Item recommendation scoring method and apparatus, device, medium, and product

Abstract

The application relates to a commodity item recommendation scoring method and device, equipment, medium and product, the method comprising: acquiring feature encoding information of each historical commodity item in a historical behavior sequence of a user and feature encoding information of a target commodity item, and converting each feature encoding information into a corresponding hash fingerprint; converting each hash fingerprint in the historical behavior sequence into a hash mapping table, the hash mapping table containing a plurality of key-value pairs, the value of each key-value pair storing feature encoding information corresponding to the hash fingerprint carrying the corresponding key; performing hash collision on the hash fingerprint of the target commodity item and the key-value pairs in the hash mapping table, determining a plurality of matched key-value pairs, and constructing the feature encoding information stored in the plurality of key-value pairs into a user interest vector; and performing classification mapping on the user interest vector to determine the correlation score of the target commodity item. Thus, not only the operation effect is ensured, but also the operation performance is improved.

Description

Technical Field

[0001] This application relates to e-commerce information processing technology, and more particularly to a product item recommendation and rating method and its apparatus, equipment, medium, and product. Background Technology

[0002] A user's historical behavior sequence refers to a series of user behavior trajectories generated after a user interacts with products on a website (clicking / purchasing / viewing), which can effectively reflect the user's interests and preferences. Using a deep learning model that incorporates an autocorrelation mechanism, weights are extracted from candidate products and products previously visited by the user based on the historical behavior sequence, essentially a relevance score. Since candidate products require attention operations with each historical behavior product, the time complexity of the entire process is O(L*B*d), where L represents the length of the user behavior, B represents the number of candidate products scored each time, and d represents the size of the feature vector of the product item. Therefore, the corresponding deep learning model can only model relatively short behavior sequences.

[0003] Contrary to popular belief, the longer the user's historical behavior sequence, the better the model performs. To model long-term user interests, long-period historical behavior sequences are necessary. However, if the sampling period for the historical behavior sequence is too long, the computational load will increase significantly due to the limitations of the autocorrelation mechanism, making it unsuitable for online real-time prediction.

[0004] One improved approach borrows from the two-stage method of search systems, which involves recalling and then ranking users. The first stage filters user behaviors related to candidate products from a long list of user behaviors. The second stage extracts the interest relationship between the behavioral sequences obtained in the first stage and the candidate products, thereby modeling user interests. This method can handle the increased sequence length, but it suffers from inconsistent objectives and different data update frequencies. The inconsistency arises because the feature vectors of the product items in the two stages have no direct relationship, leading to different optimization goals. Furthermore, online models need to capture real-time changes in user interests, but the indexes used are built offline, resulting in different data update frequencies.

[0005] Other industry-wide improvements have attempted to further reduce the dimensionality of the feature vectors of product items to decrease computational complexity, reducing the time complexity to O(L*B*1). However, these efforts have ultimately failed to optimize the sequence length L, and the complexity still increases linearly with L, limiting the sequence length to the thousands. To ensure computational efficiency, in these situations, it is necessary to segment and truncate the user's historical behavior sequence, potentially leading to suboptimal solutions that still affect computational performance.

[0006] It is evident that traditional techniques for scoring the relevance of candidate products based on users' historical behavior sequences have shortcomings in terms of effectiveness and performance, and need to be improved. Summary of the Invention

[0007] The purpose of this application is to solve the above-mentioned problems by providing a product recommendation and rating method and corresponding apparatus, devices, non-volatile readable storage media, and computer program products.

[0008] According to one aspect of this application, a product item recommendation and rating method is provided, comprising the following steps:

[0009] Obtain the feature encoding information of each historical product item and the feature encoding information of the target product item in the user's historical behavior sequence, and convert each feature encoding information into a corresponding hash fingerprint;

[0010] Each hash fingerprint in the historical behavior sequence is converted into a hash mapping table, which contains multiple key-value pairs. The value of each key-value pair stores the feature encoding information corresponding to the hash fingerprint carrying its corresponding key.

[0011] The hash fingerprint of the target product item is hashed and collided with the key-value pairs in the hash mapping table to determine multiple matching key-value pairs. The feature encoding information stored in the multiple key-value pairs is used to construct a user interest vector.

[0012] The user interest vector is classified and mapped to determine the relevance score of the target product item.

[0013] Optionally, the encoded information of each feature is converted into a corresponding hash fingerprint, including:

[0014] Obtain the corresponding number of hyperplane matrices based on the preset number of bits in the hash fingerprint;

[0015] Calculate the dot product between each of the aforementioned feature encoding information and each of the aforementioned hyperplane matrices, and binarize the dot product results to obtain the corresponding binarized values ​​of each hyperplane matrix;

[0016] The binary values ​​of each hyperplane matrix corresponding to each of the aforementioned feature encoding information are ordered to construct the corresponding hash fingerprint.

[0017] Optionally, converting each hash fingerprint in the historical behavior sequence into a hash mapping table includes:

[0018] The hash fingerprints of each historical product item in the historical behavior sequence are segmented according to a preset word length to obtain multiple substrings;

[0019] Using the common substring of each segment as the key, the feature encoding information corresponding to the hash fingerprints of the corresponding segments with the same substring is concatenated into a feature encoding sequence;

[0020] The feature encoding sequence is normalized and stored in the value range corresponding to the key, forming key-value pairs of the corresponding substrings, which are then stored in the hash mapping table.

[0021] Optionally, the hash fingerprint of the target product item is hashed against the key-value pairs in the hash mapping table to determine multiple matching key-value pairs. The feature encoding information stored in the multiple key-value pairs is then used to construct a user interest vector, including:

[0022] The hash fingerprint of the target product item is segmented according to a preset word length to obtain multiple substrings;

[0023] Retrieve the normalized feature encoding sequences corresponding to the multiple substrings from the hash mapping table and concatenate them into the same feature encoding sequence;

[0024] The user interest vector is obtained by performing an average pooling operation on the feature encoding sequence.

[0025] Optionally, the user interest vector is classified and mapped to determine the relevance score of the target product item, including:

[0026] Obtain the user's profile feature information and / or access scenario feature information as external feature information;

[0027] The external feature information is concatenated with the user interest vector to form user feature information;

[0028] The user feature information is fully connected and then classified and mapped to a specific category. The classification probability corresponding to the specific category is used as the relevance score of the target product item.

[0029] Optionally, after classifying and mapping the user interest vector to determine the relevance score of the target product item, the process includes:

[0030] The product items are sorted in reverse order based on the relevance scores obtained by taking each product item in the candidate product list as the target product item in turn.

[0031] The promotional information for each product item is constructed in reverse order and pushed to the user's terminal device for display.

[0032] Optionally, the product item recommendation and rating method is executed by a deep learning model based on a neural network.

[0033] According to another aspect of this application, a product item recommendation and rating device is provided, comprising:

[0034] The hash representation module is configured to obtain the feature encoding information of each historical product item in the user's historical behavior sequence and the feature encoding information of the target product item, and convert each feature encoding information into a corresponding hash fingerprint;

[0035] The hash mapping module is configured to convert each hash fingerprint in the historical behavior sequence into a hash mapping table. The hash mapping table contains multiple key-value pairs, and the value of each key-value pair stores the feature encoding information corresponding to the hash fingerprint carrying its corresponding key.

[0036] The interest construction module is configured to perform hash collision between the hash fingerprint of the target product item and the key-value pairs in the hash mapping table to determine multiple matching key-value pairs, and construct a user interest vector from the feature encoding information stored in the multiple key-value pairs.

[0037] The relevant rating module is configured to classify and map the user interest vector to determine the relevance rating of the target product item.

[0038] According to another aspect of this application, a product item recommendation and rating device is provided, including a central processing unit and a memory, wherein the central processing unit is used to invoke and run a computer program stored in the memory to perform the steps of the product item recommendation and rating method described in this application.

[0039] According to another aspect of this application, a non-volatile readable storage medium is provided, which stores, in the form of computer-readable instructions, a computer program implemented according to the product item recommendation and rating method, wherein the computer program, when invoked by a computer, performs the steps included in the method.

[0040] According to another aspect of this application, a computer program product is provided, including a computer program / instructions that, when executed by a processor, implement the steps of the method described in any embodiment of this application.

[0041] Compared to existing technologies, this application maps the target product item and the product item representation in the user's historical behavior sequence to hash fingerprints, and then normalizes the collision of the two representations to obtain the final user interest vector. The relevance score is determined based on the user interest vector. This significantly reduces the computational load, eliminates the need to truncate the user's historical behavior sequence, and avoids the problem of obtaining suboptimal solutions after truncating the original sequence. At the same time, it does not require setting a truncation length, which not only ensures the computational effect but also improves the computational performance, achieving a balance between effect and performance. It is suitable for quickly and efficiently determining the product recommendation score for a large number of products to be recommended in e-commerce platforms, thereby enabling accurate ranking of these products and improving the user experience related to product recommendations. Attached Figure Description

[0042] To more clearly illustrate the technical solutions in the embodiments of this application, 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 application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0043] Figure 1 This is a schematic diagram of the network architecture in an exemplary e-commerce platform application scenario of this application;

[0044] Figure 2 A flowchart illustrating one embodiment of the product item recommendation and rating method of this application;

[0045] Figure 3 A schematic diagram of the network architecture of a deep learning model implemented according to the product item recommendation and scoring method of this application;

[0046] Figure 4 This is a schematic diagram illustrating the implementation principle of the interest extraction network in the deep learning model of this application;

[0047] Figure 5 This is a schematic diagram of the process for generating hash fingerprints with feature encoding information in an embodiment of this application;

[0048] Figure 6 This is a schematic diagram illustrating the process of hashing and storing the feature encoding information of each historical item in the user's historical behavior sequence in an embodiment of this application.

[0049] Figure 7 This is a schematic diagram illustrating the process of retrieving feature encoding sequences and constructing user interest vectors based on hash collisions using the hash fingerprint of the target product item in an embodiment of this application.

[0050] Figure 8 This is a schematic diagram illustrating the process of determining the relevance score of a target product item by superimposing external feature information and user interest vector in an embodiment of this application;

[0051] Figure 9 A schematic diagram of the network architecture of another deep learning model implemented according to the product item recommendation and scoring method of this application;

[0052] Figure 10 This is a schematic diagram illustrating the process of pushing promotional information for product items to users in an embodiment of this application;

[0053] Figure 11 A schematic diagram of the product item recommendation and scoring device for this application;

[0054] Figure 12 This is a schematic diagram of the structure of a product item recommendation and scoring device used in this application. Detailed Implementation

[0055] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain this application, and should not be construed as limiting this application.

[0056] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this application means the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof. It should be understood that when we say an element is “connected” or “coupled” to another element, it can be directly connected or coupled to the other element, or there may be intermediate elements. Furthermore, “connected” or “coupled” as used herein can include wireless connections or wireless coupling. The term “and / or” as used herein includes all or any units and all combinations of one or more associated listed items.

[0057] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless specifically defined as herein.

[0058] Those skilled in the art will understand that the terms "client," "terminal," and "terminal device" as used herein include both devices that receive wireless signals, devices that only possess wireless signal receiver capabilities without transmission capabilities, and devices with receiving and transmitting hardware, devices that have receiving and transmitting hardware capable of bidirectional communication over a bidirectional communication link. Such devices may include: cellular or other communication devices such as personal computers or tablets, having single-line displays, multi-line displays, or cellular or other communication devices without multi-line displays; PCS (Personal Communications Service) that can combine voice, data processing, fax, and / or data communication capabilities; PDA (Personal Digital Assistant) that may include a radio frequency receiver, pager, internet / intranet access, web browser, notepad, calendar, and / or GPS (Global Positioning System) receiver; and conventional laptops and / or handheld computers or other devices that have and / or include radio frequency receivers. As used herein, "client," "terminal," and "terminal device" can be portable, transportable, installed in a means of transportation (air, sea, and / or land), or suitable and / or configured to operate locally and / or in a distributed manner, operating in any other location on Earth and / or in space. "Client," "terminal," and "terminal device" as used herein can also be a communication terminal, an internet access terminal, or a music / video playback terminal, such as a PDA, a MID (Mobile Internet Device), and / or a mobile phone with music / video playback capabilities, or a smart TV, set-top box, etc.

[0059] The hardware referred to by the names "server," "client," and "service node" in this application is essentially an electronic device with the equivalent capabilities of a personal computer. It is a hardware device with the necessary components revealed by the von Neumann architecture, such as a central processing unit (including an arithmetic logic unit and a control unit), memory, input devices, and output devices. The computer program is stored in its memory, and the central processing unit loads the program stored in the secondary storage into the main memory to run it, execute the instructions in the program, and interact with the input and output devices to complete specific functions.

[0060] It should be noted that the concept of "server" used in this application can also be extended to the case of server clusters. Based on the network deployment principles understood by those skilled in the art, the servers should be logically divided. Physically, these servers can be independent of each other but accessible through interfaces, or they can be integrated into a single physical computer or a computer cluster. Those skilled in the art should understand this flexibility and should not use it to constrain the implementation of the network deployment method in this application.

[0061] One or more of the technical features of this application, unless explicitly specified herein, can be deployed on a server and accessed by a client remotely calling the online service interface provided by the server, or can be directly deployed and run on a client for access.

[0062] Unless otherwise specified, all data involved in this application may be stored remotely on a server or on a local terminal device, as long as it is suitable for use by the technical solution of this application.

[0063] Those skilled in the art will understand that although the various methods in this application are described based on the same concept and thus present commonality among them, they can be performed independently unless otherwise specified. Similarly, the various embodiments disclosed in this application are all based on the same inventive concept; therefore, concepts expressed in the same way, as well as concepts that are appropriately changed for convenience but are expressed differently, should be understood equivalently.

[0064] Unless otherwise expressly stated, the various embodiments disclosed in this application can be combined in a cross-cutting manner to flexibly construct new embodiments, as long as such combination does not depart from the inventive spirit of this application and can meet the needs of the prior art or solve a certain deficiency in the prior art. Those skilled in the art should be aware of such modifications.

[0065] Please see Figure 1The network architecture adopted in an exemplary application scenario of this application includes a terminal device 80, an independent website server 81, and an application server 82. The application server 82 can be used to deploy a product recommendation service, which is implemented by running a computer program product based on the product recommendation scoring method of this application and opening corresponding interfaces. The independent website server 81, referred to as an independent website, can be used to deploy and open an online store. Users on the terminal device 80 can access a product item in the online store through the independent website, including adding the product item to the shopping cart, accessing the product details page of the product item, etc. The independent website server 81 further calls the product recommendation service to determine the product items to be recommended, takes these product items to be recommended as target product items, and determines the corresponding product recommendation score for each target product item by executing the various steps of the product recommendation method. Then, a product recommendation list is obtained by sorting the product recommendation scores and pushed to the user's terminal device for parsing and display.

[0066] The application server can be a service cluster that runs various services through a microservice architecture. Accordingly, the product recommendation and rating service can run in one or more containers provided by the microservice architecture, concurrently providing services to a massive number of independent websites.

[0067] The product recommendation and scoring method of this application can be implemented by a deep learning model based on a neural network. The deep learning model is trained in advance using a corresponding training dataset to achieve convergence and then put into online inference, so that it is suitable for outputting corresponding relevance scores based on the input data.

[0068] The exemplary application scenario of this application is the implementation of product recommendation in an e-commerce platform. For example, when a consumer user visits the details page of a product item in an online store on an independent website, promotional information for multiple product items that match the long-term interests of the consumer user will be displayed in an advertising area on the details page. When the consumer user becomes interested in a certain product item, they can directly access the details page of the corresponding product item by clicking on the relevant promotional information, thereby stimulating the user to make a purchase.

[0069] Similarly, when a consumer clicks on a "You May Also Like" page, promotional information for multiple product items that match their long-term interests can be recommended to that consumer. Alternatively, when a consumer adds a product item to their shopping cart, promotional information for the aforementioned multiple product items can be displayed. These methods can all be implemented by the product item recommendation service of this application. The product item recommendation service will recall product items to be recommended based on various preset recall strategies. Then, it will use the product item recommendation scoring method of this application to score each product item to be recommended to determine its relevance score. Finally, based on the relevance score, the product items to be recommended will be sorted, and their promotional information will be pushed to the consumer's terminal device for display.

[0070] In addition to the exemplary application scenarios mentioned above, the product item recommendation and scoring method of this application can be used to implement any other scenario that requires associating with the user's long-term interests to determine the product items to be recommended, i.e., the relevance score of the candidate product items. Those skilled in the art can apply it as needed.

[0071] Based on the principles revealed above, please refer to Figure 2 According to a product item recommendation and rating method provided in this application, in one embodiment, it includes the following steps:

[0072] Step S1100: Obtain the feature encoding information of each historical product item and the feature encoding information of the target product item in the user's historical behavior sequence, and convert each feature encoding information into a corresponding hash fingerprint;

[0073] To characterize a user's long-term interests, the user's previously visited product items, i.e., historical product items, can be extracted in advance based on the user's behavioral records in the current online store or e-commerce platform. The feature information of these historical product items can be obtained, and the feature encoding information of each historical product item can be constructed through encoding. Then, the feature encoding information of all historical product items can be constructed into a historical behavior sequence to characterize the user's historical interest features.

[0074] Because the product item recommendation and rating method of this application has the efficiency advantage of low computational load, in one embodiment, when constructing the historical behavior sequence, the historical behavior sequence can include feature encoding information corresponding to all historical product items of the user since the creation of their account. In another embodiment, a preset historical time period can be determined as needed, and all historical product items within the historical time period range can be selected to construct the user's historical behavior sequence accordingly. In summary, based on the advantages achieved by this application, the historical behavior sequence in the case of a single user can contain a large number of product items. When implementing multi-user parallel processing on the same e-commerce platform, it can theoretically achieve stable computing power at the level of tens of thousands, hundreds of thousands, or even millions.

[0075] The feature information of each historical product item in the historical behavior sequence can be pre-defined and set, for example, including but not limited to obtaining the product ID, category ID, attribute data, etc. of each product item. Those skilled in the art can flexibly set this. The historical behavior sequence can be input into the corresponding encoding layer of the product item recommendation and scoring model of this application for feature extraction in order to obtain its corresponding feature encoding information.

[0076] Similarly, for target product items that need to be rated, their corresponding feature information can be obtained in the same way, and then input into the corresponding encoding layer of the product item recommendation rating model to construct their corresponding feature encoding information. The principle is the same as that of obtaining the feature encoding information of historical product items.

[0077] An exemplary first type of product item recommendation rating model is as follows: Figure 3 The network architecture shown has a first encoding layer for processing the feature encoding information of each historical item in the user behavior sequence, a second encoding layer for processing the feature information of the target item to obtain its feature encoding information, an interest extraction network for constructing user interest vectors based on the outputs of the first and second encoding layers, and a classifier for performing classification mapping to determine relevance scores.

[0078] based on Figure 3 The network architecture shown has a first encoding layer that outputs the feature encoding information of each historical item in the historical behavior sequence, a second encoding layer that outputs the feature encoding information of the target item, and then enters the interest extraction network to construct the user's interest vector.

[0079] In the interest extraction network, hash fingerprints are first extracted from the feature encoding information of the historical product items and the target product items. This involves hashing each feature encoding information to create a hash value, the number of bits in which can be preset as needed. In one embodiment, the SimHash algorithm can be used to convert the feature encoding information into a corresponding hash fingerprint.

[0080] Each historical product item in the historical behavior sequence, as well as the target product item, has its feature encoding information converted into a corresponding hash fingerprint. Since the hash fingerprint is binary data, operations performed on the hash fingerprint have a significant advantage in computational efficiency. The number of bits in the hash value of the hash fingerprint can be set as needed by those skilled in the art according to the complexity of the feature information, for example, set to 4, 6, 8, 12, etc.

[0081] Step S1200: Convert each hash fingerprint in the historical behavior sequence into a hash mapping table. The hash mapping table contains multiple key-value pairs, and the value of each key-value pair stores the feature encoding information corresponding to the hash fingerprint carrying its corresponding key.

[0082] The hash fingerprints of each historical product item in the historical behavior sequence can be organized in an ordered manner to improve retrieval efficiency. By hashing the hash fingerprints, the hash value in each hash fingerprint is divided into multiple segments corresponding to substrings. Each substring can serve as the content of the key field in a key-value pair, thus providing an index for the hash fingerprint. The value field of each key-value pair can be used to store the feature encoding information corresponding to the hash fingerprint of the substring in the key field of the corresponding segment. Therefore, the value field of the same key-value pair may store the feature encoding information of multiple historical product items. In one embodiment, when storing the feature encoding information of the historical product item in the value field of the key-value pair, the feature encoding information can first be normalized, represented as Norm(). Figure 3 In the network architecture shown, the above process can be implemented in its interest extraction network.

[0083] When segmenting a hash fingerprint, the segment length can be flexibly set, for example, every 2 bits per segment, so that each substring has 2 bits. Of course, if the hash fingerprint itself has a large number of bits, such as 16 bits, it can also be divided into segments of four bits.

[0084] It is easy to understand that by using hash storage, the feature encoding information of historical product items in the historical behavior sequence is stored in a hash mapping table. This hash mapping table consists of a finite number of key-value pairs. In each key-value pair, the content of its key field is a substring corresponding to a certain value of a segment. The hash fingerprint carrying this substring and its corresponding feature encoding information are all stored in the value field of the key-value pair. This achieves the orderly organization of the feature encoding information in the historical behavior sequence and establishes an index relationship with hash substrings, which facilitates efficient retrieval based on hash values ​​with extremely low retrieval computation.

[0085] Step S1300: Perform hash collision between the hash fingerprint of the target product item and the key-value pairs in the hash mapping table to determine multiple matching key-value pairs, and construct a user interest vector from the feature encoding information stored in the multiple key-value pairs.

[0086] The feature encoding information of each product item is localized; the more similar the product items are in terms of features, the higher the probability of their hash fingerprints being consistent. Therefore, the probability of hash collisions can be used instead of the traditional Attention mechanism for filtering related information to determine the feature encoding information of historical product items that match the feature encoding information of the target product item. Specifically, since a hash fingerprint is the feature extraction result of feature encoding information, for the target product item, its hash fingerprint can be searched and matched with the key field content of each key-value pair in the hash mapping table to achieve hash collisions, thereby obtaining multiple key-value pairs that match the hash fingerprint of the target product item. The feature encoding information stored in the value field of these key-value pairs is the feature encoding information whose hash fingerprint matches the hash fingerprint of the target product item, that is, objects similar to the feature encoding information of the target product item. Similarly, the above process can be implemented in... Figure 3 The network architecture shown is used for interest extraction in the network.

[0087] In some embodiments, the feature encoding information stored in the key-value pairs is pre-normalized, so the normalized feature encoding information stored in the value fields of each matched key-value pair can be directly concatenated into a user interest vector. In other embodiments, if the feature encoding information in the key-value pairs is not normalized, the feature encoding information retrieved from the key-value pairs can be normalized first.

[0088] In some embodiments, for all the feature encoding information extracted from the plurality of key-value pairs, a pooling layer can be further applied to perform average pooling operations to achieve feature compression. For example, combining... Figure 4 The data content shown, for the target product item with hash fingerprint "0010", the feature encoding information Norm(item2+item3) and Norm(item3) of the historical product items retrieved from the hash mapping table are pooled using the following formula example: [Norm(item3)+Norm(item2+item3)] / 2. The result can be used as a user interest vector to achieve a comprehensive representation of the feature encoding information of historical product items that the user is interested in and that are associated with the target product item.

[0089] Step S1400: Classify and map the user interest vector to determine the relevance score of the target product item.

[0090] Please combine Figure 3The network architecture shown allows for the input of the user interest vector into a classifier for classification mapping after the user interest vector is determined. In the classifier, the vector is fully connected and activated through one or more fully connected layers, ultimately mapping to a preset classification space. This classification space contains two or more categories, one of which is a positive category trained under supervised conditions on the corresponding positive sample. The classification probability of mapping to this positive category is obtained and used as the relevance score corresponding to the target product item, thus quantifying the degree to which the target product item matches the user's long-term interests.

[0091] In some embodiments, a large number of candidate product items associated with the user are first recalled from multiple preset recall channels through a product item recommendation service, forming a candidate product list. Then, for each candidate product item in the candidate product list, it can be used as the target product item of this application, and its corresponding relevance score is determined according to the above process. Then, all candidate product items in the candidate product list can be sorted using the relevance score, and a portion of the candidate product items are selected based on the sorting result to form a product recommendation list. The product recommendation list is then pushed to the user's terminal device for display, so that the user can obtain promotional information for product items that match their long-term interests. Of course, in some embodiments, considering the high-efficiency computing advantage of this application, the candidate product list can also be the full range of product items of an online store on an independent website. Therefore, it is not necessary to perform the recall operation described above, but to directly determine the corresponding relevance score for each product item in the online store.

[0092] Based on the above process, the modeling of the user's full historical behavior sequence was completed, and the time complexity was reduced from O(L*B*d) to O(B*m*log(d)), meeting the performance requirements of online inference. Since no dot product operation is needed on the user sequence, L is removed from the complexity. When using the hash fingerprint collision method, it is necessary to calculate the dot product with m hyperplanes. The complexity of simhash is log(d), so the final time complexity is O(B*m*log(d)). The improvement in computational efficiency achieved by the technical solution of this application is thus evident.

[0093] As can be seen from the above embodiments, this application maps the representations of target product items and product items in the user's historical behavior sequence to hash fingerprints, and then normalizes the representations of the collision between the two to obtain the final user interest vector. The relevance score is determined based on the user interest vector. The computational workload is greatly reduced. It does not require sequence truncation of the user's historical behavior sequence, avoiding the problem of obtaining suboptimal solutions after truncation of the original sequence. At the same time, it does not require setting the truncation length, which not only ensures the computational effect but also improves the computational performance, achieving a balance between effect and performance. It is suitable for quickly and efficiently determining the product recommendation score of a large number of product items to be recommended in e-commerce platforms, thereby enabling accurate ranking of these product items and improving the user experience related to product recommendations.

[0094] Based on any embodiment of this application, please refer to Figure 5 The various feature encoding information is converted into corresponding hash fingerprints, including:

[0095] Step S1110: Obtain the corresponding number of hyperplane matrices according to the preset number of bits of the hash fingerprint;

[0096] To extract hash fingerprints, multiple hyperplane matrices can be pre-set randomly. For each bit of the hash fingerprint (i.e., word length), a corresponding hyperplane matrix is ​​created. The hyperplane matrix is ​​a two-dimensional plane.

[0097] Step S1120: Calculate the dot product between each of the aforementioned feature encoding information and each of the aforementioned hyperplane matrices, and binarize the dot product results to obtain the corresponding binarized values ​​of each hyperplane matrix.

[0098] For each historical product item in the user's historical behavior sequence and the target product item's feature encoding information, for each feature encoding information, a similarity can be calculated with each hyperplane matrix to convert it into a binarized value corresponding to each hyperplane matrix, so that the binarized values ​​of each hyperplane matrix can be organized according to the preset sorting of the hyperplane matrices.

[0099] In this embodiment, a more efficient method is used to calculate the similarity between each feature encoding information and each hyperplane matrix. Specifically, the dot product of the feature encoding information and the hyperplane matrix is ​​calculated to obtain a numerical value representing the similarity. It is then determined whether this numerical value is greater than a preset threshold. This preset threshold is used to measure whether the two are similar. It can be a value of 0. When it is greater than the preset threshold, the value obtained by the feature encoding information corresponding to the hyperplane matrix is ​​represented by a binary value of 1, indicating that the two are similar. Otherwise, the value obtained by the feature encoding information corresponding to the hyperplane matrix is ​​represented by a binary value of 0, indicating that the two are not similar.

[0100] Step S1130: Construct the corresponding hash fingerprint by orderly constructing the binarized values ​​of each hyperplane matrix corresponding to each feature encoding information.

[0101] After obtaining corresponding binary values ​​for different hyperplane matrices corresponding to the same feature encoding information, these binary values ​​are arranged in order according to the preset sorting between different hyperplane matrices to obtain a binary value. This binary value can be used as the hash value of the hash fingerprint to realize the hash representation of the feature encoding information.

[0102] It's easy to understand that each feature encoding information can generate its hash fingerprint using the above process. All hash fingerprints are hash values ​​represented as binary data, and the number of bits in the hash values ​​is the same for different hash fingerprints.

[0103] As can be seen from the above embodiments, by performing similarity operations on hyperplane matrices and feature encoding information, generating binary values ​​based on whether the two are similar, and arranging the binary values ​​corresponding to multiple hyperplane matrices in an orderly manner to form hash values ​​as hash fingerprints, the random representation of the feature encoding information can be quickly achieved. This lays the foundation for subsequent efficient operations and hash collisions based on the hash fingerprints of target product items and historical product items to retrieve relevant historical product items.

[0104] Based on any embodiment of this application, please refer to Figure 6 Converting each hash fingerprint in the historical behavior sequence into a hash mapping table includes:

[0105] Step S1210: Divide the hash fingerprint of each historical product item in the historical behavior sequence into segments according to a preset word length to obtain multiple substrings;

[0106] To facilitate hash collision implementation, the hash fingerprints of each historical item in the historical behavior sequence can be converted and stored in a hash mapping table. To this end, based on the total number of bits in the hash value of the hash fingerprint, the segment word length is set, and each hash fingerprint is segmented, such as... Figure 4 As shown, each hash fingerprint consists of 4 binary bits, and each segment is set to a word length of 2 bits. Therefore, each hash fingerprint can be divided into two substrings. For example, the hash value "1110" can be divided into two substrings, "11" and "10", after being segmented using the above principle. Of course, if the hash fingerprint has more bits, there may be three or more substrings; similarly, the word length of each segment can be flexibly set as needed.

[0107] Step S1220: Using the same substring in each segment as the key, concatenate the feature encoding information corresponding to the hash fingerprints of the corresponding segments with the same substring into a feature encoding sequence;

[0108] To index the hash fingerprints of all historical product items within a historical behavior sequence, key-value pairs can be used. For each segment, a key-value pair is created for a substring of a certain value within that segment. This substring serves as the key field of the key-value pair, while the value field stores the feature encoding information corresponding to the hash fingerprint of the substring in the key field within the corresponding segment. When the substring in the key field of a key-value pair is carried by corresponding segments of the hash fingerprints of multiple historical product items, its value field will store the feature encoding information of those multiple historical product items. In this case, this feature encoding information can be concatenated and stored in the value field.

[0109] Therefore, the value range in each key-value pair may be composed of one or more feature encoding information, thus forming a feature encoding sequence, which can be directly called for further processing.

[0110] As mentioned above, in some embodiments, when it is necessary to store the feature encoding information in the corresponding value range, it can be normalized first, for example, each feature value can be normalized to the floating-point value range of [0,1] to facilitate subsequent standardization operations. Therefore, the feature encoding sequence can be represented in the normalized numerical form.

[0111] Step S1230: After normalizing the feature encoding sequence, store it in the value range corresponding to the corresponding key, construct key-value pairs of the corresponding substrings, and store them in the hash mapping table.

[0112] As mentioned earlier, the normalized feature encoding sequence of each key-value pair is stored in the value field of the corresponding key-value pair, thereby completing the construction of the key-value pair. The key-value pair then uses the substring of its key field as an index to establish a mapping relationship between the feature encoding information of the historical product item to which the hash fingerprint of the substring belongs in the corresponding segment. By storing these key-value pairs in the hash mapping table, the hash mapping storage of the feature encoding information of all historical product items in the user's historical behavior list can be completed, facilitating high-speed access.

[0113] As can be seen from the above embodiments, the hash storage technology is highly efficient and convenient to access by hashing and storing the feature encoding information of all historical product items in the user's historical behavior sequence. Its retrieval computation is low, which is suitable for the needs of big data with low computation.

[0114] Based on any embodiment of this application, please refer to Figure 7 The hash fingerprint of the target product item is hashed against the key-value pairs in the hash mapping table to determine multiple matching key-value pairs. The feature encoding information stored in the multiple key-value pairs is used to construct a user interest vector, including:

[0115] Step S1310: Segment the hash fingerprint of the target product item according to the preset word length to obtain multiple substrings;

[0116] To retrieve historical product items associated with the target product item from various historical product items in a user's historical behavior sequence, hash collisions based on hash fingerprints can be used for retrieval. Accordingly, following the method of segmenting the hash fingerprint in the historical behavior sequence described above, the hash fingerprint of the target product item is also segmented accordingly, resulting in multiple substrings.

[0117] Step S1320: Retrieve the normalized feature coding sequences corresponding to the multiple substrings from the hash mapping table and concatenate them into the same feature coding sequence;

[0118] After obtaining the substrings of each segment of the hash fingerprint of the target product item, the key-value pairs are retrieved in the hash mapping table on a segment-by-segment basis. Specifically, for each segment, the corresponding key-value pair is retrieved from the hash storage table, and then the feature encoding sequence within the value range is obtained from these key-value pairs. After obtaining the corresponding feature encoding sequences for different segments, they are concatenated together to form a single feature encoding sequence. This concatenated feature encoding sequence actually contains the feature encoding information of all relevant historical product items associated with the feature encoding information of the target product item.

[0119] Step S1330: Perform average pooling on the feature encoding sequence to obtain the user interest vector.

[0120] Considering that different segments may contain the same historical product item's feature encoding information, a pooling layer can be used to perform average pooling on the concatenated feature encoding sequence for standardization purposes, and the pooling result can be used as the user interest vector.

[0121] As can be seen from the above process, the feature encoding information of historical product items associated with the target product item is retrieved based on hash collision to construct the user interest vector. The entire retrieval is based on binary hash values, which has a very low computational load. This allows the hash fingerprint to be used to quickly match even if the user's historical behavior sequence contains a massive number of historical product items, and quickly identify the historical product items that are associated with the target product item, thus achieving an effective representation of the user's long-term interests.

[0122] Based on any embodiment of this application, please refer to Figure 8 The user interest vector is classified and mapped to determine the relevance score of the target product item, including:

[0123] Step S1410: Obtain the user's profile feature information and / or access scene feature information as external feature information;

[0124] Please see Figure 9 The extended network architecture of the product item recommendation rating model shown can further introduce one or two feature information through a third and / or fourth encoding layer, or similarly extend to include more feature information, as external feature information for determining the relevance score of the target product item. Specifically, this can be the user's profile feature information and / or the user's access scenario feature information when performing the access behavior. The profile feature information can be composed of the encoded features of the user's personal information such as age, gender, and interest tags, and can be flexibly set. The scenario feature information can be composed of information such as the ID of the page the user accessed and the geographical location of the terminal device, and can also be determined as needed. Through the profile feature information and / or access scenario feature information, further auxiliary information can be provided for the relevance score of the target product item, so that the determined relevance score can be more comprehensively considered to obtain a more accurate quantitative effect.

[0125] Step S1420: Concatenate the external feature information with the user interest vector to form user feature information;

[0126] To ensure that the external feature information plays its proper role, in the product recommendation rating model, the user interest vector obtained in the previous embodiments can be concatenated with the external feature information to form the same user feature information, which is then sent to the classifier for classification mapping.

[0127] Step S1430: After fully connecting the user feature information, classify and map it to a specific category to obtain the classification probability corresponding to the specific category as the relevance score of the target product item.

[0128] In the classifier, the user feature information is fully connected and activated, and then mapped to a preset classification space in the output layer. Finally, in a specific category, that is, the positive category, the corresponding classification probability is obtained. This classification probability can then be used as the relevance score of the target product item.

[0129] It is easy to understand that in the above embodiments, the external feature information superimposed on the user interest vector enriches the semantics required for the classifier to perform classification mapping, thereby enabling the acquisition of a more accurate relevance score.

[0130] Based on any embodiment of this application, please refer to Figure 10 After classifying and mapping the user interest vectors to determine the relevance score of the target product item, the process includes:

[0131] Step S1500: Sort the product items in reverse order based on the relevance scores obtained by taking each product item in the candidate product list as the target product item in turn;

[0132] The target product item can be a product item in a candidate product list. The candidate product list can be obtained by recalling all product items in the online store of an independent website through one or more recall channels. Each recall channel uses the user's relevant information, such as their feature encoding information or personal feature information, to recall all product items in the online store according to a certain strategy to obtain its corresponding recall product set. Multiple recall product sets are merged after deduplication to obtain the candidate product list. If necessary, the candidate product list can also be filtered and / or truncated according to a certain preset strategy.

[0133] For each item in the candidate product list, any of the embodiments described above can be applied to obtain its corresponding relevance score, which characterizes its relevance to the user's long-term interests. Therefore, the items in the candidate product list can be sorted in reverse order based on the relevance score, so that items with higher relevance scores are ranked higher.

[0134] In one embodiment, considering the necessity of truncating the candidate product list, the first few items can be obtained according to a preset quantity, or the first few items with high relevance scores can be selected according to a preset threshold, so that only the recommended product items are retained in the candidate product list.

[0135] Step S1600: The promotional information of each product item is constructed in order according to the reverse sorted product items and pushed to the user's terminal device for display.

[0136] After optimizing the composition and sorting relationship of the product items in the candidate product list according to the previous step, corresponding promotional information can be constructed for each product item. For example, the promotional information may include the product image, product title, product price, product link, etc. of the corresponding product item. These promotional information are constructed into a promotional list according to the same sorting relationship, and then pushed to the user's terminal device for parsing and display.

[0137] After browsing the promotional information of each product item in the promotional list, the user can access the product item according to their wishes and directly reach the corresponding details page.

[0138] As can be seen from the above embodiments, since the relevance score of each product item can be quickly determined according to the various embodiments of this application, product items that match the user's long-term interests can be quickly identified from a massive number of product items, thus achieving effective product promotion. For e-commerce platforms, when implementing product recommendations, they can better match user needs, help increase the total transaction volume, and achieve economies of scale.

[0139] Based on any embodiment of this application, the product item recommendation and rating method described in this application is executed by a deep learning model implemented based on a neural network. The deep learning model can be designed with its corresponding network architecture as needed, such as... Figure 3 and Figure 9 As an example, each deep learning model of a network architecture should be pre-trained using the corresponding training dataset. After training it to a convergent state, it should learn the ability to obtain the relevance score corresponding to the target product item based on the given user's historical behavior sequence and the feature encoding information of the specified target product item. Then, it can be deployed and run to participate in the implementation of product item recommendation services.

[0140] Please see Figure 11 According to one aspect of this application, a product item recommendation and rating device includes a hash representation module 1100, a hash mapping module 1200, an interest construction module 1300, and a correlation rating module 1400. The hash representation module 1100 is configured to acquire feature encoding information of each historical product item and the feature encoding information of a target product item from a user's historical behavior sequence, and convert each feature encoding information into a corresponding hash fingerprint. The hash mapping module 1200 is configured to convert each hash fingerprint in the historical behavior sequence into a hash mapping table, the hash mapping table containing multiple key-value pairs, each key-value pair storing feature encoding information corresponding to the hash fingerprint carrying its corresponding key. The interest construction module 1300 is configured to perform hash collision between the hash fingerprint of the target product item and the key-value pairs in the hash mapping table to determine multiple matching key-value pairs, and construct a user interest vector from the feature encoding information stored in the multiple key-value pairs. The correlation rating module 1400 is configured to classify and map the user interest vector to determine the correlation rating of the target product item.

[0141] Based on any embodiment of this application, the hash representation module 1100 includes: a hyperplane acquisition unit, configured to acquire a corresponding number of hyperplane matrices according to a preset number of bits of the hash fingerprint; a similarity calculation unit, configured to calculate the dot product of each of the aforementioned feature encoding information and each of the aforementioned hyperplane matrices, and binarize the dot product result to obtain the binarized values ​​corresponding to each hyperplane matrix; and a fingerprint generation unit, configured to construct the corresponding hash fingerprint by orderly constructing the binarized values ​​of each of the aforementioned feature encoding information corresponding to each of the aforementioned hyperplane matrices.

[0142] Based on any embodiment of this application, the hash mapping module 1200 includes: a substring segmentation unit, configured to segment the hash fingerprints of each historical product item in the historical behavior sequence according to a preset word length to obtain multiple substrings; a substring aggregation unit, configured to use the same substring in each segment as a key to concatenate the feature encoding information corresponding to the hash fingerprints of the corresponding segments with the same substring into a feature encoding sequence; and a key-value pair construction unit, configured to normalize the feature encoding sequence and store it in the value field corresponding to the corresponding key to construct key-value pairs for the corresponding substrings, and store them in the hash mapping table.

[0143] Based on any embodiment of this application, the interest construction module 1300 includes: a fingerprint segmentation unit, configured to segment the hash fingerprint of the target product item according to a preset word length to obtain multiple substrings; a fingerprint retrieval unit, configured to retrieve the normalized feature encoding sequences corresponding to the multiple substrings from the hash mapping table and concatenate them into the same feature encoding sequence; and a vector construction unit, configured to perform an average pooling operation on the feature encoding sequence to obtain the user interest vector.

[0144] Based on any embodiment of this application, the relevant scoring module 1400 includes: an external information acquisition unit, configured to acquire the user's profile feature information and / or access scenario feature information as external feature information; a vector concatenation unit, configured to concatenate the external feature information with the user's interest vector to obtain user feature information; and a classification scoring unit, configured to fully connect the user feature information and classify it to a specific category, and obtain the classification probability corresponding to the specific category as the relevance score of the target product item.

[0145] Based on any embodiment of this application, the product item recommendation and scoring device of this morning shift includes: a sorting processing module, configured to sort the product items in reverse order according to the relevance scores obtained by each product item in the candidate product list as the target product item; and a push display module, configured to construct the promotion information of each product item in order according to the reverse-sorted product items and push it to the user's terminal device for display.

[0146] Based on any embodiment of this application, the product item recommendation and rating device is constructed and implemented by running a deep learning model based on a neural network.

[0147] Another embodiment of this application also provides a product item recommendation and rating device. For example... Figure 12 The diagram shows the internal structure of a product recommendation and rating device. This device includes a processor, a computer-readable storage medium, a memory, and a network interface connected via a system bus. The computer-readable, non-volatile storage medium stores an operating system, a database, and computer-readable instructions. The database stores information sequences, and when executed by the processor, these computer-readable instructions enable the processor to implement a product recommendation and rating method.

[0148] The processor of the product recommendation and rating device provides computing and control capabilities, supporting the operation of the entire device. The device's memory can store computer-readable instructions, which, when executed by the processor, cause the processor to perform the product recommendation and rating method of this application. The device's network interface is used for communication with a terminal.

[0149] Those skilled in the art will understand that Figure 12 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the product item recommendation and scoring device to which the present application is applied. A specific product item recommendation and scoring device may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0150] In this embodiment, the processor is used to execute... Figure 11 The specific functions of each module are described, and the memory stores the program code and various data required to execute the above modules or sub-modules. The network interface is used to realize data transmission between user terminals or servers. The non-volatile readable storage medium in this embodiment stores the program code and data required to execute all modules in the product item recommendation and rating device of this application, and the server can call the server's program code and data to execute the functions of all modules.

[0151] This application also provides a non-volatile readable storage medium storing computer-readable instructions, which, when executed by one or more processors, cause the one or more processors to perform the steps of the product item recommendation rating method of any embodiment of this application.

[0152] This application also provides a computer program product, including a computer program / instructions that, when executed by one or more processors, implement the steps of the method described in any embodiment of this application.

[0153] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments of this application can be implemented by a computer program instructing related hardware. This computer program can be stored in a non-volatile readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The aforementioned storage medium can be a computer-readable storage medium such as a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM).

[0154] In summary, this application can process all historical user behavior data, effectively modeling long-term user interests while ensuring inference efficiency, and balancing performance and effectiveness metrics. Because it uses full behavioral data, sequence truncation is unnecessary, effectively avoiding the pitfalls of local optima. Furthermore, it eliminates the need to manually set truncation lengths to balance performance and effectiveness, providing strong assurance for online service stability. When applied to product ranking in search scenarios, this application's technical solution can identify user click intent when a user enters the site for a search, allowing for refined ranking of products retrieved by the engine, placing products of user interest at the top and improving business metrics.

Claims

1. A product item recommendation and rating method, characterized in that, include: Obtain the feature encoding information of each historical product item and the feature encoding information of the target product item in the user's historical behavior sequence, and convert each feature encoding information into a corresponding hash fingerprint; Each hash fingerprint in the historical behavior sequence is converted into a hash mapping table, which contains multiple key-value pairs. The value of each key-value pair stores the feature encoding information corresponding to the hash fingerprint carrying its corresponding key. The hash fingerprint of the target product item is hash-collided with the key-value pairs in the hash mapping table to determine multiple matching key-value pairs. The feature encoding information stored in the multiple key-value pairs is then used to construct a user interest vector. This process includes: segmenting the hash fingerprint of the target product item according to a preset word length to obtain multiple substrings; retrieving the normalized feature encoding sequences corresponding to the multiple substrings from the hash mapping table and concatenating them into a single feature encoding sequence; and performing an average pooling operation on the feature encoding sequence to obtain the user interest vector. The process of classifying and mapping the user interest vector to determine the relevance score of the target product item includes: obtaining the user's profile feature information and / or access scenario feature information as external feature information; concatenating the external feature information with the user interest vector to form user feature information; and fully connecting and classifying the user feature information to a specific category to obtain the classification probability corresponding to the specific category as the relevance score of the target product item.

2. The product item recommendation and rating method according to claim 1, characterized in that, Each feature encoding information is converted into a corresponding hash fingerprint, including: Obtain the corresponding number of hyperplane matrices based on the preset number of bits in the hash fingerprint; Calculate the dot product between each of the aforementioned feature encoding information and each of the aforementioned hyperplane matrices, and binarize the dot product results to obtain the corresponding binarized values ​​of each hyperplane matrix; The binary values ​​of each hyperplane matrix corresponding to each of the aforementioned feature encoding information are ordered to construct the corresponding hash fingerprint.

3. The product item recommendation and rating method according to claim 1, characterized in that, Converting each hash fingerprint in the historical behavior sequence into a hash mapping table includes: The hash fingerprints of each historical product item in the historical behavior sequence are segmented according to a preset word length to obtain multiple substrings; Using the common substring of each segment as the key, the feature encoding information corresponding to the hash fingerprints of the corresponding segments with the same substring is concatenated into a feature encoding sequence; The feature encoding sequence is normalized and stored in the value range corresponding to the key, forming key-value pairs of the corresponding substrings, which are then stored in the hash mapping table.

4. The product item recommendation and rating method according to any one of claims 1 to 3, characterized in that, After classifying and mapping the user interest vectors to determine the relevance score of the target product item, the process includes: The product items are sorted in reverse order based on the relevance scores obtained by taking each product item in the candidate product list as the target product item in turn. The promotional information for each product item is constructed in reverse order and pushed to the user's terminal device for display.

5. The product item recommendation and rating method according to any one of claims 1 to 3, characterized in that, The product recommendation and rating method is executed by a deep learning model based on a neural network.

6. A product item recommendation and rating device, characterized in that, include: The hash representation module is configured to obtain the feature encoding information of each historical product item and the feature encoding information of the target product item in the user's historical behavior sequence, and convert each feature encoding information into a corresponding hash fingerprint; The hash mapping module is configured to convert each hash fingerprint in the historical behavior sequence into a hash mapping table. The hash mapping table contains multiple key-value pairs, and the value of each key-value pair stores the feature encoding information corresponding to the hash fingerprint carrying its corresponding key. The interest construction module is configured to perform hash collisions between the hash fingerprint of the target product item and the key-value pairs in the hash mapping table to determine multiple matching key-value pairs, and construct a user interest vector from the feature encoding information stored in the multiple key-value pairs. This includes: segmenting the hash fingerprint of the target product item according to a preset word length to obtain multiple substrings; retrieving the normalized feature encoding sequences corresponding to the multiple substrings from the hash mapping table and concatenating them into a single feature encoding sequence; and performing average pooling on the feature encoding sequence to obtain the user interest vector. The relevant scoring module is configured to classify and map the user interest vector to determine the relevance score of the target product item, including: obtaining the user's profile feature information and / or access scenario feature information as external feature information; concatenating the external feature information with the user interest vector to form user feature information; and fully connecting the user feature information and classifying and mapping it to a specific category to obtain the classification probability corresponding to the specific category as the relevance score of the target product item.

7. The product item recommendation and scoring device according to claim 6, characterized in that, The hash representation module includes: The hyperplane acquisition unit is configured to acquire a corresponding number of hyperplane matrices based on a preset number of bits in the hash fingerprint; The similarity calculation unit is configured to calculate the dot product between each of the aforementioned feature encoding information and each of the aforementioned hyperplane matrices, and binarize the dot product results to obtain the corresponding binarized values ​​of each hyperplane matrix. The fingerprint generation unit is configured to construct a corresponding hash fingerprint by sequentially constructing the binarized values ​​of each hyperplane matrix corresponding to each feature encoding information.

8. The product item recommendation and scoring device according to claim 6 or 7, characterized in that, The hash mapping module includes: The substring segmentation unit is configured to segment the hash fingerprint of each historical product item in the historical behavior sequence according to a preset word length to obtain multiple substrings; The substring aggregation unit is set up to use the same substring in each segment as the key, and concatenate the feature encoding information corresponding to the hash fingerprints of the corresponding segments with the same substring into a feature encoding sequence; The key-value pair construction unit is configured to normalize the feature encoding sequence and store it in the value range corresponding to the key, construct key-value pairs of the corresponding substrings, and store them in the hash mapping table.

9. A product item recommendation and rating device, comprising a central processing unit and a memory, characterized in that, The central processing unit is used to invoke and run a computer program stored in the memory to perform the steps of the method as described in any one of claims 1 to 5.

10. A non-volatile readable storage medium, characterized in that, It stores a computer program in the form of computer-readable instructions, which, when invoked by a computer, performs the steps of the method as described in any one of claims 1 to 5.

Images