on a shopping website, news on a news website, short videos on a video website, or works such as movies and TV dramas, etc. The recommended object data set may be the sum of all resources on an operation platform website, etc., or a part of all resources. In most cases, the platform or website can select the required part from all resources to form the recommended object data set.

[0149] In step S220, the historical behavior sequence of the target object is obtained. In this step, regarding the historical behavior sequence, it may include all the historical operation behaviors of the target object in the past, or it may be the historical operation behaviors of the target object within a predefined sufficient long time in the past, the purpose of which is to reflect the long-term behavior characteristics and interests of the target object. For example, the historical behavior sequence may include the historical operation behaviors in the past few years or months, rather than necessarily all the behaviors since the creation of the target object.

[0150] In step S230, according to the feature recognition method of the target object according to the embodiments of the present invention, the features of the target object are obtained based on the historical behavior sequence of the target object, where the features of the target object are used to describe the relative degree of interest of the target object in the recommended objects in the recommended object data set. The operation behaviors whose relevance to the target recommended object in the historical behavior sequence meets the preset conditions are filtered out as the reference behavior sequence of the target object regarding the target recommended object. This step can be implemented using the feature recognition method of the target object provided according to the embodiments of the present application.

[0151] For its specific implementation manner and beneficial effects, reference can be made to the foregoing description of the feature recognition method of the target object, which will not be elaborated here.

[0152] In step S240, one or more recommended objects are selected from the recommended object data set according to the features of the target object and data is recommended to the target object.

[0153] Refer to Figure 1 the recommended framework shown. Selecting one or more recommended objects from the recommended object data set according to the features of the target object and recommending data to the target object can be specifically reflected in the recall and ranking stages.

[0154] In the recall stage, according to the features of the target object, for example, according to the interest scores of the target object for each recommended object, the embedding vectors of the recommended objects are weighted and averaged to obtain the comprehensive features of the target object, and this comprehensive feature is used for recall. For example, the method in the related art can be used to implement the recall through the inner product operation between the embedding vector of the comprehensive feature of the target object and the embedding vector of the recommended object, so as to obtain a subset of recommended objects with a relatively high degree of relevance to the comprehensive feature of the target object. In the ranking stage, in addition to using the embedding vector of the comprehensive feature of the target object, the interest score information of the target object for the recommended object can be further directly used to select the recommended objects with high interest scores from the recalled subset of recommended objects with a relatively high probability for data recommendation.

[0155] It should be noted that the above description of the usage method of the features of the target object is only exemplary. In the related art, there are already various different strategies for data recommendation, and the feature description of the target object will be used in these strategies. The technical solutions formed by applying the feature recognition method of the target object of the present application and its recognized results to these recommendation strategies will also fall within the protection scope of the present application.

[0156] According to the data recommendation series technical solutions of the embodiments of the present application, by first screening the historical behavior sequence of the target object, the behaviors most relevant to the target recommendation object are selected to form a reference behavior sequence, which is used as the reference basis for interest recognition and data recommendation. Compared with directly using the ultra-long historical behavior sequence as the reference behavior sequence, the computational amount of later processing can be greatly reduced. At the same time, compared with directly selecting several most recent behaviors from the historical behavior sequence as the reference behavior sequence, the operation behaviors selected according to the solutions of the embodiments of the present application preserve long-term historical information, are more complete in information, and can better represent the true intention of the target object's interest in the target recommendation object.

[0157] Based on the reference behavior sequence screened by the behavior sequence automatic screening series technical solutions of the present application, better data processing effects can be obtained whether for model training, interest recognition or data recommendation, including but not limited to: less computational amount, faster computational speed, faster model training, better model training results, recommendation results more in line with the interest characteristics of the target object, etc.

[0158] The embodiments of the present application also provide a behavior sequence automatic screening device. Refer to Figure 4 , Figure 4 which is a structural schematic diagram of a behavior sequence automatic screening device provided by the embodiments of the present application. The behavior sequence automatic screening device 100 may include: a behavior sequence acquisition module 110, a correlation calculation module 120, and an operation behavior screening module 130.

[0159] The behavior sequence acquisition module 110 is used to acquire the numerical expression of the behavior sequence of the target object, where the numerical expression of the behavior sequence includes the numerical expressions of multiple operation behaviors of the target object, and the numerical expression of the operation behavior is used to characterize the recommendation object involved in the operation behavior.

[0160] The interest recognition module 120 is used to acquire the relative degree of interest of the target object in the target recommendation object. The interest recognition module specifically includes:

[0161] The correlation calculation unit 121 is used to calculate the correlation between each operation behavior in the behavior sequence and the target recommendation object according to the numerical expression of the behavior sequence, specifically including: calculating the correlation between the category to which the operation behavior belongs and the category to which the target recommendation object belongs as the first component, and calculating the correlation between the numerical expression of the operation behavior and the numerical expression of the target recommendation object as the second component, and obtaining the correlation between each operation behavior in the behavior sequence and the target recommendation object according to the first component and the second component;

[0162] An operation behavior screening unit 122 is configured to screen out operation behaviors that meet a preset condition in terms of the relevance to a target recommended object from the behavior sequence according to the relevance of each operation behavior in the behavior sequence to the target recommended object, and use them as a reference behavior sequence.

[0163] An interest estimation unit 123 is configured to obtain the relative interest degree of a target object in a target recommended object according to the reference behavior sequence.

[0164] A feature recognition module 130 is configured to obtain the features of a target object according to the relative interest degree of the target object in each target recommended object.

[0165] In some embodiments, the behavior sequence acquisition module 110 acquires a numerical representation of the behavior sequence of a target object, which may include: respectively obtaining vector representations of each recommended object in the object data set; and determining the numerical representation of an operation behavior according to the vector representations of the recommended objects involved in each operation behavior in the behavior sequence.

[0166] In some embodiments, determining the numerical representation of an operation behavior according to the vector representations of the recommended objects involved in each operation behavior in the behavior sequence includes: performing a dense vector representation on the vector representations of the recommended objects involved in the operation behavior, and determining the numerical representation of the operation behavior according to the dense vector representations of the recommended objects involved in the operation behavior.

[0167] In some embodiments, a relevance calculation unit 121 is configured to calculate the relevance between each operation behavior in the behavior sequence and a target recommended object according to the numerical representation of the behavior sequence, and calculate the relevance between the category to which the operation behavior belongs and the category to which the target recommended object belongs as a first component. Specifically, it includes: obtaining the category to which the recommended object involved in the operation behavior belongs; determining the category to which the operation behavior belongs according to the category to which the recommended object involved in the operation behavior belongs; and comparing the category to which the operation behavior belongs with the category to which the target recommended object belongs, and calculating the relevance between the category to which the operation behavior belongs and the category to which the target recommended object belongs as the first component.

[0168] In some embodiments, the correlation calculation unit 121 calculates the correlation between the numerical representation of the operation behavior and the numerical representation of the target recommended object as the second component, including calculating one or more of the following distances, and obtaining the correlation between the numerical representation of the operation behavior and the numerical representation of the target recommended object as the second component according to the calculated distance. The distances include: the weighted inner product distance between the numerical representation of the operation behavior and the numerical representation of the target recommended object; the Euclidean distance between the numerical representation of the operation behavior and the numerical representation of the target recommended object; the Manhattan distance between the numerical representation of the operation behavior and the numerical representation of the target recommended object; the Chebyshev distance between the numerical representation of the operation behavior and the numerical representation of the target recommended object; the standardized Euclidean distance between the numerical representation of the operation behavior and the numerical representation of the target recommended object; the Mahalanobis distance between the numerical representation of the operation behavior and the numerical representation of the target recommended object; the cosine of the angle between the numerical representation of the operation behavior and the numerical representation of the target recommended object; the Hamming distance between the numerical representation of the operation behavior and the numerical representation of the target recommended object; the Jaccard distance between the numerical representation of the operation behavior and the numerical representation of the target recommended object; the correlation distance between the numerical representation of the operation behavior and the numerical representation of the target recommended object.

[0169] In some embodiments, the correlation calculation unit 121 calculates the correlation between the numerical representation of the operation behavior and the numerical representation of the target recommended object as the second component, including: performing vector mapping on the numerical representation of the operation behavior according to a preset hash function matrix to obtain a hash vector of the operation behavior; performing vector mapping on the numerical representation of the target recommended object according to the preset hash function matrix to obtain a hash vector of the target recommended object; and obtaining a hash distance for calculating the correlation between the numerical representation of the operation behavior and the numerical representation of the target recommended object according to the hash vector of the operation behavior and the hash vector of the target recommended object; calculating the correlation between the numerical representation of the operation behavior and the numerical representation of the target recommended object as the second component according to the hash distance.

[0170] In some embodiments, the operation behavior screening unit 122 screens out operation behaviors that meet the preset conditions for the correlation with the target recommended object from the behavior sequence according to the correlation between each operation behavior in the behavior sequence and the target recommended object, specifically including: performing weighted summation according to the first component and the second component to obtain the importance score of each operation behavior relative to the target recommended object; and screening out operation behaviors that meet the preset conditions for the correlation with the target recommended object from the behavior sequence according to the importance score.

[0171] In some embodiments, the interest estimation unit 123 obtains the relative interest degree of the target object in the target recommended object according to the reference behavior sequence, including: obtaining the numerical expression of the reference behavior sequence; inputting the numerical expression of the reference behavior sequence and the numerical expression of the target recommended object into the trained self-attention neural network, and obtaining the interest recognition result of the target object output by the self-attention neural network; and obtaining the interest score of the target object for the target recommended object according to the interest recognition result to represent the relative interest degree of the target object in the target recommended object. Among them, the self-attention neural network is trained to calculate the interest recognition result of the target object according to the input numerical expressions of the reference behavior sequence and the target recommended object based on the multi-head self-attention mechanism.

[0172] In some embodiments, the interest estimation unit 123 obtains the interest score of the target object for the target recommended object according to the interest recognition result to represent the relative interest degree of the target object in the target recommended object, including: inputting the interest recognition result of the target object, the numerical expression of the non-behavioral features outside the behavior sequence of the target object, and the numerical expression of the target recommended object into the trained interest score prediction neural network, and obtaining the interest score of the target object for the target recommended object output by the interest score prediction neural network; where the model structure of the interest score prediction neural network is a multi-layer perceptron, and the model parameters are obtained through supervised training according to the historical behavior sequence or the reference behavior sequence of the target object, and are trained to: when the historical behavior sequence or the reference behavior sequence of the target object includes an operation behavior involving the target recommended object, output a relatively high interest score; when the historical behavior sequence or the reference behavior sequence of the target object does not include an operation behavior involving the target recommended object, output a relatively low interest score.

[0173] In some embodiments, the feature recognition module 130 obtains the features of the target object according to the relative interest degree of the target object in each target recommended object, including: obtaining the numerical expressions of the respective target recommended objects; setting the interest weights of the respective target recommended objects according to the relative interest degree of the target object in each target recommended object, and performing weighted averaging on the numerical expressions of the respective target recommended objects according to the interest weights; and obtaining the features of the target object according to the weighted average of the numerical expressions of the respective target recommended objects.

[0174] The embodiments of the present application also provide a data recommendation device. Refer to Figure 5 , Figure 5 which is a schematic structural diagram of a data recommendation device provided by an embodiment of the present application. The data recommendation device 200 includes a historical behavior sequence acquisition module 210, a recommended object set acquisition module 220, a feature recognition module 230, and a data recommendation module 240.

[0175] A historical behavior sequence acquisition module 210 is configured to acquire the historical behavior sequence of a target object.

[0176] A recommended object set acquisition module 220 is configured to acquire a recommended object data set including multiple recommended objects.

[0177] A feature recognition module 230 is configured to acquire the features of the target object according to the feature recognition method of the target object in the first aspect embodiment of the present application, where the features of the target object are used to describe the relative interest degree of the target object in the recommended objects in the recommended object data set.

[0178] A data recommendation module 240 selects one or more recommended objects from the recommended object data set according to the features of the target object and performs data recommendation to the target object.

[0179] It should be noted that the device in the embodiment of the present application can execute the method provided in the embodiment of the present application, and its implementation principle is similar. The actions performed by each module in the device in the embodiments of the present application correspond to the steps in the method in the embodiments of the present application. For the detailed function description and technical effects of each module of the device, reference can specifically be made to the description in the corresponding method shown in the foregoing text, and details are not described herein again.

[0180] An electronic device is provided in an embodiment of the present application, including a memory, a processor, and a computer program stored on the memory. When the computer program is executed by the processor, it implements the steps of the feature recognition method of the target object according to the embodiment of the present application, or when the computer program is executed by the processor, it implements the steps of the data recommendation method according to the embodiment of the present application.

[0181] Compared with the related art, it can be achieved that:

[0182] For the series of technical solutions for feature recognition of the target object in the present application, by comprehensively using a variety of relevance evaluation methods, the relevance between each operation behavior in the historical behavior sequence of the target object and the target recommended object is calculated, and the behavior most relevant to the target recommended object can be screened out from the ultra-long behavior sequence as a reference basis for interest recognition and data recommendation. Compared with directly processing the ultra-long historical behavior sequence, the amount of calculation for later processing can be greatly reduced. At the same time, compared with directly selecting several most recent behaviors from the historical behavior sequence, the operation behaviors selected according to the solution in the embodiment of the present application, because they preserve the historical information for a long time, have more complete information and can better represent the true intention of the target object's interest in the target commodity.

[0183] In the data recommendation series technical solutions of this application, by first screening the historical behavior sequence of the target object, the behaviors most relevant to the target recommendation object are selected to form a reference behavior sequence, which serves as the reference basis for interest recognition and data recommendation. Compared with directly using the extremely long historical behavior sequence as the reference behavior sequence, the computational workload of subsequent processing can be greatly reduced. At the same time, compared with directly selecting several of the most recent behaviors from the historical behavior sequence as the reference behavior sequence, the operation behaviors selected according to the solution of the embodiment of this application save long-term historical information, have more complete information, and can better represent the true intention of the target object's interest in the target commodity.

[0184] Based on the reference behavior sequence screened by the behavior sequence automatic screening series technical solutions of this application, better data processing effects can be obtained whether for model training, interest recognition, or data recommendation, including but not limited to: less computational workload, faster computational speed, faster model training, better model training results, recommendation results more in line with the interest characteristics of the target object, etc.

[0185] In an optional embodiment, an electronic device is provided, as Figure 6 shown. The electronic device 4000 includes: a processor 4001 and a memory 4003. Among them, the processor 4001 and the memory 4003 are connected, such as through a bus 4002. Optionally, the electronic device 4000 may further include a transceiver 4004, and the transceiver 4004 can be used for data interaction between this electronic device and other electronic devices, such as data sending and / or data receiving, etc. It should be noted that in actual applications, the transceiver 4004 is not limited to one, and the structure of the electronic device 4000 does not constitute a limitation to the embodiments of this application.

[0186] The processor 4001 can be a CPU (Central Processing Unit, central processor), a general-purpose processor, a DSP (Digital Signal Processor, data signal processor), an ASIC (Application Specific Integrated Circuit, application-specific integrated circuit), an FPGA (Field Programmable Gate Array, field programmable gate array), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. It can implement or execute various exemplary logic blocks, modules, and circuits described in combination with the disclosure of this application. The processor 4001 can also be a combination that realizes computing functions, such as a combination including one or more microprocessors, a combination of a DSP and a microprocessor, etc.

[0187] The bus 4002 may include a path for transmitting information between the above components. The bus 4002 may be a PCI (Peripheral Component Interconnect) bus, an EISA (Extended Industry Standard Architecture) bus, or the like. The bus 4002 may be divided into an address bus, a data bus, a control bus, etc. For the sake of convenience of representation, Figure 6 it is only represented by a thick line in the figure, but it does not mean that there is only one bus or one type of bus.

[0188] The memory 4003 may be a ROM (Read Only Memory), or other types of static storage devices that can store static information and instructions, a RAM (Random Access Memory), or other types of dynamic storage devices that can store information and instructions. It may also be an EEPROM (Electrically Erasable Programmable Read Only Memory), a CD-ROM (Compact Disc Read Only Memory), or other optical disc storage, optical disc storage (including compact discs, laser discs, optical discs, digital versatile discs, Blu-ray discs, etc.), magnetic disk storage media, other magnetic storage devices, or any other medium that can be used to carry or store computer programs and can be read by a computer, which is not limited herein.

[0189] The memory 4003 is used to store the computer program for implementing the embodiments of the present application and is controlled by the processor 4001 to execute. The processor 4001 is used to execute the computer program stored in the memory 4003 to implement the steps shown in the foregoing method embodiments.

[0190] Among them, the electronic device includes but is not limited to: mobile terminals, computers, servers, cloud-based computing and storage devices, etc.

[0191] The embodiments of the present application provide a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps and corresponding contents of the foregoing method embodiments can be implemented.

[0192] The embodiments of the present application also provide a computer program product, including a computer program. When the computer program is executed by a processor, the steps and corresponding contents of the foregoing method embodiments can be implemented.

[0193] The terms "first", "second", "third", "fourth", "1", "2", etc. (if any) in the description, claims and the above-mentioned drawings of this application are used to distinguish similar objects and do not necessarily describe a specific order or sequence. It should be understood that the data used in this way can be interchanged under appropriate circumstances so that the embodiments of this application described herein can be implemented in an order other than that shown or described in words.

[0194] It should be understood that although the flowchart of the embodiments of this application indicates each operation step by an arrow, the execution order of these steps is not limited to the order indicated by the arrow. Unless there is a clear description in this article, in some implementation scenarios of the embodiments of this application, the implementation steps in each flowchart can be executed in other orders according to requirements. In addition, some or all of the steps in each flowchart may include multiple sub-steps or multiple stages based on the actual implementation scenario. Some or all of these sub-steps or stages can be executed at the same time, and each sub-step or stage in these sub-steps or stages can also be executed at different times respectively. In the scenario where the execution times are different, the execution order of these sub-steps or stages can be flexibly configured according to requirements, and the embodiments of this application do not limit this.

[0195] The above are only optional implementation manners of some implementation scenarios of this application. It should be noted that for those of ordinary skill in the art, without departing from the technical concept of the solution of this application, using other similar implementation means based on the technical idea of this application also belongs to the protection scope of the embodiments of this application.

Claims

1. A method for feature recognition of a target object, characterized in that, Including: Obtaining a numerical representation of the behavior sequence of the target object, where the numerical representation of the behavior sequence includes the numerical representations of multiple operation behaviors of the target object, and the numerical representation of the operation behavior is used to characterize the recommended object involved in the operation behavior; For multiple target recommended objects in the object data set composed of recommended objects, respectively obtaining the relative degree of interest of the target object in each target recommended object according to the numerical representation of the behavior sequence; Obtaining the characteristics of the target object according to the relative degree of interest of the target object in each target recommended object; Among them, obtaining the relative degree of interest of the target object in each target recommended object specifically includes: According to the numerical representation of the behavior sequence, calculating the correlation between each operation behavior in the behavior sequence and the target recommended object, specifically including: calculating the correlation between the category to which the operation behavior belongs and the category to which the target recommended object belongs as the first component, and calculating the correlation between the numerical representation of the operation behavior and the numerical representation of the target recommended object as the second component, and obtaining the correlation between each operation behavior in the behavior sequence and the target recommended object according to the first component and the second component; According to the correlation between each operation behavior in the behavior sequence and the target recommended object, screening out the operation behaviors that meet the preset conditions for the correlation with the target recommended object from the behavior sequence as the reference behavior sequence; According to the reference behavior sequence, obtaining the relative degree of interest of the target object in the target recommended object.

2. The method for identifying the characteristics of the target object according to claim 1, wherein The obtaining of the numerical representation of the behavior sequence of the target object includes: For the recommended objects in the object data set, respectively obtaining the vector representations of each recommended object; Determining the numerical representation of the operation behavior according to the vector representations of the recommended objects involved in each operation behavior in the behavior sequence.

3. The method for identifying the characteristics of the target object according to claim 2, wherein The determining the numerical representation of the operation behavior according to the vector representations of the recommended objects involved in each operation behavior in the behavior sequence includes: Performing a dense vector representation on the vector representations of the recommended objects involved in the operation behavior, and determining the numerical representation of the operation behavior according to the dense vector representations of the recommended objects involved in the operation behavior.

4. The method for identifying the characteristics of the target object according to claim 1, wherein, The calculating the correlation between the category to which the operation behavior belongs and the category to which the target recommended object belongs as the first component specifically includes: Obtaining the category to which the recommended object involved in the operation behavior belongs; Determining the category to which the operation behavior belongs according to the category to which the recommended object involved in the operation behavior belongs; and Comparing the category to which the operation behavior belongs with the category to which the target recommended object belongs, and calculating the correlation between the category to which the operation behavior belongs and the category to which the target recommended object belongs as the first component according to the comparison result.

5. The method for identifying the characteristics of the target object according to claim 1, wherein Calculating the correlation between the numerical representation of the operation behavior and the numerical representation of the target recommended object as the second component, including calculating at least one of the following distances, and obtaining the correlation between the numerical representation of the operation behavior and the numerical representation of the target recommended object as the second component according to the calculated distance, where the distances include: The weighted inner product distance between the numerical representation of the operation behavior and the numerical representation of the target recommended object; The Euclidean distance between the numerical representation of the operation behavior and the numerical representation of the target recommended object; The Manhattan distance between the numerical representation of the operation behavior and the numerical representation of the target recommended object; The Chebyshev distance between the numerical representation of the operation behavior and the numerical representation of the target recommended object; The standardized Euclidean distance between the numerical representation of the operation behavior and the numerical representation of the target recommended object; The Mahalanobis distance between the numerical representation of the operation behavior and the numerical representation of the target recommended object; The cosine of the angle between the numerical representation of the operation behavior and the numerical representation of the target recommended object; The Hamming distance between the numerical representation of the operation behavior and the numerical representation of the target recommended object; The Jaccard distance between the numerical representation of the operation behavior and the numerical representation of the target recommended object; The correlation distance between the numerical representation of the operation behavior and the numerical representation of the target recommended object.

6. The method for identifying the characteristics of the target object according to claim 1 or 5, characterized in that, Calculating the correlation between the numerical representation of the operation behavior and the numerical representation of the target recommended object as the second component includes: Performing vector mapping on the numerical representation of the operation behavior according to a preset hash function matrix to obtain the hash vector of the operation behavior; Performing vector mapping on the numerical representation of the target recommended object according to a preset hash function matrix to obtain the hash vector of the target recommended object; and Obtaining the hash distance between the numerical representation of the operation behavior and the numerical representation of the target recommended object according to the hash vector of the operation behavior and the hash vector of the target recommended object; Calculating the correlation between the numerical representation of the operation behavior and the numerical representation of the target recommended object as the second component according to the hash distance.

7. The method for identifying the characteristics of the target object according to claim 1, wherein, Filtering out the operation behaviors in the behavior sequence whose correlation with the target recommended object meets the preset conditions according to the correlation between each operation behavior in the behavior sequence and the target recommended object, specifically including: Performing weighted summation according to the first component and the second component to obtain the importance score of each operation behavior relative to the target recommended object; and Filtering out the operation behaviors in the behavior sequence whose correlation with the target recommended object meets the preset conditions according to the importance score.

8. The method for identifying the characteristics of the target object according to claim 1, wherein Obtaining the relative degree of interest of the target object in the target recommended object according to the reference behavior sequence, including: Obtaining the numerical representation of the reference behavior sequence; Inputting the numerical representation of the reference behavior sequence and the numerical representation of the target recommended object into a pre-trained self-attention neural network, and obtaining the interest recognition result of the target object output by the self-attention neural network; and Based on the interest recognition result, obtain the interest score of the target object for the target recommended object to represent the relative degree of interest of the target object in the target recommended object; Among them, the self-attention neural network is trained to calculate the interest recognition result of the target object according to the numerical expression of the input reference behavior sequence and the numerical expression of the target recommended object based on the multi-head self-attention mechanism.

9. The method for identifying the characteristics of the target object according to claim 8, wherein, The obtaining the interest score of the target object for the target recommended object based on the interest recognition result to represent the relative degree of interest of the target object in the target recommended object includes: Input the interest recognition result of the target object, the numerical expression of non-behavior features other than the behavior sequence of the target object, and the numerical expression of the target recommended object into the trained interest score prediction neural network, and obtain the interest score of the target object for the target recommended object output by the interest score prediction neural network; Among them, the model structure of the interest score prediction neural network is a multi-layer perceptron, and the model parameters are obtained through supervised training according to the historical behavior sequence or reference behavior sequence of the target object, and are trained to: when the historical behavior sequence or reference behavior sequence of the target object includes an operation behavior related to the target recommended object, output a relatively high interest score; when the historical behavior sequence or reference behavior sequence of the target object does not include an operation behavior related to the target recommended object, output a relatively low interest score.

10. The method for identifying characteristics of a target object according to claim 1, wherein, The obtaining the features of the target object according to the relative degree of interest of the target object in each target recommended object includes: Obtain the numerical expressions of each target recommended object; Set the interest weights of each target recommended object according to the relative degree of interest of the target object in each target recommended object, and perform weighted averaging on the numerical expressions of each target recommended object according to the interest weights; Obtain the features of the target object according to the weighted average of the numerical expressions of each target recommended object.

11. A data recommendation method, characterized in that, including: Obtain a recommended object data set including multiple recommended objects; Obtain the historical behavior sequence of the target object; According to the target object feature recognition method according to any one of claims 1-10, obtain the features of the target object according to the historical behavior sequence of the target object, where the features of the target object are used to describe the relative degree of interest of the target object in the recommended objects in the recommended object data set; Select one or more recommended objects from the recommended object data set according to the features of the target object and perform data recommendation to the target object.

12. A feature recognition device for a target object, characterized in that, including: A behavior sequence acquisition module, configured to obtain the numerical expression of the behavior sequence of the target object, where the numerical expression of the behavior sequence includes the numerical expressions of multiple operation behaviors of the target object, and the numerical expression of the operation behavior is used to characterize the recommended object involved in the operation behavior; An interest recognition module, configured to obtain the relative degree of interest of the target object in the target recommended object; the interest recognition module specifically includes: A correlation calculation unit, configured to calculate the correlation between each operation behavior in the behavior sequence and the target recommended object according to the numerical expression of the behavior sequence, specifically including: calculating the correlation between the category to which the operation behavior belongs and the category to which the target recommended object belongs as a first component, and calculating the correlation between the numerical expression of the operation behavior and the numerical expression of the target recommended object as a second component, and obtaining the correlation between each operation behavior in the behavior sequence and the target recommended object according to the first component and the second component; An operation behavior screening unit, configured to screen out operation behaviors that meet a preset condition with the target recommended object from the behavior sequence as a reference behavior sequence according to the correlation between each operation behavior in the behavior sequence and the target recommended object; An interest estimation unit, configured to obtain the relative interest degree of the target object in the target recommended object according to the reference behavior sequence; and A feature recognition module, configured to obtain the features of the target object according to the relative interest degree of the target object in each target recommended object.

13. An electronic device, comprising a memory, a processor, and a computer program stored on the memory, characterized in that, When the computer program is executed by a processor, it implements the steps of the feature recognition method of the target object described in any one of claims 1-10; or when the computer program is executed by a processor, it implements the steps of the data recommendation method described in claim 11.

14. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the feature recognition method of the target object described in any one of claims 1-10; or when the computer program is executed by a processor, it implements the steps of the data recommendation method described in claim 11.

15. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the feature recognition method of the target object described in any one of claims 1-10; or when the computer program is executed by a processor, it implements the steps of the data recommendation method described in claim 11.

Images