WO2023138381A1 - 基于物品属性与时序模式耦合关系的序列推荐方法和系统 - Google Patents

基于物品属性与时序模式耦合关系的序列推荐方法和系统 Download PDF

Info

Publication number
WO2023138381A1
WO2023138381A1 PCT/CN2023/070574 CN2023070574W WO2023138381A1 WO 2023138381 A1 WO2023138381 A1 WO 2023138381A1 CN 2023070574 W CN2023070574 W CN 2023070574W WO 2023138381 A1 WO2023138381 A1 WO 2023138381A1
Authority
WO
WIPO (PCT)
Prior art keywords
item
sequence
attribute
relationship
interaction
Prior art date
Application number
PCT/CN2023/070574
Other languages
English (en)
French (fr)
Inventor
张全贵
李鑫
冯勰
罗代忠
马新强
黄羿
Original Assignee
重庆文理学院
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 重庆文理学院 filed Critical 重庆文理学院
Publication of WO2023138381A1 publication Critical patent/WO2023138381A1/zh

Links

Images

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/90Details of database functions independent of the retrieved data types
    • G06F16/95Retrieval from the web
    • G06F16/953Querying, e.g. by the use of web search engines
    • G06F16/9535Search customisation based on user profiles and personalisation
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/20Information retrieval; Database structures therefor; File system structures therefor of structured data, e.g. relational data
    • G06F16/24Querying
    • G06F16/245Query processing
    • G06F16/2457Query processing with adaptation to user needs
    • G06F16/24578Query processing with adaptation to user needs using ranking
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/20Information retrieval; Database structures therefor; File system structures therefor of structured data, e.g. relational data
    • G06F16/24Querying
    • G06F16/245Query processing
    • G06F16/2458Special types of queries, e.g. statistical queries, fuzzy queries or distributed queries
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/20Information retrieval; Database structures therefor; File system structures therefor of structured data, e.g. relational data
    • G06F16/24Querying
    • G06F16/245Query processing
    • G06F16/2458Special types of queries, e.g. statistical queries, fuzzy queries or distributed queries
    • G06F16/2474Sequence data queries, e.g. querying versioned data
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/90Details of database functions independent of the retrieved data types
    • G06F16/95Retrieval from the web
    • G06F16/953Querying, e.g. by the use of web search engines
    • G06F16/9537Spatial or temporal dependent retrieval, e.g. spatiotemporal queries
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/04Architecture, e.g. interconnection topology
    • G06N3/044Recurrent networks, e.g. Hopfield networks
    • G06N3/0442Recurrent networks, e.g. Hopfield networks characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU]
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/04Architecture, e.g. interconnection topology
    • G06N3/045Combinations of networks
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/08Learning methods
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/08Learning methods
    • G06N3/088Non-supervised learning, e.g. competitive learning
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/08Learning methods
    • G06N3/09Supervised learning
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06QINFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q30/00Commerce
    • G06Q30/01Customer relationship services
    • G06Q30/015Providing customer assistance, e.g. assisting a customer within a business location or via helpdesk
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06QINFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q30/00Commerce
    • G06Q30/02Marketing; Price estimation or determination; Fundraising
    • G06Q30/0201Market modelling; Market analysis; Collecting market data
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06QINFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q30/00Commerce
    • G06Q30/02Marketing; Price estimation or determination; Fundraising
    • G06Q30/0282Rating or review of business operators or products
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06QINFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q30/00Commerce
    • G06Q30/06Buying, selling or leasing transactions
    • G06Q30/0601Electronic shopping [e-shopping]
    • G06Q30/0631Item recommendations
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/04Architecture, e.g. interconnection topology
    • G06N3/048Activation functions

Definitions

  • the invention relates to the technical field of intelligent recommendation, in particular to a sequential recommendation method and system based on the coupling relationship between item attributes and time series patterns.
  • sequence-based recommendation has gradually become an important subfield in the field of recommender systems.
  • traditional recommendation methods which are mostly based on modeling user information
  • sequential recommendation methods do not pay too much attention to user information, but generate recommendations for users by mining the relationship between items in session sequence data.
  • the existing sequence recommendation methods often focus too much on the sequence data itself, while ignoring the importance of some auxiliary information, and do not make full use of the impact of item attribute information and the coupling relationship between item attribute information and time series on the recommendation results, so the resulting recommendation results are often not accurate enough.
  • the technical task of the present invention is to provide a sequence recommendation method and system based on the coupling relationship between item attributes and timing patterns, to solve the problem of how to make full use of the coupling relationship between the user and the interaction session sequence mode of the item and the item attribute information to effectively improve the accuracy of recommendation, and then improve the user's satisfaction.
  • Divide the data set Divide the user behavior sequence data set into training set, test set and verification set; among them, the verification set is used to adjust the hyperparameters of the sequence recommendation model during the training phase.
  • Training sequence recommendation model input the training set into the sequence recommendation model, and the sequence recommendation model learns the interactive sequence pattern representation that combines the coupling relationship between item attributes and timing patterns, and then trains the learnable parameters of the sequence recommendation model;
  • Sequence recommendation model prediction Input the test set into the sequence recommendation model, the sequence recommendation model predicts the interaction probability of all candidate items and sorts them, and selects the top K items as the recommendation list.
  • each user behavior sequence data set includes a plurality of sub-sequence data sets
  • the sub-sequence data sets include item ID sub-sequence data sets and item attribute sub-sequence data sets
  • the item attribute sub-sequence data sets are constructed according to each attribute of the item.
  • sequence recommendation model More preferably, the details of constructing the sequence recommendation model are as follows:
  • Obtain the temporal relationship of item attributes form multiple attribute sequences for items in the user-item interaction sequence according to each attribute of the item, and use the GRU network to learn the temporal relationship representation vector of each item attribute sequence;
  • Obtain the coupling relationship between the item attribute and the timing mode use the implicit interaction relationship representation vector q between the user and the item as the query vector, and use the attention mechanism network to learn the coupling relationship between the timing relationship representation vector and the query vector of each item attribute sequence, thereby generating the final sequence representation vector based on coupling relationship analysis;
  • Candidate Item Interaction Prediction After the final sequence representation vector based on the coupling relationship analysis and the embedding vector of the candidate items are subjected to similarity calculation, they are input into the fully connected network to finally generate the interaction probabilities between each candidate item and the user, and the Top-K recommendations are generated according to the interaction probability sorting results.
  • timing relationship of obtaining item attributes is specifically as follows:
  • Construct item attribute sequence According to the time series relationship between users and items, each attribute value of the item in the sequence data is respectively formed into an attribute sequence, such as item brand sequence, item category sequence, etc.;
  • the attribute values in each attribute sequence are represented by one-hot encoding, and converted into low-dimensional dense vectors, namely embedding vectors, through a single-layer fully connected network;
  • Each attribute sequence is input into the GRU network respectively, and the representation vector s ai of each attribute sequence is learned respectively.
  • the coupling relationship between obtaining item attributes and timing mode is as follows:
  • the implicit interaction relationship representation vector q between the user and the item is used as the query vector, and the implicit interaction relationship representation vector q between the user and the item is calculated with the representation vector s ai of each attribute sequence to obtain the weight through dot product similarity calculation, and the normalized weight ⁇ i is obtained by using the Softmax function.
  • the formula is as follows:
  • the weight ⁇ i and the corresponding attribute sequence representation vector are weighted and summed to obtain the vector s,
  • W q , W k and W v are transformation matrices, which belong to learnable parameters
  • the vector s is further learned through one or more layers of fully connected networks to obtain the final item attribute and timing pattern summation relationship representation c;
  • Candidate item interaction prediction is as follows:
  • Similarity calculation convert the ID of the candidate item i into a low-dimensional dense vector through a single-layer fully connected network, and then multiply the corresponding elements with the final item attribute and timing mode summation relationship representation c to calculate the similarity vector d of the candidate item and the final item attribute and timing mode summation relationship representation c;
  • the similarity vector d is further learned through one or more layers of fully connected networks
  • Sigmoid activation function uses the Sigmoid activation function to compress the output to the range of [0,1] to obtain the probability of the candidate item as the item interacted with by the user at time t+1, that is, the final prediction result.
  • a sequential recommendation system based on the coupling relationship between item attributes and temporal patterns includes,
  • the data set construction unit is used to clean the historical data of interaction between users and items, and construct user behavior sequence data sets according to the order of interaction time between users and items;
  • the data set division unit is used to divide the user behavior sequence data set into a training set, a test set and a verification set;
  • the model construction unit is used to learn the interaction timing relationship between users and items and the timing relationship of item attributes through the GRU network, and learn the coupling relationship between item attributes and timing patterns through the attention mechanism network, and then combine the candidate item interaction prediction layer to construct a sequence recommendation model;
  • the model training unit is used to input the training set into the sequence recommendation model, and the sequence recommendation model learns the interactive sequence pattern representation that fuses the coupling relationship between item attributes and time sequence patterns, and then trains the learnable parameters of the sequence recommendation model;
  • the prediction unit is used to input the test set into the sequence recommendation model, and the sequence recommendation model predicts and sorts the interaction probabilities of all candidate items, and selects the top K items as the recommendation list.
  • the model building unit includes,
  • the acquisition module of the time series relationship between the user and the item is used to construct the sequence data using the item ID of the item that the user has interacted with at time t, and use the GRU network to learn the implicit interaction relationship representation vector q between the user and the item;
  • the item attribute temporal relationship acquisition module is used to form multiple attribute sequences according to each attribute of the items in the user-item interaction sequence, and use the GRU network to learn the temporal relationship representation vectors of each item attribute sequence;
  • the acquisition module of the coupling relationship between item attributes and timing patterns is used to use the implicit interaction relationship representation vector q between the user and the item as a query vector, and use the attention mechanism network to learn the coupling relationship between the timing relationship representation vector and the query vector of each item attribute sequence, so as to generate the final sequence representation vector based on coupling relationship analysis;
  • the candidate item interaction prediction module is used to perform similarity calculations on the final sequence representation vector based on coupling relationship analysis and the embedding vector of the candidate item, and then input it into the fully connected network to finally generate the interaction probability between each candidate item and the user, and generate Top-K recommendations based on the interaction probability sorting results.
  • the acquisition module of the time series relationship of item attributes includes,
  • the item attribute sequence construction sub-module is used to form each attribute value of the item in the sequence data into an attribute sequence according to the time sequence relationship between the user and the item, such as the item brand sequence, the item category sequence, etc.;
  • the encoding and conversion sub-module is used to express the attribute values in each attribute sequence using one-hot encoding, and convert them into low-dimensional dense vectors, namely embedding vectors, through a single-layer fully connected network;
  • Learning sub-module 1 is used to input each attribute sequence into the GRU network respectively, and learn the representation vector s ai of each attribute sequence respectively;
  • the acquisition module of the coupling relationship between the item attribute and the timing mode includes,
  • the dot product sub-module is used to use the implicit interaction relationship representation vector q between the user and the item as the query vector, and the implicit interaction relationship representation vector q between the user and the item and the representation vector s ai of each attribute sequence to calculate the similarity through the dot product to obtain the weight, and use the Softmax function to calculate the normalized weight ⁇ i , the formula is as follows:
  • the weighted summation sub-module is used to weight the weight ⁇ i and the corresponding attribute sequence representation vector to obtain the vector s,
  • W q , W k and W v are transformation matrices, which belong to learnable parameters
  • Learning sub-module 2 is used to further learn the vector s through one or more layers of fully-connected networks to obtain the final item attribute and temporal pattern summation relationship representation c.
  • the prediction unit includes,
  • the similarity calculation sub-module is used to convert the ID of the candidate item i into a low-dimensional dense vector through a single-layer fully connected network, and then multiply the corresponding elements with the final item attribute and timing mode summation relationship representation c to calculate the similarity vector d of the candidate item and the final item attribute and timing mode summation relationship representation c;
  • the learning sub-module three is used to further learn the similarity vector d through one or more layers of fully connected networks
  • the compression sub-module is used to use the Sigmoid activation function to compress the output to the range of [0,1] to obtain the probability of the candidate item as the item interacted with by the user at time t+1, that is, the final prediction result.
  • An electronic device comprising: memory and at least one processor
  • a computer program is stored on the memory
  • the at least one processor executes the computer program stored in the memory, so that the at least one processor executes the sequence recommendation method based on the coupling relationship between item attributes and timing patterns as described above.
  • a computer-readable storage medium wherein a computer program is stored in the computer-readable storage medium, and the computer program can be executed by a processor to implement the above-mentioned sequence recommendation method based on the coupling relationship between item attributes and timing patterns.
  • the present invention integrates the interaction sequence mode between the user and the item and the coupling relationship between each attribute of the item into the sequence recommendation model, and makes full use of the coupling relationship between the user and the interaction session sequence mode of the item and the item attribute information to effectively improve the accuracy of recommendation, thereby improving user satisfaction;
  • the present invention learns the interaction sequence mode representation of the coupling relationship between the item attribute and the timing mode by constructing and training the network model composed of the item timing relationship learning network, the item attribute timing relationship learning network, the coupling relationship learning network, and the candidate item interaction prediction layer, so as to generate a prediction of the interaction probability of the candidate items, thereby realizing Top-K recommendation;
  • the present invention considers the coupling relationship between the user-item interaction sequence mode and the item attributes, and improves the accuracy of recommendation;
  • the present invention separately considers the impact of the subsequences formed by each attribute of the item on the final recommendation result
  • the present invention effectively combines the recurrent neural network, attention mechanism and collaborative filtering to improve the interpretability of sequence recommendation.
  • this embodiment provides a sequence recommendation method based on the coupling relationship between item attributes and timing patterns.
  • the method is specifically as follows:
  • each user behavior sequence data set includes multiple sub-sequence data sets, and the sub-sequence data sets include item ID sub-sequence data sets and item attribute sub-sequence data sets, and the item attribute sub-sequence data sets are constructed according to the attributes of the items.
  • the following table shows an example of the historical data of interaction between users and items:
  • Training sequence recommendation model input the training set into the sequence recommendation model, and the sequence recommendation model learns the interactive sequence pattern representation that combines the coupling relationship between item attributes and time sequence patterns, and then trains the learnable parameters of the sequence recommendation model;
  • Sequence recommendation model prediction input the test set into the sequence recommendation model, the sequence recommendation model predicts the interaction probabilities of all candidate items and sorts them, and selects the top K items as the recommendation list.
  • the construction sequence recommendation model in step S3 of this embodiment is specifically as follows:
  • step S302 of this embodiment is specifically as follows:
  • each attribute value of the item in the sequence data is respectively formed into an attribute sequence, such as an item brand sequence, an item category sequence, etc.;
  • the coupling relationship between the acquired item attributes and the timing mode in step S303 of this embodiment is specifically as follows:
  • the implicit interaction relationship representation vector q between the user and the item as the query vector, the implicit interaction relationship representation vector q between the user and the item and the representation vector s ai of each attribute sequence are calculated by dot product similarity to obtain the weight, and the normalized weight ⁇ i is obtained by using the Softmax function.
  • the formula is as follows:
  • W q , W k and W v are transformation matrices, which belong to learnable parameters
  • the interaction prediction of candidate items in step S304 of this embodiment is specifically as follows:
  • Similarity calculation convert the ID of the candidate item i into a low-dimensional dense vector through a single-layer fully connected network, and then multiply the corresponding elements with the final item attribute and timing mode summation relationship representation c to calculate the similarity vector d of the candidate item and the final item attribute and timing mode summation relationship representation c;
  • this embodiment provides a sequential recommendation system based on the coupling relationship between item attributes and timing patterns, the system includes:
  • the data set construction unit is used to clean the historical data of interaction between users and items, and construct user behavior sequence data sets according to the order of interaction time between users and items;
  • the data set division unit is used to divide the user behavior sequence data set into a training set, a test set and a verification set;
  • the model construction unit is used to learn the interaction timing relationship between users and items and the timing relationship of item attributes through the GRU network, and learn the coupling relationship between item attributes and timing patterns through the attention mechanism network, and then combine the candidate item interaction prediction layer to construct a sequence recommendation model;
  • the model training unit is used to input the training set into the sequence recommendation model, and the sequence recommendation model learns the interactive sequence pattern representation that fuses the coupling relationship between item attributes and time sequence patterns, and then trains the learnable parameters of the sequence recommendation model;
  • the prediction unit is used to input the test set into the sequence recommendation model, and the sequence recommendation model predicts and sorts the interaction probabilities of all candidate items, and selects the top K items as the recommendation list.
  • the model construction unit in this embodiment includes,
  • the acquisition module of the time series relationship between the user and the item is used to construct the sequence data using the item ID of the item that the user has interacted with at time t, and use the GRU network to learn the implicit interaction relationship representation vector q between the user and the item;
  • the item attribute temporal relationship acquisition module is used to form multiple attribute sequences according to each attribute of the items in the user-item interaction sequence, and use the GRU network to learn the temporal relationship representation vectors of each item attribute sequence;
  • the acquisition module of the coupling relationship between item attributes and timing patterns is used to use the implicit interaction relationship representation vector q between the user and the item as a query vector, and use the attention mechanism network to learn the coupling relationship between the timing relationship representation vector and the query vector of each item attribute sequence, so as to generate the final sequence representation vector based on coupling relationship analysis;
  • the candidate item interaction prediction module is used to perform similarity calculations on the final sequence representation vector based on coupling relationship analysis and the embedding vector of the candidate item, and then input it into the fully connected network to finally generate the interaction probability between each candidate item and the user, and generate Top-K recommendations based on the interaction probability sorting results.
  • the acquisition module of the time series relationship of item attributes in this embodiment includes,
  • the item attribute sequence construction sub-module is used to form each attribute value of the item in the sequence data into an attribute sequence according to the time sequence relationship between the user and the item, such as the item brand sequence, the item category sequence, etc.;
  • the encoding and conversion sub-module is used to express the attribute values in each attribute sequence using one-hot encoding, and convert them into low-dimensional dense vectors, namely embedding vectors, through a single-layer fully connected network;
  • Learning sub-module 1 is used to input each attribute sequence into the GRU network respectively, and learn the representation vector s ai of each attribute sequence respectively;
  • the dot product sub-module is used to use the implicit interaction relationship representation vector q between the user and the item as the query vector, and the implicit interaction relationship representation vector q between the user and the item and the representation vector s ai of each attribute sequence to calculate the similarity through the dot product to obtain the weight, and use the Softmax function to calculate the normalized weight ⁇ i , the formula is as follows:
  • the weighted sum submodule is used to weight the weight ⁇ i and the corresponding attribute sequence representation vector to obtain the vector s,
  • W q , W k and W v are transformation matrices, which belong to learnable parameters
  • Learning sub-module 2 is used to further learn the vector s through one or more layers of fully-connected networks to obtain the final item attribute and temporal pattern summation relationship representation c.
  • the similarity calculation sub-module is used to convert the ID of the candidate item i into a low-dimensional dense vector through a single-layer fully connected network, and then multiply the corresponding elements with the final item attribute and timing mode summation relationship representation c to calculate the similarity vector d of the candidate item and the final item attribute and timing mode summation relationship representation c;
  • the learning sub-module three is used to further learn the similarity vector d through one or more layers of fully connected networks
  • the compression sub-module is used to use the Sigmoid activation function to compress the output to the range of [0,1] to obtain the probability of the candidate item as the item interacted with by the user at time t+1, that is, the final prediction result.
  • This embodiment also provides an electronic device, including: a memory and a processor;
  • the memory stores computer-executable instructions
  • the processor executes the computer-executed instructions stored in the memory, so that the processor executes the sequence recommendation method based on the coupling relationship between item attributes and temporal patterns in any embodiment of the present invention.
  • the processor may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), off-the-shelf programmable gate arrays (FPGAs) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, and the like.
  • the processor may be a microprocessor or the processor may be any conventional processor or the like.
  • the memory can be used to store computer programs and/or modules, and the processor implements various functions of the electronic device by running or executing the computer programs and/or modules stored in the memory and calling data stored in the memory.
  • the memory may mainly include a program storage area and a data storage area, wherein the program storage area may store an operating system, an application program required by at least one function, etc.; the data storage area may store data created according to the use of the terminal, etc.
  • the memory can also include high-speed random access memory, and can also include non-volatile memory, such as a hard disk, internal memory, plug-in hard disk, memory card (SMC), secure digital (SD) card, flash memory card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.
  • SMC plug-in hard disk
  • SD secure digital
  • This embodiment also provides a computer-readable storage medium, in which a plurality of instructions are stored, and the instructions are loaded by a processor, so that the processor executes the sequence recommendation method based on the coupling relationship between item attributes and timing patterns in any embodiment of the present invention.
  • a system or device equipped with a storage medium may be provided, on which a software program code for realizing the functions of any of the above embodiments is stored, and the computer (or CPU or MPU) of the system or device is allowed to read and execute the program code stored in the storage medium.
  • the program code itself read from the storage medium can realize the function of any one of the above-mentioned embodiments, so the program code and the storage medium storing the program code constitute a part of the present invention.
  • Examples of storage media for providing program code include floppy disks, hard disks, magneto-optical disks, optical disks (such as CD-ROM, CD-R, CD-RW, DVD-ROM, DVD-RYM, DVD-RW, DVD+RW), magnetic tape, non-volatile memory cards, and ROM.
  • the program code can be downloaded from a server computer via a communication network.
  • the program code read from the storage medium is written into the memory provided in the expansion board inserted into the computer or written into the memory provided in the expansion unit connected to the computer, and then based on the instructions of the program code, the CPU installed on the expansion board or the expansion unit executes part or all of the actual operations, thereby realizing the functions of any of the above embodiments.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Business, Economics & Management (AREA)
  • Data Mining & Analysis (AREA)
  • Databases & Information Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Accounting & Taxation (AREA)
  • Finance (AREA)
  • Development Economics (AREA)
  • Computational Linguistics (AREA)
  • Strategic Management (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Computing Systems (AREA)
  • Artificial Intelligence (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • General Business, Economics & Management (AREA)
  • Evolutionary Computation (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Health & Medical Sciences (AREA)
  • Marketing (AREA)
  • Economics (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Fuzzy Systems (AREA)
  • Probability & Statistics with Applications (AREA)
  • Game Theory and Decision Science (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

本发明公开了基于物品属性与时序模式耦合关系的序列推荐方法和系统,属于智能推荐技术领域,本发明要解决的技术问题为如何充分利用用户以及物品的交互会话序列模式与物品属性信息之间存在的耦合关系有效提高推荐的准确性,采用的技术方案为:该方法具体如下:构建数据集;划分数据集;构建序列推荐模型:通过GRU网络学习用户与物品交互时序关系、物品属性时序关系,并通过注意力机制网络学习物品属性与时序模式耦合关系,再结合候选物品交互预测层构建序列推荐模型;训练序列推荐模型:将训练集输入到序列推荐模型中,序列推荐模型学习融合物品属性与时序模式耦合关系的交互序列模式表示,进而训练出序列推荐模型的可学习参数;序列推荐模型预测。

Description

基于物品属性与时序模式耦合关系的序列推荐方法和系统 技术领域
本发明涉及智能推荐技术领域,具体地说是一种基于物品属性与时序模式耦合关系的序列推荐方法和系统。
背景技术
随着推荐系统的普遍使用,基于序列的推荐逐渐成为推荐系统领域的一个重要子领域。与传统的推荐方法大多基于对用户信息建模不同,序列推荐方法并不过多关注用户的信息,而是通过挖掘会话序列数据中各个物品之间的关系为用户产生推荐。但现有的序列推荐方法往往都过度聚焦于序列数据本身,而忽略了一些辅助信息的重要性,没有充分利用物品属性信息以及物品属性信息与时间序列之间的耦合关系对推荐结果的影响,因而产生的推荐结果往往是不够准确的。实际上,用户与物品的交互会话序列模式与物品属性信息之间是存在耦合关系的,故如何充分利用用户以及物品的交互会话序列模式与物品属性信息之间存在的耦合关系有效提高推荐的准确性,进而提高用户的满意度是目前亟待解决的技术问题。
发明内容
本发明的技术任务是提供一种基于物品属性与时序模式耦合关系的序列推荐方法和系统,来解决如何充分利用用户以及物品的交互会话序列模式与物品属性信息之间存在的耦合关系有效提高推荐的准确性,进而提高用户的满意度的问题。
本发明的技术任务是按以下方式实现的,一种基于物品属性与时序模式耦合关系的序列推荐方法,该方法具体如下:
构建数据集:对用户与物品交互历史数据进行清洗,并按照用户与物品交互时间的先后顺序构建用户行为序列数据集;
划分数据集:将用户行为序列数据集划分为训练集、测试集和验证集;其中,验证集用于训练阶段调整序列推荐模型的超参数。
构建序列推荐模型:通过GRU网络学习用户与物品交互时序关系、物品属性时序关系,并通过注意力机制网络学习物品属性与时序模式耦合关系,再结合候选物品交互预测层构建序列推荐模型;
训练序列推荐模型:将训练集输入到序列推荐模型中,序列推荐模型学习融合物品属性与时序模式耦合关系的交互序列模式表示,进而训练出序列推荐模型的可学习参数;
序列推荐模型预测:将测试集输入到序列推荐模型中,序列推荐模型预测所有候选物品的交互概率并进行排序,选取前K个物品作为推荐列表。
作为优选,每个用户行为序列数据集包括多个子序列数据集,子序列数据集包括物品ID子序列数据集和物品属性子序列数据集,物品属性子序列数据集按照物品各个属性构建。
更优地,构建序列推荐模型具体如下:
获取用户与物品交互时序关系:将用户在t个时刻交互过的物品使用物品ID构建成序列数据,并使用GRU网络学习用户与物品的隐式交互关系表示向量q;
获取物品属性时序关系:将用户与物品交互序列中的物品按照物品每个属性形成多个属性序列,并使用GRU网络学习各个物品属性序列的时序关系表示向量;
获取物品属性与时序模式耦合关系:将用户与物品的隐式交互关系表示向量q作为查询向量,使用注意力机制网络学习各个物品属性序列的时序关系表示向量与查询向量的耦合关系,从而生成最终的基于耦合关系分析的序列表示向量;
候选物品交互预测:最终的基于耦合关系分析的序列表示向量与候选物品的嵌入向量进行相似性运算后,输入全连接网络中最终产生各个候选物品与用户的交互概率,根据交互概率排序结果产生Top-K推荐。
更优地,获取物品属性时序关系具体如下:
构建物品属性序列:根据用户与物品交互时序关系将序列数据中物品的各个属性值分别形成属性序列,如物品品牌序列、物品类别序列等;
将各个属性序列中的属性值使用one-hot编码进行表示,并分别通过单层全连接网络转换为低维度稠密向量,即嵌入向量;
将每个属性序列分别输入到GRU网络中,分别学习各个属性序列的表示向量s ai
更优地,获取物品属性与时序模式耦合关系具体如下:
将用户与物品的隐式交互关系表示向量q作为查询向量,用户与物品的隐式交互关系表示向量q与各个属性序列的表示向量s ai通过点积进行相似度计算从而得到权重,并使用Softmax函数计算得到归一化后的权重α i,公式如下:
α i=softmax(dot(W qq,W ks ai));
将权重α i和相应的属性序列表示向量加权求和从而得到向量s,
s=Σ iα i(W vs ai);
其中,W q、W k和W v分别为变换矩阵,属于可学习的参数;
将向量s经过一层或多层全连接网络进一步学习得到最终的物品属性与时序模式谋和关系表示c;
候选物品交互预测具体如下:
相似性计算:将候选物品i的ID通过单层全连接网络转换为低维度稠密向量后与最终的物品属性与时序模式谋和关系表示c进行对应元素相乘计算候选物品与最终的物品属性与时序模式谋和关系表示c的相似性向量d;
将相似性向量d通过一层或多层全连接网络进一步学习;
使用Sigmoid激活函数将输出压缩到[0,1]范围内得到候选物品作为t+1时刻用户交互的物品的概率,即最终的预测结果。
一种基于物品属性与时序模式耦合关系的序列推荐系统,该系统包括,
数据集构建单元,用于对用户与物品交互历史数据进行清洗,并按照用户与物品交互时间的先后顺序构建用户行为序列数据集;
数据集划分单元,用于将用户行为序列数据集划分为训练集、测试集和验证集;
模型构建单元,用于通过GRU网络学习用户与物品交互时序关系、物品属性时序关系,并通过注意力机制网络学习物品属性与时序模式耦合关系,再结合候选物品交互预测层构建序列推荐模型;
模型训练单元,用于将训练集输入到序列推荐模型中,序列推荐模型学习融合物品属性与时序模式耦合关系的交互序列模式表示,进而训练出序列推荐模型的可学习参数;
预测单元,用于将测试集输入到序列推荐模型中,序列推荐模型预测所有候选物品的交互概率并进行排序,选取前K个物品作为推荐列表。
作为优选,所述模型构建单元包括,
用户与物品交互时序关系获取模块,用于将用户在t个时刻交互过的物品使用物品ID构建成序列数据,并使用GRU网络学习用户与物品的隐式交互关系表示向量q;
物品属性时序关系获取模块,用于将用户与物品交互序列中的物品按照物品每个属性形成多个属性序列,并使用GRU网络学习各个物品属性序列的时序关系表示向量;
物品属性与时序模式耦合关系获取模块,用于将用户与物品的隐式交互关系表示向量q作为查询向量,使用注意力机制网络学习各个物品属性序列的时序关系表示向量与查询向量的耦合关系,从而生成最终的基于耦合关系分析的序列表示向量;
候选物品交互预测模块,用于最终的基于耦合关系分析的序列表示向量与候选物品的嵌入向量进行相似性运算后,输入全连接网络中最终产生各个候选物品与用户的交互概率,根据交互概率排序结果产生Top-K推荐。
更优地,所述物品属性时序关系获取模块包括,
物品属性序列构建子模块,用于根据用户与物品交互时序关系将序列数据中物品的各个属性值分别形成属性序列,如物品品牌序列、物品类别序列等;
编码及转换子模块,用于将各个属性序列中的属性值使用one-hot编码进行表示,并分别通过单层全连接网络转换为低维度稠密向量,即嵌入向量;
学习子模块一,用于将每个属性序列分别输入到GRU网络中,分别学习各个属性序列的表示向量s ai
所述物品属性与时序模式耦合关系获取模块包括,
点积子模块,用于将用户与物品的隐式交互关系表示向量q作为查询向量,用户与物品的隐式交互关系表示向量q与各个属性序列的表示向量s ai通过点积进行相似度计算从而得到权重,并使用Softmax函数计算得到归一化后的权重α i,公式如下:
α i=softmax(dot(W qq,W ks ai));
加权求和子模块,用于将权重α i和相应的属性序列表示向量加权求和从而得到向量s,
s=Σ iα i(W vs ai);
其中,W q、W k和W v分别为变换矩阵,属于可学习的参数;
学习子模块二,用于将向量s经过一层或多层全连接网络进一步学习得到最终的物品属性与时序模式谋和关系表示c。
所述预测单元包括,
相似性计算子模块,用于将候选物品i的ID通过单层全连接网络转换为低维度稠密向量后与最终的物品属性与时序模式谋和关系表示c进行对应元素相乘计算候选物品与最终的物品属性与时序模式谋和关系表示c的相似性向量d;
学习子模块三,用于将相似性向量d通过一层或多层全连接网络进一步学习;
压缩子模块,用于使用Sigmoid激活函数将输出压缩到[0,1]范围内得到候选物品作为t+1时刻用户交互的物品的概率,即最终的预测结果。
一种电子设备,包括:存储器和至少一个处理器;
其中,所述存储器上存储有计算机程序;
所述至少一个处理器执行所述存储器存储的计算机程序,使得所述至少一 个处理器执行如上述的基于物品属性与时序模式耦合关系的序列推荐方法。
一种计算机可读存储介质,所述计算机可读存储介质中存储有计算机程序,所述计算机程序可被处理器执行以实现如上述的基于物品属性与时序模式耦合关系的序列推荐方法。
本发明的基于物品属性与时序模式耦合关系的序列推荐方法和系统具有以下优点:
(一)本发明将用户与物品的交互序列模式和物品各个属性之间的耦合关系融入到序列推荐模型中,充分利用用户以及物品的交互会话序列模式与物品属性信息之间存在的耦合关系有效提高推荐的准确性,进而提高用户的满意度;
(二)本发明通过构建并训练由物品时序关系学习网络、物品属性时序关系学习网络、耦合关系学习网络,候选物品交互预测层构成的网络模型学习融合物品属性与时序模式耦合关系的交互序列模式表示,产生对候选物品交互概率的预测,从而实现Top-K推荐;
(三)本发明考虑了用户与物品交互序列模式与物品属性之间的耦合关系,提高了推荐的准确性;
(四)本发明分别考虑物品各个属性形成的子序列对最终推荐结果的影响;
(五)本发明将循环神经网络、注意力机制以及协同过滤有效结合,以提高序列推荐的可解释性。
附图说明
下面结合附图对本发明进一步说明。
附图1为基于物品属性与时序模式耦合关系的序列推荐方法的流程框图;
附图2为构建序列推荐模型的流程框图;
附图3为基于物品属性与时序模式耦合关系的序列推荐系统的结构框图;
附图4为序列推荐模型的示意图。
具体实施方式
参照说明书附图和具体实施例对本发明的基于物品属性与时序模式耦合关系的序列推荐方法和系统作以下详细地说明。
实施例1:
如附图1所示,本实施例提供了一种基于物品属性与时序模式耦合关系的序列推荐方法,该方法具体如下:
S1、构建数据集:对用户与物品交互历史数据进行清洗,并按照用户与物品交互时间的先后顺序构建用户行为序列数据集;其中,每个用户行为序列数据集包括多个子序列数据集,子序列数据集包括物品ID子序列数据集和物品属性子序列数据集,物品属性子序列数据集按照物品各个属性构建。
以Amazon数据集为例,用户与物品交互历史数据示例如下表所示:
Figure PCTCN2023070574-appb-000001
物品属性信息示例如下表所示:
Figure PCTCN2023070574-appb-000002
S2、划分数据集:将用户行为序列数据集划按照6:2:2的比例分为训练集、测试集和验证集;
S3、构建序列推荐模型:通过GRU网络学习用户与物品交互时序关系、物品属性时序关系,并通过注意力机制网络学习物品属性与时序模式耦合关系,再结合候选物品交互预测层构建序列推荐模型;
S4、训练序列推荐模型:将训练集输入到序列推荐模型中,序列推荐模型学习融合物品属性与时序模式耦合关系的交互序列模式表示,进而训练出序列推荐模型的可学习参数;
S5、序列推荐模型预测:将测试集输入到序列推荐模型中,序列推荐模型预测所有候选物品的交互概率并进行排序,选取前K个物品作为推荐列表。
如附图2所示,本实施例步骤S3中的构建序列推荐模型具体如下:
S301、获取用户与物品交互时序关系:将用户在t个时刻交互过的物品使用物品ID构建成序列数据,并使用GRU网络学习用户与物品的隐式交互关系表示向量q;
S302、获取物品属性时序关系:将用户与物品交互序列中的物品按照物品每个属性形成多个属性序列,并使用GRU网络学习各个物品属性序列的时序关系表示向量;
S303、获取物品属性与时序模式耦合关系:将用户与物品的隐式交互关系表示向量q作为查询向量,使用注意力机制网络学习各个物品属性序列的时序关系表示向量与查询向量的耦合关系,从而生成最终的基于耦合关系分析的序列表示向量;
S304、候选物品交互预测:最终的基于耦合关系分析的序列表示向量与候选物品的嵌入向量进行相似性运算后,输入全连接网络中最终产生各个候选物品与用户的交互概率,根据交互概率排序结果产生Top-K推荐。
如附图4所示,本实施例步骤S302中的获取物品属性时序关系具体如下:
S30201、构建物品属性序列:根据用户与物品交互时序关系将序列数据中物品的各个属性值分别形成属性序列,如物品品牌序列、物品类别序列等;
S30202、将各个属性序列中的属性值使用one-hot编码进行表示,并分别通过单层全连接网络转换为低维度稠密向量,即嵌入向量;
S30203、将每个属性序列分别输入到GRU网络中,分别学习各个属性序列的表示向量s ai
本实施例步骤S303中的获取物品属性与时序模式耦合关系具体如下:
S30301、将用户与物品的隐式交互关系表示向量q作为查询向量,用户与物品的隐式交互关系表示向量q与各个属性序列的表示向量s ai通过点积进行相似度计算从而得到权重,并使用Softmax函数计算得到归一化后的权重α i,公式如下:
α i=softmax(dot(W qq,W ks ai));
S30302、将权重α i和相应的属性序列表示向量加权求和从而得到向量s,
s=Σ iα i(W vs ai);
其中,W q、W k和W v分别为变换矩阵,属于可学习的参数;
S30303、将向量s经过一层或多层全连接网络进一步学习得到最终的物品属性与时序模式谋和关系表示c;
本实施例步骤S304中的候选物品交互预测具体如下:
S30401、相似性计算:将候选物品i的ID通过单层全连接网络转换为低维度稠密向量后与最终的物品属性与时序模式谋和关系表示c进行对应元素相乘计算候选物品与最终的物品属性与时序模式谋和关系表示c的相似性向量d;
S30402、将相似性向量d通过一层或多层全连接网络进一步学习;
S30403、使用Sigmoid激活函数将输出压缩到[0,1]范围内得到候选物品作为t+1时刻用户交互的物品的概率,即最终的预测结果。
实施例2:
如附图3所示,本实施例提供了一种基于物品属性与时序模式耦合关系的序列推荐系统,该系统包括,
数据集构建单元,用于对用户与物品交互历史数据进行清洗,并按照用户与物品交互时间的先后顺序构建用户行为序列数据集;
数据集划分单元,用于将用户行为序列数据集划分为训练集、测试集和验证集;
模型构建单元,用于通过GRU网络学习用户与物品交互时序关系、物品属性时序关系,并通过注意力机制网络学习物品属性与时序模式耦合关系,再结合候选物品交互预测层构建序列推荐模型;
模型训练单元,用于将训练集输入到序列推荐模型中,序列推荐模型学习融合物品属性与时序模式耦合关系的交互序列模式表示,进而训练出序列推荐模型的可学习参数;
预测单元,用于将测试集输入到序列推荐模型中,序列推荐模型预测所有候选物品的交互概率并进行排序,选取前K个物品作为推荐列表。
本实施例中的模型构建单元包括,
用户与物品交互时序关系获取模块,用于将用户在t个时刻交互过的物品使用物品ID构建成序列数据,并使用GRU网络学习用户与物品的隐式交互关系表示向量q;
物品属性时序关系获取模块,用于将用户与物品交互序列中的物品按照物品每个属性形成多个属性序列,并使用GRU网络学习各个物品属性序列的时序关系表示向量;
物品属性与时序模式耦合关系获取模块,用于将用户与物品的隐式交互关系表示向量q作为查询向量,使用注意力机制网络学习各个物品属性序列的时序关系表示向量与查询向量的耦合关系,从而生成最终的基于耦合关系分析的序列表示向量;
候选物品交互预测模块,用于最终的基于耦合关系分析的序列表示向量与候选物品的嵌入向量进行相似性运算后,输入全连接网络中最终产生各个候选物品与用户的交互概率,根据交互概率排序结果产生Top-K推荐。
如附图4所示,本实施例中的物品属性时序关系获取模块包括,
物品属性序列构建子模块,用于根据用户与物品交互时序关系将序列数据中物品的各个属性值分别形成属性序列,如物品品牌序列、物品类别序列等;
编码及转换子模块,用于将各个属性序列中的属性值使用one-hot编码进行表示,并分别通过单层全连接网络转换为低维度稠密向量,即嵌入向量;
学习子模块一,用于将每个属性序列分别输入到GRU网络中,分别学习各个属性序列的表示向量s ai
本实施例中的物品属性与时序模式耦合关系获取模块包括,
点积子模块,用于将用户与物品的隐式交互关系表示向量q作为查询向量,用户与物品的隐式交互关系表示向量q与各个属性序列的表示向量s ai通过点积进行相似度计算从而得到权重,并使用Softmax函数计算得到归一化后的权重α i,公式如下:
α i=softmax(dot(W qq,W ks ai));
加权求和子模块,用于将权重α i和相应的属性序列表示向量加权求和从而得到向量s,
s=Σ iα i(W vs ai);
其中,W q、W k和W v分别为变换矩阵,属于可学习的参数;
学习子模块二,用于将向量s经过一层或多层全连接网络进一步学习得到最终的物品属性与时序模式谋和关系表示c。
本实施例中的预测单元包括,
相似性计算子模块,用于将候选物品i的ID通过单层全连接网络转换为低维度稠密向量后与最终的物品属性与时序模式谋和关系表示c进行对应元素相乘计算候选物品与最终的物品属性与时序模式谋和关系表示c的相似性向量d;
学习子模块三,用于将相似性向量d通过一层或多层全连接网络进一步学习;
压缩子模块,用于使用Sigmoid激活函数将输出压缩到[0,1]范围内得到候选物品作为t+1时刻用户交互的物品的概率,即最终的预测结果。
实施例3:
本实施例还提供了一种电子设备,包括:存储器和处理器;
其中,存储器存储计算机执行指令;
处理器执行所述存储器存储的计算机执行指令,使得处理器执行本发明任一实施例中的基于物品属性与时序模式耦合关系的序列推荐方法。
处理器可以是中央处理单元(CPU),还可以是其他通用处理器、数字信号处理器(DSP)、专用集成电路(ASIC)、现成可编程门阵列(FPGA)或者其他可编程逻辑器件、分立门或者晶体管逻辑器件、分立硬件组件等。通过处理器可以是微处理器或者该处理器也可以是任何常规的处理器等。
存储器可用于储存计算机程序和/或模块,处理器通过运行或执行存储在存储器内的计算机程序和/或模块,以及调用存储在存储器内的数据,实现电子设备的各种功能。存储器可主要包括存储程序区和存储数据区,其中,存储程序区可存储操作系统、至少一个功能所需的应用程序等;存储数据区可存储根据终端的使用所创建的数据等。此外,存储器还可以包括高速随机存取存储器,还可以包括非易失性存储器,例如硬盘、内存、插接式硬盘,只能存储卡(SMC),安全数字(SD)卡,闪存卡、至少一个磁盘存储期间、闪存器件、或其他易失性固态存储器件。
实施例4:
本实施例还提供了一种计算机可读存储介质,其中存储有多条指令,指令由处理器加载,使处理器执行本发明任一实施例中的基于物品属性与时序模式耦合关系的序列推荐方法。具体地,可以提供配有存储介质的系统或者装置,在该存储介质上存储着实现上述实施例中任一实施例的功能的软件程序代码,且使该系统或者装置的计算机(或CPU或MPU)读出并执行存储在存储介质中的程序代码。
在这种情况下,从存储介质读取的程序代码本身可实现上述实施例中任何一项实施例的功能,因此程序代码和存储程序代码的存储介质构成了本发明的一部分。
用于提供程序代码的存储介质实施例包括软盘、硬盘、磁光盘、光盘(如CD-ROM、CD-R、CD-RW、DVD-ROM、DVD-RYM、DVD-RW、DVD+RW)、磁带、非易失性存储卡和ROM。可选择地,可以由通信网络从服务器计算机上下载程序代码。
此外,应该清楚的是,不仅可以通过执行计算机所读出的程序代码,而且可以通过基于程序代码的指令使计算机上操作的操作系统等来完成部分或者全 部的实际操作,从而实现上述实施例中任意一项实施例的功能。
此外,可以理解的是,将由存储介质读出的程序代码写到插入计算机内的扩展板中所设置的存储器中或者写到与计算机相连接的扩展单元中设置的存储器中,随后基于程序代码的指令使安装在扩展板或者扩展单元上的CPU等来执行部分和全部实际操作,从而实现上述实施例中任一实施例的功能。
最后应说明的是:以上各实施例仅用以说明本发明的技术方案,而非对其限制;尽管参照前述各实施例对本发明进行了详细的说明,本领域的普通技术人员应当理解:其依然可以对前述各实施例所记载的技术方案进行修改,或者对其中部分或者全部技术特征进行等同替换;而这些修改或者替换,并不使相应技术方案的本质脱离本发明各实施例技术方案的范围。

Claims (10)

  1. 一种基于物品属性与时序模式耦合关系的序列推荐方法,其特征在于,该方法具体如下:
    构建数据集:对用户与物品交互历史数据进行清洗,并按照用户与物品交互时间的先后顺序构建用户行为序列数据集;
    划分数据集:将用户行为序列数据集划分为训练集、测试集和验证集;
    构建序列推荐模型:通过GRU网络学习用户与物品交互时序关系、物品属性时序关系,并通过注意力机制网络学习物品属性与时序模式耦合关系,再结合候选物品交互预测层构建序列推荐模型;
    训练序列推荐模型:将训练集输入到序列推荐模型中,序列推荐模型学习融合物品属性与时序模式耦合关系的交互序列模式表示,进而训练出序列推荐模型的可学习参数;
    序列推荐模型预测:将测试集输入到序列推荐模型中,序列推荐模型预测所有候选物品的交互概率并进行排序,选取前K个物品作为推荐列表。
  2. 根据权利要求1所述的基于物品属性与时序模式耦合关系的序列推荐方法,其特征在于,每个用户行为序列数据集包括多个子序列数据集,子序列数据集包括物品ID子序列数据集和物品属性子序列数据集,物品属性子序列数据集按照物品各个属性构建。
  3. 根据权利要求1或2所述的基于物品属性与时序模式耦合关系的序列推荐方法,其特征在于,构建序列推荐模型具体如下:
    获取用户与物品交互时序关系:将用户在t个时刻交互过的物品使用物品ID构建成序列数据,并使用GRU网络学习用户与物品的隐式交互关系表示向量q;
    获取物品属性时序关系:将用户与物品交互序列中的物品按照物品每个属 性形成多个属性序列,并使用GRU网络学习各个物品属性序列的时序关系表示向量;
    获取物品属性与时序模式耦合关系:将用户与物品的隐式交互关系表示向量q作为查询向量,使用注意力机制网络学习各个物品属性序列的时序关系表示向量与查询向量的耦合关系,从而生成最终的基于耦合关系分析的序列表示向量;
    候选物品交互预测:最终的基于耦合关系分析的序列表示向量与候选物品的嵌入向量进行相似性运算后,输入全连接网络中最终产生各个候选物品与用户的交互概率,根据交互概率排序结果产生Top-K推荐。
  4. 根据权利要求3所述的基于物品属性与时序模式耦合关系的序列推荐方法,其特征在于,获取物品属性时序关系具体如下:
    构建物品属性序列:根据用户与物品交互时序关系将序列数据中物品的各个属性值分别形成属性序列;
    将各个属性序列中的属性值使用one-hot编码进行表示,并分别通过单层全连接网络转换为低维度稠密向量,即嵌入向量;
    将每个属性序列分别输入到GRU网络中,分别学习各个属性序列的表示向量s ai
  5. 根据权利要求4所述的基于物品属性与时序模式耦合关系的序列推荐方法,其特征在于,获取物品属性与时序模式耦合关系具体如下:
    将用户与物品的隐式交互关系表示向量q作为查询向量,用户与物品的隐式交互关系表示向量q与各个属性序列的表示向量s ai通过点积进行相似度计算从而得到权重,并使用Softmax函数计算得到归一化后的权重α i,公式如下:
    α i=softmax(dot(W qq,W ks ai));
    将权重α i和相应的属性序列表示向量加权求和从而得到向量s,
    s=Σ iα i(W vs ai);
    其中,W q、W k和W v分别为变换矩阵,属于可学习的参数;
    将向量s经过一层或多层全连接网络进一步学习得到最终的物品属性与时序模式谋和关系表示c;
    候选物品交互预测具体如下:
    相似性计算:将候选物品i的ID通过单层全连接网络转换为低维度稠密向量后与最终的物品属性与时序模式谋和关系表示c进行对应元素相乘计算候选物品与最终的物品属性与时序模式谋和关系表示c的相似性向量d;
    将相似性向量d通过一层或多层全连接网络进一步学习;
    使用Sigmoid激活函数将输出压缩到[0,1]范围内得到候选物品作为t+1时刻用户交互的物品的概率,即最终的预测结果。
  6. 一种基于物品属性与时序模式耦合关系的序列推荐系统,其特征在于,该系统包括,
    数据集构建单元,用于对用户与物品交互历史数据进行清洗,并按照用户与物品交互时间的先后顺序构建用户行为序列数据集;
    数据集划分单元,用于将用户行为序列数据集划分为训练集、测试集和验证集;
    模型构建单元,用于通过GRU网络学习用户与物品交互时序关系、物品属性时序关系,并通过注意力机制网络学习物品属性与时序模式耦合关系,再结合候选物品交互预测层构建序列推荐模型;
    模型训练单元,用于将训练集输入到序列推荐模型中,序列推荐模型学习融合物品属性与时序模式耦合关系的交互序列模式表示,进而训练出序列推荐模型的可学习参数;
    预测单元,用于将测试集输入到序列推荐模型中,序列推荐模型预测所有 候选物品的交互概率并进行排序,选取前K个物品作为推荐列表。
  7. 根据权利要求6所述的基于物品属性与时序模式耦合关系的序列推荐系统,其特征在于,所述模型构建单元包括,
    用户与物品交互时序关系获取模块,用于将用户在t个时刻交互过的物品使用物品ID构建成序列数据,并使用GRU网络学习用户与物品的隐式交互关系表示向量q;
    物品属性时序关系获取模块,用于将用户与物品交互序列中的物品按照物品每个属性形成多个属性序列,并使用GRU网络学习各个物品属性序列的时序关系表示向量;
    物品属性与时序模式耦合关系获取模块,用于将用户与物品的隐式交互关系表示向量q作为查询向量,使用注意力机制网络学习各个物品属性序列的时序关系表示向量与查询向量的耦合关系,从而生成最终的基于耦合关系分析的序列表示向量;
    候选物品交互预测模块,用于最终的基于耦合关系分析的序列表示向量与候选物品的嵌入向量进行相似性运算后,输入全连接网络中最终产生各个候选物品与用户的交互概率,根据交互概率排序结果产生Top-K推荐。
  8. 根据权利要求7所述的基于物品属性与时序模式耦合关系的序列推荐系统,其特征在于,所述物品属性时序关系获取模块包括,
    物品属性序列构建子模块,用于根据用户与物品交互时序关系将序列数据中物品的各个属性值分别形成属性序列;
    编码及转换子模块,用于将各个属性序列中的属性值使用one-hot编码进行表示,并分别通过单层全连接网络转换为低维度稠密向量,即嵌入向量;
    学习子模块一,用于将每个属性序列分别输入到GRU网络中,分别学习各个属性序列的表示向量s ai
    所述物品属性与时序模式耦合关系获取模块包括,
    点积子模块,用于将用户与物品的隐式交互关系表示向量q作为查询向量,用户与物品的隐式交互关系表示向量q与各个属性序列的表示向量s ai通过点积进行相似度计算从而得到权重,并使用Softmax函数计算得到归一化后的权重α i,公式如下:
    α i=softmax(dot(W qq,W ks ai));
    加权求和子模块,用于将权重α i和相应的属性序列表示向量加权求和从而得到向量s,
    s=Σ iα i(W vs ai);
    其中,W q、W k和W v分别为变换矩阵,属于可学习的参数;
    学习子模块二,用于将向量s经过一层或多层全连接网络进一步学习得到最终的物品属性与时序模式谋和关系表示c。
    所述预测单元包括,
    相似性计算子模块,用于将候选物品i的ID通过单层全连接网络转换为低维度稠密向量后与最终的物品属性与时序模式谋和关系表示c进行对应元素相乘计算候选物品与最终的物品属性与时序模式谋和关系表示c的相似性向量d;
    学习子模块三,用于将相似性向量d通过一层或多层全连接网络进一步学习;
    压缩子模块,用于使用Sigmoid激活函数将输出压缩到[0,1]范围内得到候选物品作为t+1时刻用户交互的物品的概率,即最终的预测结果。
  9. 一种电子设备,其特征在于,包括:存储器和至少一个处理器;
    其中,所述存储器上存储有计算机程序;
    所述至少一个处理器执行所述存储器存储的计算机程序,使得所述至少一个处理器执行如权利要求1至5任一项所述的基于物品属性与时序模式耦合关 系的序列推荐方法。
  10. 一种计算机可读存储介质,其特征在于,所述计算机可读存储介质中存储有计算机程序,所述计算机程序可被处理器执行以实现如权利要求1至5中任一项所述的基于物品属性与时序模式耦合关系的序列推荐方法。
PCT/CN2023/070574 2022-09-26 2023-01-05 基于物品属性与时序模式耦合关系的序列推荐方法和系统 WO2023138381A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202211171763.3A CN115601098A (zh) 2022-09-26 2022-09-26 基于物品属性与时序模式耦合关系的序列推荐方法和系统
CN202211171763.3 2022-09-26

Publications (1)

Publication Number Publication Date
WO2023138381A1 true WO2023138381A1 (zh) 2023-07-27

Family

ID=84844082

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2023/070574 WO2023138381A1 (zh) 2022-09-26 2023-01-05 基于物品属性与时序模式耦合关系的序列推荐方法和系统

Country Status (3)

Country Link
CN (1) CN115601098A (zh)
LU (1) LU503730B1 (zh)
WO (1) WO2023138381A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116703529A (zh) * 2023-08-02 2023-09-05 山东省人工智能研究院 基于特征空间语义增强的对比学习推荐方法

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20200311585A1 (en) * 2019-03-31 2020-10-01 Palo Alto Networks Multi-model based account/product sequence recommender
CN112084407A (zh) * 2020-09-08 2020-12-15 辽宁工程技术大学 一种融合图神经网络和注意力机制的协同过滤推荐方法
CN112269927A (zh) * 2020-10-22 2021-01-26 辽宁工程技术大学 一种基于会话序列动态行为偏好耦合关系分析的推荐方法
CN112328893A (zh) * 2020-11-25 2021-02-05 重庆理工大学 一种基于记忆网络和协同注意力的推荐方法

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20200311585A1 (en) * 2019-03-31 2020-10-01 Palo Alto Networks Multi-model based account/product sequence recommender
CN112084407A (zh) * 2020-09-08 2020-12-15 辽宁工程技术大学 一种融合图神经网络和注意力机制的协同过滤推荐方法
CN112269927A (zh) * 2020-10-22 2021-01-26 辽宁工程技术大学 一种基于会话序列动态行为偏好耦合关系分析的推荐方法
CN112328893A (zh) * 2020-11-25 2021-02-05 重庆理工大学 一种基于记忆网络和协同注意力的推荐方法

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
ZHANG XIN YI: "Researh on Sequential Recommendation Method by Jointing Multi-Intention Perception and Sequentia Patterns", MASTER'S THESIS, no. 12, 30 May 2021 (2021-05-30), CN, pages 1 - 73, XP009547480, DOI: 10.27272/d.cnki.gshdu.2021.003101 *

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116703529A (zh) * 2023-08-02 2023-09-05 山东省人工智能研究院 基于特征空间语义增强的对比学习推荐方法
CN116703529B (zh) * 2023-08-02 2023-10-20 山东省人工智能研究院 基于特征空间语义增强的对比学习推荐方法

Also Published As

Publication number Publication date
LU503730B1 (en) 2023-10-27
CN115601098A (zh) 2023-01-13

Similar Documents

Publication Publication Date Title
He et al. STCNN: A spatio-temporal convolutional neural network for long-term traffic prediction
Chalupka et al. Visual causal feature learning
Liu et al. Incdet: In defense of elastic weight consolidation for incremental object detection
Qu et al. Learning to selectively transfer: Reinforced transfer learning for deep text matching
CN111324774A (zh) 一种视频去重方法和装置
WO2021208535A1 (zh) 基于自动特征分组的推荐方法及装置
Liu et al. Visual question answering with dense inter-and intra-modality interactions
CN113761359B (zh) 数据包推荐方法、装置、电子设备和存储介质
WO2023138381A1 (zh) 基于物品属性与时序模式耦合关系的序列推荐方法和系统
CN113609337A (zh) 图神经网络的预训练方法、训练方法、装置、设备及介质
CN116227624A (zh) 面向异构模型的联邦知识蒸馏方法和系统
CN115168721A (zh) 融合协同变换与时态感知的用户兴趣推荐方法及系统
CN113704565B (zh) 基于全局区间误差的学习型时空索引方法、装置及介质
Vaiyapuri Deep Learning Enabled Autoencoder Architecture for Collaborative Filtering Recommendation in IoT Environment.
CN111626827A (zh) 基于序列推荐模型的物品推荐方法、装置、设备及介质
Zhao et al. Traffic inflow and outflow forecasting by modeling intra-and inter-relationship between flows
Gao et al. Generalized pyramid co-attention with learnable aggregation net for video question answering
CN110232397A (zh) 一种结合支持向量机和投影矩阵的多标签分类方法
Steinki et al. Introduction to ensemble learning
CN117035089A (zh) 基于多关系循环事件图注意力的时态知识图谱推理方法
CN112434734A (zh) 基于动态分类器序列组合的选择性集成方法
CN116110232B (zh) 一种基于分层动态残差图卷积网络的交通流量预测方法
CN116306969A (zh) 基于自监督学习的联邦学习方法和系统
CN115907144A (zh) 一种事件的预测方法、装置、终端设备以及存储介质
He et al. Interest HD: An interest frame model for recommendation based on HD image generation

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 23742697

Country of ref document: EP

Kind code of ref document: A1