Item recommendation method and system

Abstract

The item recommendation method and system provided in the specification, after obtaining attribute information and user interaction information of a target item, extracting similarity features in the attribute information to obtain explicit similarity features, determining at least one item similar to the target item based on the user interaction information to obtain implicit similarity features, and determining initialization parameters of a recommendation model corresponding to the target item based on the explicit similarity features and the implicit similarity features, and using the recommendation model to recommend the target item based on the initialization parameters; this scheme can improve the accuracy of item recommendation under cold start.

Description

Technical Field

[0001] This manual relates to the field of item recommendation, and in particular to an item recommendation method and system. Background Technology

[0002] In recent years, with the rapid development of internet technology, deep learning has been increasingly widely applied in the field of item recommendation. Item recommendation methods using deep learning often have high data requirements. When items added to the candidate set have little or no historical data, overfitting can occur, affecting recommendation performance; this is known as the cold start problem. To address the cold start problem, existing item recommendation methods often employ content-based approaches.

[0003] In the process of researching and practicing existing technologies, the inventors of this invention discovered that content-based methods only enhance the representation of cold-start items by using item attributes, resulting in a relatively simple representation of cold-start items. This approach cannot accurately represent the similarity between the cold-start item and other items, thus leading to a low accuracy rate in item recommendations. Summary of the Invention

[0004] This manual provides a more accurate method and system for recommending items.

[0005] Firstly, this specification provides an item recommendation method, comprising: acquiring attribute information and user interaction information of a target item, wherein the target item includes items whose number of historical recommendation data is less than a preset threshold; extracting similarity features from the attribute information to obtain explicit similarity features, and determining at least one item similar to the target item based on the user interaction information to obtain implicit similarity features, wherein the similarity features include feature information characterizing the similarity between the target item and other items; and determining initialization parameters of a recommendation model corresponding to the target item based on the explicit similarity features and the implicit similarity features, and recommending the target item using the recommendation model based on the initialization parameters.

[0006] In some embodiments, the step of extracting similarity features from the attribute information to obtain explicit similarity features includes: classifying the attribute information to obtain target attribute information corresponding to each type; extracting similarity features from the target attribute information to obtain initial explicit similarity features corresponding to each type; and concatenating the initial explicit similarity features corresponding to each type to obtain explicit similarity features of the target item.

[0007] In some embodiments, the step of extracting similarity features from the target attribute information to obtain initial explicit similarity features corresponding to each type includes: extracting similarity features from the target attribute information to obtain multiple discrete similarity features; and performing summation pooling on the multiple discrete similarity features to obtain initial explicit similarity features corresponding to each type.

[0008] In some embodiments, determining at least one item similar to the target item based on the user interaction information to obtain implicit similarity features includes: determining a set of candidate items corresponding to the target item based on the user interaction information; selecting at least one item similar to the target item from the set of candidate items to obtain a set of similar items; and extracting features from the items in the set of similar items to obtain implicit similarity features of the target item.

[0009] In some embodiments, determining the candidate item set corresponding to the target item based on the user interaction information includes: extracting at least one user who interacts with the target item from the user interaction information to obtain a user set; obtaining historical item interaction records for each user in the user set; and identifying items other than the target item from the historical item interaction records to obtain a candidate item set.

[0010] In some embodiments, selecting at least one item similar to the target item from the candidate item set to obtain a similar item set includes: determining the user interaction weight of each user in the user set based on the historical item interaction records; calculating the item similarity between the target item and each item in the candidate item set based on the user interaction weight; and selecting at least one item similar to the target item from the candidate item set based on the item similarity to obtain the similar item set.

[0011] In some embodiments, determining the user interaction weight of each user in the user set based on the historical item interaction records includes: counting the number of historical item interactions for each user in the user set in the historical item interaction records; and determining the user interaction weight of each user based on the number of historical item interactions, wherein the user interaction weight is inversely proportional to the number of historical item interactions.

[0012] In some embodiments, an item is selected as a candidate item from the candidate item set, and at least one user pair that interacts with the candidate item is selected from the user set; based on the historical item interaction records, the number of historical interactions of the items jointly interacted by each user pair in the at least one user pair is determined; based on the number of historical interactions and the user interaction weight, the item similarity between the target item and the candidate item is calculated; and the step of selecting an item as a candidate item from the candidate item set is returned to be executed until all items in the candidate item set are selected, so as to obtain the item similarity between the target item and each item in the candidate item set.

[0013] In some embodiments, calculating the item similarity between the target item and the candidate item based on the number of historical interactions and the user interaction weights includes: obtaining a smoothing coefficient corresponding to the target item, and smoothing the number of historical interactions based on the smoothing coefficient to obtain an initial item similarity for each user pair; selecting a target user interaction weight for each user pair from the user interaction weights, and weighting the initial item similarity based on the target user interaction weights; and summing the weighted initial item similarities to obtain the item similarity between the target item and the candidate item.

[0014] In some embodiments, selecting at least one item similar to the target item from the candidate item set to obtain a similar item set includes: determining the interaction relationship between each user in the user set and each item in the candidate item set based on the historical item interaction records; constructing an item interaction graph network with each user and each item as nodes, the item interaction graph network representing the interaction relationship between users and items; and using the item interaction graph network to select at least one item similar to the target item from the candidate item set to obtain a similar item set.

[0015] In some embodiments, the step of extracting features from items in the set of similar items to obtain implicit similarity features of the target item includes: obtaining item information for each item in the set of similar items; extracting features from the item information to obtain item features for each item; and performing average pooling on the item features to obtain implicit similarity features of the target item.

[0016] In some embodiments, determining the initialization parameters of the recommendation model corresponding to the target item based on the explicit similarity features and the implicit similarity features includes: concatenating the explicit similarity features and the implicit similarity features to obtain the target similarity features of the target item; clustering the target item based on the target similarity features to obtain the category-sharing features of the target item; and determining the initialization parameters of the recommendation model corresponding to the target item based on the target similarity features and the category-sharing features.

[0017] In some embodiments, clustering the target items based on the target similarity features to obtain the category-sharing features of the target items includes: obtaining the cluster center features of each preset task cluster; determining the allocation probability of the recommended task corresponding to the target item being assigned to each preset task cluster based on the target similarity features; and fusing the allocation probability with the corresponding cluster center features to obtain the category-sharing features of the target items.

[0018] In some embodiments, determining the allocation probability of the recommended task corresponding to the target item to each preset task cluster based on the target similarity feature includes: determining the allocation value of the recommended task corresponding to the target item to each preset task cluster using a task allocation network based on the target similarity feature; and normalizing the allocation value of each preset task cluster to obtain the allocation probability corresponding to each preset task cluster.

[0019] In some embodiments, the task allocation network includes a first allocation network layer and a second allocation network layer; and

[0020] The step of determining the allocation value of the recommended task corresponding to the target item to each preset task cluster using a task allocation network based on the target similarity feature includes: using the first allocation network layer to perform cluster allocation of the recommended tasks based on the target similarity feature to obtain the initial allocation value corresponding to each preset task cluster; using the second allocation network layer to weight the initial allocation value to obtain the weighted initial allocation value; multiplying the initial allocation value by a preset correction parameter to obtain a correction value of the weighted initial allocation value; and adding the correction value to the weighted initial allocation value to obtain the allocation value.

[0021] In some embodiments, determining the initialization parameters of the recommendation model corresponding to the target item based on the target similarity features and category sharing features includes: concatenating the target similarity features and the category sharing features to obtain the initialization parameter features corresponding to the target item; using a modulation network to identify the modulation weights corresponding to the target item in the initialization parameter features; and obtaining the current initialization parameters of the recommendation model, and weighting the current initialization parameters based on the modulation weights to obtain the initialization parameters of the recommendation model corresponding to the target item.

[0022] In some embodiments, before obtaining the current initialization parameters of the recommendation model, the method further includes: obtaining sample item information corresponding to the item sample set, and extracting sample attribute features, user features, and context features for each item sample in the item sample set from the sample item information; concatenating the sample attribute features, user features, and context features to obtain the target recommendation features for each item sample; predicting the predicted recommended user for each item sample using a preset recommendation model based on the target recommendation features; and converging the preset recommendation model based on the predicted recommended user to obtain the recommendation model.

[0023] In some embodiments, recommending the target item using the recommendation model based on the initialization parameters includes: initializing the recommendation model based on the initial parameters; predicting the conversion rate of each user in a preset user set for the target item using the initialized recommendation model; and selecting at least one target user in the preset user set based on the conversion rate, and recommending the target item to the at least one target user.

[0024] In some embodiments, the recommendation model is a meta-learning recommendation model.

[0025] Secondly, this specification also provides an item recommendation system, comprising: at least one storage medium storing at least one instruction set for item recommendation; and at least one processor communicatively connected to the at least one storage medium, wherein, when the item recommendation system is running, the at least one processor reads the at least one instruction set and executes the item recommendation method described in the first aspect of this specification according to the instructions of the at least one instruction set.

[0026] As can be seen from the above technical solutions, the item recommendation method and system provided in this specification, after obtaining the attribute information and user interaction information of the target item, extracts similarity features from the attribute information to obtain explicit similarity features, and determines at least one item similar to the target item based on the user interaction information to obtain implicit similarity features. Based on the explicit and implicit similarity features, the initialization parameters of the recommendation model corresponding to the target item are determined, and the recommendation model is used to recommend the target item based on the initialization parameters. Since this solution can not only extract the explicit similarity features of the target item from the attribute information, but also mine the similarity between items in the user interaction information to obtain the implicit similarity features of the target item, it supplements the content needed for the cold start of the target item, making the representation of the target item more accurate. Therefore, it can improve the accuracy of item recommendation.

[0027] Other functions of the article recommendation methods and systems provided in this specification will be partially listed in the following description. The figures and examples described below will be readily apparent to those skilled in the art. The inventive aspects of the article recommendation methods and systems provided in this specification can be fully understood through practice or use of the methods, apparatus, and combinations described in the detailed examples below. Attached Figure Description

[0028] To more clearly illustrate the technical solutions in the embodiments of this specification, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this specification. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0029] Figure 1 A schematic diagram illustrating an application scenario of an item recommendation system provided according to an embodiment of this specification is shown.

[0030] Figure 2 A hardware structure diagram of a computing device provided according to an embodiment of this specification is shown;

[0031] Figure 3 A flowchart illustrating an article recommendation method according to embodiments of this specification is shown; and

[0032] Figure 4 A schematic diagram of the architecture of a task similarity adaptive meta-learning framework provided according to an embodiment of this specification is shown. Detailed Implementation

[0033] The following description provides specific application scenarios and requirements for this specification, intended to enable those skilled in the art to make and use the contents of this specification. Various partial modifications to the disclosed embodiments will be apparent to those skilled in the art, and the general principles defined herein can be applied to other embodiments and applications without departing from the spirit and scope of this specification. Therefore, this specification is not limited to the embodiments shown, but rather to the widest scope consistent with the claims.

[0034] The terminology used herein is for the purpose of describing particular exemplary embodiments only and is not restrictive. For example, unless the context clearly indicates otherwise, the singular forms “a,” “an,” and “the” used herein may also include the plural forms. When used in this specification, the terms “comprising,” “including,” and / or “containing” mean that the associated integers, steps, operations, elements, and / or components are present, but do not exclude the presence of one or more other features, integers, steps, operations, elements, components, and / or groups, or that other features, integers, steps, operations, elements, components, and / or groups may be added to the system / method.

[0035] Considering the following description, these and other features of this specification, as well as the operation and function of the related components of the structure, and the economy of assembly and manufacture of the parts, can be significantly improved. All of these form part of this specification with reference to the accompanying drawings. However, it should be clearly understood that the drawings are for illustrative and descriptive purposes only and are not intended to limit the scope of this specification. It should also be understood that the drawings are not drawn to scale.

[0036] The flowcharts used in this specification illustrate operations implemented according to some embodiments of this specification. It should be clearly understood that the operations in the flowcharts may not be implemented in a sequential order. Instead, the operations may be implemented in reverse order or simultaneously. Furthermore, one or more additional operations may be added to the flowcharts. One or more operations may be removed from the flowcharts.

[0037] For ease of description, the terms that will appear in the following descriptions will be explained as follows:

[0038] Item cold start: When a new item is added to the item candidate pool, there is not enough historical data to accurately recommend it to suitable users.

[0039] Meta-learning, often understood as "learning-to-learn," refers to the process of improving a learning algorithm across multiple learning stages. In the basic learning process, the internal (or lower-level / basic) learning algorithm solves the task defined by the dataset and the objective. In the meta-learning process, the external (or upper-level / meta) algorithm updates the internal learning algorithm, causing its learned model to improve upon the external objective.

[0040] Recommender system: A recommender system is an information filtering system used to predict a user's rating or preference for items.

[0041] Zero sample: A new sample with absolutely no historical data.

[0042] Small sample: A sample with very little historical data, or a sample with less than a preset threshold of historical data.

[0043] Conversion Rate (CRV): A recommendation metric used in recommendation systems to evaluate the effectiveness of recommendations. For example, in an advertising scenario, it can be a metric for measuring the effectiveness of CPA (cost per action), which is simply the conversion rate from a user clicking on an ad to becoming an active, registered, or even paying user.

[0044] AUC (Area Under Curve): The area under the ROC curve and the coordinate axis is widely used in machine learning classification models. It is an evaluation metric for measuring the performance of binary classification models and represents the probability that a predicted positive example is ranked before a negative example.

[0045] Before describing the specific embodiments in this specification, the application scenarios of this specification will be introduced as follows:

[0046] The item recommendation method provided in this manual can be applied to any item recommendation scenario. For example, in e-commerce scenarios, the item recommendation algorithm provided in this manual can be used to recommend at least one product to the user so that the user can purchase it directly. Or, in advertising scenarios, the item recommendation algorithm provided in this manual can be used to recommend advertisements to the potential customer groups corresponding to the advertisements. Or, it can be applied to any other item recommendation scenario, which will not be elaborated here.

[0047] Those skilled in the art should understand that the recommended methods and systems for the items described in this specification are also within the scope of protection of this specification when applied to other use scenarios.

[0048] Figure 1 This diagram illustrates an application scenario of an item recommendation system 001 provided according to an embodiment of this specification. The item recommendation system 001 (hereinafter referred to as System 001) can be applied to item recommendation in any scenario, such as item recommendation in e-commerce scenarios and item recommendation in advertising scenarios, etc. Figure 1 As shown, system 001 may include target user 100, client 200, server 300 and network 400.

[0049] Target user 100 can be the user who triggers the recommendation of the target item. Target user 100 can perform the item recommendation operation in client 200.

[0050] Client 200 can be a device that recommends target items in response to an item recommendation operation by target user 100. In some embodiments, the item recommendation method can be executed on client 200. In this case, client 200 may store data or instructions for executing the item recommendation method described herein, and may execute or be used to execute said data or instructions. In some embodiments, client 200 may include a hardware device with data processing capabilities and the necessary programs required to drive the hardware device. Figure 1 As shown, client 200 can communicate with server 300. In some embodiments, server 300 can communicate with multiple clients 200. In some embodiments, client 200 can interact with server 300 through network 400 to receive or send messages, such as receiving or sending target items or target item attribute information and user interaction information. In some embodiments, client 200 may include mobile devices, tablets, laptops, built-in devices in motor vehicles, or similar content, or any combination thereof. In some embodiments, the mobile device may include smart home devices, smart mobile devices, virtual reality devices, augmented reality devices, or similar devices, or any combination thereof. In some embodiments, the smart home device may include smart TVs, desktop computers, etc., or any combination thereof. In some embodiments, the smart mobile device may include smartphones, personal digital assistants, gaming devices, navigation devices, etc., or any combination thereof. In some embodiments, the virtual reality device or augmented reality device may include virtual reality headsets, virtual reality glasses, virtual reality patches, augmented reality headsets, augmented reality glasses, augmented reality patches, or similar content, or any combination thereof. For example, the virtual reality device or the augmented reality device may include Google Glass, head-mounted displays, VR, etc. In some embodiments, the built-in device in the motor vehicle may include an onboard computer, an onboard television, etc. In some embodiments, the client 200 may be a device with positioning technology for locating the position of the client 200.

[0051] In some embodiments, the client 200 may have one or more applications (APPs) installed. The APPs provide the target user 110 with the ability and interface to interact with the outside world via the network 400. The APPs include, but are not limited to: web browser APPs, search APPs, chat APPs, shopping APPs, video APPs, financial management APPs, instant messaging tools, email clients, social media platform software, etc. In some embodiments, the client 200 may have a target APP installed. The target APP can obtain attribute information and user interaction information of the target item for the client 200. In some embodiments, the target object 100 can also trigger an item recommendation request through the target APP. The target APP can respond to the item recommendation request and execute the item recommendation method described in this specification. The item recommendation method will be described in detail later.

[0052] Server 300 may be a server providing various services, such as a backend server supporting attribute information and user interaction information of target items obtained from client 200. In some embodiments, the item recommendation method can be executed on server 300. In this case, server 300 may store data or instructions for executing the item recommendation method described herein, and may execute or be used to execute said data or instructions. In some embodiments, server 300 may include hardware devices with data processing capabilities and necessary programs to drive the hardware devices. Server 300 can communicate with multiple clients 200 and receive data sent by clients 200.

[0053] Network 400 serves as a medium to provide a communication connection between client 200 and server 300. Network 400 facilitates the exchange of information or data. For example... Figure 1 As shown, client 200 and server 300 can connect to network 400 and transmit information or data to each other through network 400. In some embodiments, network 400 can be any type of wired or wireless network, or a combination thereof. For example, network 400 may include cable networks, wired networks, fiber optic networks, telecommunications networks, intranets, the Internet, local area networks (LANs), wide area networks (WANs), wireless local area networks (WLANs), metropolitan area networks (MANs), public switched telephone networks (PSTNs), and Bluetooth networks. TM ZigBee TMA network, a near-field communication (NFC) network, or a similar network. In some embodiments, network 400 may include one or more network access points. For example, network 400 may include wired or wireless network access points, such as base stations or internet exchange points, through which one or more components of client 200 and server 300 can connect to network 400 to exchange data or information.

[0054] It should be understood that Figure 1 The number of clients 200, servers 300, and networks 400 shown is merely illustrative. Depending on implementation needs, there can be any number of clients 200, servers 300, and networks 400.

[0055] It should be noted that the item recommendation method can be executed entirely on the client 200, entirely on the server 300, or partially on the client 200 and partially on the server 300.

[0056] Figure 2 A hardware structure diagram of a computing device 600 provided according to an embodiment of this specification is shown. The computing device 600 can execute the item recommendation method described in this specification. The item recommendation method is described in other parts of this specification. When the item recommendation method is executed on a client 200, the computing device 600 can be the client 200. When the item recommendation method is executed on a server 300, the computing device 600 can be the server 300. When the item recommendation method can be executed partly on the client 200 and partly on the server 300, the computing device 600 can be both the client 200 and the server 300.

[0057] like Figure 2 As shown, the computing device 600 may include at least one storage medium 630 and at least one processor 620. In some embodiments, the computing device 600 may also include a communication port 650 and an internal communication bus 610. Additionally, the computing device 600 may include I / O components 660.

[0058] The internal communication bus 610 can connect different system components, including storage medium 630, processor 620 and communication port 650.

[0059] I / O component 660 supports input / output between computing device 600 and other components.

[0060] Communication port 650 is used for data communication between computing device 600 and external sources. For example, communication port 650 can be used for data communication between computing device 600 and network 400. Communication port 650 can be a wired communication port or a wireless communication port.

[0061] Storage medium 630 may include a data storage device. The data storage device may be a non-transitory storage medium or a temporary storage medium. For example, the data storage device may include one or more of a disk 632, a read-only storage medium (ROM) 634, or a random access storage medium (RAM) 636. Storage medium 630 also includes at least one set of instructions stored in the data storage device. The instructions are computer program code, which may include programs, routines, objects, components, data structures, processes, modules, etc., that perform the recommended methods for carrying out the items provided in this specification.

[0062] At least one processor 620 can be communicatively connected to at least one storage medium 630 and a communication port 650 via an internal communication bus 610. The at least one processor 620 is used to execute the at least one instruction set described above. When the computing device 600 is running, the at least one processor 620 reads the at least one instruction set and, according to the instructions of the at least one instruction set, executes the article recommendation method provided in this specification. The processor 620 can execute all the steps included in the article recommendation method. The processor 620 can be in the form of one or more processors. In some embodiments, the processor 620 may include one or more hardware processors, such as a microcontroller, microprocessor, reduced instruction set computer (RISC), application-specific integrated circuit (ASIC), application-specific instruction set processor (ASIP), central processing unit (CPU), graphics processing unit (GPU), physical processing unit (PPU), microcontroller unit, digital signal processor (DSP), field-programmable gate array (FPGA), advanced RISC machine (ARM), programmable logic device (PLD), any circuit or processor capable of performing one or more functions, or any combination thereof. For illustrative purposes only, only one processor 620 is described in this specification for the computing device 600. However, it should be noted that the computing device 600 in this specification may also include multiple processors. Therefore, the operation and / or method steps disclosed in this specification may be executed by one processor as described in this specification, or they may be executed jointly by multiple processors. For example, if the processor 620 of the computing device 600 in this specification executes steps A and B, it should be understood that steps A and B may also be executed jointly or separately by two different processors 620 (e.g., the first processor executes step A, the second processor executes step B, or the first and second processors jointly execute steps A and B).

[0063] Figure 3A flowchart of an item recommendation method P100 provided according to an embodiment of this specification is shown. As previously described, the computing device 600 can execute the item recommendation method P100 of this specification. Specifically, the processor 620 can read an instruction set stored in its local storage medium and then execute the item recommendation method P100 of this specification according to the instructions in the instruction set. Figure 3 As shown, method P100 may include:

[0064] S110: Obtain the target item's attribute information and user interaction information.

[0065] The target items include items whose historical recommendation count is less than a preset threshold. For example, if the preset threshold is 0, the target item can be a zero-sample, meaning a sample with no historical recommendation data. If the preset threshold is a threshold corresponding to a small sample and is not 0, the target item can be a small sample, meaning a sample with a small amount of historical recommendation data. Therefore, the target item can be a cold-start item for the recommendation model, and a cold-start problem will exist when recommending the target item through the trained recommendation model. The target item can be of various types, such as including at least one of the following: goods or content containing goods. The goods can be of various types, such as physical goods or virtual goods, such as various virtual services, etc. Content containing goods can be understood as advertisements or promotional videos containing the goods, etc.

[0066] Among them, attribute information can be information that characterizes the attributes of the target item. There can be various types of attribute information. For example, taking the target item as a physical commodity, the attribute information can include at least one of the following: name, industry, manufacturer, brand, production address, sales platform, or sales region.

[0067] The user interaction information can be information about a user's interaction with the target product. The types of interaction can be varied, such as browsing, purchasing, rating, clicking, adding to favorites, or forwarding. User interaction information can also include information about the user interacting with the target product; for example, taking a purchase as an example, it could include the user's name, purchase time, purchase address, purchase price, and purchase quantity, etc.

[0068] There are several ways to obtain the attribute information and user interaction information of the target item, including the following:

[0069] For example, the processor 620 can directly receive the attribute information and user interaction information of the target item uploaded by the target user 100 through the client 200; or, it can receive the item identifier of the target item uploaded by the target user 100 through the client 200, obtain the item information of the target item based on the item identifier, and extract the attribute information and user interaction information from the item information; or, it can obtain the item information of the target item from the network or item database, and identify the attribute information and user interaction information of the target item from the item information; or, it can obtain a cold start item set, arbitrarily select one or more items from the cold start item set as target items, and obtain the attribute information and user interaction information of the target items; or, when there are many target items or the item information memory is large, it can also receive an item recommendation request sent by the terminal or client 200, which includes the information storage address of the target item, and obtain the attribute information and user interaction information of the target item based on the storage address, and so on.

[0070] S120: Extract similarity features from attribute information to obtain explicit similarity features, and determine at least one item similar to the target item based on user interaction information to obtain implicit similarity features.

[0071] The similarity features include characteristic information that characterizes the similarity between a target item and other items. These similarity features allow for the determination of the similarity or degree of similarity between the target item and other items. Explicit similarity features are those that can be directly obtained from the information of the target item itself. Implicit similarity features, on the other hand, cannot be directly obtained from the information of the target item itself. They require mining the similarities between items through user interaction relationships to obtain the implicit similarity features of the target item. These implicit similarity features can supplement the item representation of target items with zero or small sample sizes, thereby improving the similarity representation of the target item and ultimately increasing the recommendation accuracy of the target item.

[0072] S121: Extract similarity features from attribute information to obtain explicit similarity features.

[0073] For example, the processor 620 can classify attribute information to obtain target attribute information corresponding to each type, extract similarity features from the target attribute information to obtain initial explicit similarity features corresponding to each type, and concatenate the initial explicit similarity features corresponding to each type to obtain the explicit similarity features of the target item.

[0074] There are multiple ways to classify attribute information. For example, the processor 620 can directly classify attribute information according to the information type in the attribute information to obtain the target attribute information corresponding to each type. For example, attribute information can be divided into multiple types such as name, industry, manufacturer, brand, production address, sales platform or sales region; or, the attribute information can be clustered to obtain the target attribute information corresponding to each type.

[0075] After classifying the attribute information, the processor 620 can extract similarity features from the target attribute information corresponding to each type. There are multiple ways to extract target attribute information. For example, the processor 620 can extract similarity features from the target attribute information to obtain multiple discrete similarity features, and then perform summation and pooling on the multiple discrete similarity features to obtain the initial explicit similarity features corresponding to each type.

[0076] Discrete similarity features can be discrete similarity features. These discrete similarity features are discrete variables. There are various ways to sum and pool multiple discrete similarity features. For example, the processor 620 can use an embedding layer to transform multiple discrete similarity features and then sum and pool the transformed continuous similarity features to obtain the initial explicit similarity features corresponding to each type. The embedding value of the continuous similarity feature after embedding transformation using the i-th discrete feature among multiple discrete similarity features is... For example, the method for summing and pooling the transformed continuous similarity features can be shown in formula (1):

[0077]

[0078] in, Let be the initial explicit similarity feature, n be the number of discrete similarity features, and i be the i-th discrete similarity feature.

[0079] After extracting similarity features from the target attribute information, the processor 620 can concatenate the initial explicit similarity features corresponding to each type to obtain the explicit similarity features of the target item. There are multiple ways to concatenate the initial explicit similarity features. For example, the processor 620 can concatenate the initial explicit similarity features corresponding to each type and use a preset activation function to activate the concatenated initial explicit similarity features to obtain the explicit similarity features of the target item. The specific process is shown in formula (2):

[0080]

[0081] in, For explicit similarity features, For activation function, For hyperparameters, The initial explicit similarity features corresponding to Class 1, The initial explicit similarity features corresponding to the second category are: Let be the initial explicit similarity feature corresponding to the m-th class.

[0082] S122: Based on user interaction information, identify at least one item that is similar to the target item to obtain implicit similarity features.

[0083] For example, the processor 620 can determine a set of candidate items corresponding to the target item based on user interaction information, select at least one item similar to the target item from the candidate item set to obtain a set of similar items, and extract features from the items in the set of similar items to obtain the implicit similarity features of the target item. Specifically, this can be done as follows:

[0084] (1) Based on user interaction information, determine the set of candidate items corresponding to the target item.

[0085] The candidate item set can be a collection of other items that users interact with when interacting with the target item. For example, taking a purchase as an example, if both user A and user B purchase the target item, but also purchase other items, then the other items purchased by user A and user B can form the candidate item set.

[0086] There are several ways to determine the set of candidate items corresponding to a target item based on user interaction information, as follows:

[0087] For example, the processor 620 can extract at least one user who interacts with the target item from the user interaction information to obtain a user set, obtain the historical item interaction records of each user in the user set, and identify items other than the target item in the historical item interaction records to obtain a candidate item set.

[0088] Among them, the historical item interaction record can be a record of interactions between the user and the item in historical periods before the current time. Taking the interaction type as a purchase as an example, it can be the user's historical purchase record.

[0089] (2) Select at least one item that is similar to the target item from the candidate item set to obtain a similar item set.

[0090] For example, the processor 620 can determine the user interaction weight of each user in the user set based on historical item interaction records, calculate the item similarity between the target item and each item in the candidate item set based on the user interaction weight, and select at least one item similar to the target item in the candidate item set based on the item similarity to obtain a similar item set.

[0091] Among these, user interaction weights can be used to extract weight information when mining item similarity in user interaction relationships. For example, multiple users may only interact with item i and item j simultaneously. These two items may be related, but a single user's interaction with item j may be different from the similarity between the two items.<u, v> The more interactions there are between items, the less dense the relationships between them. Therefore, user interaction weight can be used as a weighting factor to penalize active users. Here, active users can be understood as users who interact with a large number of items. Based on historical item interaction records, there are multiple ways to determine the user interaction weight of each user in the user set. For example, the processor 620 can count the number of historical item interactions of each user in the user set in the historical item interaction records, and determine the user interaction weight of each user based on the number of historical item interactions. The user interaction weight is inversely proportional to the number of historical item interactions, as shown in formula (3):

[0092]

[0093] in, User interaction weights for user u A collection of items that user u has interacted with historically. This refers to the number of historical item interactions. For example, if user u has interacted with 3 items in the past, then... It is 3.

[0094] After determining the user interaction weight for each user in the user set, the processor 620 can calculate the item similarity between the target item and each item in the candidate item set based on the user interaction weight. There are several ways to calculate the item similarity between the target item and each item in the candidate item set. For example, the processor 620 can select an item from the candidate item set as a candidate item, select at least one user pair from the user set that interacts with the candidate item, determine the number of historical interactions of the items jointly interacted by each user pair based on historical item interaction records, calculate the item similarity between the target item and the candidate item based on the historical item interaction records and user interaction weights, and return to execute the step of selecting an item from the candidate item set as a candidate item, until all items in the candidate item set are selected, thus obtaining the item similarity between the target item and each item in the candidate item set.

[0095] The historical interaction count can be defined as the number of items that the two users in a user pair interacted with during a historical period. For example, taking user pair u and user v as an example, and the interaction being purchasing, user u historically purchased items 1, 2, and 3, while user v historically purchased items 1, 3, and 4. In this case, the number of historical interactions for the user pair...<u,v> The number of historical interactions can be 2, meaning that both user u and user v purchased product 1 and product 3 in the historical phase. Based on the number of historical interactions and user interaction weights, there are multiple ways to calculate the similarity between the target item and the candidate item. For example, the processor 620 can obtain the smoothing coefficient corresponding to the target item, and based on the smoothing coefficient, smooth the number of historical interactions to obtain the initial item similarity for each user pair. The target user interaction weight for each user pair is selected from the user interaction weights, and the initial item similarity is weighted based on the target user interaction weights. The weighted initial item similarity is then accumulated to obtain the item similarity between the target item and the candidate item, as shown in formula (4).

[0096]

[0097] in, Item similarity between the target item and candidate items For the target item, For the remaining items, u and v form the user pair<u,v> , The set of users who interact with the target item. For the set of users who interact with candidate items, User interaction weights for user u User interaction weights for user v For smoothing coefficients, This is a collection of items that user u has interacted with historically. The collection of items that user v has interacted with in the past. For users<u,v> The number of historical interactions.

[0098] After calculating the item similarity between the target item and each item in the candidate item set, the processor 620 can select at least one item in the candidate item set that is similar to the target item based on the item similarity, thus obtaining a set of similar items. There are several ways to select at least one item similar to the target item from the candidate item set. For example, the processor 620 can sort the item similarity and select the top n items with the highest similarity from the candidate item set as items similar to the target item, thus obtaining a set of similar items. Alternatively, it can select at least one item in the candidate item set whose item similarity exceeds a preset similarity threshold, thus obtaining a set of similar items, and so on.

[0099] In some embodiments, the processor 620 can also construct an item interaction network graph and mine the similarity between items in user interaction relationships through a graph neural network. For example, the processor 620 can determine the interaction relationship between each user in the user set and each item in the candidate item set based on historical item interaction records, construct an item interaction graph network with each user and each item as nodes, the item interaction graph network representing the interaction relationship between users and items, and use the item interaction graph network to select at least one item similar to the target item in the candidate item set to obtain a similar item set.

[0100] (3) Extract features from items in the set of similar items to obtain the implicit similarity features of the target item.

[0101] For example, the processor 620 can obtain the item information of each item in the set of similar items, extract features from the item information to obtain the item features of each item, and perform average pooling on the item features to obtain the implicit similarity features of the target item.

[0102] The item information can include the item's ID, attributes, historical recommendation information, and so on. There are various ways to extract features from the item information. For example, the processor 620 can directly extract features from the item information to obtain discrete features, and then use an embedding layer to convert the discrete features into continuous features, thereby obtaining the item features of each item. Alternatively, an item feature extraction network can be used to directly extract item features from the item information, and so on.

[0103] After extracting features from the item information, the processor 620 can perform average pooling on the item features to obtain the implicit similarity features of the target item. There are various ways to perform average pooling on the item features. For example, the processor 620 can perform average pooling on the item features to obtain the pooled item features, and then use an activation function to activate the pooled item features to obtain the implicit similarity features of the target item. Specifically, it can be shown in formula (5):

[0104]

[0105] in, These are implicit similarity features of the target items. For activation function, For hyperparameters, Let n be the item characteristics of the i-th item in the set of similar items, and n be the number of items in the set of similar items.

[0106] S130: Based on explicit and implicit similarity features, determine the initialization parameters of the recommendation model corresponding to the target item, and recommend the target item using the recommendation model based on the initialization parameters.

[0107] The recommendation model is a model for recommending items. This recommendation model can be a meta-learning recommendation model, which can be a recommendation model using a meta-learning network structure.

[0108] The initialization parameters can be used to initialize the recommendation model. After initializing the recommendation model with these parameters, cold-start recommendations for the target item can be performed.

[0109] Among them, based on explicit and implicit similarity features, the initialization parameters of the recommendation model corresponding to the target item are determined, and based on the initialization parameters, there are several ways to use the recommendation model to recommend the target item, as follows:

[0110] S131: Determine the initialization parameters of the recommendation model corresponding to the target item based on explicit and implicit similarity features.

[0111] For example, processor 620 can concatenate explicit similarity features and implicit similarity features to obtain target similarity features of the target item. Based on the target similarity features, it can cluster the target items to obtain category-shared features of the target items. Based on the target similarity features and category-shared features, it can determine the initialization parameters of the recommendation model corresponding to the target item, as follows:

[0112] (1) The explicit similarity features and the implicit similarity features are concatenated to obtain the target similarity features of the target item.

[0113] For example, the processor 620 can directly concatenate explicit similarity features and implicit similarity features to obtain the current similarity features of the target item, obtain the concatenation weights of the preset concatenation network, and weight the current similarity features based on the concatenation weights to obtain the target similarity features of the target item, as shown in formula (6):

[0114]

[0115] in, For target similarity features of target items, To concatenate weights, For the preset splicing network, For explicit similarity features, This is an implicit similarity feature.

[0116] In some embodiments, the processor 620 may also obtain a preset splicing weight, and based on the preset splicing weight, weight the explicit similarity features and the implicit similarity features respectively, and splice the weighted explicit similarity features and the weighted implicit similarity features to obtain the target similarity features of the target item.

[0117] (2) Based on the target similarity features, cluster the target items to obtain the category sharing features of the target items.

[0118] Among them, category-shared features can include feature information of shared knowledge of tasks similar to the recommendation task corresponding to the target item.

[0119] There are several ways to cluster target items based on target similarity features, as follows:

[0120] For example, the processor 620 obtains the cluster center features of each preset task cluster, determines the allocation probability of the recommended task corresponding to the target item to each preset task cluster based on the target similarity features, and accumulates the allocation probability with the corresponding cluster center features to obtain the category sharing features of the target item.

[0121] The allocation probability can be the probability information of assigning the recommendation task corresponding to the target item to each preset task cluster. The recommendation task corresponding to the target item can be understood as treating the item as a task in the meta-learning algorithm. This recommendation task can be equivalent to the target item, or it can be understood that there is a mapping relationship between the recommendation task and the target item. Based on the target similarity features, there are multiple ways to determine the allocation probability of the recommendation task corresponding to the target item to each preset task cluster. For example, the processor 620 uses a task allocation network based on the target similarity features to determine the allocation value of the recommendation task corresponding to the target item to each preset task cluster, and normalizes the allocation value of each preset task cluster to obtain the allocation probability corresponding to each preset task cluster.

[0122] The allocation value represents the probability of a recommended task being assigned to each preset task cluster. The task allocation network may include a first allocation network layer and a second allocation network layer. Based on target similarity features, the task allocation network can determine the allocation value of the recommended task corresponding to the target item for each preset task cluster in various ways. For example, the processor 620 may use the first allocation network layer to perform cluster allocation of recommended tasks based on target similarity features to obtain the initial allocation value corresponding to each preset task cluster, use the second allocation network layer to weight the initial allocation value to obtain the weighted initial allocation value, multiply the initial allocation value by a preset correction parameter to obtain the correction value of the weighted initial allocation value, and add the correction value to the weighted initial allocation value to obtain the allocation value.

[0123] Among them, based on the target similarity features, the first allocation network layer is used to cluster the recommendation tasks to obtain the initial allocation value corresponding to each preset task cluster. There are multiple ways to obtain the initial allocation value corresponding to each preset task cluster. For example, the processor 620 can obtain the first allocation weight of the first classification network layer, and weight the target similarity features based on the first allocation weight. Then, the activation function of the first allocation network layer is used to activate the weighted target similarity features to obtain the initial allocation value corresponding to each preset task cluster. Specifically, it can be shown in formula (7):

[0124]

[0125] in, It is a set that includes the initial allocation values ​​corresponding to each preset task cluster. Assign the activation function to the first network layer. Assign weights to the first one. The target similarity features are those of the target items.

[0126] After the processor 620 uses the first allocation network to weight the initial allocation value, it also multiplies the preset correction parameter by the initial allocation value to obtain the corrected value of the weighted initial allocation value. The corrected value is then added to the weighted initial allocation value to obtain the allocation value, as shown in formula (8).

[0127]

[0128] in, Assigning value, Assign weights to the second group. For initial assignment features, For preset correction parameters, This is a correction value.

[0129] It should be noted that in this scheme, when calculating the allocation probability using the task allocation network, a two-layer allocation network with skip connections is used to correct the initial allocation value, thereby ensuring the accuracy of obtaining the allocation value corresponding to each preset task cluster, and thus improving the clustering accuracy of the target items.

[0130] After determining the allocation value of the recommended task corresponding to the target item to each preset task cluster, the processor 620 can normalize the allocation value of each preset task cluster to obtain the allocation probability corresponding to each preset task cluster. There are many ways to normalize the allocation value. For example, the processor 620 can use the softmax function (a regression function) to normalize the allocation value of each preset task cluster to obtain the allocation probability corresponding to each preset task cluster, as shown in formula (9):

[0131]

[0132] in, Let be the allocation probability corresponding to the i-th preset task cluster among K preset task clusters. Assigning value, K is a hyperparameter, where K is the number of preset task clusters.

[0133] In some embodiments, the processor 620 may also use other normalization algorithms to normalize the allocation values, thereby obtaining the allocation probability corresponding to each preset task cluster.

[0134] After determining the allocation probability of each preset task cluster corresponding to the recommended task allocation value of the target item, the processor 620 can fuse the allocation probability with the corresponding cluster center feature to obtain the category shared feature of the target item. There are many ways to fuse them. For example, the processor 620 can select the cluster center feature corresponding to each allocation probability from the cluster center features, and then multiply the allocation probability with the cluster center feature to obtain the shared feature of each preset task cluster pair. The shared feature is then accumulated to obtain the category shared feature of the target item, as shown in formula (10).

[0135]

[0136] in, To share features for the target item's category, Let i be the allocation probability corresponding to the i-th preset task cluster. Let K be the cluster center vector of the i-th task cluster, and K be the number of preset task clusters.

[0137] It should be noted that the category-shared features include the shared features of the cluster to which the recommendation task corresponding to the target item belongs. This clustering method can be soft clustering, where each recommendation task can obtain a clustering enhancement feature (embedding). This clustering enhancement feature can be understood as the category-shared feature. Moreover, this category-shared feature can include pre-defined shared knowledge information of each preset task cluster, thereby increasing the accuracy of the feature representation of the target item.

[0138] In this process, the task assignment network can be trained end-to-end, which incorporates the information from the labels. Specifically, assuming there are K task clusters, a cluster center vector Ec is first assigned to each task cluster. Each cluster center is randomly initialized at the beginning and continuously updated, thus obtaining the trained task assignment network.

[0139] (3) Based on the target similarity features and category sharing features, determine the initialization parameters of the recommendation model corresponding to the target item.

[0140] For example, the processor 620 can concatenate target similarity features and category-shared features to obtain the initialization parameter features of the target item, use a modulation network to identify the modulation weights corresponding to the target item in the initialization parameter features, obtain the current initialization parameters of the recommendation model, and weight the current initialization parameters based on the modulation weights to obtain the initialization parameters of the recommendation model corresponding to the target item.

[0141] The initialization parameter features can be the feature information of the initialization parameters of the recommendation model corresponding to the target item. There are several ways to concatenate the target similarity features and the category-shared features. For example, the processor 620 can directly concatenate the target similarity features and the category-shared features to obtain the initialization parameter features of the target item. Alternatively, it can obtain the feature weights, and based on the feature weights, weight the target similarity features and the category-shared features respectively, and then concatenate the weighted target similarity features and the weighted category-shared features to obtain the initialization parameter features of the target item.

[0142] After concatenating the target similarity features and category-shared features, the processor 620 can use a modulation network to identify the modulation weights corresponding to the target item in the initialization parameter features. The modulation network can be a network that modulates the initialization parameters of the recommendation model corresponding to the target item. The modulation weights can be the weight information between the initialization parameters corresponding to the target item and the current initialization parameters of the recommendation model. The processor 620 can use a modulation network to identify the modulation weights corresponding to the target item in the initialization parameter features in various ways. For example, the processor 620 can input the initialization parameter features into the modulation network for modulation processing, and activate the modulated initialization parameter features through an activation function to obtain the modulation weights corresponding to the target item, as shown in formula (11):

[0143]

[0144] in, For modulation weights, For modulation networks, The network parameters of the modulation network, whose network structure can be a fully connected layer activated by the sigmoid function, are then considered. The activation function is sigmoid. For target similarity features, For category-shared features, The initialization parameter characteristics corresponding to the target item.

[0145] The current initialization parameters can be the initialization parameters included in the trained recommendation model. There are several ways to obtain the current initialization parameters of the recommendation model. For example, the processor 620 can obtain the trained recommendation model and extract the current initialization parameters from it, or it can receive the current initialization parameters of the recommendation model uploaded by the target user 110 through the client 200, and so on.

[0146] After identifying the modulation weights and obtaining the current initialization parameters of the recommendation model, the processor 620 can weight the current initialization parameters based on the modulation weights to obtain the initialization parameters of the recommendation model corresponding to the target item. The current initialization parameters can be the initialization parameters included in the trained recommendation model. The processor 620 can weight the current initialization parameters based on the modulation weights in various ways. For example, the processor 620 can directly weight the current initialization parameters based on the modulation weights to obtain the initialization parameters of the recommendation model corresponding to the target item, as shown in formula (12):

[0147]

[0148] in, For initialization parameters, For the current initialization parameters, For modulation weights.

[0149] It should be noted that by modulating the initialization parameters of the network, not only can the knowledge of specific tasks be preserved, but also the specific attributes of the cluster can be preserved, thereby improving the accuracy of the initialization parameters.

[0150] In some embodiments, before obtaining the current initialization parameters of the recommendation model, the processor 620 can also train a preset recommendation model to obtain a trained recommendation model, and extract the current recommendation model from the trained recommendation model. The processor 620 can train the preset recommendation model in various ways. For example, the processor 620 can obtain sample item information corresponding to the item sample set, extract sample attribute features, user features, and context features for each item sample in the item sample set from the sample item information, concatenate the sample item features, user features, and context features to obtain the target recommendation features for each item sample, use the preset recommendation model to predict the predicted recommendation object for each item sample based on the target recommendation features, and converge the preset recommendation model based on the predicted recommendation object to obtain the recommendation model.

[0151] Among them, sample attribute features can be the attribute features of the item sample. User features can be the feature information of the user interacting with the item sample, such as user profile information, user granular statistics (statistical values ​​of various user behaviors within a time window), and user preference information. Context features can be other feature information besides attribute features and user features, such as geographical location, access device, weather, time, and whether it is a holiday. The processor 620 can extract the sample attribute features, user features, and context features of each item sample in the item sample set from the sample item information in various ways. For example, the processor 620 can use an embedding layer to extract the sample attribute features, user features, and context features of each item sample in the item sample set from the sample item information. The specific extraction process can be found above and will not be repeated here.

[0152] After extracting the sample attribute features, user features, and context features of the item sample, the processor 620 can concatenate these features. There are several ways to do this. For example, the processor 620 can directly concatenate the sample attribute features, user features, and context features. Alternatively, it can obtain the sample concatenation weight and use this weight to weight the sample attribute features, user features, and context features respectively. Then, it concatenates the weighted sample attribute features, weighted user features, and weighted context features to obtain the target recommendation features for the item sample.

[0153] After concatenating the sample attribute features, user features, and context features, the processor 620 can predict the recommended object for each item sample based on the target recommendation features and using a preset recommendation model. The processor 620 can predict the recommended object for each item sample using the preset recommendation model in several ways. For example, the processor 620 can predict the conversion rate of each user sample in the user sample set to the item sample based on the target recommendation features and using the preset recommendation model, and then select at least one user sample in the user sample set as the predicted recommended object based on this conversion rate.

[0154] After predicting the recommended user for each item sample using a preset recommendation model, the processor 620 can converge the preset recommendation model based on the predicted recommended user. There are several convergence methods. For example, the processor 620 can extract the recommendation tag for each item sample from the sample item information, and based on this recommendation tag, determine the target recommended user corresponding to each item sample. Based on the target recommended user and the predicted recommended user, the processor determines the recommendation loss information for each item sample, and converges the preset recommendation model based on this recommendation loss information, thus obtaining the recommendation model.

[0155] S132: Based on the initialization parameters, a recommendation model is used to recommend target items.

[0156] For example, the processor 620 can initialize the recommendation model based on initialization parameters, use the initialized recommendation model to predict the conversion rate of each user in the preset user set for the target item, and based on the conversion rate, select at least one target user in the preset user set and recommend the target item to at least one target user.

[0157] The processor 620 can initialize the recommendation model in several ways based on the initialization parameters. For example, the processor 620 can replace the network parameters of the recommendation model with the initialization parameters to initialize the recommendation model. Alternatively, it can configure the recommendation model based on the initialization parameters to obtain the initialized recommendation model, and so on.

[0158] After initializing the recommendation model, the processor 620 can use the initialized recommendation model to predict the conversion rate of each user in the preset user set for the target item. There are various ways to predict the conversion rate of users for the target item. For example, the processor 620 can use the feature extraction network of the recommendation model to extract recommendation features from the target item, and based on the recommendation features, use the conversion rate prediction network of the recommendation model to predict the conversion rate of each user in the preset user set for the target item, and so on.

[0159] After predicting the conversion rate of each user in the preset user set for the target item, the processor 620 can select at least one target user from the preset user set based on the conversion rate. There are several ways to select a target user. For example, the processor 620 can sort the conversion rates and select the users corresponding to the top N conversion rates in the preset user set to obtain at least one target user. Alternatively, it can select at least one user in the preset user set whose conversion rate exceeds a preset conversion rate threshold to obtain at least one target user.

[0160] After selecting at least one target user, the processor 620 can recommend the target item to that target user. There are several ways to do this. For example, the processor 620 can directly obtain the user address of at least one target user and recommend the target item to that user based on that address. Alternatively, it can send the user information of at least one target user to a recommendation platform so that the platform can recommend the target item to that user, and so on.

[0161] It should be noted that this solution addresses the cold start problem of target items by combining content-based methods with meta-learning methods, proposing a Task Similarity Aware Meta Learning (TSAML) framework. The framework structure is as follows: Figure 4 As shown. The overall process can include mining the similarities between items and learning a good representation for each item, then clustering the items to learn knowledge between the classes, and using the learned shared knowledge to generate the same initialization parameters for similar tasks. In this way, the learned parameters have generalizability for similar tasks and specificity for dissimilar tasks. The framework mainly includes three parts: task similarity representation, automatic soft clustering module, and initialization modulation network, which can be described as follows:

[0162] (a) Task similarity representation

[0163] Since items with the same attributes exhibit explicit similarity and can play a more significant role in zero-sample testing, the processor 620 extracts features from the attribute information of the target item to obtain explicit similarity features (representations). Furthermore, if multiple users interact with only items i and j simultaneously, these two items may be related. Therefore, to mine the similarity between items, the Swing algorithm (a recall algorithm) can be introduced. The processor 620 can use the Swing algorithm to calculate the similarity between the target item and other interacted items, and define the embedding representations of the n items with the highest similarity scores as the implicit similarity features of the target item. Combining the explicit and implicit similarity feature embeddings represents the item task of the target item, thereby providing a more comprehensive description of the similarity between items at the feature level, and ultimately improving the accuracy of target item recommendation.

[0164] (b) Automatic soft clustering module

[0165] The processor 620 can input the target similarity features (item tasks) of the target item into an automatic assignment function to calculate the clustering probability. Specifically, each target similarity feature (item task) can be assigned to the k-th task cluster using a two-layer neural network with skip connections, thus obtaining the assignment value for each task cluster. Then, it is normalized using a Softmax function to obtain the assigned clustering probability (assignment probability) for each task cluster. Finally, a weighted sum aggregation function is used to obtain the cluster representation (category shared features) of the item tasks corresponding to the target item. Through automatic soft clustering, each item task can obtain a clustering augmentation embedding, which contains shared knowledge with similar tasks.

[0166] (c) Initialize the modulation network

[0167] The processor 620 can combine task embedding (similar features) and cluster embedding (category-sharing features) to adjust the initialization parameters of different item tasks (target items), thereby realizing adaptive initialization of task similarity, as shown in formulas (11) and (12), which will not be elaborated here. By modulating the network, not only the knowledge of specific tasks (target items) is preserved, but also the knowledge shared within the cluster is preserved.

[0168] The processor 620 uses the initial parameters of the recommendation model corresponding to the target item obtained at the end to apply to the recommendation model trained normally, thereby achieving cold-start recommendation of the target item and obtaining better recommendation results. Recommendation tests were conducted on this scheme (TSAML) and other cold-start methods using two test sets. The test results are shown in Table 1 below:

[0169]

[0170] Among them, DCN, DCN-F, MetaEmb, MeLU, MMAML, HSML, and CBML are other cold start solutions, while TSAML is the proposed solution. Therefore, the test results show that the proposed solution achieves a higher AUC than other methods, and its log loss is smaller than other methods.

[0171] In summary, the item recommendation method P100 and system 001 provided in this specification, after obtaining the attribute information and user interaction information of the target item, extract similarity features from the attribute information to obtain explicit similarity features, and based on the user interaction information, determine at least one item similar to the target item to obtain implicit similarity features. Furthermore, based on the explicit and implicit similarity features, determine the initialization parameters of the recommendation model corresponding to the target item, and based on the initialization parameters, use the recommendation model to recommend the target item. Since this scheme can not only extract the explicit similarity features of the target item from the attribute information, but also mine the similarity between items in the user interaction information to obtain the implicit similarity features of the target item, it supplements the content needed for the cold start of the target item, making the representation of the target item more accurate. Therefore, it can improve the accuracy of item recommendation.

[0172] This specification, in another aspect, provides a non-transitory storage medium storing at least one set of executable instructions for making item recommendations. When the executable instructions are executed by a processor, they instruct the processor to perform the steps of the item recommendation method P100 described in this specification. In some possible embodiments, various aspects of this specification can also be implemented as a program product comprising program code. When the program product is run on a computing device 600, the program code causes the computing device 600 to perform the steps of the item recommendation method P100 described in this specification. The program product for implementing the above method may employ a portable compact disc read-only memory (CD-ROM) containing program code and may run on the computing device 600. However, the program product of this specification is not limited thereto. In this specification, a readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system. The program product may employ any combination of one or more readable media. A readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of readable storage media include: electrical connections having one or more wires, portable disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. The computer-readable storage medium may include data signals propagated in baseband or as part of a carrier wave, carrying readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A readable storage medium may also be any readable medium other than a readable storage medium that can send, propagate, or transmit programs for use by or in connection with an instruction execution system, apparatus, or device. Program code contained on a readable storage medium may be transmitted using any suitable medium, including but not limited to wireless, wired, optical fiber, RF, etc., or any suitable combination thereof. Program code for performing the operations described herein can be written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Java and C++, and conventional procedural programming languages ​​such as C or similar languages. The program code can be executed entirely on computing device 600, partially on computing device 600, as a standalone software package, partially on computing device 600 and partially on a remote computing device, or entirely on a remote computing device.

[0173] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0174] In summary, after reading this detailed disclosure, those skilled in the art will understand that the foregoing detailed disclosure is presented by way of example only and is not restrictive. Although not explicitly stated herein, those skilled in the art will understand that this specification requires various reasonable changes, improvements, and modifications to the embodiments. These changes, improvements, and modifications are intended to be made by this specification and are within the spirit and scope of the exemplary embodiments described herein.

[0175] Furthermore, certain terms in this specification have been used to describe embodiments of this specification. For example, "an embodiment," "an embodiment," and / or "some embodiments" mean that a particular feature, structure, or characteristic described in connection with that embodiment may be included in at least one embodiment of this specification. Therefore, it is to be emphasized and understood that two or more references to "an embodiment" or "an embodiment" or "alternative embodiment" in various parts of this specification do not necessarily refer to the same embodiment. Moreover, specific features, structures, or characteristics may be suitably combined in one or more embodiments of this specification.

[0176] It should be understood that in the foregoing description of the embodiments in this specification, various features are combined in a single embodiment, drawing, or description for the purpose of simplifying the description and aiding in the understanding of a feature. However, this does not mean that the combination of these features is necessary, and those skilled in the art may readily identify some of the devices as separate embodiments when reading this specification. That is, the embodiments in this specification can also be understood as an integration of multiple secondary embodiments. It is also valid when each secondary embodiment contains fewer than all the features of a single foregoing disclosed embodiment.

[0177] Each patent, patent application, publication of the patent application, and other materials such as articles, books, specifications, publications, documents, articles, etc., cited herein may be incorporated by reference hereafter. For example, in the event of any inconsistency or conflict between the description, definition, and / or use of terms related to any of the included materials and those related to this document, the terms used herein shall prevail.

[0178] Finally, it should be understood that the embodiments disclosed herein are illustrative of the principles of the embodiments described in this specification. Other modified embodiments are also within the scope of this specification. Therefore, the embodiments disclosed in this specification are merely examples and not limitations. Those skilled in the art can implement the applications described in this specification using alternative configurations based on the embodiments in this specification. Therefore, the embodiments in this specification are not limited to the embodiments precisely described in the applications.

Claims

1. A method for recommending items, including: Obtain attribute information and user interaction information of the target item, wherein the target item includes items whose number of historical recommendation data is less than a preset threshold; Similarity features are extracted from the attribute information to obtain explicit similarity features. Based on the user interaction information, at least one item similar to the target item is determined to obtain implicit similarity features. The similarity features include feature information characterizing the similarity between the target item and other items. Based on the explicit similarity features and the implicit similarity features, the allocation probability of the recommended task corresponding to the target item to each preset task cluster is determined, and the allocation probability is fused with the cluster center feature of the corresponding preset task cluster to obtain the category sharing feature of the target item. as well as Based on the explicit similarity features, the implicit similarity features, and the category sharing features, the initialization parameters of the recommendation model corresponding to the target item are determined, and the recommendation model is used to recommend the target item based on the initialization parameters.

2. The item recommendation method according to claim 1, wherein, The step of extracting similarity features from the attribute information to obtain explicit similarity features includes: The attribute information is classified to obtain the target attribute information corresponding to each type; Similarity features are extracted from the target attribute information to obtain initial explicit similarity features corresponding to each type; and The initial display similarity features corresponding to each type are concatenated to obtain the explicit similarity features of the target item.

3. The item recommendation method according to claim 2, wherein, The step of extracting similarity features from the target attribute information to obtain initial explicit similarity features corresponding to each type includes: Similarity features are extracted from the target attribute information to obtain multiple discrete similarity features; and The discrete similarity features are summed and pooled to obtain the initial explicit similarity features corresponding to each type.

4. The item recommendation method according to claim 1, wherein, The step of determining at least one item similar to the target item based on the user interaction information to obtain implicit similarity features includes: Based on the user interaction information, a set of candidate items corresponding to the target item is determined; Select at least one item similar to the target item from the candidate item set to obtain a similar item set; and Feature extraction is performed on the items in the set of similar items to obtain the implicit similarity features of the target item.

5. The item recommendation method according to claim 4, wherein, The step of determining the candidate item set corresponding to the target item based on the user interaction information includes: Extract at least one user who interacted with the target item from the user interaction information to obtain a user set; Obtain the historical item interaction records of each user in the user set; and Items other than the target item are identified in the historical item interaction records to obtain a candidate item set.

6. The item recommendation method according to claim 5, wherein, The step of selecting at least one item similar to the target item from the candidate item set to obtain a similar item set includes: Based on the historical item interaction records, determine the user interaction weight of each user in the user set; Based on the user interaction weights, calculate the item similarity between the target item and each item in the candidate item set; and Based on the item similarity, at least one item similar to the target item is selected from the candidate item set to obtain the similar item set.

7. The item recommendation method according to claim 6, wherein, The step of determining the user interaction weight of each user in the user set based on the historical item interaction records includes: The number of historical item interactions for each user in the user set is counted in the historical item interaction records; and Based on the number of interactions with historical items, a user interaction weight is determined for each user, and the user interaction weight is inversely proportional to the number of interactions with historical items.

8. The item recommendation method according to claim 6, wherein, The step of calculating the item similarity between the target item and each item in the candidate item set based on the user interaction weight includes: One item is selected from the set of candidate items as a candidate item, and at least one user pair that interacts with the candidate item is selected from the set of users. Based on the historical item interaction records, determine the number of historical interactions of each user pair with common items in the at least one user pair; Based on the historical interaction count and the user interaction weight, calculate the item similarity between the target item and the candidate items; and Return to the step of selecting an item as a candidate item from the candidate item set, and continue until all items in the candidate item set have been selected, to obtain the item similarity between the target item and each item in the candidate item set.

9. The item recommendation method according to claim 8, wherein, The step of calculating the item similarity between the target item and the candidate items based on the number of historical interactions and the user interaction weight includes: Obtain the smoothing coefficient corresponding to the target item, and based on the smoothing coefficient, smooth the number of historical interactions to obtain the similarity of each user to the corresponding initial item; Select the target user interaction weight for each user pair from the user interaction weights, and weight the initial item similarity based on the target user interaction weights; and The weighted initial item similarities are summed to obtain the item similarity between the target item and the candidate item.

10. The item recommendation method according to claim 5, wherein, The step of selecting at least one item similar to the target item from the candidate item set to obtain a similar item set includes: Based on the historical item interaction records, determine the interaction relationship between each user in the user set and each item in the candidate item set; Each user and each item are treated as nodes to construct an item interaction graph network, which represents the interaction relationship between users and items; and Using the item interaction graph network, at least one item similar to the target item is selected from the candidate item set to obtain a similar item set.

11. The item recommendation method according to claim 4, wherein, The step of extracting features from the items in the set of similar items to obtain the implicit similarity features of the target item includes: Obtain the item information for each item in the set of similar items; Feature extraction is performed on the item information to obtain the item features of each item; and The item features are subjected to average pooling to obtain the implicit similarity features of the target item.

12. The item recommendation method according to claim 1, wherein, The step of determining the allocation probability of the recommended task corresponding to the target item to each preset task cluster based on the explicit similarity features and the implicit similarity features includes: The explicit similarity features and the implicit similarity features are concatenated to obtain the target similarity features of the target item. Based on the target similarity features, determine the allocation probability of the recommended task corresponding to the target item to each preset task cluster; and The initialization parameters of the recommendation model corresponding to the target item, determined based on the explicit similarity features, the implicit similarity features, and the category-sharing features, include: Based on the target similarity features and the category sharing features, the initialization parameters of the recommendation model corresponding to the target item are determined.

13. The item recommendation method according to claim 12, wherein, Before fusing the allocation probability with the cluster center feature of the corresponding preset task cluster to obtain the category-sharing feature of the target item, the method further includes: Obtain the cluster center characteristics of each preset task cluster.

14. The item recommendation method according to claim 13, wherein, The step of determining the allocation probability of the recommended task corresponding to the target item to each preset task cluster based on the target similarity features includes: Based on the target similarity features, a task allocation network is used to determine the allocation value of the recommended task corresponding to the target item to each preset task cluster; and The allocation value of each preset task cluster is normalized to obtain the allocation probability corresponding to each preset task cluster.

15. The item recommendation method according to claim 14, wherein, The task allocation network includes a first allocation network layer and a second allocation network layer; as well as The step of determining the allocation value of the recommended task corresponding to the target item to each preset task cluster based on the target similarity feature using a task allocation network includes: Based on the target similarity features, the first allocation network layer is used to allocate the recommended tasks into clusters, obtaining the initial allocation value corresponding to each preset task cluster. The initial allocation value is weighted using the second allocation network layer to obtain a weighted initial allocation value, and Multiply the preset correction parameter by the initial allocation value to obtain the corrected value of the weighted initial allocation value, and add the corrected value to the weighted initial allocation value to obtain the allocation value.

16. The item recommendation method according to claim 12, wherein, The step of determining the initialization parameters of the recommendation model corresponding to the target item based on the target similarity features and category sharing features includes: The target similarity features and the category-shared features are concatenated to obtain the initialization parameter features corresponding to the target item; The modulation network is used to identify the modulation weights corresponding to the target item from the initialization parameter features; and Obtain the current initialization parameters of the recommendation model, and weight the current initialization parameters based on the modulation weights to obtain the initialization parameters of the recommendation model corresponding to the target item.

17. The item recommendation method according to claim 16, wherein, Before obtaining the current initialization parameters of the recommendation model, the method further includes: Obtain sample item information corresponding to the item sample set, and extract sample attribute features, user features, and context features for each item sample in the item sample set from the sample item information; The sample attribute features, user features, and context features are concatenated to obtain the target recommendation features for each item sample; Based on the target recommendation features, a preset recommendation model is used to predict the recommended user for each item sample; and The recommended user is converged to the preset recommendation model based on the predicted recommendation, and the recommendation model is obtained.

18. The item recommendation method according to claim 1, wherein, The step of recommending the target item using the recommendation model based on the initialization parameters includes: The recommendation model is initialized based on the initialization parameters; The initialized recommendation model is used to predict the conversion rate of each user in the preset user set for the target item; and Based on the conversion rate, at least one target user is selected from the preset user set, and the target item is recommended to the at least one target user.

19. The item recommendation method according to claim 1, wherein, The recommendation model is a meta-learning recommendation model.

20. An item recommendation system, comprising: At least one storage medium storing at least one instruction set for item recommendation; as well as At least one processor is communicatively connected to the at least one storage medium. When the item recommendation system is running, the at least one processor reads the at least one instruction set and executes the item recommendation method according to any one of claims 1-19.

Images