Commodity recommendation method and device

Abstract

The invention provides a commodity recommendation method and device, and relates to the field of commodity recommendation. Performing sequence modeling on the user behavior sequence data under a plurality of time spans to obtain a plurality of generated user behavior sequence sub-features, and performing sequence modeling on the commodity identification information to obtain generated basic sequence sub-features; fusing the plurality of user behavior sequence sub-features and the basic sequence sub-features to obtain user behavior sequence features; and obtaining commodity features corresponding to the candidate recommended commodities, inputting the commodity features and the user behavior sequence features into a click rate estimation model, obtaining a click rate estimation score corresponding to each candidate recommended commodity, and generating a commodity recommendation list based on the click rate estimation scores. According to the method, the evolution and change trend of user behaviors can be captured, and the evolution of interests and preferences of the users can be better understood, so that a more accurate recommendation result is provided, and the customer satisfaction is improved.

Description

Technical Field

[0001] The present application relates to the fields of product recommendation, artificial intelligence, and computer technology, and in particular to a product recommendation method and device thereof. Background Art

[0002] With the rapid development of e-commerce, people's demand for online shopping is increasing. However, faced with a huge number of product choices, consumers often feel confused and exhausted. Therefore, in order to help consumers better find the products they are interested in, a fast and accurate product recommendation system becomes crucial. Summary of the invention

[0003] The present disclosure provides a product recommendation method and a device thereof, so as to at least solve the problem of quickly and accurately recommending products to users in the related art.

[0004] The first aspect of the present application proposes a product recommendation method, including: obtaining user behavior sequence data, wherein the user behavior sequence data includes product identification information, product attribute information, timestamp information and behavior type information; performing sequence modeling on the user behavior sequence data at multiple time spans, obtaining multiple user behavior sequence sub-features generated after sequence modeling, and performing sequence modeling on the product identification information, obtaining basic sequence sub-features generated after sequence modeling; fusing the multiple user behavior sequence sub-features and the basic sequence sub-features to obtain user behavior sequence features; obtaining product features corresponding to candidate recommended products, inputting the product features and the user behavior sequence features into a click-through rate prediction model, obtaining a click-through rate prediction score corresponding to each candidate recommended product, and generating a product recommendation list based on the click-through rate prediction score.

[0005] According to one embodiment of the present application, the product recommendation method also includes: dividing the click-through rate prediction model into a computationally intensive module and a storage-intensive module; deploying the computationally intensive module on a graphics processing unit (GPU) cluster, and the computationally intensive module is used to perform matrix calculations and vector operations; deploying the storage-intensive module on a central processing unit (CPU) cluster, and the storage-intensive module is used to store sparse parameters corresponding to the click-through rate prediction model.

[0006] According to one embodiment of the present application, the product recommendation method also includes: dividing the model-related services of the click-through rate prediction model into a main graph service, a computing subgraph service, and a data subgraph service; loading the click-through rate prediction model based on the main graph service, and performing hash operations on the product features and user behavior sequence features input into the click-through rate prediction model to obtain their respective corresponding hash codes, and sending the hash codes to the computing subgraph service to obtain the click-through rate prediction score returned by the computing subgraph service; loading the computing-intensive module based on the computing subgraph service, and querying the data subgraph service according to the hash code to obtain the sparse parameters of the click-through rate prediction model, and performing click-through rate estimation based on the sparse parameters, and sending the click-through rate prediction score to the main graph service; and loading the sparse parameters in a distributed manner based on the data subgraph service, and the sparse parameters are divided into multiple shards and stored in the storage-intensive module.

[0007] According to an embodiment of the present application, the product recommendation method also includes: dividing the click-through rate prediction model into multiple compiled sub-graphs; determining whether the product features and user behavior sequence features input into the click-through rate prediction model need to be pre-compiled; if pre-compilation is required, dividing each compiled sub-graph into multiple feature buckets according to the feature dimension; for any feature bucket, performing feature filling on the input corresponding to the feature bucket based on the feature dimension corresponding to the feature bucket, and performing compilation operations corresponding to the filled features based on the compiled sub-graph to which the feature bucket belongs.

[0008] According to an embodiment of the present application, the product recommendation method further includes: if pre-compilation is not required, performing a compilation operation based on the compiled original image corresponding to the click-through rate prediction model.

[0009] According to one embodiment of the present application, the click-through rate prediction model adopts a teacher-student model as a framework; the teacher model adopts a PPnet model, and the teacher model is used to obtain the weight matrix generated after the training based on the sample training set and pass the weight matrix to the student model; the student model splits the weight matrix according to the feature domain, and performs online click-through rate prediction based on the weight sub-matrix generated after the split.

[0010] According to one embodiment of the present application, the GPU adopts a fixed memory method, and during the processing of the GPU, execution operations are performed based on multiple computing unified device architecture CUDA streams within the GPU, wherein each CUDA stream corresponds to a CUDA context.

[0011] According to one embodiment of the present application, the product recommendation method further includes: during the processing of the GPU, calculations are performed based on the tensor cores in the GPU using a high bandwidth and mixed precision method.

[0012] The second aspect of the present application proposes a product recommendation device, including: an acquisition module, used to acquire user behavior sequence data, wherein the user behavior sequence data includes product identification information, product attribute information, timestamp information and behavior type information; a modeling module, used to perform sequence modeling on the user behavior sequence data at multiple time spans, and acquire multiple user behavior sequence sub-features generated after sequence modeling, and to perform sequence modeling on the product identification information and acquire basic sequence sub-features generated after sequence modeling; a generation module, used to fuse multiple user behavior sequence sub-features and basic sequence sub-features to obtain user behavior sequence features; a recommendation module, used to acquire product features corresponding to candidate recommended products, input the product features and user behavior sequence features into a click-through rate prediction model, acquire the click-through rate prediction score corresponding to each candidate recommended product, and generate a product recommendation list based on the click-through rate prediction score.

[0013] According to one embodiment of the present application, the product recommendation device also includes a model acceleration module, which is used to: divide the click-through rate prediction model into a computationally intensive module and a storage-intensive module; deploy the computationally intensive module in a graphics processing unit GPU cluster, and the computationally intensive module is used to perform matrix calculations and vector operations; deploy the storage-intensive module in a central processing unit CPU cluster, and the storage-intensive module is used to store sparse parameters corresponding to the click-through rate prediction model.

[0014] According to one embodiment of the present application, the model acceleration module is also used to: divide the model-related services of the click-through rate prediction model into a main graph service, a computing subgraph service, and a data subgraph service; load the click-through rate prediction model based on the main graph service, and perform hash operations on the product features and user behavior sequence features input into the click-through rate prediction model to obtain their respective corresponding hash codes, and send the hash codes to the computing subgraph service to obtain the click-through rate prediction score returned by the computing subgraph service; load the computing-intensive module based on the computing subgraph service, and query the data subgraph service according to the hash code to obtain the sparse parameters of the click-through rate prediction model, perform click-through rate estimation based on the sparse parameters, and send the click-through rate prediction score to the main graph service; load the sparse parameters in a distributed manner based on the data subgraph service, and the sparse parameters are divided into multiple shards and stored in the storage-intensive module.

[0015] According to one embodiment of the present application, the model acceleration module is also used to: divide the click-through rate prediction model into multiple compiled sub-graphs; determine whether the product features and user behavior sequence features input into the click-through rate prediction model need to be pre-compiled; if pre-compilation is required, divide each compiled sub-graph into multiple feature buckets according to the feature dimension; for any feature bucket, perform feature filling on the input corresponding to the feature bucket based on the feature dimension corresponding to the feature bucket, and perform compilation operations corresponding to the filled features based on the compiled sub-graph to which the feature bucket belongs.

[0016] According to one embodiment of the present application, the model acceleration module is further used to: if pre-compilation is not required, perform a compilation operation based on the compiled original image corresponding to the click-through rate prediction model.

[0017] According to one embodiment of the present application, the click-through rate prediction model adopts a teacher-student model as a framework; the teacher model adopts a PPnet model, and the teacher model is used to obtain the weight matrix generated after the training based on the sample training set and pass the weight matrix to the student model; the student model splits the weight matrix according to the feature domain, and performs online click-through rate prediction based on the weight sub-matrix generated after the split.

[0018] According to one embodiment of the present application, the GPU adopts a fixed memory method. During the processing of the GPU, execution operations are performed based on multiple computing unified device architecture CUDA streams within the GPU, wherein each CUDA stream corresponds to a CUDA context.

[0019] According to one embodiment of the present application, the model acceleration module is also used to: during the processing of the GPU, perform calculations based on the tensor cores in the GPU using a high bandwidth and mixed precision method.

[0020] The third aspect embodiment of the present application proposes an electronic device, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to implement the product recommendation method as described in the first aspect embodiment of the present application.

[0021] The fourth aspect embodiment of the present application proposes a non-transitory computer-readable storage medium storing computer instructions, wherein the computer instructions are used to implement the product recommendation method as described in the first aspect embodiment of the present application.

[0022] The fifth aspect embodiment of the present application proposes a computer program product, including a computer program, which, when executed by a processor, implements the product recommendation method as described in the first aspect embodiment of the present application.

[0023] The technical solution provided by the embodiments of the present disclosure brings at least the following beneficial effects: by performing sequence modeling on user behavior sequence data in multiple time spans, the present application can capture the evolution and change trends of user behavior, and can better understand the evolution of user interests and preferences, thereby providing more accurate recommendation results; through sequence modeling in multiple time spans, data from different time periods can be integrated, the integrity and coverage of the data can be improved, and the inaccuracy of recommendations caused by data sparsity can be reduced.

[0024] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] The above and / or additional aspects and advantages of the present application will become apparent and easily understood from the following description of the embodiments in conjunction with the accompanying drawings, in which:

[0026] Figure 1 It is a schematic diagram of an exemplary implementation of a product recommendation method shown in an embodiment of the present application.

[0027] Figure 2 This is a schematic diagram of performing sequence modeling on user behavior sequence data, shown in one embodiment of the present application.

[0028] Figure 3 It is a schematic diagram of a main graph service, a computing subgraph service and a data subgraph service shown in an embodiment of the present application.

[0029] Figure 4 It is a schematic diagram of dividing a compilation subgraph shown in an embodiment of the present application.

[0030] Figure 5 It is a schematic diagram of a compilation solution shown in an embodiment of the present application.

[0031] Figure 6 This is a schematic diagram of multi-CUDA stream computing shown in an embodiment of the present application.

[0032] Figure 7 It is a schematic diagram of a commodity recommendation device shown in an embodiment of the present application.

[0033] Figure 8 It is a schematic diagram of an electronic device shown in one embodiment of the present application. DETAILED DESCRIPTION

[0034] The embodiments of the present application are described in detail below, and examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present application, and should not be construed as limiting the present application.

[0035] It should be noted that the information (including but not limited to user device information, user personal information, etc.), data (including but not limited to data used for analysis, stored data, displayed data, etc.) and signals involved in this application are all authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant laws, regulations and standards.

[0036] Figure 1 is a schematic diagram of an exemplary implementation of a commodity recommendation method shown in the present application, such as Figure 1 As shown, the product recommendation method includes the following steps:

[0037] S101, obtaining user behavior sequence data, wherein the user behavior sequence data includes commodity identification information, commodity attribute information, timestamp information and behavior type information.

[0038] The product identification information refers to a unique identifier for each product, which is used to distinguish different products.

[0039] Among them, product attribute information includes product category, brand, price and other attributes, which are used to describe the characteristics of the product. Product attribute information also includes source scene information, such as search scene, recommendation scene or flash sale channel page and other information.

[0040] Among them, the timestamp information records the time when the user operates the product and is used to analyze the time characteristics of user behavior.

[0041] Among them, the behavior type information includes different types of user behaviors towards products, such as browsing, collecting, adding to shopping cart, purchasing, etc., which is used to analyze user behavior patterns.

[0042] S102, performing sequence modeling on the user behavior sequence data in multiple time spans respectively, obtaining multiple user behavior sequence sub-features generated after the sequence modeling, and performing sequence modeling on the product identification information, obtaining basic sequence sub-features generated after the sequence modeling.

[0043] Figure 2 This is a schematic diagram of a method for performing sequence modeling on user behavior sequence data shown in the present application, such as Figure 2As shown, since the interests and needs of users often change over time, in this application, the time span is divided into short-term, medium-term and long-term. First, the user behavior sequence data is preprocessed, and the preprocessed user behavior sequence data is modeled into short-term behavior sequence, medium-term behavior sequence and long-term behavior sequence, respectively, to obtain short-term user behavior sequence sub-features, medium-term user behavior sequence sub-features and long-term user behavior sequence sub-features. In addition, in this application, the commodity identification information in the preprocessed user behavior sequence data is used as the basic sequence for basic sequence modeling, and the basic sequence sub-features generated after sequence modeling are obtained.

[0044] Optionally, models such as Recurrent Neural Network (RNN) or Long Short-Term Memory Network (LSTM) can be used for sequence modeling.

[0045] Among them, in the present application, after obtaining short-term user behavior sequence sub-features, medium-term user behavior sequence sub-features and long-term user behavior sequence sub-features, it is also necessary to unify the storage formats of short-term user behavior sequence sub-features, medium-term user behavior sequence sub-features and long-term user behavior sequence sub-features, so as to realize the rapid reuse of features and one-click migration capability of feature modeling when needed.

[0046] S103, fusing multiple user behavior sequence sub-features and basic sequence sub-features to obtain user behavior sequence features.

[0047] Optionally, multiple user behavior sequence sub-features and basic sequence sub-features may be fused, and various common fusion methods (such as weighted averaging, concatenation, addition, multiplication, etc.) may be used to obtain the final user behavior sequence feature generated after fusion.

[0048] Optionally, multiple user behavior sequence sub-features and basic sequence sub-features are aligned according to the time window, and then concatenated or merged element by element to form a final user behavior sequence feature, so as to integrate feature information of different time spans and different types.

[0049] S104, obtaining product features corresponding to the candidate recommended products, inputting the product features and user behavior sequence features into a click-through rate prediction model, obtaining a click-through rate prediction score corresponding to each candidate recommended product, and generating a product recommendation list based on the click-through rate prediction score.

[0050] Among them, click-through rate refers to the ratio of the number of times a specified content is clicked and exposed on a website platform or APP. Click-through rate is usually an important indicator measured in the recommendation system.

[0051] Among them, the click-through rate prediction model can be trained and generated based on the feature log and the click order log. The feature log contains product features and user behavior sequence features.

[0052] Among them, product features may include product attribute features (such as product name, price, category, label, etc.), product context features (such as product popularity (such as hot-selling products, trending products, etc.), product inventory status, product delivery information), etc.

[0053] The product features and user behavior sequence features are input into the click-through rate prediction model to obtain the click-through rate prediction score corresponding to each candidate recommended product.

[0054] Optionally, you can sort the products in descending order of their estimated click rate scores, putting products with higher scores at the front.

[0055] Optionally, you can set a threshold according to business needs and only select products with scores higher than the threshold as recommendation results.

[0056] The embodiment of the present application provides a method for product recommendation, including obtaining user behavior sequence data, wherein the user behavior sequence data includes product identification information, product attribute information, timestamp information and behavior type information; performing sequence modeling on the user behavior sequence data at multiple time spans, obtaining multiple user behavior sequence sub-features generated after sequence modeling, and performing sequence modeling on the product identification information, obtaining basic sequence sub-features generated after sequence modeling; fusing multiple user behavior sequence sub-features and basic sequence sub-features to obtain user behavior sequence features; obtaining product features corresponding to candidate recommended products, inputting product features and user behavior sequence features into a click-through rate prediction model, obtaining click-through rate prediction scores corresponding to each candidate recommended product, and generating a product recommendation list based on the click-through rate prediction scores. The present application can capture the evolution and change trends of user behavior by performing sequence modeling on user behavior sequence data at multiple time spans, can better understand the evolution of user interests and preferences, thereby providing more accurate recommendation results and improving customer satisfaction; through sequence modeling at multiple time spans, data from different time periods can be integrated, the integrity and coverage of the data can be improved, and the recommendation inaccuracy caused by data sparsity can be reduced.

[0057] Furthermore, after determining the product features corresponding to the candidate recommended products, the product features and user behavior sequence features may be input into a conversion rate estimation model to obtain a conversion rate estimation score corresponding to each candidate recommended product.

[0058] After obtaining the click rate estimated score and conversion rate estimated score corresponding to the candidate recommended products as mentioned above, as a feasible method, the click rate estimated score and the conversion rate estimated score can be weighted and summed to obtain the target score corresponding to the candidate recommended products, thereby determining the product recommendation list according to the target score.

[0059] After obtaining the click-through rate estimate score and conversion rate estimate score corresponding to the candidate recommended products as described above, as another feasible method, based on the conversion rate estimate score corresponding to each candidate recommended product, obtain the merchant promotion cost corresponding to each candidate recommended product; for each candidate recommended product, obtain the product of the merchant promotion cost corresponding to the candidate recommended product and the click-through rate estimate score corresponding to the candidate recommended product, and use the product as the target score corresponding to the candidate recommended product, thereby determining the product recommendation list based on the target score. In this way, the candidate products with higher target scores will be ranked higher, so that the merchant's promotion resources can be more accurately placed on those products that are more likely to be converted into actual sales, thereby improving the promotion effect. After the promotion is completed, it is also convenient to bill the merchant for the promotion.

[0060] Furthermore, the click-through rate prediction model adopts a teacher-student model as a framework; the teacher model is usually responsible for training and providing guidance signals to the student model (such as generating a weight matrix), while the student model is used to perform actual online predictions. In this application, the teacher model adopts a permutation propagation network (PPnet) model, which can learn the personalized expressions of different users. The teacher model is used to obtain the weight matrix generated after the training based on the sample training set and pass the weight matrix to the student model; the student model splits the weight matrix according to the feature domain, thereby reducing the complexity of the matrix multiplication calculation, and performs online click-through rate estimation based on the weight submatrix obtained after the split, thereby effectively reducing the time consumption for large-scale MLP networks.

[0061] Furthermore, with the increase in model calculation complexity and the increase in model parameter scale, the current single-machine central processing unit (CPU) architecture has limited computing power on the one hand, and cannot give full play to the high-performance computing advantages of the graphics processing unit (GPU) to achieve online reasoning acceleration. On the other hand, the CPU storage capacity is limited and cannot load models with ultra-large-scale parameters, which is not conducive to the expansion of the model scale and limits the expressive power of the model. In this application, the click-through rate prediction model is divided into a computing-intensive module and a storage-intensive module; the computing-intensive module is deployed in a GPU cluster, and the computing-intensive module is used to perform matrix calculations and vector operations; the storage-intensive module is deployed in a CPU cluster, and the storage-intensive module is used to store sparse parameters corresponding to the click-through rate prediction model to achieve model acceleration.

[0062] Furthermore, the model-related services of the click rate prediction model are divided into main graph service, computing sub-graph service and data sub-graph service. Figure 3 It is a schematic diagram of a main graph service, a computing subgraph service and a data subgraph service shown in this application.

[0063] like Figure 3 As shown, in the present application, a click-through rate prediction model is loaded based on the main graph service, and hash operations are performed on the product features and user behavior sequence features input into the click-through rate prediction model to obtain their respective corresponding hash codes, and the hash codes are sent to the calculation subgraph service to obtain the click-through rate prediction score returned by the calculation subgraph service.

[0064] like Figure 3 As shown, in the present application, computationally intensive modules are loaded based on the computational subgraph service, including complex computing networks such as Transformer and MLP, and the data subgraph service is queried according to the hash code to obtain the sparse parameters of the click-through rate prediction model, and the click-through rate is estimated based on the sparse parameters, and the click-through rate prediction score is sent to the main graph service.

[0065] like Figure 3 As shown, in this application, the loading of sparse parameters is realized in a distributed manner based on the data subgraph service. The sparse parameters are divided into multiple shards and deployed as remote procedure call (RPC) services and stored in the storage-intensive module. The data subgraph service provides basic functions such as data query, data update, and data deletion.

[0066] Furthermore, when computing subgraph services on GPUs, a large amount of kernel launch overhead and memory access overhead are generated. The deep learning compiler is usually used to optimize operator fusion to optimize the above problems. The compiler's underlying optimization is used, and the applicable model structure is more flexible, and there is no need to implement new operators each time. The performance can be greatly improved through operator fusion and graph structure optimization, but the input dimension of the model substructure is required to be fixed. Since the input dimension of the recommendation scene feature is not fixed, it will trigger the real-time compilation of the deep learning compiler, and the compilation time is intolerable for online reasoning.

[0067] Figure 4 This is a schematic diagram of the division of a compilation subgraph shown in this application. In this application, the click-through rate prediction model is divided into multiple compilation subgraphs to solve the compilation time-consuming problem in online reasoning scenarios. XLA (Accelerated Linear Algebra) is a high-performance machine learning compiler. AD refers to Attributed Deep Embedding, which is a method of embedding learning by coupling node attributes and graph structure information. Figure 4 As shown in the figure, the product features and user behavior sequence features are compiled separately to make them of the same dimension. In addition to the compiled subgraphs corresponding to the product features and user behavior sequence features, the deep computing layer of the model can also be divided into compiled subgraphs to accelerate computing. Each compiled subgraph corresponds to an execution unit. During execution, the XLA engine optimizes the computing subgraph, significantly improving efficiency.

[0068] Figure 5 is a schematic diagram of a compilation solution shown in this application, such as Figure 5 As shown, before compilation, determine whether the product features and user behavior sequence features of the click-through rate prediction model need to be precompiled; if precompilation has not been performed before and the current scenario supports precompilation, it is deemed that precompilation is required, and each compiled subgraph is divided into multiple feature buckets according to the feature dimension (for example, according to the type of feature); for any feature bucket, the input corresponding to the feature bucket is feature-filled to a fixed shape based on the feature dimension padding corresponding to the feature bucket, and the compilation operation corresponding to the filled feature is performed based on the compiled subgraph to which the feature bucket belongs, and the calculation subgraph is optimized through the XLA engine after compilation.

[0069] If pre-compilation is not required (for example, in the high real-time and high concurrency scenarios of online reasoning, the feature dimension is not covered by the pre-compilation scope), the compilation operation is performed based on the compiled original graph corresponding to the click-through rate prediction model and written back to the compilation cache.

[0070] Furthermore, in this application, the GPU adopts a fixed memory method. The fixed memory is fixed in the physical memory and can directly access the physical memory without going through the memory management of the operating system. Therefore, the performance in data transmission is better and lower data transmission delay can be provided.

[0071] Furthermore, the Compute Unified Device Architecture (CUDA) stream is a mechanism for executing GPU operations in parallel. TensorFlow uses only one CUDA Stream by default, which may cause competition between computing and data transmission, thereby reducing computing efficiency and failing to give full play to the GPU computing power. Therefore, in this application, in order to improve the concurrent utilization of GPU cores, this solution expands a single CUDA Stream to multiple CUDA Streams through a multi-stream mechanism and a multi-context mechanism, and creates a CUDA context for each CUDA stream to solve the problem of kernel launch lock competition and improve GPU utilization. Among them, the CUDA context refers to a set of status information related to a CUDA stream. It includes information such as resource allocation, memory management, and device configuration of the program. Each CUDA stream is associated with an independent CUDA context to better manage computing tasks.

[0072] Figure 6 This is a schematic diagram of multi-CUDA stream computing shown in this application, such as Figure 6 As shown, in the present application, during the processing of the GPU, execution operations are performed based on multiple computing unified device architecture CUDA streams within the GPU, wherein each CUDA stream corresponds to a CUDA context.

[0073] Furthermore, during the GPU processing, the Tensor Cores in the GPU perform operations through high bandwidth and mixed precision to speed up the calculation of matrix multiplication. Among them, Tensor Cores: is a special hardware unit that is specifically used to accelerate tensor (matrix) operations, especially in deep learning and high-performance computing tasks, and can perform operations such as matrix multiplication with higher efficiency.

[0074] Figure 7 is a schematic diagram of a commodity recommendation device shown in this application, such as Figure 7 As shown, the product recommendation device 700 includes an acquisition module 701, a modeling module 702, a generation module 703 and a recommendation module 704, wherein:

[0075] The acquisition module 701 is used to acquire user behavior sequence data, wherein the user behavior sequence data includes commodity identification information, commodity attribute information, timestamp information and behavior type information.

[0076] The modeling module 702 is used to perform sequence modeling on the user behavior sequence data in multiple time spans, obtain multiple user behavior sequence sub-features generated after sequence modeling, and perform sequence modeling on the product identification information to obtain basic sequence sub-features generated after sequence modeling.

[0077] The generating module 703 is used to fuse multiple user behavior sequence sub-features and basic sequence sub-features to obtain user behavior sequence features.

[0078] The recommendation module 704 is used to obtain product features corresponding to candidate recommended products, input the product features and user behavior sequence features into the click-through rate prediction model, obtain the click-through rate prediction score corresponding to each candidate recommended product, and generate a product recommendation list based on the click-through rate prediction score.

[0079] By performing sequence modeling on user behavior sequence data in multiple time spans, this device can capture the evolution and changing trends of user behavior, better understand the evolution of user interests and preferences, and thus provide more accurate recommendation results; through sequence modeling in multiple time spans, data from different time periods can be integrated, which can improve data integrity and coverage and reduce recommendation inaccuracies caused by data sparsity.

[0080] Furthermore, the product recommendation device 700 also includes a model acceleration module, which is used to: divide the click-through rate prediction model into a computationally intensive module and a storage-intensive module; deploy the computationally intensive module in a graphics processing unit GPU cluster, and the computationally intensive module is used to perform matrix calculations and vector operations; deploy the storage-intensive module in a central processing unit CPU cluster, and the storage-intensive module is used to store sparse parameters corresponding to the click-through rate prediction model.

[0081] Furthermore, the model acceleration module is also used to: divide the model-related services of the click-through rate prediction model into the main graph service, the computing subgraph service and the data subgraph service; load the click-through rate prediction model based on the main graph service, and perform hash operations on the product characteristics and user behavior sequence characteristics input into the click-through rate prediction model to obtain their respective corresponding hash codes, and send the hash codes to the computing subgraph service to obtain the click-through rate prediction score returned by the computing subgraph service; load the computing-intensive module based on the computing subgraph service, and query the data subgraph service according to the hash code to obtain the sparse parameters of the click-through rate prediction model, and perform click-through rate estimation based on the sparse parameters, and send the click-through rate prediction score to the main graph service; load the sparse parameters in a distributed manner based on the data subgraph service, and the sparse parameters are divided into multiple shards and stored in the storage-intensive module.

[0082] Furthermore, the model acceleration module is also used to: divide the click-through rate prediction model into multiple compilation sub-graphs; determine whether the product features and user behavior sequence features input into the click-through rate prediction model need to be pre-compiled; if pre-compilation is required, divide each compilation sub-graph into multiple feature buckets according to the feature dimension; for any feature bucket, fill the input corresponding to the feature bucket based on the feature dimension corresponding to the feature bucket, and perform the compilation operation corresponding to the filled feature based on the compilation sub-graph to which the feature bucket belongs.

[0083] Furthermore, the model acceleration module is also used to: if pre-compilation is not required, perform a compilation operation based on the compiled original image corresponding to the click-through rate prediction model.

[0084] Furthermore, the click-through rate prediction model adopts the teacher-student model as the framework; the teacher model adopts the PPnet model, and the teacher model is used to obtain the weight matrix generated after the training based on the sample training set and pass the weight matrix to the student model; the student model splits the weight matrix according to the feature domain, and performs online click-through rate estimation based on the weight sub-matrix generated after the split.

[0085] Furthermore, the GPU adopts a fixed memory method. During the processing of the GPU, the operation processing is performed based on multiple computing unified device architecture CUDA streams in the GPU, wherein each CUDA stream corresponds to a CUDA context.

[0086] Furthermore, the model acceleration module is also used to perform calculations using high bandwidth and mixed precision based on the Tensor Cores in the GPU during GPU processing.

[0087] In order to implement the above embodiment, the present application embodiment also proposes an electronic device 800, such as Figure 8As shown, the electronic device 800 includes: a processor 801 and a memory 802 communicatively connected to the processor, the memory 802 stores instructions executable by at least one processor, and the instructions are executed by at least one processor 801 to implement the product recommendation method shown in the above embodiment.

[0088] In order to implement the above embodiment, the embodiment of the present application also proposes a non-transitory computer-readable storage medium storing computer instructions, wherein the computer instructions are used to enable a computer to implement the product recommendation method shown in the above embodiment.

[0089] In order to implement the above embodiments, the embodiments of the present application also provide a computer program product, including a computer program, which implements the product recommendation method shown in the above embodiments when executed by a processor.

[0090] In the description of the present application, it should be understood that the terms "center", "longitudinal", "lateral", "length", "width", "thickness", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", "clockwise", "counterclockwise", "axial", "radial", "circumferential" and the like indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the referred device or element must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as a limitation on the present application.

[0091] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the features. In the description of this application, the meaning of "plurality" is two or more, unless otherwise clearly and specifically defined.

[0092] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present application. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, without contradiction.

[0093] Although the embodiments of the present application have been shown and described above, it can be understood that the above embodiments are exemplary and cannot be understood as limitations on the present application. Ordinary technicians in the field can change, modify, replace and modify the above embodiments within the scope of the present application.

Claims

1. A product recommendation method, characterized in that: include: Acquire user behavior sequence data, wherein the user behavior sequence data includes commodity identification information, commodity attribute information, timestamp information and behavior type information; Performing sequence modeling on the user behavior sequence data in multiple time spans respectively, and obtaining multiple user behavior sequence sub-features generated after the sequence modeling, and performing sequence modeling on the product identification information, and obtaining basic sequence sub-features generated after the sequence modeling; fusing the multiple user behavior sequence sub-features and the basic sequence sub-feature to obtain a user behavior sequence feature; Obtain product features corresponding to candidate recommended products, input the product features and the user behavior sequence features into a click-through rate prediction model, obtain a click-through rate prediction score corresponding to each candidate recommended product, and generate a product recommendation list based on the click-through rate prediction score.

2. The method according to claim 1, characterized in that The method further comprises: Dividing the click rate prediction model into a computation-intensive module and a storage-intensive module; Deploy the computation-intensive module on a graphics processing unit (GPU) cluster, wherein the computation-intensive module is used to perform matrix calculations and vector operations; The storage-intensive module is deployed in a central processing unit (CPU) cluster, and the storage-intensive module is used to store sparse parameters corresponding to the click-through rate prediction model.

3. The method according to claim 2, characterized in that The method further comprises: Dividing the model-related services of the click rate prediction model into a main graph service, a computing sub-graph service and a data sub-graph service; Based on the main graph service, the click-through rate prediction model is loaded, and hash operations are performed on the product features and the user behavior sequence features input into the click-through rate prediction model to obtain their respective corresponding hash codes, and the hash codes are sent to the computational subgraph service to obtain the click-through rate prediction score returned by the computational subgraph service; Loading the computationally intensive module based on the computational subgraph service, querying the data subgraph service according to the hash code to obtain the sparse parameters of the click-through rate prediction model, performing click-through rate prediction based on the sparse parameters, and sending the click-through rate prediction score to the main graph service; Based on the data subgraph service, sparse parameters are loaded in a distributed manner, and the sparse parameters are divided into multiple slices and stored in the storage-intensive module.

4. The method according to any one of claims 1 to 3, characterized in that The method further comprises: Dividing the click rate prediction model into a plurality of compiled subgraphs; Determining whether the product features and the user behavior sequence features input into the click rate prediction model need to be precompiled; If pre-compilation is required, each of the compiled subgraphs is divided into a plurality of feature buckets according to the feature dimension; For any of the feature buckets, feature filling is performed on the input corresponding to the feature bucket based on the feature dimension corresponding to the feature bucket, and a compilation operation corresponding to the filled feature is performed based on the compilation subgraph to which the feature bucket belongs.

5. The method according to claim 4, characterized in that The method further comprises: If pre-compilation is not required, a compilation operation is performed based on the compiled original image corresponding to the click-through rate prediction model.

6. The method according to claim 1, characterized in that The method further comprises: The click rate prediction model adopts the teacher-student model as the framework; The teacher model adopts a PPnet model, and the teacher model is used to obtain a weight matrix generated after the training is completed based on a sample training set and pass the weight matrix to the student model; The student model splits the weight matrix according to feature domains, and performs online click-through rate estimation based on the weight sub-matrices generated after the splitting.

7. The method according to claim 2, characterized in that The GPU adopts a fixed memory mode, and in the processing process of the GPU, execution operation processing is performed based on multiple computing unified device architecture CUDA streams in the GPU, wherein each of the CUDA streams corresponds to a CUDA context.

8. The method according to claim 7, characterized in that The method further comprises: During the processing of the GPU, operations are performed based on the Tensor Cores in the GPU using high bandwidth and mixed precision.

9. A commodity recommendation device, characterized in that: include: An acquisition module, used to acquire user behavior sequence data, wherein the user behavior sequence data includes commodity identification information, commodity attribute information, timestamp information and behavior type information; A modeling module, configured to perform sequence modeling on the user behavior sequence data at multiple time spans, obtain multiple user behavior sequence sub-features generated after the sequence modeling, and perform sequence modeling on the product identification information, and obtain basic sequence sub-features generated after the sequence modeling; A generating module, configured to fuse the plurality of user behavior sequence sub-features and the basic sequence sub-feature to obtain a user behavior sequence feature; The recommendation module is used to obtain the product features corresponding to the candidate recommended products, input the product features and the user behavior sequence features into the click-through rate prediction model, obtain the click-through rate prediction score corresponding to each of the candidate recommended products, and generate a product recommendation list based on the click-through rate prediction score.

10. An electronic device, comprising: at least one processor; as well as, a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the method according to any one of claims 1 to 8.

11. A non-transitory computer-readable storage medium storing computer instructions, wherein: The computer instructions are used to cause the computer to execute the method according to any one of claims 1-8.

12. A computer program product comprising a computer program, which, when executed by a processor, implements the steps of the method according to any one of claims 1 to 8.

Images