Feature processing method and device applied to commodity recommendation system

Abstract

This invention provides a feature processing method and apparatus for a product recommendation system. The method includes: constructing a feature decision tree, determining the significance coefficients of the root node, each child node, and each leaf node; selecting leaf nodes with significance coefficients greater than a first threshold, constructing paths from the selected leaf nodes to the root node, selecting paths whose average significance coefficient is greater than a second threshold, obtaining all nodes of each path to form a second path node set for each path, and selecting target nodes with significance coefficients greater than a third threshold from this set; combining the user behavior features corresponding to the target nodes selected for each path to form new synthetic features; wherein the new synthetic features are used to train the product recommendation system. By setting three thresholds, the filtering of leaf nodes, paths from leaf nodes to the root node, and feature nodes on the paths is introduced, effectively and significantly reducing the computational processing and storage overhead when the feature dimension is very large.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a feature processing method and apparatus for use in a product recommendation system. Background Technology

[0002] With the development of artificial intelligence technology, in the field of e-commerce, recommendation systems help e-commerce platforms push products or recommendations that customers are interested in, thereby providing users with a better shopping experience and enhancing user stickiness. Product recommendation systems are typically recommendation models, which generally consist of two stages: model training and model prediction. During the model training process for a recommendation system, the feature data in the sample data needs to undergo various processing steps. The processed feature data and label data are then used by the model algorithm to calculate, set, and adjust parameters, ultimately obtaining the recommendation model used for prediction. The process of processing these features is called feature engineering. Current feature engineering techniques typically select and combine multiple features to form new feature data for use.

[0003] Currently, existing technologies for selecting and combining multiple feature data include: Method 1, which involves multiplying a feature by itself or other features using vector products to form a new feature. Method 2, when a feature data is a continuous numerical value with a large range, discretizes or reduces the dimensionality of the value using a bucketing approach to form a new feature. This method uses bucketing to discretize continuous numerical values, representing them as a new synthetic feature vector. The synthetic feature vector is represented using one-hot encoding to facilitate vector inner product multiplication.

[0004] However, the inventors discovered that the existing technology has at least the following technical problems: when the feature dimension is very large, it results in high computational and storage overhead for the server, affecting the processing speed. Summary of the Invention

[0005] This invention provides a feature processing method and apparatus for a product recommendation system. By setting specific thresholds to filter nodes and paths, new synthetic features are generated by combining the features corresponding to the filtered nodes. This can effectively and significantly reduce the computational and storage overhead when the feature dimension is very large, and speed up the processing.

[0006] In a first aspect, the present invention provides a feature processing method for a product recommendation system, comprising:

[0007] Obtain a sample dataset of a product recommendation system, and construct a feature decision tree based on the user behavior feature data and corresponding user behavior result label data in the sample dataset of the product recommendation system.

[0008] The significance coefficients of the root node, each child node, and each leaf node in the feature decision tree are determined respectively.

[0009] Leaf nodes with a significance coefficient greater than a first threshold are selected to form a leaf node set; by traversing each leaf node in the leaf node set, the path from each leaf node to the root node is constructed to obtain an initial path set;

[0010] By traversing all paths in the initial path set, a first path node set is obtained for each path. Paths whose average significance coefficient of the first path node set is greater than a second threshold are selected from the initial path set to form a target path set.

[0011] By traversing all paths in the target path set, a second path node set is obtained for each path. Target nodes with a significance coefficient greater than a third threshold are selected from the second path node set of each path. The user behavior features corresponding to the target nodes selected for each path are combined to form new synthetic features. The new synthetic features are used to train the product recommendation system.

[0012] In one possible design, determining the significance coefficients of the root node, each child node, and each leaf node in the feature decision tree includes:

[0013] For any node among the child nodes and leaf nodes, the ratio of the positive and negative sample ratio of the given node to the positive and negative sample ratio of the given node's parent node is determined as the significance coefficient of the given node.

[0014] The significance coefficient of the root node is set to a preset value.

[0015] In one possible design, constructing a feature decision tree based on the user behavior feature data and the corresponding user behavior result label data includes:

[0016] Using either the Iterative Dichotomiser 3 (ID3) algorithm or the Classification and Regression Trees (CART) algorithm, an initial feature decision tree is constructed based on the user behavior feature data and the corresponding user behavior result label data.

[0017] The initial feature decision tree is pruned using a pre-pruning and post-pruning strategy to obtain the feature decision tree.

[0018] In one possible design, the step of selecting paths from the initial path set whose average significance coefficient of the first path node set is greater than a second threshold includes: determining the average significance coefficient of the first path node set: obtaining the significance coefficients of each node in the first path node set of each path, including the root node, each child node, and each leaf node, and summing the significance coefficients of each node; and determining the average significance coefficient by taking the ratio of the sum of the significance coefficients of each node to the number of nodes in the first path node set, including the root node, each child node, and each leaf node.

[0019] In one possible design, after combining the user behavior features corresponding to the target nodes selected from each path to form synthetic features, the method further includes:

[0020] Based on the evaluation function, the optimal synthetic feature is selected from the user behavior features corresponding to the target nodes selected from all paths.

[0021] The product recommendation system is trained based on the optimal synthetic features.

[0022] Secondly, the present invention provides a feature processing apparatus for a product recommendation system, comprising:

[0023] The first construction module is used to obtain a sample dataset of the product recommendation system and construct a feature decision tree based on the user behavior feature data and the corresponding user behavior result label data in the sample dataset of the product recommendation system.

[0024] The determination module is used to determine the significance coefficients of the root node, each child node, and each leaf node in the feature decision tree, respectively.

[0025] The second construction module is used to select leaf nodes with a significance coefficient greater than the first threshold to form a leaf node set; by traversing each leaf node in the leaf node set, the path from each leaf node to the root node is constructed to obtain an initial path set.

[0026] The first filtering module is used to obtain a first path node set for each path by traversing all paths in the initial path set, and to filter paths from the initial path set whose average significance coefficient of the first path node set is greater than a second threshold to form a target path set.

[0027] The second filtering module is used to obtain a second path node set for each path by traversing all paths in the target path set, and to filter target nodes with a significance coefficient greater than a third threshold from the second path node set of each path.

[0028] The combination module is used to combine the user behavior features corresponding to the target nodes selected from each path to form new synthetic features; wherein the new synthetic features are used to train the product recommendation system.

[0029] In one possible design, the determining module is specifically used to determine the significance coefficient of any node by taking the ratio of the positive and negative sample ratio of any node to the positive and negative sample ratio of the parent node of any node among the child nodes and leaf nodes; and to set the significance coefficient of the root node to a preset value.

[0030] In one possible design, the first building module is specifically used to construct an initial feature decision tree based on the user behavior feature data and the corresponding user behavior result label data using the ID3 algorithm or the CART algorithm, and to prune the initial feature decision tree using pre-pruning and post-pruning strategies to obtain the feature decision tree.

[0031] Thirdly, the present invention provides a server, comprising: at least one processor and a memory;

[0032] The memory stores computer-executed instructions;

[0033] The at least one processor executes computer execution instructions stored in the memory, causing the at least one processor to perform the feature processing method for a product recommendation system as described in the first aspect and various possible designs of the first aspect.

[0034] Fourthly, the present invention provides a computer storage medium storing computer execution instructions, wherein when a processor executes the computer execution instructions, the feature processing method for a product recommendation system described in the first aspect and various possible designs of the first aspect is implemented.

[0035] The feature processing method and apparatus for product recommendation systems provided by this invention effectively and significantly reduce the computational and storage overhead when the feature dimension is very large by introducing leaf node filtering, path filtering from leaf node to root node, and feature node filtering on the path of the feature decision tree, thereby effectively improving the processing speed of the product recommendation system. Attached Figure Description

[0036] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0037] Figure 1 A flowchart illustrating the feature processing method for a product recommendation system provided in this embodiment of the invention. Figure 1 ;

[0038] Figure 2 This is a schematic diagram of a feature decision tree model provided in an embodiment of the present invention;

[0039] Figure 3 This is a schematic diagram of the feature processing device applied to a product recommendation system provided in an embodiment of the present invention;

[0040] Figure 4 This is a schematic diagram of the hardware structure of a server provided in an embodiment of the present invention. Detailed Implementation

[0041] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, 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 scope of protection of the present invention.

[0042] The collection, storage, use, processing, transmission, provision, and disclosure of user data and other information involved in the technical solution of this invention all comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0043] Due to the development of informatization and digitalization, as user data in systems such as e-commerce systems increases, the corresponding data on user purchasing behavior also surges. To provide users with more accurate and high-quality services, recommendation systems have been proposed. In the e-commerce field, recommendation systems help e-commerce platforms push product information and suggestions that customers are interested in, thereby increasing sales. Personalized product recommendation systems can be even more specific to the individual user's behavioral data, providing customers with fully personalized decision support and information services based on massive data mining. Recommendation systems generally consist of two stages: model training and prediction. During model training, various feature data from the sample data, such as user profiles, behavioral characteristics, and item characteristics, need to be processed in various ways. These features are then combined with the user's final behavioral data, such as the tag data representing the purchase results, and the model algorithm is used to calculate, set, and adjust parameters to ultimately obtain the recommendation model used for prediction. The process of processing these features is called feature engineering.

[0044] In the prior art, to reduce the processing overhead of feature engineering, multiple features are combined to form new feature data for use. In existing feature engineering, the implementation schemes for combining multi-dimensional features are as follows: Technique 1: New features are formed by multiplying a feature by itself or other features. For example, if the feature data is represented as x1 and x2, the newly formed feature, called the synthetic feature x3, where x1 has m values and x2 has n values, and x3 is or a vector of dimension, then the calculation relationship can be expressed as: x3 = x1 2 or x3 = . When the values of the feature data are very large, such as m and n being greater than 10000, the representation of the synthetic feature x3 is greater than 100000000, which is extremely large. When performing feature selection, to determine the relationship between the synthetic vector and the label data, at this time, the performance of vector calculation will be seriously affected. In addition to increasing the interference with the relationship of the label data, invalid feature combinations will also exacerbate the duration of model training. Technique 2: Since a certain feature data is continuous numerical and the value representation range is too large, by adopting the method of bucketing, the numerical value is discretized or dimension-reduced to form a new feature. For example, the continuous feature data is x1, and its value range is (m, n), m < n. Then, bucketing can be performed by using equal-frequency or equal-distance methods. For example, when using the equal-distance bucketing method, the width of the bucket is d, then x1 can be divided into (n - m) / d buckets, and the newly formed feature x3 can be represented as a one-dimensional vector containing d elements. For example, when a synthetic feature x3 formed by equal-distance bucketing is 4-dimensional, its representation form is as follows: [0, 1, 0, 0], and this vector indicates that the synthetic feature x3 falls into the second bucket. This technique represents the continuous numerical value as a new synthetic feature vector through the discretization method of bucketing, and the synthetic feature vector will be represented in the form of one-hot encoding to facilitate vector inner product multiplication processing. However, different bucketing methods will be affected by the number of feature samples in the bucket, which will bring great changes to the output of the label data. Second, the use of one-hot encoding makes the result of vector inner product a sparse matrix, which will occupy more storage.

[0045] To solve the above technical problems, the embodiments of the present invention propose the following technical solutions: By introducing the selection of leaf nodes of the feature decision tree, the selection of paths from leaf nodes to root nodes, and the selection of feature nodes on the paths, the screened features are combined to form new synthetic features, achieving effective and significant reduction of the calculation processing and storage overhead when the feature dimension is extremely large during feature processing.

[0046] Figure 1 Schematic diagram of the feature processing method applied to the commodity recommendation system provided by the embodiments of the present invention Figure 1In this embodiment, the executing entity can be a server; however, no special restrictions are imposed here. Figure 1 As shown, the method includes:

[0047] S101: Obtain the sample dataset of the product recommendation system, and construct a feature decision tree based on the user behavior feature data and the corresponding user behavior result label data in the sample dataset of the product recommendation system.

[0048] In this embodiment of the invention, user behavior feature data includes user profile features, behavioral features, and other information, such as the user's age, education level, and income. User behavior result tag data can be specific user behavior results, such as whether the user clicked on a push notification or whether the user purchased a pushed product.

[0049] Define a sample dataset D containing user behavior characteristics data related to a product recommendation system. sample The user behavior feature data includes user profile data and user behavior feature data, and the sample dataset D sample Divided into training set D train and verification set D valid Let the user behavior feature data of the sample dataset be X = {x1, x2, x3, ... x...} m};Sample user behavior result label data Y = {y1,y2,…y n The algorithm for the product recommendation system is defined as RS, and for the sample dataset, Y = RS(D). sample Let R(X) represent the product recommendation system algorithm RS as a model function of the user behavior feature data X in the sample dataset, and the result dataset of this function be Y. Therefore, the functional form of the recommendation system for the training set can be expressed as follows:

[0050] Y train = RS(D train , X),

[0051] The functional form of the recommender system for the validation set can be expressed as:

[0052] Y valid = RS(D valid , X).

[0053] In this embodiment of the invention, the ID3 algorithm or the CART algorithm is used to construct an initial feature decision tree based on user behavior feature data and corresponding user behavior result label data. The initial feature decision tree is then pruned using pre-pruning and post-pruning strategies to obtain the final feature decision tree.

[0054] Specifically, the feature decision tree obtained in this embodiment of the invention is a multi-layered tree structure diagram composed of root nodes, child nodes, and leaf nodes represented by different user behavior feature data and corresponding user behavior result label data. Figure 2 This is a schematic diagram of the feature decision tree model in this embodiment. Let the root node be represented as RNode, the child nodes as SNode, and the leaf nodes as LNode. For example... Figure 2 As shown, in a feature decision tree, the root node RNode is unique, while there can be multiple child nodes SNode and leaf nodes LNode. Furthermore, there must exist a connected path from LNode to RNode, denoted as Path. The set of all nodes along a path from a leaf node to the root node can be represented as: NSet = {LNode, Path}.

[0055] S102: Determine the significance coefficients of the root node, each child node, and each leaf node in the feature decision tree.

[0056] In this embodiment of the invention, the significance coefficients of the root node, each child node, and each leaf node in the feature decision tree represent the correlation between the distribution of user behavior feature data of the current node and the distribution of user behavior feature data of the parent node of the current node.

[0057] S103: Select leaf nodes with a significance coefficient greater than the first threshold, construct the path from the leaf node to the root node, and obtain the initial path set.

[0058] In this embodiment of the invention, the specific steps of S103 are as follows: select leaf nodes with a significance coefficient greater than the first threshold to form a leaf node set; and construct the path from each leaf node to the root node by traversing each leaf node in the leaf node set to obtain an initial path set.

[0059] In this embodiment of the invention, the first threshold is the leaf node significance coefficient threshold α. The significance coefficient r(LNode) of each leaf node is compared with α, and leaf nodes with a value greater than α are retained to form the corresponding leaf node set LSet = {LNode1, LNode2, ..., LNode...} l When the leaf node set contains l nodes, l paths can be constructed, resulting in the initial path set PSet = {Path1, Path2, ..., Path...}. l When the feature data dimension is very large, the initial path selected in this embodiment is far less than the total number of paths from all leaf nodes to the root node in the feature decision tree.

[0060] S104: By traversing all paths in the initial path set, obtain the first path node set for each path. Select paths from the initial path set whose average significance coefficient of the first path node set is greater than the second threshold to form the target path set.

[0061] In this embodiment of the invention, the specific steps of S104 are as follows: by traversing all paths in the initial path set, all nodes of each path are obtained to form the first path node set of each path, and the average significance coefficient of the first path node set of each path is calculated.

[0062] The target path set is formed by selecting paths from the initial path set whose average significance coefficient of the first path node set is greater than the second threshold.

[0063] In this embodiment of the invention, the first set of path nodes is denoted as NSet, such as NSet i = {LNode i ,Path i} represents the path from the i-th leaf node to the root node. i The set of all nodes on the path. When there are l paths in the initial path set, the corresponding first path node set has l nodes, represented as {NSet1, NSet2, ..., NSet...} l}

[0064] The second threshold is the path significance coefficient threshold β. Paths with a mean significance coefficient greater than β are selected to form a target path set, denoted as SelPSet.

[0065] In this embodiment of the invention, step S104, selecting paths from the initial path set whose average significance coefficient of the first path node set is greater than a second threshold, includes: determining the average significance coefficient of the first path node set:

[0066] S1041: Obtain the significance coefficients of the root node, each child node, and each leaf node in the first path node set of each of the above paths, and sum the significance coefficients of each node.

[0067] S1042: The sum of the significance coefficients of each node is used to determine the mean significance coefficient by taking the ratio of the sum of the significance coefficients of each node to the number of nodes of the root node, each child node and each leaf node in the first path node set.

[0068] In this embodiment, the formula for calculating the mean significance coefficient is as follows:

[0069] ;

[0070] In the formula, Represents the mean of significance coefficients, ||NSet i|| represents the total number of nodes in the first path node set, including the root node, all child nodes, and all leaf nodes; This represents the significance coefficient of the k-th node.

[0071] S105: By traversing all paths in the target path set, a second path node set is obtained for each path, and target nodes with a significance coefficient greater than a third threshold are selected from the second path node set of each path.

[0072] In this embodiment of the invention, the specific step of S105 is: by traversing all paths in the target path set, all nodes of each path are obtained to form a second path node set for each path;

[0073] Select target nodes from the set of second path nodes for each path whose significance coefficient is greater than the third threshold.

[0074] In this embodiment of the invention, the second path node set is denoted as SelNSet, and the second path node set of all paths in the target path set is denoted as SelSNSet = {SelNSet}. i , SelNSet j ,…, SelNSet k}

[0075] The third threshold is the node significance coefficient threshold, denoted as γ. The features corresponding to the nodes whose significance coefficients are greater than the node significance coefficient threshold γ are combined to form new synthetic features.

[0076] S106: Combine the user behavior features corresponding to the target nodes selected from each path to form new synthetic features; the new synthetic features are used to train the product recommendation system.

[0077] In an embodiment of the present invention, for example, the second path node set SelNSet i After comparing and selecting the significance coefficients of all nodes, the synthetic feature corresponding to the second path node set is obtained, denoted as Xnew. i .

[0078] As can be seen from the above embodiments, the feature processing method for a product recommendation system provided in this embodiment, by setting a first threshold, a second threshold, and a third threshold, introduces leaf node filtering, path filtering from leaf node to root node, and feature node filtering on the path for the feature decision tree. This effectively and significantly reduces the computational processing and storage overhead when the feature dimension is very large, and can effectively improve the prediction speed of user features without increasing the overfitting of the recommendation system.

[0079] In one embodiment of the present invention, based on the above embodiments, step S102, which involves determining the significance coefficients of the root node, each child node, and each leaf node in the feature decision tree, includes:

[0080] For any node among the child nodes and leaf nodes, the ratio of the positive and negative sample ratio of any node to the positive and negative sample ratio of its parent node is determined as the significance coefficient of that node. The significance coefficient of the root node is set to a preset value.

[0081] In this embodiment of the invention, the significance coefficient r of the child node SNode and the leaf node LNode is calculated using the following formula:

[0082]

[0083] In the formula, The 'significance' value is represented by 'Node', where 'Node' represents any node and 'ParentNode' represents the parent node of any node.

[0084] In this embodiment, the positive and negative samples are represented by the positive and negative sample ratio. For example, taking product push clicks as an example, if a user clicks on a pushed product, it is recorded as a positive sample; if the user does not click on a pushed product, it is recorded as a negative sample.

[0085] In this embodiment, the preset value can be set as needed, and this embodiment of the invention does not impose any restrictions on it. For example, the preset value can be 1, 2, or 3.

[0086] As described in the above embodiments, the node significance coefficients of child nodes (SNode) and leaf nodes (LNode) in this embodiment are calculated as the quotient of the ratio of positive to negative samples of the current node to the ratio of positive to negative samples of the parent node. This calculation method can reflect the distribution of the current feature data while ensuring that the distribution of the parent node's feature data is reflected, thereby obtaining a certain correlation with the distribution of the parent node's feature data, and thus enabling the selection of more suitable features.

[0087] In one embodiment of the present invention, Figure 2 Based on the provided embodiments, the following steps are included after S106:

[0088] S107: Based on the evaluation function, select the optimal synthetic feature from the user behavior features corresponding to the target nodes selected from all paths; train the product recommendation system based on the optimal synthetic feature.

[0089] In this embodiment of the invention, the evaluation function is ε,max(ε(Y)). valid , RS(D validThe expression `Xnew` represents the search for an `Xnew` such that the evaluation function `ε` reaches its optimal value. `Xnew = {x1, x2, x3, ... x...}` m , x m+1 ,x m+2 ,…,x m+k} indicates that the optimal synthetic feature has been selected.

[0090] As can be seen from the above embodiments, by selecting the optimal synthetic features from the user behavior features corresponding to the selected target nodes, the optimal synthetic features can be obtained and used for training of the product recommendation system, which can more effectively explore the nonlinear relationship between user behavior feature data and user behavior result label data.

[0091] Figure 3 This is a schematic diagram of the feature processing device applied to a product recommendation system provided in an embodiment of the present invention. Figure 3 As shown, the feature processing device 30 applied to the product recommendation system includes: a first construction module 301, a determination module 302, a second construction module 303, a first filtering module 304, a second filtering module 305, and a combination module 306.

[0092] The first construction module 301 is used to obtain a sample dataset of the product recommendation system and construct a feature decision tree based on the user behavior feature data and the corresponding user behavior result label data in the sample dataset of the product recommendation system.

[0093] The determination module 302 is used to determine the significance coefficients of the root node, each child node, and each leaf node in the feature decision tree.

[0094] The second construction module 303 is used to select leaf nodes with a significance coefficient greater than the first threshold to form a leaf node set; by traversing each leaf node in the leaf node set, a path from each leaf node to the root node is constructed to obtain an initial path set.

[0095] The first filtering module 304 is used to obtain a first path node set for each path by traversing all paths in the initial path set, and to filter paths from the initial path set whose average significance coefficient of the first path node set is greater than a second threshold to form a target path set.

[0096] The second filtering module 305 is used to obtain a second path node set for each path by traversing all paths in the target path set, and to filter target nodes with a significance coefficient greater than a third threshold from the second path node set of each path.

[0097] The combination module 306 is used to combine the user behavior features corresponding to the target nodes selected from each path to form new synthetic features; the new synthetic features are used to train the product recommendation system.

[0098] In one possible implementation, the first construction module 301 is specifically used to construct an initial feature decision tree based on user behavior feature data and corresponding user behavior result label data using the ID3 algorithm or CART algorithm, and to prune the initial feature decision tree using pre-pruning and post-pruning strategies to obtain the feature decision tree.

[0099] In one possible implementation, the determining module 302 is specifically used to determine the significance coefficient of any node as the ratio of the positive and negative sample ratio of any node to the positive and negative sample ratio of the parent node of any node, for each child node and any leaf node; and to set the significance coefficient of the root node to a preset value.

[0100] In one possible implementation, the first filtering module 304 specifically determines the average significance coefficient of the first path node set by filtering paths from the initial path set whose average significance coefficient is greater than a second threshold: obtaining the significance coefficients of the root node, each child node, and each leaf node in the first path node set of each path, and summing the significance coefficients of each node; and determining the average significance coefficient by taking the summation result of the significance coefficients of each node and the ratio of the number of nodes in the root node, each child node, and each leaf node in the first path node set.

[0101] In one possible implementation, the feature processing device applied to the product recommendation system further includes a training module 307, which is specifically used to select the optimal synthetic feature from the user behavior features corresponding to the target nodes selected from all paths based on the evaluation function; and to train the product recommendation system based on the optimal synthetic feature.

[0102] The apparatus provided in this embodiment can be used to execute the technical solutions of the above method embodiments. Its implementation principle and technical effects are similar, and will not be described again here.

[0103] Figure 4 This is a schematic diagram of the hardware structure of a server provided in an embodiment of the present invention. Figure 4 As shown, the server 40 in this embodiment includes: at least one processor 401 and a memory 402; wherein

[0104] Memory 402 is used to store instructions executed by the computer;

[0105] Processor 401 is configured to execute computer execution instructions stored in memory to implement the various steps performed by the server in the above embodiments. For details, please refer to the relevant descriptions in the foregoing method embodiments.

[0106] Alternatively, the memory 402 can be either standalone or integrated with the processor 401.

[0107] When the memory 402 is configured independently, the server also includes a bus 403 for connecting the memory 402 and the processor 401.

[0108] This invention also provides a computer storage medium storing computer execution instructions. When the processor executes the computer execution instructions, the feature processing method applied to the product recommendation system described above is implemented.

[0109] This invention also provides a computer program product, including a computer program that, when executed by a processor, implements the feature processing method described above for use in a product recommendation system.

[0110] In the several embodiments provided by this invention, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or modules, and may be electrical, mechanical, or other forms.

[0111] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to implement the solution of this embodiment according to actual needs.

[0112] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing unit, or each module can exist physically separately, or two or more modules can be integrated into one unit. The unit composed of the above modules can be implemented in hardware or in the form of hardware plus software functional units.

[0113] The integrated modules described above, implemented as software functional modules, can be stored in a computer-readable storage medium. These software functional modules, stored in a storage medium, include several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute some steps of the methods of the various embodiments of this application.

[0114] It should be understood that the aforementioned processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. A general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly manifested as execution by a hardware processor, or execution by a combination of hardware and software modules within the processor.

[0115] The memory may include high-speed RAM, and may also include non-volatile storage (NVM), such as at least one disk storage device, and may also be a USB flash drive, external hard drive, read-only memory, disk or optical disc, etc.

[0116] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of illustration, the buses shown in the accompanying drawings are not limited to a single bus or a single type of bus.

[0117] The aforementioned storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The storage medium can be any available medium that can be accessed by a general-purpose or special-purpose computer.

[0118] An exemplary storage medium is coupled to a processor, enabling the processor to read information from and write information to the storage medium. Alternatively, the storage medium can be an integral part of the processor. Both the processor and the storage medium can reside in an Application Specific Integrated Circuit (ASIC). Alternatively, the processor and storage medium can exist as discrete components in an electronic device or host device.

[0119] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0120] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A feature processing method applied to a product recommendation system, characterized in that, Applied to servers, including: Obtain a sample dataset of a product recommendation system, and construct a feature decision tree based on the user behavior feature data and corresponding user behavior result label data in the sample dataset of the product recommendation system. The significance coefficients of the root node, each child node, and each leaf node in the feature decision tree are determined respectively; wherein, for any node in each child node and each leaf node, the ratio of the positive and negative sample ratio of the node to the positive and negative sample ratio of the parent node is determined as the significance coefficient of the node; the significance coefficient of the root node is set to a preset value. Select leaf nodes with a significance coefficient greater than the first threshold, construct the path from the leaf node to the root node, and obtain the initial path set; By traversing all paths in the initial path set, a first path node set is obtained for each path. Paths whose average significance coefficient of the first path node set is greater than a second threshold are selected from the initial path set to form a target path set. By traversing all paths in the target path set, a second path node set is obtained for each path, and target nodes with a significance coefficient greater than a third threshold are selected from the second path node set of each path. The user behavior features corresponding to the target nodes selected from each path are combined to form new synthetic features; wherein the new synthetic features are used to train the product recommendation system. Based on the evaluation function, the optimal synthetic feature is selected from the user behavior features corresponding to the target nodes selected from all paths; the product recommendation system is trained based on the optimal synthetic feature.

2. The method according to claim 1, characterized in that, The step of constructing a feature decision tree based on the user behavior feature data and the corresponding user behavior result label data includes: An initial feature decision tree is constructed based on the user behavior feature data and the corresponding user behavior result label data, using either the iterative binary search algorithm ID3 or the classification and regression tree algorithm CART. The initial feature decision tree is pruned using a pre-pruning and post-pruning strategy to obtain the feature decision tree.

3. The method according to any one of claims 1 to 2, characterized in that, The step of selecting paths from the initial path set whose mean significance coefficient of the first path node set is greater than a second threshold includes: Determine the mean significance coefficient of the first path node set: obtain the significance coefficients of the root node, each child node and each leaf node in the first path node set of each path, and sum the significance coefficients of each node; determine the mean significance coefficient by taking the sum of the significance coefficients of each node and the ratio of the number of nodes in the root node, each child node and each leaf node in the first path node set.

4. A feature processing device for a product recommendation system, characterized in that, include: The first construction module is used to obtain a sample dataset of the product recommendation system and construct a feature decision tree based on the user behavior feature data and the corresponding user behavior result label data in the sample dataset of the product recommendation system. The determination module is used to determine the significance coefficients of the root node, each child node, and each leaf node in the feature decision tree, respectively. The second construction module is used to select leaf nodes with a significance coefficient greater than the first threshold to form a leaf node set; by traversing each leaf node in the leaf node set, the path from each leaf node to the root node is constructed to obtain an initial path set. The first filtering module is used to obtain a first path node set for each path by traversing all paths in the initial path set, and to filter paths from the initial path set whose average significance coefficient of the first path node set is greater than a second threshold to form a target path set. The second filtering module is used to obtain a second path node set for each path by traversing all paths in the target path set, and to filter target nodes with a significance coefficient greater than a third threshold from the second path node set of each path. The combination module is used to combine the user behavior features corresponding to the target nodes selected from each path to form new synthetic features; The novel synthetic features are used to train the product recommendation system. The training module is used to select the optimal synthetic features from the user behavior features corresponding to the target nodes selected from all paths based on the evaluation function; and to train the product recommendation system based on the optimal synthetic features. The determining module is specifically used to determine the significance coefficient of any node by taking the ratio of positive and negative samples of any node to the ratio of positive and negative samples of the parent node of any node among the child nodes and leaf nodes; and to set the significance coefficient of the root node to a preset value.

5. The apparatus according to claim 4, characterized in that, The first construction module is specifically used to construct an initial feature decision tree based on the user behavior feature data and the corresponding user behavior result label data using the ID3 algorithm or the CART algorithm, and to prune the initial feature decision tree using pre-pruning and post-pruning strategies to obtain the feature decision tree.

6. A server, characterized in that, include: At least one processor and memory; The memory stores computer-executed instructions; The at least one processor executes computer execution instructions stored in the memory, causing the at least one processor to perform the feature processing method applied to a product recommendation system as described in any one of claims 1 to 3.

7. A computer storage medium, characterized in that, The computer storage medium stores A computer execution instruction is provided, which, when executed by the processor, implements the feature processing method for a product recommendation system as described in any one of claims 1 to 3.

Images