A commodity recommendation method and device, and a computer readable storage medium

Abstract

Embodiments of the present application disclose an article recommendation method and device, and a computer readable storage medium. The method comprises: obtaining historical dense features corresponding to an object, historical article features corresponding to the object, and article features corresponding to a plurality of to-be-recommended articles; extracting the article features and the historical article features by using an interest extraction model to determine object interest features; the interest extraction model represents the extraction of the object interest features; predicting and recommending the historical dense features, the object interest features and the article features by using a preset recommendation network to determine a recommendation result; and determining a recommended article to be recommended from the plurality of to-be-recommended articles based on the recommendation result. By using the historical article features and the article features of the to-be-recommended articles to determine the object interest features, the accuracy of the extraction of the object interest is improved, and suitable articles are recommended for the user.

Description

Technical Field

[0001] This invention relates to the field of computer applications, and more particularly to a product recommendation method and apparatus, and a computer-readable storage medium. Background Technology

[0002] Current item recommendation methods primarily focus on the items and their attributes within user behavior sequences. Taking a click sequence as an example, it includes the items the user clicked and their associated attributes such as brand, store, secondary category, and tertiary category. However, it cannot distinguish the strength of the user's interest in each action, resulting in a weak ability to extract user interests. Consequently, the correlation between extracted user interests and the items to be recommended is poor, making it impossible to recommend suitable items to customers. Summary of the Invention

[0003] The present invention aims to provide a method and apparatus for recommending items, as well as a computer-readable storage medium, which can accurately extract user interests, thereby improving the relevance between the extracted user interests and the items to be recommended, and thus recommending suitable items to the user.

[0004] The technical solution of this invention is implemented as follows:

[0005] This invention provides a method for recommending items, the method comprising:

[0006] Obtain the historical dense features corresponding to the object, the historical item features corresponding to the object, and the item features corresponding to each of the multiple items to be recommended;

[0007] An interest extraction model is used to extract features of the items to be recommended and features of historical items to determine the object's interest features; the interest extraction model represents the extracted object's interest features.

[0008] Using a pre-defined recommendation network, predictions and recommendations are made based on the historical dense features, the object interest features, and the features of the items to be recommended, and the recommendation results are determined.

[0009] Based on the recommendation results, a recommended item is determined from the plurality of items to be recommended.

[0010] In the above scheme, the characteristics of historical items include: multiple characteristics of historical sub-items;

[0011] In the above scheme, the step of extracting features of the item to be recommended and the features of historical items using an interest extraction model to determine the object's interest features includes:

[0012] The interest extraction module extracts the features of the item to be recommended and the features of the multiple historical sub-items to obtain the correlation between each pair of historical sub-item features and the correlation between the features of the item to be recommended and each historical sub-item feature.

[0013] The object interest features are determined based on the correlation between the pairwise historical sub-item features and the correlation between the features of the item to be recommended and each historical sub-item feature.

[0014] In the above scheme, the characteristics of each historical sub-item include: behavioral sequence characteristics, temporal characteristics, and fine behavioral attribute characteristics;

[0015] In the above scheme, the step of extracting the features of the item to be recommended and the features of the multiple historical sub-items through the interest extraction model to obtain the correlation between each pair of historical sub-item features and the correlation between the features of the item to be recommended and each historical sub-item feature includes:

[0016] By using the fine behavioral attribute features, behavioral sequence features, and temporal features corresponding to each historical sub-item feature, new fine behavioral attribute features and new behavioral sequence features corresponding to each historical sub-item feature are determined.

[0017] Based on the new refined behavioral attribute features corresponding to each historical sub-item feature and the new behavioral sequence features corresponding to each historical sub-item feature, pairwise correlation calculations are performed to obtain the correlation between the pairwise historical sub-item features;

[0018] The correlation between the features of the item to be recommended and each historical sub-item feature is determined by the new refined behavioral attribute features corresponding to each historical sub-item feature, the new behavioral sequence features corresponding to each historical sub-item feature, and the features of the item to be recommended.

[0019] In the above scheme, determining the new fine-grained behavioral attribute features and the new behavioral sequence features corresponding to each historical sub-item feature through the fine-grained behavioral attribute features, behavioral sequence features, and temporal features corresponding to each historical sub-item feature includes:

[0020] Using a pre-defined modeling network, the fine behavioral attribute features corresponding to each historical sub-item feature and the behavioral sequence features corresponding to each historical sub-item feature are fused to obtain new fine behavioral attribute features corresponding to each historical sub-item feature;

[0021] By performing a summation operation on the behavioral sequence feature and the temporal feature corresponding to each historical sub-item feature, a new behavioral sequence feature corresponding to each historical sub-item feature is obtained.

[0022] In the above scheme, the step of performing pairwise correlation operations based on the new refined behavioral attribute features corresponding to each historical sub-item feature and the new behavioral sequence features corresponding to each historical sub-item feature to obtain the correlation between the pairwise historical sub-item features includes:

[0023] By multiplying the new fine-grained behavioral attribute features corresponding to each pair of historical sub-item features, the correlation of fine-grained behavioral attribute features between the pairs of historical sub-item features is obtained.

[0024] By multiplying the new behavioral sequence features corresponding to each pair of historical sub-items by a dot product, the correlation of behavioral sequence features between the pairs of historical sub-items is obtained.

[0025] The correlation between the two pairs of historical item features is obtained by summing the correlation between the fine behavioral attribute features and the correlation between the behavioral sequence features of the two pairs of historical item features.

[0026] In the above scheme, determining the correlation between the features of the item to be recommended and each historical sub-item feature by using the new refined behavioral attribute features corresponding to each historical sub-item feature, the new behavioral sequence features corresponding to each historical sub-item feature, and the features of the item to be recommended includes:

[0027] Using a pre-defined modeling network, the new fine-grained behavioral attribute features are compressed to obtain the fine-grained behavioral feature correlation between the features of the item to be recommended and each historical sub-item feature.

[0028] By multiplying the new behavioral sequence feature and the feature of the item to be recommended, the correlation of behavioral sequence features between the feature of the item to be recommended and each historical sub-item feature is obtained;

[0029] The correlation between the features of the item to be recommended and each historical sub-item is obtained by summing the fine behavioral feature correlations between the features of the item to be recommended and each historical sub-item feature, and the behavioral sequence feature correlations between the features of the item to be recommended and each historical sub-item feature.

[0030] In the above scheme, before obtaining the historical dense features corresponding to the object, the historical item behavior and time features corresponding to the object, and the features of multiple items to be recommended, the method further includes:

[0031] Collect historical dense data corresponding to the object, historical item data corresponding to the object, and item data corresponding to each of the multiple items to be recommended;

[0032] Feature extraction and discretization are performed on the historical dense data, the historical item data, and the item data to be recommended to obtain the historical dense features corresponding to the object, the historical item features corresponding to the object, and the item features corresponding to each of the multiple items to be recommended.

[0033] In the above scheme, determining the recommended item from the plurality of items to be recommended based on the recommendation result includes:

[0034] Based on the recommendation results, the multiple items to be recommended are sorted to determine the recommended items.

[0035] This invention provides an item recommendation device, characterized in that the item recommendation device includes an acquisition unit and a determination unit;

[0036] The acquisition unit is used to acquire the historical dense features corresponding to the object, the historical item features corresponding to the object, and the recommended item features corresponding to each of the multiple recommended items.

[0037] The determining unit is used to extract the features of the item to be recommended and the features of the historical items using an interest extraction model to determine the object's interest features; the interest extraction model represents the extracted object's interest features; a preset recommendation network is used to predict and recommend the historical dense features, the object's interest features, and the features of the item to be recommended to determine the recommendation result; based on the recommendation result, a recommended item to be recommended is determined from the multiple items to be recommended.

[0038] This invention provides an item recommendation device, characterized in that the item recommendation device includes:

[0039] Memory, used to store executable data instructions;

[0040] A processor is configured to execute executable instructions stored in the memory, and when the executable instructions are executed, the processor executes the item recommendation method.

[0041] This invention provides a computer-readable storage medium, characterized in that it stores executable instructions, which, when executed by one or more processors, enable the processors to perform the item recommendation method.

[0042] This invention aims to provide a method and apparatus for item recommendation, as well as a computer-readable storage medium. The method includes: acquiring historical dense features corresponding to an object, historical item features corresponding to the object, and features corresponding to multiple items to be recommended; using an interest extraction model to extract the features of the items to be recommended and the historical item features to determine object interest features; the interest extraction model representing the extracted object interest features; using a preset recommendation network to predict and recommend based on the historical dense features, the object interest features, and the features of the items to be recommended, to determine a recommendation result; and based on the recommendation result, determining recommended items from the multiple items to be recommended. This solution utilizes the behavioral sequence features, temporal features, and fine-grained behavioral attribute features of an object browsing items to extract object interest features. Because the fine-grained behavioral attribute features of an object browsing items can reflect the strength of the object's interest in each browsing session, it can accurately obtain object interest features, thereby solving the problem of weak object interest extraction capabilities, improving the accuracy of object interest extraction, and increasing the relevance between object interest extraction and the items to be recommended, thus recommending suitable items to the user.

[0043] Figure 1 This invention provides an optional flowchart of an item recommendation method. Figure 1 ;

[0044] Figure 2 This invention provides an optional flowchart of an item recommendation method. Figure 2 ;

[0045] Figure 3 This invention provides an optional flowchart of an item recommendation method. Figure 3 ;

[0046] Figure 4 This invention provides an optional flowchart of an item recommendation method. Figure 4 ;

[0047] Figure 5 This is a schematic diagram of an optional model for an item recommendation method provided in an embodiment of the present invention;

[0048] Figure 6 The present invention provides an optional model flowchart for an item recommendation method.

[0049] Figure 7 Schematic diagram of the item recommendation device provided in the embodiments of the present invention Figure 1 ;

[0050] Figure 8 Schematic diagram of the item recommendation device provided in the embodiments of the present invention Figure 2 . Detailed Implementation

[0051] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0052] To enable those skilled in the art to better understand the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Figure 1 This is an optional flowchart illustrating an item recommendation method provided by an embodiment of the present invention. Figure 1 , will combine Figure 1 The steps shown are explained.

[0053] S101. Obtain the historical dense features corresponding to the object, the historical item features corresponding to the object, and the recommended item features corresponding to each of the multiple recommended items.

[0054] In some embodiments of the present invention, the item recommendation method is applicable to online shopping scenarios, the object can be a user, and the item can be a commodity; the embodiments of the present invention are not limited thereto; the historical dense feature is obtained by discretizing the obtained historical dense data; the historical item feature is obtained by discretizing the obtained historical item data; the item feature to be recommended is obtained by discretizing the obtained item to be recommended.

[0055] In some embodiments of the present invention, the terminal can obtain dense features, item features, and recommended item features corresponding to each of the multiple recommended items during the user's historical browsing of goods.

[0056] S102. Using an interest extraction model, extract the features of the items to be recommended and the features of historical items to determine the interest features of the target; the interest extraction model represents the extracted interest features of the target.

[0057] In some embodiments of the present invention, the interest extraction model is built on an attention mechanism and a deep learning transformer for feature extraction, mainly comprising an encoder and a decoder. The features of the items to be recommended include the features corresponding to each of the multiple items to be recommended.

[0058] In some embodiments of the present invention, the terminal can use an interest extraction model to extract features from the features of the items to be recommended and the features of historical items, thereby determining the user's interest features.

[0059] S103. Using a pre-set recommendation network, predict and recommend based on historical dense features, object interest features, and features of items to be recommended, and determine the recommendation results.

[0060] In some embodiments of this invention, the preset recommendation network is a multi-layer neural network, typically three layers. A multi-layer neural network is an algorithmic mathematical model that mimics the behavioral characteristics of animal neural networks to perform distributed parallel information processing. The model consists of a large number of nodes (or "neurons") and the connections between them. Each node represents a specific output function, called an activation function. The connection between any two nodes represents a weighted value for the signal passing through that connection, called a weight, which is equivalent to the memory of the artificial neural network. The network output varies depending on the network's connection method, weight values, and activation functions. The network itself is usually an approximation of a certain algorithm or function in nature, or it may be an expression of a logical strategy.

[0061] In some embodiments of the present invention, the terminal can fuse historical dense features, object interest features and features of items to be recommended, and then input them into a preset recommendation network to estimate the click-through rate of multiple items to be recommended and determine the recommendation result.

[0062] S104. Based on the recommendation results, determine the recommended items from multiple items to be recommended.

[0063] In some embodiments of the present invention, the recommendation result is the estimated click-through rate of multiple items to be recommended.

[0064] In some embodiments of the present invention, the terminal can determine the recommended item from the multiple recommended items by sorting the estimated click-through rates of multiple recommended items in the recommendation results.

[0065] It is understood that, in some embodiments of the present invention, the terminal can obtain the historical dense features corresponding to the object, the historical item features corresponding to the object, and the item features corresponding to each of the multiple items to be recommended; use an interest extraction model to extract the item features to be recommended and the historical item features to determine the object's interest features; the interest extraction model represents the extracted object's interest features; use a preset recommendation network to predict and recommend based on the historical dense features, object interest features, and item features to be recommended to determine the recommendation result; based on the recommendation result, determine the recommended item to be recommended from the multiple items to be recommended, thereby improving the accuracy of object interest extraction, improving the correlation between object interest extraction and the items to be recommended, and improving the accuracy of recommending items to the user.

[0066] In some embodiments of the present invention, step S104 can be implemented by S1041, as follows:

[0067] S1041. Based on the recommendation results, sort the multiple items to be recommended and determine the recommended items to be recommended.

[0068] In some embodiments of the present invention, the terminal can sort the expected click-through rates of each item to be recommended in the recommendation results and recommend the items with higher estimated click-through rates to the user.

[0069] It is understood that, in some embodiments of the present invention, the terminal can recommend items with a high estimated click-through rate from the recommendation results to the user, thereby improving the accuracy of the items recommended to the user.

[0070] In some embodiments of the present invention, the historical item features include: multiple historical sub-item features, each of which is different. Figure 2 This is an optional flowchart illustrating an item recommendation method provided by an embodiment of the present invention. Figure 2 Using an interest extraction model, features of the items to be recommended and features of historical items are extracted. Determining the object's interest features can be achieved through steps S201-S202, which will combine... Figure 2 The steps shown are explained.

[0071] S201. Using an interest extraction model, extract the features of the item to be recommended and the features of multiple historical sub-items to obtain the correlation between each pair of historical sub-item features and the correlation between the features of the item to be recommended and each historical sub-item feature.

[0072] In some embodiments of the present invention, the correlation between any two historical sub-item features is obtained by performing calculations on the corresponding information features generated during any two browsing sessions of the product, thus obtaining the relationship between the corresponding information features generated after any two browsing sessions. The correlation between the feature of the item to be recommended and each historical sub-item feature is obtained by performing calculations on the feature of the item to be recommended and each historical item feature, thus obtaining the relationship between the feature of the item to be recommended and each historical item feature.

[0073] In some embodiments of the present invention, the terminal can extract the features of the item to be recommended and multiple historical sub-item features using an interest extraction model, thereby obtaining the correlation between pairs of historical sub-item features and the correlation between the features of the item to be recommended and each historical sub-item feature.

[0074] S202. Based on the correlation between the features of each pair of historical sub-items and the correlation between the features of the item to be recommended and each historical sub-item feature, determine the interest features of the target.

[0075] In some embodiments of the present invention, the object interest features are extracted from information features generated when the user browses products in history and features of multiple items to be recommended, thereby extracting product features that the user is interested in.

[0076] In some embodiments of the present invention, the terminal can extract the product features that the user is interested in by the correlation between pairs of historical sub-item features and the correlation between the features of the item to be recommended and each historical sub-item feature.

[0077] It is understood that, in some embodiments of the present invention, the terminal can use an interest extraction model to extract the features of the item to be recommended and the features of multiple historical sub-items, thereby obtaining the correlation between each pair of historical sub-item features and the correlation between the features of the item to be recommended and each historical sub-item feature. Based on the correlation between each pair of historical sub-item features and the correlation between the features of the item to be recommended and each historical sub-item feature, the object's interest features can be determined, which can accurately extract the user's interest features and provide conditions for subsequently determining recommended products.

[0078] In some embodiments of the present invention, each historical sub-item feature includes: behavioral sequence features, temporal features, and fine behavioral attribute features. Figure 3 This is an optional flowchart illustrating an item recommendation method provided by an embodiment of the present invention. Figure 3 Through the interest extraction model, features of the item to be recommended and features of multiple historical sub-items are extracted. The correlation between each pair of historical sub-item features and the correlation between the features of the item to be recommended and each historical sub-item feature can be achieved by S301-S303, which will be combined with Figure 3 The steps shown are explained.

[0079] S301. By using the fine behavioral attribute features, behavioral sequence features, and temporal features corresponding to each historical sub-item feature, determine the new fine behavioral attribute features and the new behavioral sequence features corresponding to each historical sub-item feature.

[0080] In some embodiments of the present invention, taking click behavior as an example, refined behavioral attribute features are the user's behavioral patterns within the product details page. These include the number of clicks, the total number of operations, the maximum number of operations within a single click, the number of operations during the last click, the total dwell time on the product details page, the maximum dwell time within a single click, the dwell time during the last click, and the channel used to jump to the recommended position. It reflects the strength of the user's interest in each click. Behavioral sequence features are which products, brands, stores, second-level categories, and third-level categories the user has historically clicked. Temporal features are the time and order of each historical click.

[0081] In some embodiments of the present invention, the terminal can calculate new fine behavioral attribute features corresponding to each historical sub-item feature by using fine behavioral attribute features corresponding to each historical sub-item feature and temporal features corresponding to each historical sub-item feature; and calculate new behavioral sequence features corresponding to each historical sub-item feature by using behavioral sequence features corresponding to each historical sub-item feature and temporal features corresponding to each historical sub-item feature.

[0082] In some embodiments of the present invention, step S301 can be implemented by S3011 and S3012, as follows:

[0083] S3011. Using a pre-defined modeling network, the fine behavioral attribute features corresponding to each historical sub-item feature and the behavioral sequence features corresponding to each historical sub-item feature are fused to obtain new fine behavioral attribute features corresponding to each historical sub-item feature.

[0084] In some embodiments of the present invention, the preset modeling network is a multi-layer neural network, typically with three layers. A multi-layer neural network is an algorithmic mathematical model that mimics the behavioral characteristics of animal neural networks to perform distributed parallel information processing. It includes an input layer, hidden layers, and an output layer. Because user behavior patterns differ—for example, some users like to compare prices across multiple vendors and spend a relatively long time on all products—the same dwell time represents different levels of interest for different users. Furthermore, product characteristics vary; for instance, users tend to spend longer and perform more actions when purchasing high-priced, standardized products such as large home appliances. Therefore, fine-grained behavioral attribute features and behavioral sequence features are cross-referenced to obtain new fine-grained behavioral attribute features.

[0085] In some embodiments of the present invention, the terminal can use a preset modeling network to perform fusion processing on the fine behavioral attribute features corresponding to each historical sub-item feature and the behavioral sequence features corresponding to each historical sub-item feature to obtain new fine behavioral features corresponding to each historical sub-item feature.

[0086] For example, when calculating the new fine behavioral features corresponding to each historical sub-item feature using the fine behavioral attribute features and behavioral sequence features corresponding to each historical sub-item feature, the following formula (1) can be used.

[0087] embedding Interaction =MLP(embedding) behavior ,embedding sku (1)

[0088] Where embedding represents vectorized representation; embeddingInteraction It is a new, refined behavioral attribute feature corresponding to the characteristics of each historical sub-item; embedding behavior It refers to the fine-grained behavioral attribute features corresponding to the characteristics of each historical sub-item; embedding sku It refers to the behavioral sequence features corresponding to the features of each historical sub-item; MLP stands for Preset Modeling Network.

[0089] S3012. Based on the behavioral sequence features and temporal features corresponding to each historical sub-item feature, perform a summation operation to obtain a new behavioral sequence feature corresponding to each historical sub-item feature.

[0090] In some embodiments of the present invention, the influence of the time and order of each click on the behavioral sequence features is considered. Therefore, the behavioral sequence features and temporal features are processed to obtain new refined behavioral attribute features.

[0091] In some embodiments of the present invention, the terminal can add the behavioral sequence feature corresponding to each historical sub-item feature and the temporal feature corresponding to each historical sub-item feature to obtain a new behavioral sequence feature corresponding to each historical sub-item feature.

[0092] For example, when calculating the new behavior sequence feature corresponding to each historical sub-item feature using the behavior sequence feature and the temporal feature corresponding to each historical sub-item feature, the following formula (2) can be used.

[0093] embedding product =embedding sku +embedding position (2)

[0094] Where embedding represents vectorized representation; embedding product It is a new behavioral sequence feature corresponding to the features of each historical sub-item; embedding position It is the temporal feature corresponding to the feature of each historical sub-item; embedding sku It is the behavioral sequence feature corresponding to the feature of each historical sub-item.

[0095] It is understood that, in some embodiments of the present invention, the terminal can use a preset modeling network to fuse the fine behavioral attribute features corresponding to each historical sub-item feature and the behavioral sequence features corresponding to each historical sub-item feature to obtain new fine behavioral attribute features corresponding to each historical sub-item feature; and perform a summation operation based on the behavioral sequence features corresponding to each historical sub-item feature and the temporal features corresponding to each historical sub-item feature to obtain new behavioral sequence features corresponding to each historical sub-item feature, thereby updating the fine behavioral attribute features corresponding to each historical sub-item feature and the behavioral sequence features corresponding to each historical sub-item feature, and improving the accuracy of user interest extraction.

[0096] S302. Based on the new fine-grained behavioral attribute features corresponding to each historical sub-item feature and the new behavioral sequence features corresponding to each historical sub-item feature, perform pairwise correlation operations to obtain the correlation between pairs of historical sub-item features.

[0097] In some embodiments of the present invention, the terminal can calculate the correlation between fine behavioral attribute features of any two historical sub-items in the new fine behavioral attribute features corresponding to each historical sub-item feature; calculate the correlation between behavioral sequence features of any two historical sub-items in the new behavioral sequence features corresponding to each historical sub-item feature; and calculate the correlation between historical sub-item features by combining the fine behavioral attribute feature correlation between historical sub-item features and the behavioral sequence feature correlation between historical sub-item features.

[0098] In some embodiments of the present invention, step S302 can be implemented by S3021, S3022 and S3023, as follows:

[0099] S3021. By multiplying the new fine-grained behavioral attribute features corresponding to each pair of historical sub-item features, the correlation of fine-grained behavioral attribute features between each pair of historical sub-item features is obtained.

[0100] In some embodiments of the present invention, the terminal can perform a dot product operation on the new fine behavioral attribute features corresponding to each pair of historical sub-item features to obtain the correlation of fine behavioral attribute features between the pairs of historical sub-item features.

[0101] For example, when calculating the correlation between the fine behavioral attribute features of two pairs of historical item features by using the new fine behavioral attribute features corresponding to each pair of historical item features, the following formula (3) can be used.

[0102]

[0103] Where dot points to the dot product; It is a new fine-grained behavioral attribute feature corresponding to one of the historical sub-item features in a pair of historical sub-item features, where i represents one of the historical sub-item features in a pair of historical sub-item features; It is a new, refined behavioral attribute feature corresponding to another historical sub-item in a pair of historical sub-item features, where j represents the other historical sub-item feature in the pair of historical sub-item features; score behavior It is the correlation of fine behavioral attribute features between the sub-features of two historical items.

[0104] S3022. By multiplying the new behavioral sequence features corresponding to each pair of historical sub-item features, the correlation of behavioral sequence features between each pair of historical sub-items is obtained.

[0105] In some embodiments of the present invention, the terminal can perform a dot product operation on the new behavioral sequence features corresponding to each pair of historical sub-item features to obtain the behavioral sequence feature correlation between the pairs of historical sub-item features.

[0106] For example, when calculating the correlation of behavioral sequence features between two pairs of historical item features by using the new behavioral sequence features corresponding to each pair of historical item features, the following formula (4) can be used.

[0107]

[0108] Where dot points to the dot product; It is a new behavior sequence feature corresponding to one historical sub-item feature in a pair of historical sub-item features, where i represents one historical sub-item feature in a pair of historical sub-item features; It is a new behavioral sequence feature corresponding to another historical sub-item in a pair of historical sub-item features, where j represents the other historical sub-item feature in a pair of historical sub-item features; score Product It is the correlation of behavioral sequence features between the sub-features of two historical items.

[0109] S3023. Based on the correlation between the fine behavioral attribute features and the correlation between the behavioral sequence features of each pair of historical item sub-features, perform a summation operation to obtain the correlation between the features of each pair of historical item sub-features.

[0110] In some embodiments of the present invention, the terminal can add the correlation of fine behavioral attribute features between two pairs of historical sub-item features and the correlation of behavioral sequence features between two pairs of historical sub-item features to obtain the correlation between two pairs of historical sub-item features.

[0111] For example, when calculating the correlation between two historical item sub-features by means of the correlation between the fine behavioral attribute features between two historical item features and the correlation between the behavioral sequence features between two historical item sub-features, the correlation can be obtained by the following formula (5).

[0112] score ij =score behavior +score product (5)

[0113] Among them, score ij It represents the correlation between pairwise historical item sub-features, where i represents one historical item feature in a pairwise pair, and j represents the other historical item feature in a pairwise pair; score behavior It refers to the correlation of fine-grained behavioral attribute features between pairwise sub-features of historical items; score Product It is the correlation of behavioral sequence features between the sub-features of two historical items.

[0114] It is understood that, in some embodiments of the present invention, the terminal can obtain the correlation between the fine behavioral attribute features of each pair of historical sub-item features by multiplying the new fine behavioral attribute features corresponding to each pair of historical sub-item features; obtain the correlation between the behavioral sequence features of each pair of historical sub-item features by multiplying the new behavioral sequence features corresponding to each pair of historical sub-item features; and obtain the correlation between the two pairs of historical sub-item features by performing a summation operation based on the fine behavioral attribute feature correlation between the two pairs of historical sub-item features and the behavioral sequence feature correlation between the two pairs of historical sub-item features, thus laying the foundation for subsequently determining user interest features.

[0115] S303. Determine the correlation between the features of the item to be recommended and each historical sub-item feature by using the new refined behavioral attribute features corresponding to each historical sub-item feature, the new behavioral sequence features corresponding to each historical sub-item feature, and the features of the item to be recommended.

[0116] In some embodiments of the present invention, the terminal can calculate the fine behavioral feature correlation between the feature to be recommended and each historical sub-item feature by using the new fine behavioral attribute features corresponding to each historical sub-item feature; calculate the behavioral sequence feature correlation between the feature to be recommended and each historical sub-item feature by using the new behavioral sequence features corresponding to each historical sub-item feature and the feature to be recommended; and calculate the correlation between the feature to be recommended and each historical sub-item feature by using the fine behavioral feature correlation between the feature to be recommended and each historical sub-item feature and the behavioral sequence feature correlation between the feature to be recommended and each historical sub-item feature.

[0117] It is understood that, in some embodiments of the present invention, the terminal can determine new fine-grained behavioral attribute features and new behavioral sequence features corresponding to each historical sub-item feature by using the fine-grained behavioral attribute features, behavioral sequence features, and temporal features corresponding to each historical sub-item feature; based on the new fine-grained behavioral attribute features and new behavioral sequence features corresponding to each historical sub-item feature, pairwise correlation operations are performed to obtain the correlation between each pair of historical sub-item features; by using the new fine-grained behavioral attribute features, the new behavioral sequence features, and the features of the item to be recommended, the correlation between the features of the item to be recommended and each historical sub-item feature is determined, thereby improving the accuracy of user interest extraction.

[0118] In some embodiments of the present invention, step S303 can be implemented by S3031, S3032 and S3033, as follows:

[0119] S3031. Using a pre-defined modeling network, the new fine-grained behavioral attribute features are compressed to obtain the fine-grained behavioral feature correlation between the features of the item to be recommended and the features of each historical sub-item.

[0120] In some embodiments of the present invention, the terminal can compress the new fine behavioral attribute features corresponding to each historical sub-item feature through a preset modeling network to obtain the fine behavioral feature correlation between the feature of the item to be recommended and each historical sub-item feature.

[0121] For example, when calculating the correlation between the refined behavioral features of the item to be recommended and a historical sub-item feature using the new refined behavioral attribute features corresponding to a historical sub-item feature, it can be obtained by the following formula (6).

[0122]

[0123] in, It is a new fine-grained behavioral attribute feature corresponding to one historical sub-item feature in a pair of historical sub-item features, where i represents one historical sub-item feature in a pair of historical sub-item features; MLP refers to the pre-defined modeling network. It is the fine-grained behavioral feature correlation between the features of the item to be recommended and the features of a historical sub-item.

[0124] S3032. By multiplying the new behavioral sequence features and the features of the item to be recommended, the correlation of behavioral sequence features between the features of the item to be recommended and the features of each historical sub-item is obtained.

[0125] In some embodiments of the present invention, the terminal can perform a dot product operation on the new behavioral sequence features corresponding to each historical sub-item feature and the features of the item to be recommended, and then compress the data to obtain the behavioral sequence feature correlation between the features of the item to be recommended and each historical sub-item feature.

[0126] For example, when calculating the correlation between the behavioral sequence features of the item to be recommended and a historical sub-item feature using the new behavioral sequence features corresponding to a historical sub-item feature and the item to be recommended feature, the following formula (7) can be used.

[0127]

[0128] Where dot points to the dot product; It is a new behavioral sequence feature corresponding to one historical sub-item feature in each pair of historical sub-item features, where i represents one historical sub-item feature in each pair of historical sub-item features; embedding candidate These are the characteristics of the items to be recommended; It is the correlation of behavioral sequence features between the features of the item to be recommended and the features of a historical sub-item.

[0129] S3033. Based on the correlation of fine behavioral features between the features of the item to be recommended and the features of each historical sub-item, and the correlation of behavioral sequence features between the features of the item to be recommended and the features of each historical sub-item, perform a summation operation to obtain the correlation between the features of the item to be recommended and the features of each historical sub-item.

[0130] In some embodiments of the present invention, the terminal can add the fine behavioral feature correlation between the feature of the item to be recommended and each historical sub-item feature, and the behavioral sequence feature correlation between the feature of the item to be recommended and each historical sub-item feature, to obtain the correlation between the feature of the item to be recommended and each historical sub-item feature.

[0131] For example, when calculating the correlation between the feature to be recommended and a historical sub-item feature by means of the fine behavioral feature correlation between the feature to be recommended and a historical sub-item feature, and the behavioral sequence feature correlation between the feature to be recommended and a historical sub-item feature, the following formula (8) can be used.

[0132]

[0133] in, It is the fine-grained behavioral feature correlation between the features of the item to be recommended and the features of a historical sub-item; It is the correlation of behavioral sequence features between the features of the item to be recommended and the features of a historical sub-item; score iIt is the correlation between the characteristics of the item to be recommended and the characteristics of a historical sub-item.

[0134] It is understood that, in some embodiments of the present invention, the terminal can calculate the correlation between the fine behavioral features of the item to be recommended and each historical sub-item feature by using the new fine behavioral attribute features corresponding to each historical sub-item feature; calculate the correlation between the behavioral sequence features of the item to be recommended and each historical sub-item feature by using the new behavioral sequence features corresponding to each historical sub-item feature and the item to be recommended feature; and calculate the correlation between the item to be recommended and each historical sub-item feature by using the fine behavioral feature correlation between the item to be recommended and each historical sub-item feature, and the behavioral sequence feature correlation between the item to be recommended and each historical sub-item feature, thereby providing conditions for subsequently determining user interest features and improving the accuracy of user interest extraction.

[0135] In some embodiments of the present invention, see Figure 4 , Figure 4 This is an optional flowchart illustrating the item recommendation method provided in this embodiment of the invention. Figure 4 Before step S104, steps S105-S106 are also included, as follows:

[0136] S105. Collect the historical dense data corresponding to the object, the historical item data corresponding to the object, and the item data corresponding to each of the multiple items to be recommended.

[0137] In some embodiments of the present invention, historical dense data refers to statistical information on products, users, and user-product interactions during historical shopping processes; historical item data refers to which products, brands, stores, secondary categories, and tertiary categories a user has historically clicked, and records the time and order of each historical click, as well as the detailed behavior of each historical click, including the duration of stay, number of operations, and redirection channels on the product details page; and the item data to be recommended includes the product category and detailed product information, such as functions, materials, etc.

[0138] In some embodiments of the present invention, the terminal can collect the user's historical dense data, the user's historical item data, and the item data to be recommended for each of the multiple items to be recommended, based on the records generated by the user browsing products.

[0139] S106. Perform feature extraction and discretization on the historical dense data, historical item data and the data of items to be recommended, to obtain the historical dense features corresponding to the object, the historical item features corresponding to the object, and the features of the items to be recommended corresponding to each of the multiple items to be recommended.

[0140] In some embodiments of the present invention, discretization is the process of transforming continuous features into discrete features to facilitate subsequent calculations.

[0141] In some embodiments of the present invention, the terminal can extract features by performing data analysis on historical dense data, historical item data, and data of items to be recommended. The 99th percentile of all values ​​in the data arranged from smallest to largest is selected as the upper limit, and the data is truncated to control the number of buckets. Then, the logarithm of the continuous features is taken to discretize the features, and the discretized features are represented by low-dimensional dense vectors. This vector can not only represent the information of the feature, but also reflect the similarity between features through the distance between vectors. Thus, the historical dense features corresponding to the object, the historical item features corresponding to the object, and the features of the items to be recommended corresponding to each of the multiple items to be recommended are obtained in vectorized representation.

[0142] It is understood that, in some embodiments of the present invention, the terminal can collect historical dense data corresponding to the object, historical item data corresponding to the object, and recommended item data corresponding to each of the multiple recommended items; perform feature extraction and discretization processing on the historical dense data, historical item data, and recommended item data to obtain historical dense features corresponding to the object, historical item features corresponding to the object, and recommended item features corresponding to each of the multiple recommended items, thereby providing conditions for subsequently extracting user interest features and determining recommended items.

[0143] The following will describe an exemplary application of the embodiments of the present invention in a practical application scenario.

[0144] In online shopping scenarios, the terminal can recommend suitable items to users based on their historical item features, their historical dense features, and the individual features of multiple items to be recommended. Historical item features include behavioral sequence features, refined behavioral attribute features, and temporal features.

[0145] This invention provides an optional model diagram of an item recommendation method, as shown in the following embodiment. Figure 5 As shown in the figure, the features include behavioral sequence features, refined behavioral attribute features, temporal features, features of products to be ranked, dense features, and user interest features. Using an interest extraction model, features are extracted from the behavioral sequence features, refined behavioral attribute features, temporal features, and features of products to be ranked to obtain user interest features. Then, using a pre-defined recommendation network, recommendation predictions are made based on the dense features, user interest features, and features of products to be ranked to obtain click-through rate estimates, thereby determining the recommended products.

[0146] For example, an embodiment of the present invention provides an optional model flowchart of an item recommendation method, such as... Figure 6 As shown, the method for recommending this item may include the following steps:

[0147] 1. Offline feature collection.

[0148] The terminal can collect the user's dense characteristics, product characteristics, and the characteristics of each of the multiple products to be ranked by the user's browsing history.

[0149] It should be noted that the user's density features and product features are generated from the user's historical browsing of products. They are equivalent to the historical density data and historical item data of the object. The product features of multiple products to be ranked are equivalent to the item data of multiple items to be recommended.

[0150] 2. Feature preprocessing.

[0151] The terminal can perform binning on the user's dense features, product features, and the product features corresponding to each of the multiple products to be ranked. Then, the features are discretized, and finally, the discretized features are represented by vectors to obtain the vectorized dense features, product features, and product features corresponding to each of the multiple products to be ranked.

[0152] It should be noted that the vectorized dense features of users, product features, and the individual product features of multiple products to be ranked are equivalent to the historical dense features of an object, the historical item features of an object, and the individual item features of multiple items to be recommended.

[0153] 3. Extraction of user interests.

[0154] The terminal can determine user interest features by inputting the vectorized dense features of the user, the features of the product, and the features of the products to be ranked, respectively, into the interest extraction model.

[0155] 4. Click-through rate prediction.

[0156] The terminal uses a pre-set recommendation network to predict the click-through rate (CTR) of each product by analyzing user interest characteristics, density characteristics, and features of products to be ranked. The terminal then sorts the CTRs of these products and selects those with higher CTRs to push to the user.

[0157] Understandably, the terminal can acquire product features, features of products to be ranked, and dense features; then, the continuous features are binned and discretized, and the discretized features are vectorized; the preprocessed product features and features of products to be ranked are input together into the interest extraction model for coupling; finally, the user interest features, features of products to be ranked, and dense features are input into a preset recommendation network to obtain the click-through rate prediction and determine the recommended products; the above method extracts user interest features more accurately and solves the problem of insufficient accuracy in the extraction and screening of user interests.

[0158] Based on the item recommendation method of the above embodiments, this invention also provides an item recommendation device, such as... Figure 7 The above, Figure 7 This is a schematic diagram of an object re-identification device provided in an embodiment of the present invention. The device includes: an acquisition unit 701 and a determination unit 702; wherein,

[0159] The acquisition unit 701 is used to acquire the historical dense features corresponding to the object, the historical item features corresponding to the object, and the recommended item features corresponding to each of the multiple recommended items.

[0160] The determining unit 702 is used to extract the features of the item to be recommended and the features of the historical items using an interest extraction model to determine the object interest features; the interest extraction model represents the extracted object interest features; a preset recommendation network is used to predict and recommend the historical dense features, the object interest features and the features of the item to be recommended to determine the recommendation result; based on the recommendation result, a recommended item to be recommended is determined from the multiple items to be recommended.

[0161] In some embodiments of the present invention, the historical item features include: multiple historical sub-item features;

[0162] The acquisition unit 701 is used to extract the features of the item to be recommended and the features of the multiple historical sub-items through the interest extraction model, and obtain the correlation between each pair of historical sub-item features and the correlation between the features of the item to be recommended and each historical sub-item feature respectively.

[0163] The determining unit 702 is used to determine the object interest features based on the correlation between the pairwise historical sub-item features and the correlation between the features of the item to be recommended and each historical sub-item feature.

[0164] In some embodiments of the present invention, each historical sub-item feature includes: behavioral sequence features, temporal features, and refined behavioral attribute features;

[0165] The determining unit 702 is used to determine new fine behavioral attribute features and new behavioral sequence features corresponding to each historical sub-item feature by using the fine behavioral attribute features, behavioral sequence features and temporal features corresponding to each historical sub-item feature;

[0166] The acquisition unit 701 is used to perform pairwise correlation calculations based on the new fine behavioral attribute features corresponding to each historical sub-item feature and the new behavioral sequence features corresponding to each historical sub-item feature, to obtain the correlation between the pairwise historical sub-item features.

[0167] The determining unit 702 is used to determine the correlation between the feature to be recommended and each historical sub-item feature by using the new fine behavioral attribute feature corresponding to each historical sub-item feature, the new behavioral sequence feature corresponding to each historical sub-item feature, and the feature to be recommended.

[0168] In some embodiments of the present invention, the acquisition unit 701 is used to use a preset modeling network to fuse the fine behavioral attribute features corresponding to each historical sub-item feature and the behavioral sequence features corresponding to each historical sub-item feature to obtain new fine behavioral attribute features corresponding to each historical sub-item feature; and to perform a summation operation based on the behavioral sequence features corresponding to each historical sub-item feature and the temporal features corresponding to each historical sub-item feature to obtain new behavioral sequence features corresponding to each historical sub-item feature.

[0169] In some embodiments of the present invention, the acquisition unit 701 is used to obtain the correlation of fine behavioral attribute features between the pairs of historical item features by multiplying the new fine behavioral attribute features corresponding to each pair of historical item features; to obtain the correlation of behavioral sequence features between the pairs of historical item features by multiplying the new behavioral sequence features corresponding to each pair of historical item features; and to obtain the correlation between the pairs of historical item features by performing a summation operation based on the correlation of fine behavioral attribute features between the pairs of historical item features and the correlation of behavioral sequence features between the pairs of historical item features.

[0170] In some embodiments of the present invention, the acquisition unit 701 is used to compress the new fine-grained behavioral attribute features using a preset modeling network to obtain the fine-grained behavioral feature correlation between the feature of the item to be recommended and each historical sub-item feature; by multiplying the new behavioral sequence feature and the feature of the item to be recommended, the behavioral sequence feature correlation between the feature of the item to be recommended and each historical sub-item feature is obtained; based on the fine-grained behavioral feature correlation between the feature of the item to be recommended and each historical sub-item feature, and the behavioral sequence feature correlation between the feature of the item to be recommended and each historical sub-item feature, the correlation between the feature of the item to be recommended and each historical sub-item feature is obtained by performing a summation operation.

[0171] In some embodiments of the present invention, the item recommendation device further includes a collection unit 703;

[0172] The collection unit 703 is used to collect historical dense data corresponding to the object, historical item data corresponding to the object, and item data corresponding to each of the multiple items to be recommended.

[0173] The acquisition unit 701 is used to perform feature extraction and discretization processing on the historical dense data, the historical item data and the item data to be recommended, to obtain the historical dense features corresponding to the object, the historical item features corresponding to the object, and the item features corresponding to each of the multiple items to be recommended.

[0174] In some embodiments of the present invention, the determining unit 702 is used to sort the plurality of items to be recommended based on the recommendation results, and determine the recommended items to be recommended.

[0175] Based on the item recommendation method of the above embodiments, this invention also provides an item recommendation device, such as... Figure 8 As shown, Figure 8 This is a schematic diagram of an item recommendation device provided in an embodiment of the present invention. The device includes a processor 801 and a memory 802. The memory 801 stores one or more programs executable by the processor. When one or more programs are executed, the processor 802 executes any of the object re-identification methods described in the foregoing embodiments. Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of hardware embodiments, software embodiments, or embodiments combining software and hardware aspects. Moreover, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage and optical storage) containing computer-usable program code.

[0176] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0177] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0178] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0179] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention.

Claims

1. An article recommendation method characterized by comprising: The method includes: The system obtains the historical dense features corresponding to the object, the historical item features corresponding to the object, and the features of each of the multiple items to be recommended. The historical item features include multiple historical sub-item features, each of which includes behavioral sequence features and refined behavioral attribute features. The refined behavioral attribute features are the object's behavioral patterns within the item details page. The behavioral sequence features include the items, brands, stores, second-level categories, and third-level categories that the object has historically clicked. An interest extraction model is used to extract features of the items to be recommended and features of historical items to determine the object's interest features; the interest extraction model represents the extracted object's interest features. Using a pre-defined recommendation network, predictions and recommendations are made based on the historical dense features, the object interest features, and the features of the items to be recommended, and the recommendation results are determined. Based on the recommendation results, a recommended item is determined from the plurality of items to be recommended; The step of using an interest extraction model to extract features from the items to be recommended and the features from historical items to determine the object's interest features includes: The interest extraction model is used to extract the features of the item to be recommended and the features of the multiple historical sub-items, thereby obtaining the correlation between each pair of historical sub-item features and the correlation between the features of the item to be recommended and each historical sub-item feature. Based on the correlation between the pairwise historical sub-item features and the correlation between the features of the item to be recommended and each historical sub-item feature, the object interest features are determined; Each historical sub-item feature also includes: temporal features; the process of extracting the features of the item to be recommended and the multiple historical sub-item features through the interest extraction model to obtain the correlation between each pair of historical sub-item features and the correlation between the features of the item to be recommended and each historical sub-item feature, including: Using a pre-defined modeling network, the fine behavioral attribute features corresponding to each historical sub-item feature and the behavioral sequence features corresponding to each historical sub-item feature are fused to obtain new fine behavioral attribute features corresponding to each historical sub-item feature; Based on the summation operation of the behavior sequence feature and the temporal feature corresponding to each historical sub-item feature, a new behavior sequence feature corresponding to each historical sub-item feature is obtained; Based on the new refined behavioral attribute features corresponding to each historical sub-item feature and the new behavioral sequence features corresponding to each historical sub-item feature, pairwise correlation calculations are performed to obtain the correlation between the pairwise historical sub-item features; The correlation between the features of the item to be recommended and each historical sub-item feature is determined by the new refined behavioral attribute features corresponding to each historical sub-item feature, the new behavioral sequence features corresponding to each historical sub-item feature, and the features of the item to be recommended.

2. The method of claim 1, wherein, The process involves performing pairwise correlation operations based on the new refined behavioral attribute features corresponding to each historical sub-item feature and the new behavioral sequence features corresponding to each historical sub-item feature to obtain the correlation between each pair of historical sub-item features, including: By multiplying the new fine-grained behavioral attribute features corresponding to each pair of historical sub-item features, the correlation of fine-grained behavioral attribute features between the pairs of historical sub-item features is obtained. By multiplying the new behavioral sequence features corresponding to each pair of historical sub-items by a dot product, the correlation of behavioral sequence features between the pairs of historical sub-items is obtained. The correlation between the two pairs of historical sub-item features is obtained by summing the correlation of fine behavioral attribute features between the two pairs of historical sub-item features and the correlation of behavioral sequence features between the two pairs of historical sub-item features.

3. The method according to claim 1, characterized in that, The process of determining the correlation between the features of the item to be recommended and each historical sub-item feature by using the new refined behavioral attribute features corresponding to each historical sub-item feature, the new behavioral sequence features corresponding to each historical sub-item feature, and the features of the item to be recommended includes: Using a pre-defined modeling network, the new fine-grained behavioral attribute features are compressed to obtain the fine-grained behavioral feature correlation between the features of the item to be recommended and each historical sub-item feature. By multiplying the new behavioral sequence feature and the feature of the item to be recommended, the correlation of behavioral sequence features between the feature of the item to be recommended and each historical sub-item feature is obtained; The correlation between the features of the item to be recommended and each historical sub-item is obtained by summing the fine behavioral feature correlations between the features of the item to be recommended and each historical sub-item feature, and the behavioral sequence feature correlations between the features of the item to be recommended and each historical sub-item feature.

4. The method according to any one of claims 1 to 3, characterized in that, Before obtaining the historical dense features corresponding to the object, the historical item behavior and time features corresponding to the object, and multiple features of items to be recommended, the method further includes: Collect historical dense data corresponding to the object, historical item data corresponding to the object, and item data corresponding to each of the multiple items to be recommended; Feature extraction and discretization are performed on the historical dense data, the historical item data, and the item data to be recommended to obtain the historical dense features corresponding to the object, the historical item features corresponding to the object, and the item features corresponding to each of the multiple items to be recommended.

5. The method according to any one of claims 1 to 3, characterized in that, The step of determining the recommended item from the plurality of items to be recommended based on the recommendation result includes: Based on the recommendation results, the multiple items to be recommended are sorted to determine the recommended items.

6. An item recommendation device, characterized in that, include: The acquisition unit is used to acquire the historical dense features corresponding to the object, the historical item features corresponding to the object, and the item features corresponding to each of the multiple items to be recommended. The historical item features include: multiple historical sub-item features, each historical sub-item feature including: behavior sequence features and refined behavior attribute features, the refined behavior attribute features being the object's behavior pattern within the item details page, and the behavior sequence features including the items, brands, stores, second-level categories, and third-level categories that the object has historically clicked; The determining unit is used to extract the features of the item to be recommended and the features of the historical items using an interest extraction model to determine the object interest features; the interest extraction model represents the extracted object interest features; a preset recommendation network is used to predict and recommend the historical dense features, the object interest features and the features of the item to be recommended, and to determine the recommendation result; based on the recommendation result, a recommended item to be recommended is determined from the multiple items to be recommended. The determining unit is further configured to extract the features of the item to be recommended and the features of the multiple historical sub-items through the interest extraction model, and obtain the correlation between each pair of historical sub-item features and the correlation between the features of the item to be recommended and each historical sub-item feature; and determine the object interest features based on the correlation between each pair of historical sub-item features and the correlation between the features of the item to be recommended and each historical sub-item feature. Each historical sub-item feature further includes: a temporal feature; the determining unit is further configured to use a preset modeling network to fuse the fine behavioral attribute features and the behavioral sequence features corresponding to each historical sub-item feature to obtain a new fine behavioral attribute feature corresponding to each historical sub-item feature; perform a summation operation on the behavioral sequence features and the temporal features corresponding to each historical sub-item feature to obtain a new behavioral sequence feature corresponding to each historical sub-item feature; perform pairwise correlation operations on the new fine behavioral attribute features and the new behavioral sequence features corresponding to each historical sub-item feature to obtain the correlation between the pairwise historical sub-item features; and determine the correlation between the feature to be recommended and each historical sub-item feature by using the new fine behavioral attribute features, the new behavioral sequence features, and the feature to be recommended.

7. An item recommendation device, characterized in that, include: Memory, used to store executable data instructions; A processor, when executing executable instructions stored in the memory, implements the method according to any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, It stores executable instructions for causing a processor to execute, thereby implementing the method of any one of claims 1 to 5.

Images