Intention knowledge graph construction method, intention prediction method, item recommendation method and device

Abstract

This specification provides embodiments of an intent knowledge graph construction method, an intent prediction method, an item recommendation method, and an apparatus. When constructing the intent knowledge graph, intent nodes are mined from user interaction data, and compositional relationships are established between intent nodes and functional nodes and product nodes. Semantic primitive nodes for functional and product nodes are determined, and relationships are established between functional and product nodes and their semantic primitive nodes. The hierarchical and sequential relationships between intent nodes are determined, and hierarchical and sequential relationships are established between intent nodes. Based on historical user interaction data and an intent representation set, user intent representations from historical user interaction data are determined. The user intent representations from historical user interaction data, interaction location / time information representations, current interaction location / time information representations, and the intent representation set are provided to an intent prediction model to predict the next user intent.

Description

Technical Field

[0001] The embodiments in this specification generally relate to the field of life services, and in particular to methods for constructing intent knowledge graphs, methods for predicting intent based on intent knowledge graphs, and methods and apparatus for recommending items. Background Technology

[0002] With the rapid development of information technology and the accelerated popularization of internet applications, digital life is gradually becoming a reality. More and more users can solve daily life problems through the life service applications provided by life service platforms (such as Meituan, Amazon App, etc.). For life service applications, accurately predicting user intent and providing services based on the predicted user intent not only helps users obtain a better experience, but also improves the click-through rate (CTR) and conversion rate (CVR) of life service applications. Summary of the Invention

[0003] This specification provides methods and apparatus for constructing intent knowledge graphs, methods and apparatus for intent prediction based on intent knowledge graphs, and methods and apparatus for item recommendation. Using this intent prediction method and apparatus based on intent knowledge graphs, a unified representation of multi-source heterogeneous content of user intent can be achieved, while improving the accuracy of user intent prediction when providing services to users on a lifestyle service platform.

[0004] According to one aspect of the embodiments of this specification, a method for constructing an intent knowledge graph is provided, comprising: mining intent nodes of the intent knowledge graph from user interaction data between users and items on a life service platform, wherein the intent nodes are constructed as consisting of function nodes and product nodes, and constructing compositional relationships between the intent nodes and their corresponding function nodes and product nodes; determining semantic primitive nodes corresponding to the function nodes and product nodes, and constructing relationships between the function nodes and product nodes and their corresponding semantic primitive nodes; determining hierarchical relationships and sequential relationships among the intent nodes; and constructing hierarchical relationships among intent nodes with hierarchical relationships and constructing sequential relationships among intent nodes with sequential relationships.

[0005] Optionally, in one example of the above aspects, the method may further include: constructing intent nodes based on functional verbs and product entity words in a knowledge base, or manually defining intent nodes.

[0006] Optionally, in one example of the above aspects, mining intent nodes from user interaction data between users and items on a lifestyle service platform may include: segmenting the user interaction data into words; tagging the segmented words with parts of speech; and applying phrase constraints to the segmented words using an intent part-of-speech composition format to mine intent nodes.

[0007] Optionally, in one example of the above aspects, using the intent part-of-speech composition format to perform phrase constraints on the segmented words in order to mine intent nodes may include: using the intent part-of-speech composition format to perform phrase constraints on the segmented words to obtain candidate intent nodes; performing phrase scoring on the candidate intent nodes; and determining the intent node from the candidate intent nodes based on the phrase scores of the candidate intent nodes.

[0008] Optionally, in one example of the above aspects, determining the semantic primitive nodes corresponding to the functional node and the product node may include: using a semantic acquisition model to determine the semantic primitive nodes corresponding to the functional node and the product node.

[0009] Optionally, in one example of the above aspects, determining the hierarchical relationship between the intent nodes may include: determining the hierarchical relationship between the intent nodes based on the semantic hierarchical relationship of the product nodes, or determining the hierarchical relationship between the intent nodes based on the textual semantic information of the product nodes.

[0010] Optionally, in one example of the above aspects, the intent node has timestamp information, and determining the sequence relationship between the intent nodes may include: providing the intent node to a Bayesian network for relational reasoning to identify multiple candidate intent node pairs with potential sequence relationships; and determining the intent node pairs with a determined sequence relationship from the candidate intent node pairs based on the node relevance and sequence relationship sensitivity of the intent node.

[0011] Optionally, in one example of the above aspects, the intent knowledge graph further includes category nodes. The method may further include: classifying each intent node using an intent classification model to determine category nodes of the intent knowledge graph; determining hierarchical relationships between the intent nodes and category nodes, and among the category nodes; and constructing hierarchical relationships between the intent nodes and category nodes with hierarchical relationships, and among the category nodes with hierarchical relationships.

[0012] According to another aspect of the embodiments of this specification, an intent prediction method based on an intent knowledge graph is provided, comprising: determining a user intent representation corresponding to the user's historical interaction data based on user historical interaction data and an intent representation set in a first user interaction data sequence between a user and items on a life service platform, wherein the first user interaction data sequence includes user historical interaction data immediately preceding the user's current interaction data, and the intent representation set is generated by learning an intent knowledge graph through a graph neural network, the knowledge graph being constructed according to the method described above; and providing the user intent representation, interaction location information representation, interaction time information representation, current interaction location information representation, current interaction time information representation, and the intent representation set of the user historical interaction data in the first user interaction data sequence to an intent prediction model to predict the next user intent.

[0013] Optionally, in one example of the above aspects, the intent prediction model employs a Transformer-based long-order prediction model. Providing the user intent representation, interaction location information representation, interaction time information representation, current interaction location information representation, current interaction time information representation, and the intent representation set from the first user interaction data sequence to an intent prediction model to predict the next user intent may include: generating a first user intent sequence and a second user intent sequence based on the first user interaction data sequence and the second user interaction data sequence, respectively. The second user interaction data sequence includes the user's current interaction data and at least one user's historical interaction data immediately preceding the current interaction data. The first user intent in the first user intent sequence includes the user intent representation, interaction location information representation, and interaction time information representation corresponding to the user interaction data. The second user intent in the second user intent sequence includes the user intent representation, interaction location information representation, and interaction time information representation corresponding to the user interaction data, wherein the user intent representation in the second user intent corresponding to the user's current interaction data is filled with a specific value; providing the first user intent sequence to the encoder of the intent prediction model to obtain a feature map of the first user intent sequence; providing the feature map of the first user intent sequence and the second user intent sequence to the decoder of the intent prediction model to predict the feature representation of the next user intent; and determining the next user intent based on the feature representation of the next user intent and the intent representation set.

[0014] Optionally, in one example of the above aspects, determining the user intent representation of the user historical interaction data based on the user's historical interaction data and intent representation set in the first user interaction data sequence between the user and items on the life service platform may include: extracting image features of the interactive items corresponding to the user historical interaction data; concatenating the image feature representation of the extracted image features with the data feature representation of the user historical interaction data and providing it to the item representation model to obtain the item representation of the interactive item; and determining the item intent representation of the interactive item as the user intent representation corresponding to the user historical interaction data based on the item representation and the intent representation set.

[0015] According to another aspect of the embodiments of this specification, an item recommendation method is provided, comprising: recalling candidate items to be recommended to a user from an item pool provided by a life service platform; sorting the candidate items based on a next user intent prediction result, wherein the next user intent prediction result is predicted according to the intent prediction method described above; and recommending items based on the sorting result of the candidate items.

[0016] Optionally, in one example of the above aspects, sorting the candidate items based on the next user intent prediction result may include: sorting the candidate items based on the next user intent prediction result, the intent representation set, and the item intent representation.

[0017] Optionally, in one example of the above aspects, recalling candidate items recommended to the user from the item pool provided by the life service platform may include: using an intent-based item recall method to recall candidate items recommended to the user from the item pool provided by the life service platform based on the next user intent prediction result.

[0018] According to another aspect of the embodiments of this specification, an apparatus for constructing an intent knowledge graph is provided, comprising: an intent node mining unit for mining intent nodes of the intent knowledge graph from user interaction data between users and items on a life service platform, wherein the intent nodes are constructed as consisting of functional nodes and product nodes; a semantic primitive node determination unit for determining the semantic primitive nodes corresponding to the functional nodes and the product nodes; a node relationship determination unit for determining hierarchical and sequential relationships between the intent nodes; and a node relationship construction unit for constructing compositional relationships between the intent nodes and their corresponding functional nodes and product nodes, constructing relationships between the functional nodes and the product nodes and their corresponding semantic primitive nodes, constructing hierarchical relationships between intent nodes with hierarchical relationships, and constructing sequential relationships between intent nodes with sequential relationships.

[0019] Optionally, in one example of the above aspects, the intent node mining unit may include: a word segmentation module for segmenting the user interaction data into words; a part-of-speech tagging module for tagging the segmented words with part-of-speech tags; and an intent node mining module for applying phrase constraints to the segmented words using an intent part-of-speech composition format to mine intent nodes.

[0020] Optionally, in one example of the above aspects, the user intent knowledge graph further includes category nodes. The apparatus may further include: a category node determination unit, which uses an intent classification model to classify each intent node to determine the category nodes of the user intent knowledge graph. The node relationship determination unit determines hierarchical relationships between intent nodes and category nodes, and among the category nodes themselves; and the node relationship construction unit constructs hierarchical relationships between intent nodes and category nodes with hierarchical relationships, and among the category nodes with hierarchical relationships.

[0021] According to another aspect of the embodiments of this specification, an intent prediction device based on an intent knowledge graph is provided, comprising: an intent representation set generation unit, which performs graph neural network learning on the intent knowledge graph to generate an intent representation set, wherein the intent knowledge graph is constructed according to the method described above; an intent representation determination unit, which determines the user intent representation corresponding to the user's historical interaction data based on user historical interaction data in a first user interaction data sequence between the user and items on a life service platform and the intent representation set, wherein the first user interaction data sequence includes user historical interaction data immediately preceding the user's current interaction data; and an intent prediction unit, which provides the user intent representation, interaction location information representation, interaction time information representation, current interaction location information representation, current interaction time information representation, and the intent representation set of the user historical interaction data in the first user interaction data sequence to an intent prediction model to predict the next user intent.

[0022] Optionally, in one example of the above aspects, the intent prediction model employs a Transformer-based long-order prediction model. The intent prediction unit may include: an intent sequence generation module, which generates a first user intent sequence and a second user intent sequence based on the first user interaction data sequence and the second user interaction data sequence, respectively. The second user interaction data sequence includes the user's current interaction data and at least one user historical interaction data immediately preceding the current user interaction data. The first user intent in the first user intent sequence includes a user intent representation, interaction location information representation, and interaction time information representation corresponding to the user interaction data. The second user intent in the second user intent sequence includes a user intent representation, interaction location information representation, and interaction time information representation corresponding to the user interaction data, wherein the user intent representation in the second user intent corresponding to the user's current interaction data is filled with specific values; an intent sequence feature map determination module, which provides the first user intent sequence to the encoder of the intent prediction model to obtain a feature map of the first user intent sequence; an intent feature representation prediction module, which provides the feature map of the first user intent sequence and the second user intent sequence to the decoder of the intent prediction model to predict the feature representation of the next user intent; and an intent determination module, which determines the next user intent based on the feature representation of the next user intent and the intent representation set.

[0023] Optionally, in one example of the above aspects, the intent representation determination unit may include: an item feature extraction module, which extracts image features of interactive items corresponding to user historical interaction data; an item representation generation module, which concatenates the image feature representation of the extracted image features with the data feature representation of the user interaction data and provides it to an item representation model to obtain the item representation of the interactive item; and an intent representation determination module, which determines the item intent representation of the interactive item based on the item representation and the intent representation set, as the user intent representation corresponding to the user historical interaction data.

[0024] According to another aspect of the embodiments of this specification, an intent prediction system based on an intent knowledge graph is provided, comprising: an apparatus for constructing an intent knowledge graph as described above; and an intent prediction apparatus as described above.

[0025] According to another aspect of the embodiments of this specification, an item recommendation device is provided, comprising: a candidate item recall unit for recalling candidate items to be recommended to a user from an item pool provided by a life service platform; a candidate item sorting unit for sorting the candidate items based on a next user intent prediction result, wherein the next user intent prediction result is predicted according to the intent prediction method described above; and an item recommendation unit for recommending items based on the sorting result of the candidate items.

[0026] According to another aspect of the embodiments of this specification, an item recommendation system is provided, comprising: an apparatus for constructing an intent knowledge graph as described above; an intent prediction apparatus as described above; and an item recommendation apparatus as described above.

[0027] According to another aspect of the embodiments of this specification, an apparatus for constructing an intent knowledge graph is provided, comprising: at least one processor, a memory coupled to the at least one processor, and a computer program stored in the memory, wherein the at least one processor executes the computer program to implement the method for constructing an intent knowledge graph as described above.

[0028] According to another aspect of the embodiments of this specification, an intent prediction apparatus based on an intent knowledge graph is provided, comprising: at least one processor, a memory coupled to the at least one processor, and a computer program stored in the memory, wherein the at least one processor executes the computer program to implement the intent prediction method based on the intent knowledge graph as described above.

[0029] According to another aspect of the embodiments of this specification, an item recommendation apparatus is provided, comprising: at least one processor, a memory coupled to the at least one processor, and a computer program stored in the memory, wherein the at least one processor executes the computer program to implement the item recommendation method as described above. Attached Figure Description

[0030] A further understanding of the nature and advantages of this specification can be achieved by referring to the following figures. In the figures, similar components or features may have the same reference numerals.

[0031] Figure 1 An example schematic diagram of an application scenario of an intent prediction system based on an intent knowledge graph according to an embodiment of this specification is shown.

[0032] Figure 2 An example architecture diagram of an intent prediction system based on an intent knowledge graph according to an embodiment of this specification is shown.

[0033] Figure 3 An example flowchart of a method for constructing an intent knowledge graph according to an embodiment of this specification is shown.

[0034] Figure 4 An example schematic diagram of the intent node mining process according to an embodiment of this specification is shown.

[0035] Figure 5 An example flowchart of a lexical rule-based hyponym determination process according to an embodiment of this specification is shown.

[0036] Figure 6 An example flowchart of a process for determining hierarchical relationships based on textual semantic information according to an embodiment of this specification is shown.

[0037] Figure 7 An example flowchart illustrating the process for determining the sequence relationship between intent nodes according to an embodiment of this specification is shown.

[0038] Figure 8 An example schematic diagram of an intent knowledge graph according to an embodiment of this specification is shown.

[0039] Figure 9 An example schematic diagram of an intent knowledge graph construction process according to an embodiment of this specification is shown.

[0040] Figure 10 An example flowchart of an intent prediction method based on an intent knowledge graph according to an embodiment of this specification is shown.

[0041] Figure 11 An example flowchart of a user intent representation determination process according to an embodiment of this specification is shown.

[0042] Figure 12 An example flowchart of a user intent prediction process according to an embodiment of this specification is shown.

[0043] Figure 13 An example schematic diagram of an intent prediction process based on an intent knowledge graph according to an embodiment of this specification is shown.

[0044] Figure 14 Example schematic diagrams of a first user intent sequence and a second user intent sequence according to embodiments of this specification are shown.

[0045] Figure 15 An example flowchart of an item recommendation method according to an embodiment of this specification is shown.

[0046] Figure 16 An example schematic diagram of an item recommendation process according to an embodiment of this specification is shown.

[0047] Figure 17 An example block diagram of an intent knowledge graph construction apparatus according to an embodiment of this specification is shown.

[0048] Figure 18 An example block diagram of an intent node mining unit according to an embodiment of this specification is shown.

[0049] Figure 19 An example block diagram of an intent prediction device according to an embodiment of this specification is shown.

[0050] Figure 20 An example block diagram of an intent characterization determination unit according to an embodiment of this specification is shown.

[0051] Figure 21 An example block diagram of an intent prediction unit according to an embodiment of this specification is shown.

[0052] Figure 22 An example block diagram of an item recommendation system according to an embodiment of this specification is shown.

[0053] Figure 23 An example block diagram of an article recommendation device according to an embodiment of this specification is shown.

[0054] Figure 24 An example schematic diagram of an intent knowledge graph construction apparatus implemented using a computer system according to an embodiment of this specification is shown.

[0055] Figure 25 An example schematic diagram of an intent prediction device based on a computer system according to an embodiment of this specification is shown.

[0056] Figure 26 An example schematic diagram of a computer-based item recommendation device according to an embodiment of this specification is shown. Detailed Implementation

[0057] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed merely to enable those skilled in the art to better understand and implement the subject matter described herein, and are not intended to limit the scope, applicability, or examples set forth in the claims. The function and arrangement of the elements discussed may be changed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the various examples. For example, the described methods may be performed in a different order than described, and steps may be added, omitted, or combined. Furthermore, features described in some examples may be combined in other examples.

[0058] As used herein, the term "comprising" and its variations are open terms meaning "including but not limited to". The term "based on" means "at least partially based on". The terms "one embodiment" and "an embodiment" mean "at least one embodiment". The term "another embodiment" means "at least one other embodiment". The terms "first", "second", etc., may refer to different or the same objects. Other definitions, whether explicit or implicit, may be included below. Unless explicitly indicated by the context, the definition of a term shall remain consistent throughout the specification.

[0059] The flowcharts used in this specification illustrate operations implemented according to some embodiments of this specification. It should be clearly understood that the operations in the flowcharts may not be implemented in a sequential order. Instead, the operations may be implemented in reverse order or simultaneously. Furthermore, one or more additional operations may be added to the flowcharts. One or more operations may be removed from the flowcharts.

[0060] Local service platforms such as Meituan and Amazon App offer a variety of items, including mini-program services, stores, coupons, and more. These items provide one or more functions, such as buying movie tickets or ordering takeout. Users express their needs by interacting with these items on the platform. The platform captures (or understands) the user's intent and provides matching services based on the functions offered by the items.

[0061] In the lifestyle services sector, platforms typically need to predict user intent. For example, on digital lifestyle service platforms like Meituan, after purchasing movie tickets through a ticketing app (corresponding intent: "buy movie tickets"), users often buy snacks at the cinema (corresponding intent: "buy snacks"). This means that the intent to "buy movie tickets" may lead to the subsequent intent to "buy snacks." Therefore, based on the user's historical interaction behavior when purchasing movie tickets through ticketing apps, the platform can predict that the user's next intent is "buy snacks" and provide corresponding services accordingly.

[0062] In user intent prediction, historical user intent is typically abstracted from historical user interactions, and then used to predict the user's next intent. User intent can usually be represented as clustering patterns of user behavior, which are often hidden in the user interaction content between the user and items provided by the lifestyle service application. This user interaction content can include, for example, queries, mini-program services, bills, tickets, shops, reviews, etc. In this specification, the terms "user interaction content" and "user interaction data" are used interchangeably.

[0063] However, in the lifestyle services sector, user interactions are highly fragmented and not directly observable, making it challenging to abstract user intent from these interactions. Furthermore, unlike e-commerce services, which primarily involve shopping intent, user interactions in lifestyle services are diverse, encompassing activities such as shopping, travel, and payments, further complicating the representation of user intent.

[0064] Furthermore, in the realm of lifestyle services, user intent is not only related to user profiles and preferences, but also to the context in which the user interacts. For example, the intent to "buy movie tickets" often occurs on weekends, while the intent to "register for an appointment" typically takes place in a hospital. The user's context can include, for example, the items the user interacts with, the location where the interaction occurs, and the time of the interaction.

[0065] This specification presents an intent prediction scheme based on an intent knowledge graph. This scheme proposes an intent knowledge graph based on an intent architecture. This intent knowledge graph extracts intent nodes from multi-source user interaction behaviors and constructs relationships between these nodes, thereby achieving a unified representation of user intent across multiple user interaction behaviors. Furthermore, when predicting user intent, it considers not only the user's historical intent sequence determined based on the intent knowledge graph and historical user interaction behaviors, but also the spatiotemporal information of the user's historical interaction behaviors and the spatiotemporal information of the user's current interaction, thereby improving the accuracy of user intent prediction.

[0066] Figure 1 An example schematic diagram of an application scenario 100 of an intent prediction system based on an intent knowledge graph according to an embodiment of this specification is shown.

[0067] like Figure 1 As shown, User A engaged in three user interaction behaviors: a ride-hailing request initiated at 9 PM on day T-1 from the office (intention: "hailing a ride"); a coffee purchase request initiated at 3 PM on day T from a coffee shop (intention: "ordering coffee"); and a movie ticket purchase request initiated at 5 PM on day T from a movie theater (intention: "buying movie tickets"). At this time, User A is in the movie theater, for example, at 7 PM on day T. Once User A initiates an interaction, an intent prediction model can be used to predict the user's next intent (e.g., "buying snacks") based on the user's historical interaction behaviors and current location / time information. This predicted next intent ("buying snacks") is then provided to downstream applications, such as recommendation systems, search systems, and transaction risk systems. Subsequently, downstream applications provide corresponding services based on the predicted next intent.

[0068] Figure 2 An example architecture diagram of an intent prediction system 200 based on an intent knowledge graph according to an embodiment of this specification is shown.

[0069] like Figure 2As shown, the intent prediction system 200 includes an intent knowledge graph construction device 210, an intent knowledge graph storage device 220, and an intent prediction device 230. The intent knowledge graph construction device 210, the intent knowledge graph storage device 220, and the intent prediction device 230 can communicate with each other via a network 240 to transmit data. In some embodiments, the network 240 can be any one or more of a wired network or a wireless network. Examples of the network 240 may include, but are not limited to, cable networks, fiber optic networks, telecommunications networks, enterprise intranets, the Internet, local area networks (LANs), wide area networks (WANs), wireless local area networks (WLANs), metropolitan area networks (MANs), public switched telephone networks (PSTNs), Bluetooth networks, ZigZee networks, near field communication (NFC), device internal buses, device internal lines, etc., or any combination thereof. In some embodiments, some or all of the components in the intent knowledge graph construction device 210, the intent knowledge graph storage device 220, and the intent prediction device 230 can communicate directly without the network 240.

[0070] The intent knowledge graph construction device 210 constructs an intent knowledge graph based on the user's historical interaction data between the user and items provided by the life service platform. The constructed intent knowledge graph is stored in the intent knowledge graph storage device 220 for use by the intent prediction device 230.

[0071] The intent prediction device 230 uses user historical interaction data, current interaction location information, current interaction time information, and intent knowledge graph to predict the user's next intent.

[0072] Optionally, in some embodiments, the intent knowledge graph construction apparatus 210 may have storage capabilities, thereby the constructed intent knowledge graph may be stored locally on the intent knowledge graph construction apparatus 210, thus eliminating the need for the intent knowledge graph storage apparatus 220.

[0073] Figure 3 An example flowchart of a method 300 for constructing an intent knowledge graph according to an embodiment of this specification is shown.

[0074] like Figure 3 As shown in Figure 310, intent nodes of the intent knowledge graph are mined from user interaction data between users and items on the lifestyle service platform. The mined intent nodes are constructed as functional nodes and product nodes.

[0075] Examples of lifestyle service platforms include Meituan, Didi, Amazon, and JD.com. These platforms offer a variety of items. Each item provides one or more functions to meet user needs. Examples of items include, but are not limited to, mini-programs, store apps, coupons, etc. In this specification, the term "item" may also be referred to as "supply." Users can interact with items provided by the lifestyle service platform to initiate user requests, thereby fulfilling various services required by the user. For example, a user clicks on a mini-program to complete the services provided by the mini-program, a user communicates with a store app via text to purchase goods in the store, or a user clicks on a coupon to use it. User interaction behaviors between users and items may include, for example, user clicks, user touches, text interactions between users and items, and audio / video interactions between users and items. In this specification, the term "user interaction data" refers to the content of the item that the user interacts with when a user interacts with the item. For example, user interaction data may be a purchase invoice generated when a user interacts with a store app to purchase goods. User interaction data may also be the content displayed on a coupon when a user clicks on it. User interaction data may have location information attributes and / or time information attributes. For example, if user A's user interaction behavior is "ordering a coffee worth 35 yuan at Starbucks in the office at 3 pm on February 21, 2023", then the user interaction data is "Starbucks coffee order amount: 35 yuan", the location information is "office", and the time information is "3 pm on February 21, 2023".

[0076] User interaction data used for intent node mining can include user interactions with different items on a local services platform within a given time period and / or location range. User-item interaction data directly reflects user needs. While the items offered on local services platforms are typically multi-source and heterogeneous, user interaction data across different items shares the same key information, thus serving as an input data source for intent knowledge graphs for knowledge mining.

[0077] In this specification, a function or function node is represented by a verb, and a product or product node is represented by a product entity. Examples of function nodes may include "call," "order," "buy," "find," "watch," etc. Examples of product nodes may include "ride-hailing," "coffee," "restaurant," "movie ticket," "movie," etc. The term "intent / user intent" is used to map different user needs and the functions provided by items to the same semantic space. Specifically, an "intent" is constructed as a function verb and a product entity, describing the detailed action steps behind a user need, such as "call a ride-hailing service," "order coffee," and "find a restaurant." In some embodiments, an "intent" may have an intent part-of-speech composition format, also known as an intent part-of-speech composition constraint format. For example, an "intent" may have an intent part-of-speech composition constraint format of "verb + object," such as a part-of-speech composition format of "verb + noun," or a part-of-speech composition format of "verb + adjective + noun," etc.

[0078] In some embodiments, user interaction data includes text data. In this case, phrase mining can be used to extract intent nodes from the user interaction data.

[0079] Figure 4 An example schematic diagram of an intent node mining process 400 according to an embodiment of this specification is shown.

[0080] In section 410, word segmentation is performed on the user interaction data. For example, a user generates a coffee bill at the office at 3 PM on February 21, 2023. Here, "bill" is equivalent to "user interaction data" in this specification, and the title of the bill is "Starbucks coffee order bill amount: 35 yuan". For the above user interaction behavior of user A, the corresponding user interaction data is "Starbucks coffee order bill amount: 35 yuan". After word segmentation of the user interaction data "Starbucks coffee order bill amount: 35 yuan", we get "Starbucks|order|coffee|bill|amount:|35|yuan".

[0081] In section 420, the segmented words are tagged with their parts of speech. For example, “Starbucks (noun)|point (verb)|coffee (noun)|bill (noun)|amount (noun):|35 (quantifier)|yuan (noun)”.

[0082] After completing the part-of-speech tagging as described above, the segmented words can be constrained using the intent part-of-speech composition format to extract intent nodes. For example, assuming the intent part-of-speech composition constraint format is "verb + noun", the intent node "order coffee" can be extracted.

[0083] For example, in some embodiments, after part-of-speech tagging as described above, at 430, phrase constraints can be applied to the segmented words using an intent part-of-speech composition format to mine candidate intent nodes. The mined candidate intent nodes may include one or more candidate intent nodes.

[0084] At 440, phrase scoring is performed on candidate intent nodes. For example, the tf-idf algorithm can be used to perform phrase scoring on candidate intent nodes.

[0085] At 450, an intent node is determined from the candidate intent nodes based on their phrase scores. For example, the phrase score of a candidate intent node can be compared with a predetermined threshold. If it is not less than the predetermined threshold, the candidate intent node is determined to be an intent node. If it is less than the predetermined threshold, the candidate intent node is determined not to be an intent node.

[0086] In some embodiments, intent nodes can be constructed based on functional verbs and product entities from a knowledge base. For example, functional verbs can be combined with product entities from a knowledge base using heuristic rules to automatically generate intent nodes. For instance, "order coffee" contains "order [functional verb] + coffee [product entity]". If "milk tea" from the knowledge base is used instead of "coffee", a new intent node named "order milk tea" can be generated.

[0087] In some embodiments, intent nodes can be manually defined. For example, experts can manually define intent nodes to ensure that the obtained intent has high quality and timeliness, such as "see a flower show".

[0088] Optionally, in some embodiments, after the intent node is determined as described above, a pre-trained intent recognition model can be used to identify the determined intent node, thereby determining whether the intent node should be recognized as an intent node. The intent recognition model can be implemented using a binary classification model, such as a binary classification model employing a Deep & Cross network model. Using this model, cross-coding features for intent recognition can be obtained more automatically, thereby improving the intent node recognition performance. Following the above method, the intent recognition model is used to re-identify the determined intent node, thereby improving the accuracy of the constructed intent node.

[0089] After identifying the intent nodes as described above, in step 320, we establish a composition relationship (Consist) between each intent node and its corresponding functional and product nodes.

[0090] In step 330, the semantic primitives corresponding to the functional nodes and product nodes in the intent node are determined, and a relation (Has) is constructed between the functional nodes and product nodes and their corresponding semantic primitive nodes.

[0091] Semantic primitives are basic semantic units. Product nodes and function nodes can be represented using finer-grained semantic primitives (e.g., CNKI semantic primitives). For example, the product "movie ticket" can be further represented as the semantic primitive "{ticket, watch, performance}".

[0092] In some embodiments, a semantic acquisition model can be used to determine the semantic primitives corresponding to functional nodes and product nodes. For example, a multi-label classification model can be trained using a manually annotated corpus (e.g., the CNKI corpus) to automatically acquire the semantic information of functions and products, thereby obtaining the corresponding semantic primitive representations. Furthermore, due to the existence of aliases and ambiguities in entity names, semantic disambiguation can be further performed using models such as the Bert Int alignment model for intent nodes, functional nodes, and product nodes.

[0093] In 340, the hierarchical relationship (isA) and the sequential relationship (Consequence) between intent nodes are determined.

[0094] In some embodiments, the hierarchical relationship between intent nodes can be determined based on the semantic hierarchical relationship of product nodes. The term "hierarchical relationship" belongs to the linguistic category. A hierarchical relationship refers to the semantic relationship between a hyponym of a subtype and a hypernym of a parent type. In a semantic field, a hyponym is contained within a hypernym; for example, pigeon, crow, eagle, and seagull are all hyponyms of birds, and birds are their hypernyms; while birds themselves are hyponyms of animals, and animals are hypernyms of birds. Nodes "pigeon," "crow," "eagle," and "seagull" have a hierarchical relationship with the node "birds," and the node "birds" has a hierarchical relationship with the node "animals." However, the node "birds" does not have a hierarchical relationship with the nodes "pigeon," "crow," "eagle," and "seagull," and the node "animals" does not have a hierarchical relationship with the node "birds."

[0095] In this specification, the term "sequential relationship" refers to a series of actions or related situations arranged in a temporal, spatial, or logical order, exhibiting a logical sequential relationship. Sequential relationships focus on the logical order of events and do not necessarily involve causality. For example, the node "buying a house" can be a sequence of the node "renovating a house," and the node "renovating a house" can be a sequence of the node "buying home appliances," etc.

[0096] Figure 5 An example flowchart of a lexical rule-based hyponym determination process 500 according to an embodiment of this specification is shown.

[0097] like Figure 5 As shown, for two intent nodes with the same function, in 510, the semantics of the product nodes in the two intent nodes are compared, and in 520, it is determined whether the semantics of the product nodes in the two intent nodes have a hierarchical relationship.

[0098] If a hierarchical relationship is determined in step 520, then in step 530, the hierarchical relationship between the two intent nodes is confirmed. If a hierarchical relationship is determined in step 520, then in step 540, the hierarchical relationship between the two intent nodes is confirmed.

[0099] For example, the intent nodes "buy iPhone13" and "buy mobile phone" have the same function node "buy", and from the regular knowledge graph, we can know that the product node "iPhone13" and the product node "mobile phone" have a hierarchical relationship. Therefore, the intent nodes "buy iPhone13" and "buy mobile phone" have a hierarchical relationship.

[0100] In some embodiments, the hierarchical relationship between intent nodes can be determined based on the textual semantic information of the product node.

[0101] Figure 6 An example flowchart of a hierarchical relationship determination process 600 based on textual semantic information according to an embodiment of this specification is shown.

[0102] like Figure 6 As shown, for two intent nodes with the same function, in step 610, a semantic representation model is used to generate a semantic representation of the product node in the two intent nodes. The semantic representation model can, for example, be a StructBERT model pre-trained based on a corpus.

[0103] At 620, the distance between the semantic representations of the product nodes in the two intent nodes is calculated. For example, the semantic representation can be a semantic embedding in vector form, thus allowing the calculation of the vector distance between the semantic representations of the product nodes in the two intent nodes. Examples of distances could include Euclidean distance, Manhattan distance, cosine distance, etc.

[0104] At step 630, it is determined whether the distance between the calculated semantic representations of the product nodes meets a predetermined condition. The predetermined condition could be, for example, not exceeding a predetermined threshold.

[0105] If the predetermined conditions are met, then at 640, it is determined that there is a hierarchical relationship between the two intent nodes. If the predetermined conditions are not met, then at 650, it is determined that there is no hierarchical relationship between the two intent nodes.

[0106] In some embodiments, for each intent node, multiple candidate intent nodes with the same functionality can be summoned. Then, for that intent node and each candidate intent node, according to... Figure 6 The illustrated process for determining hierarchical relationships is used to calculate the distance between each candidate intent node and the intent node in terms of hierarchical relationship. This distance reflects the synonymy between the candidate intent node and the intent node. Then, the K candidate intent nodes with the highest synonymy among the top K are determined as synonym intent nodes of the intent node.

[0107] Figure 7 An example flowchart of a process 700 for determining the sequence relationship between intent nodes according to an embodiment of this specification is shown. Figure 7 In the example, the intent node has a timestamp attribute. The timestamp attribute indicates the time when the user interaction corresponding to the schematic node occurred. Optionally, the intent node may also have a location attribute. The location attribute indicates the location where the user interaction corresponding to the schematic node occurred.

[0108] like Figure 7 As shown in 710, the intent nodes are provided to the Bayesian network for relational reasoning to identify multiple candidate intent node pairs with potential sequential relationships.

[0109] In step 720, based on the node relevance and sequence sensitivity of the intent nodes, intent node pairs with a definite sequence relationship are identified from the candidate intent node pairs. When using Bayesian networks for relational reasoning, corresponding scores are output for the node relevance and sequence sensitivity of the intent nodes. Therefore, intent node pairs with a definite sequence relationship can be identified from the candidate intent node pairs based on the node relevance score and the sequence sensitivity score.

[0110] After determining the hierarchical and sequential relationships as described above, at 350, hierarchical relationships are constructed between intent nodes with hierarchical relationships, and sequential relationships are constructed between intent nodes with sequential relationships, thereby constructing an intent knowledge graph. Figure 8 An example schematic diagram of an intent knowledge graph according to an embodiment of this specification is shown.

[0111] Optionally, the user intent knowledge graph may also include category nodes. In this case, an intent classification model can be used to classify the intent nodes to determine the category nodes of the intent knowledge graph. Furthermore, hierarchical relationships are determined between intent nodes and category nodes in the intent knowledge graph, as well as between individual category nodes. After determining the category node pairs and / or intent node-category node pairs with hierarchical relationships, hierarchical relationships are constructed between intent nodes and category nodes with hierarchical relationships, and / or between category nodes with hierarchical relationships.

[0112] Figure 9 An example schematic diagram of an intent knowledge graph construction process according to an embodiment of this specification is shown.

[0113] like Figure 9 As shown, firstly, intent nodes are mined from user interaction data using lexical rule matching, part-of-speech tagging, and short text matching models, and these intent nodes are constructed as function nodes and product nodes. Then, a multi-label classification model is used to automatically obtain the semantic information of functions and products, thereby obtaining the corresponding semantic primitive nodes. Next, a hierarchical relationship determination method based on lexical rules or text semantic information is used to determine the hierarchical relationship between intent nodes. A Bayesian network is used to deduce the relationships between intent nodes, thereby determining the sequential relationships between them. Finally, based on the determined node relationships, hierarchical and sequential relationships are constructed between nodes in the intent knowledge graph, thus building the intent knowledge graph.

[0114] Following the above process of constructing an intent knowledge graph, by extracting intent nodes, product nodes, function nodes, and semantic primitive nodes from multi-source user interaction data (multi-source item content), and constructing the relationships between the nodes of the intent knowledge graph, a unified representation of multi-source user interaction data can be achieved for user intent, thereby establishing a correlation between user needs and items.

[0115] After constructing the intent knowledge graph as described above, the intent knowledge graph can be used for intent prediction.

[0116] Figure 10 An example flowchart of an intent prediction method 1000 based on an intent knowledge graph according to an embodiment of this specification is shown.

[0117] like Figure 10 As shown in Figure 1010, an intent representation set is generated by learning from the intent knowledge graph using a graph neural network. The intent representation set includes intent representations of each intent in the intent knowledge graph. Intent representations can be, for example, intent embeddings in vector form with a specified dimension.

[0118] In step 1020, based on the user's historical interaction data and intent representation set in the first user interaction data sequence between the user and items on the life service platform, the user intent representation of the user's historical interaction data in the first user interaction data sequence is determined. The first user interaction data sequence may include user historical interaction data immediately preceding the user's current interaction data. The user historical interaction data has interaction location attribute information and interaction time attribute information. The interaction location attribute information refers to the location where the user's interaction behavior occurs. The interaction time attribute information refers to the time when the user's interaction behavior occurs. It should be noted that, in some embodiments, the next user intent prediction process may be triggered based on the user's current actual interaction behavior. For example, in a scenario recommended on the homepage, the user clicks on an application, thereby triggering user intent prediction. In this case, the user's current interaction data may include the actual user interaction data corresponding to the user's current actual interaction behavior. The current interaction time is the time when the user's current interaction behavior occurs. The current interaction location is the location where the user's current interaction behavior occurs. In some embodiments, the next user intent prediction process may be initiated actively by the intent prediction device. In this case, the data content of the user's current interaction data is empty. The current interaction time is the time when the intent prediction device initiates intent prediction. The current interaction location can be the user's previous interaction location or the user's current actual location. For example, it can be obtained by collecting the user's current location information.

[0119] Figure 11 An example flowchart of a user intent representation determination process 1100 according to an embodiment of this specification is shown.

[0120] like Figure 11 As shown, image features of interactive items corresponding to historical user interaction data are extracted. For example, assuming the user interacted with an application-type item provided on a lifestyle service platform, image features of that item can be extracted, such as screenshots of various interfaces. The extracted image features can uniquely identify the item.

[0121] After extracting the image features of the interactive item, an image feature representation (e.g., an embedding representation) of the extracted image features is generated. For example, the image features of the interactive item are provided to a ResNet network to generate the image feature representation. Furthermore, a data feature representation (e.g., an embedding representation) of the user's historical interaction data (the item content of the interactive item) is generated, and the image feature representation of the interactive item and the data feature representation of the user's historical interaction data are concatenated and provided to the item representation model, thereby obtaining the item representation of the interactive item. Examples of user historical interaction data may include, but are not limited to, text data, audio data, video data, image data, etc. When the user historical interaction data is text data, the data feature representation of the user historical interaction data is a text feature representation. In some embodiments, when the user historical interaction data is non-text data, the user historical interaction data can be first converted into text data, and then the text feature representation can be determined based on the converted text data. Alternatively, the data feature representation of the user interaction data can be obtained directly based on non-text data.

[0122] Then, based on the item representation and intent representation set, the item intent representation of the interactive item is determined as the user intent representation corresponding to the user's historical interaction data. For example, the item representation can be multiplied by each intent representation in the intent representation set to obtain a prediction score of the item representation relative to each intent representation; for example, a corresponding prediction score can be obtained for each intent. Then, the intent representation corresponding to the item representation is determined based on the prediction scores of each intent, thus obtaining the item intent representation of the interactive item. The obtained item intent representation can reflect the matching relationship between items and intents, thereby establishing the item-intent relationship. In some embodiments, the item intent representation with the highest prediction score can be determined as the item intent representation of the interactive item. In some embodiments, the item intent representation with a prediction score exceeding a predetermined threshold or the item intent representation with a prediction score in the Top K can be determined as the item intent representation of the interactive item. In this way, the item intent representation can be represented as a multi-dimensional intent representation vector (U1, U2, ..., Un), where each intent representation has a prediction probability or prediction score.

[0123] After determining the user intent representation of the user's historical interaction data as described above, at 1030, the user intent representation, interaction location information representation, interaction time information representation, current interaction location information representation, current interaction time information representation, and intent representation set of the user's historical interaction data in the first user interaction data sequence are provided to the intent prediction model to predict the next user intent.

[0124] Figure 12 An example flowchart of a user intent prediction process 1200 according to an embodiment of this specification is shown, and Figure 13A schematic diagram illustrating an example of an intent prediction process based on an intent knowledge graph according to an embodiment of this specification is shown. In this embodiment, a Transformer-based long-order prediction model, such as an Informer model, is employed as the intent prediction model. This Transformer-based long-order prediction model may have an encoder and a decoder.

[0125] like Figure 12 As shown, in step 1210, a first user intent sequence and a second user intent sequence are generated based on a first user interaction data sequence and a second user interaction data sequence, respectively. The first user interaction data sequence includes at least one user historical interaction data immediately preceding the user's current interaction data. The second user interaction data sequence includes the user's current interaction data and at least one user historical interaction data immediately preceding the user's current interaction data. The user historical interaction data has interaction location attribute information and interaction time attribute information. Optionally, the user historical interaction data in the second user interaction data sequence is usually a subset of the user historical interaction data in the first user interaction data sequence. In the generated first user intent sequence, each first user intent includes a user intent representation corresponding to the user interaction data, an interaction location information representation generated based on the interaction location attribute information, and an interaction time information representation generated based on the interaction time attribute information. In the generated second user intent sequence, each second user intent includes a user intent representation corresponding to the user interaction data, an interaction location information representation, and an interaction time information representation. In the second user intent sequence, the user intent representation corresponding to the user's historical interaction data is the corresponding user intent representation in the intent representation set. The user intent representation corresponding to the user's current interaction data is filled with a specific value (e.g., 0 value), that is, it is assigned a specific value (0 value).

[0126] Figure 14 Example schematic diagrams of a first user intent sequence and a second user intent sequence according to embodiments of this specification are shown.

[0127] like Figure 14 As shown, the first user intent sequence includes eight user intents corresponding to, for example, eight historical user interaction data points of user A (i.e., the first user interaction data sequence). These eight historical user interaction data points occur consecutively, and each historical user interaction data point has corresponding interaction location information and interaction time information. Each column in the first user intent sequence represents a user intent corresponding to one historical user interaction data point. The user intent includes a user intent representation U. i Interactive location information representation P i and interaction time information representation T i User intent representation U iGenerated based on user historical interaction data, this represents the item intent of the interactive items within the user's historical interaction data. Interaction location information representation P i and interaction time information representation T i They are generated based on the corresponding interaction location information and interaction time information, respectively.

[0128] The second user intent sequence includes user intents corresponding to seven user interaction data points of user A. These seven user interaction data points occur consecutively, wherein one user interaction data point is the user's current interaction data, and the other six are the user's historical interaction data immediately preceding the current interaction data. Each user interaction data point has corresponding interaction location information and interaction time information. In some embodiments, the user's current interaction data may contain actual interaction data. In some embodiments, the user's current interaction data may be empty data. In some embodiments, the current interaction time is the time when the user's current interaction behavior occurs. The current interaction location is the location where the user's current interaction behavior occurs. In some embodiments, the current interaction time may be the time when the intent prediction device initiates intent prediction. The current interaction location may be the user's previous interaction location or the user's current actual location. Each column in the second user intent sequence is a user intent corresponding to one user interaction data point. The user intent includes a user intent representation U. i Interactive location information representation P i and interaction time information representation T i For a user's historical interaction behavior, the user intent representation is the corresponding user intent representation in the intent representation set. For the user's current interaction behavior, the user intent representation is assigned a specific value (e.g., 0).

[0129] It should be noted that, Figure 14 The illustrated second user intent sequence includes only one user intent to be predicted. In other embodiments, the second user intent sequence may include multiple user intents to be predicted, for example, user intents to be predicted corresponding to several future times.

[0130] At 1220, the first user intent sequence is fed to the encoder of the intent prediction model to obtain a feature map of the first user intent sequence. The obtained feature map is a feature vector with a specified dimension, for example, a 200-dimensional vector.

[0131] At 1230, the feature map of the first user intent sequence and the second user intent sequence are provided to the decoder of the intent prediction model to predict the feature representation of the next user intent. For example... Figure 13 As shown, for each second user intent in the second user intent sequence, a feature representation can be generated. The generated feature representation has the same vector dimension as the feature map.

[0132] In step 1240, the next user intent is determined based on its feature representation and intent representation set. For example, the feature representation of the next user intent can be multiplied by each intent representation in the intent representation set to obtain a prediction score for the feature representation of the next user intent relative to each intent representation; for example, a corresponding prediction score can be obtained for each intent. Then, the intent representation corresponding to the feature representation of the next user intent is determined based on the prediction scores of each intent, thereby predicting the next user intent. In some embodiments, the intent representation with the highest prediction score can be determined as the user intent representation corresponding to the user's historical interaction data. In some embodiments, intent representations with prediction scores exceeding a predetermined threshold or those with prediction scores in the Top K can be determined as the user intent representation corresponding to the user's historical interaction data.

[0133] The intent prediction method according to embodiments of this specification has been described above. This intent prediction method improves the accuracy of user intent prediction by combining the user's historical interaction sequence, the spatiotemporal information of the user's historical interaction, the user's current spatiotemporal information, and an intent knowledge graph when predicting user intent.

[0134] The aforementioned intent prediction method can be applied to scenarios such as item recommendation, item search, and transaction risk management.

[0135] Figure 15 An example flowchart of an article recommendation method 1500 according to an embodiment of this specification is shown.

[0136] like Figure 15 As shown, at step 1510, candidate items recommended to the user are recalled from the item pool provided by the life service platform. In some embodiments, an intent-based item recall method can be used to recall candidate items recommended to the user from the item pool provided by the life service platform based on the next user intent prediction result predicted according to the intent prediction method described above. In some embodiments, a location-based recall method or a representation-based recall method can be used to recall candidate items recommended to the user from the item pool provided by the life service platform.

[0137] In step 1520, candidate items are ranked based on the next user intent prediction result. In some embodiments, the candidate items can be ranked based on the next user intent prediction result, as well as the intent representation set and item intent representation obtained in the above intent prediction process. For example, suppose the next user intent prediction result is "order coffee", and there are multiple candidate items "Starbucks Mini Program" and "Luckin Coffee Mini Program". Item-intent relationships can be constructed based on the item intent representation. The existence of "Starbucks Mini Program - Order Coffee" and "Luckin Coffee Mini Program - Order Coffee" is found in the item-intent relationships, and each item-intent relationship has a relationship metric value, so the items "Starbucks Mini Program" and "Luckin Coffee Mini Program" can be ranked according to the item-intent relationships.

[0138] In step 1530, item recommendations are made based on the ranking of candidate items. For example, the top K candidate items can be recommended to the user, such as displaying these top K candidate items on the user's client.

[0139] Figure 16 An example schematic diagram of an item recommendation process according to an embodiment of this specification is shown.

[0140] like Figure 16 As shown, an intent knowledge graph is constructed based on user historical interaction data, and a graph neural network is used to learn the intent representation of each intent on the intent knowledge graph. The item text data, image feature data, and intent knowledge graph corresponding to the user's historical interaction data are provided to the intent understanding model to obtain the user intent representation corresponding to the user's historical interaction data, thereby constructing a user historical intent sequence. Then, the constructed user historical intent sequence, the user's current spatiotemporal information, and the intent knowledge graph are provided to the intent prediction device to predict the next user intent.

[0141] The obtained next user intent prediction results can be provided to the recall engine to recall candidate items. Then, the obtained next user intent prediction results, intent representations of each intent, and item-intent relationships are provided to the ranking engine to rank the recalled candidate items. Finally, item recommendations are made based on the ranking results of the candidate items.

[0142] Figure 17 An example block diagram of an intent knowledge graph construction apparatus 1700 according to an embodiment of this specification is shown. Figure 17 As shown, the intent knowledge graph construction device 1700 includes an intent node mining unit 1710, an intent primitive node determination unit 1720, a node relationship determination unit 1730, and a node relationship construction unit 1740.

[0143] The intent node mining unit 1710 is configured to mine intent nodes from user interaction data between users and items on the life service platform, forming an intent knowledge graph. The mined intent nodes are constructed from functional nodes and product nodes. The operation of the intent node mining unit 1710 can be referenced above. Figure 3 The operation described in 310.

[0144] The semantic origin node determination unit 1720 is configured to determine the semantic origin nodes corresponding to functional nodes and product nodes. The operation of the semantic origin node determination unit 1720 can be referenced above. Figure 3 The operation described in 330.

[0145] The node relationship determination unit 1730 is configured to determine hierarchical and sequential relationships between intent nodes. The operation of the node relationship determination unit 1730 can be referenced above. Figure 3 The operation described in 340.

[0146] The node relationship construction unit 1740 is configured to construct compositional relationships between intent nodes and their corresponding functional and product nodes, construct hierarchical relationships between functional and product nodes and their corresponding semantic nodes, construct hierarchical relationships between intent nodes with hierarchical relationships, and construct sequential relationships between intent nodes with sequential relationships. The operation of the node relationship construction unit 1740 can be referenced above. Figure 3 The operations described in 320 / 330 and 350.

[0147] Optionally, the intent knowledge graph may also include category nodes. Accordingly, the intent knowledge graph construction apparatus 1700 includes a category node determination unit 1750. The category node determination unit 1750 is configured to classify each intent node using an intent classification model to determine the category nodes of the intent knowledge graph.

[0148] After identifying the category nodes, the node relationship determination unit 1730 determines the hierarchical relationships between the intent nodes and category nodes in the intent knowledge graph, as well as among the category nodes themselves. Subsequently, the node relationship construction unit 1740 constructs hierarchical relationships between the intent nodes and category nodes with hierarchical relationships, and also constructs hierarchical relationships among category nodes with hierarchical relationships.

[0149] Figure 18 An example block diagram of an intent node mining unit 1800 according to an embodiment of this specification is shown. Figure 18 As shown, the intent node mining unit 1800 includes a word segmentation module 1810, a part-of-speech tagging module 1820, and an intent node mining module 1830.

[0150] The word segmentation module 1810 is configured to segment words in user interaction data. The operation of the word segmentation module 1810 can be referenced above. Figure 4 The operation described in 410.

[0151] The part-of-speech tagging module 1820 is configured to perform part-of-speech tagging on the segmented words. The operation of the part-of-speech tagging module 1820 can be referenced above. Figure 4 The operation described in 420.

[0152] The intent node mining module 1830 is configured to use intent part-of-speech composition format to perform phrase constraints on the segmented words in order to mine intent nodes.

[0153] In some embodiments, the intent node mining module 1830 may include a phrase constraint module, a phrase scoring module, and an intent node determination module. After part-of-speech tagging is completed as described above, the phrase constraint module can use the intent part-of-speech composition format to perform phrase constraints on the segmented words to mine candidate intent nodes. The mined candidate intent nodes may include one or more candidate intent nodes. Subsequently, the phrase scoring module performs phrase scoring on the candidate intent nodes. For example, the TF-IDF algorithm can be used to perform phrase scoring on the candidate intent nodes. Then, the intent node determination module determines the intent node from the candidate intent nodes based on the phrase scores of the candidate intent nodes. For example, the phrase score of the candidate intent node can be compared with a predetermined threshold. If it is not less than the predetermined threshold, the candidate intent node is determined to be an intent node. If it is less than the predetermined threshold, the candidate intent node is determined not to be an intent node.

[0154] In some embodiments, the intent node mining unit can construct intent nodes based on functional verbs and product entity words in a knowledge base. For example, functional verbs can be combined with product entity words in a knowledge base using heuristic rules to automatically generate intent nodes. In some embodiments, intent nodes can be manually defined by human experts.

[0155] Optionally, in some embodiments, the intent node mining unit may further include an intent node identification module. The intent node identification module can use a pre-trained intent recognition model to identify the determined intent nodes, thereby determining whether the intent node should be recognized as an intent node. The intent recognition model can, for example, be implemented using a binary classification model, such as a binary classification model employing a Deep&Cross network model.

[0156] Figure 19 An example block diagram of an intent prediction device 1900 according to an embodiment of this specification is shown. Figure 19As shown, the intent prediction device 1900 includes an intent representation set generation unit 1910, an intent representation determination unit 1920, and an intent prediction unit 1930.

[0157] The intent representation set generation unit 1910 is configured to perform graph neural network learning on the intent knowledge graph to generate an intent representation set. The operation of the intent representation set generation unit 1910 can be referenced above. Figure 10 The operation described in 1010.

[0158] The intent representation determination unit 1920 is configured to determine the user intent representation corresponding to the user's historical interaction data based on the user's historical interaction data and intent representation set in the first user interaction data sequence between the user and items on the life service platform. The first user interaction data sequence includes the user's historical interaction data immediately preceding the user's current interaction data. The user's historical interaction data has interaction location information and interaction time attribute information. The operation of the intent representation determination unit 1920 can be referenced above. Figure 10 The operation described in 1020.

[0159] The intent prediction unit 1930 is configured to provide the intent prediction model with user intent representations, interaction location information representations, interaction time information representations, current interaction location information representations, current interaction time information representations, and intent representation sets from the user's historical interaction data in the first user interaction data sequence to predict the next user intent. The operation of the intent prediction unit 1930 can be referenced above. Figure 10 The operation described in 1030.

[0160] Figure 20 An example block diagram of an intent characterization determination unit 2000 according to an embodiment of this specification is shown. Figure 20 As shown, the intent representation determination unit 2000 includes an item feature extraction module 2010, an item representation generation module 2020, and an intent representation determination module 2030.

[0161] The item feature extraction module 2010 is configured to extract image features of interactive items corresponding to the user's historical interaction data. After extracting the image features, the item representation generation module 2020 concatenates the extracted image feature representations with the text feature representations of the interactive items and provides them to the item representation model to obtain the item representation of the interactive item. The intent representation determination module 2030 determines the item intent representation of the interactive item based on the item representation and the intent representation set, which serves as the user intent representation corresponding to the user's historical interaction data.

[0162] Figure 21 An example block diagram of an intent prediction unit 2100 according to an embodiment of this specification is shown. Figure 21In the illustrated embodiment, the intent prediction model may employ a Transformer-based long-order prediction model. This Transformer-based long-order prediction model includes an encoder and a decoder. For example... Figure 21 As shown, the intent prediction unit 2100 includes an intent sequence generation module 2110, an intent sequence feature map determination module 2120, an intent feature representation prediction module 2130, and an intent determination module 2140.

[0163] The intent sequence generation module 2110 is configured to generate a first user intent sequence and a second user intent sequence based on a first user interaction data sequence and a second user interaction data sequence, respectively. The first user interaction data sequence includes at least one historical user interaction data immediately preceding the user's current interaction data. The user interaction data has interaction location attribute information and interaction time attribute information. The second user interaction data sequence includes the user's current interaction data and at least one historical user interaction data immediately preceding the user's current interaction data. The first user intent in the first user intent sequence includes a user intent representation, interaction location information representation, and interaction time information representation corresponding to the corresponding user interaction data. The second user intent in the second user intent sequence includes a user intent representation, interaction location information representation, and interaction time information representation corresponding to the corresponding user interaction data. In the second user intent sequence, the user intent representation corresponding to the user's historical interaction data is the corresponding user intent representation in the intent representation set. The user intent representation corresponding to the user's current interaction data is filled with a specific value. The operation of the intent sequence generation module 2110 can be referenced above. Figure 12 The operation described in 1210.

[0164] The intent sequence feature map determination module 2120 is configured to provide a first user intent sequence to the encoder of the intent prediction model to obtain a feature map of the first user intent sequence. The operation of the intent sequence feature map determination module 2120 can be referenced above. Figure 12 The operation described in 1220.

[0165] The intent feature representation prediction module 2130 is configured to provide the feature map of the first user intent sequence and the second user intent sequence to the decoder of the intent prediction model to predict the feature representation of the next user intent. The operation of the intent feature representation prediction module 2130 can be referenced above. Figure 12 The operation described in 1230.

[0166] The intent determination module 2140 is configured to determine the next user intent based on the feature representation and intent representation set of the next user intent. The operation of the intent determination module 2140 can be referenced above. Figure 12 The operation described in 1240.

[0167] Figure 22 An example block diagram of an item recommendation system 2200 according to an embodiment of this specification is shown.

[0168] like Figure 22 As shown, the item recommendation system 2200 includes an intent knowledge graph construction device 2210, an intent knowledge graph storage device 2220, an intent prediction device 2230, and an item recommendation device 2240. The intent knowledge graph construction device 2210, intent knowledge graph storage device 2220, intent prediction device 2230, and item recommendation device 2240 can communicate with each other via a network 2250 to transmit data. Examples of the network 2250 may include, but are not limited to, cable networks, fiber optic networks, telecommunications networks, enterprise intranets, the Internet, local area networks (LANs), wide area networks (WANs), wireless local area networks (WLANs), metropolitan area networks (MANs), public switched telephone networks (PSTNs), Bluetooth networks, ZigZee networks, near field communication (NFC), device internal buses, device internal lines, etc., or any combination thereof. In some embodiments, some or all of the components in the intent knowledge graph construction device 2210, intent knowledge graph storage device 2220, intent prediction device 2230, and item recommendation device 2240 can communicate directly without the network 2250.

[0169] The intent knowledge graph construction device 2210 constructs an intent knowledge graph based on the user's historical interaction data between the user and items provided by the life service platform. The constructed intent knowledge graph is stored in the intent knowledge graph storage device 2220 for use by the intent prediction device 2230.

[0170] The intent prediction device 2230 uses user historical interaction data, current interaction location information, current interaction time information, and intent knowledge graph to predict the user's next intent. The item recommendation device 2240 recommends items to the user based on the predicted next intent.

[0171] Figure 23 An example block diagram of an article recommendation device 2300 according to an embodiment of this specification is shown. Figure 23 As shown, the item recommendation device 2300 includes a candidate item recall unit 2310, a candidate item sorting unit 2320, and an item recommendation unit 2340.

[0172] The candidate item recall unit 2310 is configured to recall candidate items recommended to users from the item pool provided by the lifestyle service platform. The operation of the candidate item recall unit 2310 can be referenced above. Figure 15 The operation described in 1510.

[0173] The candidate item sorting unit 2320 is configured to sort candidate items based on the next user intent prediction result. In some embodiments, the candidate item sorting unit 2320 may sort candidate items based on the next user intent prediction result, the intent representation set obtained in the above intent prediction process, and the item intent representation. The operation of the candidate item sorting unit 2320 can be referred to the above reference. Figure 15 The operation described in 1520.

[0174] The item recommendation unit 2330 is configured to recommend items based on the ranking of candidate items. The operation of the item recommendation unit 2330 can be referenced above. Figure 15 The operation described in 1530.

[0175] As referred above Figures 1 to 23 This specification describes the intent knowledge graph construction method and apparatus, intent prediction method and apparatus, and item recommendation method and apparatus according to embodiments thereof. The intent knowledge graph construction apparatus, intent prediction apparatus, and item recommendation apparatus described above can be implemented in hardware, software, or a combination of hardware and software.

[0176] Figure 24 An example schematic diagram of an intent knowledge graph construction apparatus 2400 implemented using a computer system according to an embodiment of this specification is shown. Figure 24 As shown, the intent knowledge graph construction apparatus 2400 may include at least one processor 2410, a memory (e.g., non-volatile memory) 2420, a RAM 2430, and a communication interface 2440, and the at least one processor 2410, memory 2420, RAM 2430, and communication interface 2440 are connected together via a bus 2460. At least one processor 2410 executes at least one computer-readable instruction (i.e., the elements implemented in software described above) stored or encoded in the memory.

[0177] In one embodiment, computer-executable instructions are stored in memory, which, when executed, cause at least one processor 2410 to: mine intent nodes of an intent knowledge graph from user interaction data between users and items on a life service platform, wherein the intent nodes are constructed as consisting of functional nodes and product nodes, and construct compositional relationships between intent nodes and their corresponding functional nodes and product nodes; determine the semantic primitive nodes corresponding to the functional nodes and product nodes, and construct relationships between the functional nodes and product nodes and their corresponding semantic primitive nodes; determine hierarchical and sequential relationships between intent nodes; construct hierarchical relationships between intent nodes with hierarchical relationships, and construct sequential relationships between intent nodes with sequential relationships.

[0178] It should be understood that the computer-executable instructions stored in memory, when executed, cause at least one processor 2410 to perform the above-described combinations in the various embodiments of this specification. Figures 1-9 and Figures 17-18 The description includes various operations and functions.

[0179] Figure 25 An example schematic diagram of an intent prediction device 2500 implemented based on a computer system according to an embodiment of this specification is shown. Figure 25 As shown, the intent prediction device 2500 may include at least one processor 2510, a memory (e.g., non-volatile memory) 2520, a RAM 2530, and a communication interface 2540, and the at least one processor 2510, memory 2520, RAM 2530, and communication interface 2540 are connected together via a bus 2560. At least one processor 2510 executes at least one computer-readable instruction (i.e., the elements implemented in software described above) stored or encoded in the memory.

[0180] In one embodiment, computer-executable instructions are stored in memory, which, when executed, cause at least one processor 2510 to: determine user intent representations of the user historical interaction data based on user historical interaction data and intent representation sets in a first user interaction data sequence between the user and items on the life service platform, the first user interaction data sequence including user historical interaction data immediately preceding the user's current interaction data, the graph representation set being generated by learning an intent knowledge graph using a graph neural network; and provide the user intent representations, interaction location information representations, interaction time information representations, current interaction location information representations, current interaction time information representations, and intent representation sets of the user historical interaction data in the first user interaction data sequence to an intent prediction model to predict the next user intent.

[0181] It should be understood that the computer-executable instructions stored in memory, when executed, cause at least one processor 2510 to perform the above-described combinations in the various embodiments of this specification. Figures 10-14 and Figures 19-21 The description includes various operations and functions.

[0182] Figure 26 An example schematic diagram of a computer-based item recommendation device 2600 implemented according to an embodiment of this specification is shown. Figure 26As shown, the item recommendation device 2600 may include at least one processor 2610, a memory (e.g., non-volatile memory) 2620, a RAM 2630, and a communication interface 2640, and the at least one processor 2610, memory 2620, RAM 2630, and communication interface 2640 are connected together via a bus 2660. The at least one processor 2610 executes at least one computer-readable instruction (i.e., the elements implemented in software described above) stored or encoded in the memory.

[0183] In one embodiment, computer-executable instructions are stored in memory that, when executed, cause at least one processor 2610 to: recall candidate items to be recommended to a user from a pool of items provided by a lifestyle service platform; rank the candidate items based on a next user intent prediction; and recommend items based on the ranking of the candidate items.

[0184] It should be understood that the computer-executable instructions stored in memory, when executed, cause at least one processor 2610 to perform the above-described combinations in the various embodiments of this specification. Figures 15-16 and Figures 22-23 The description includes various operations and functions.

[0185] According to one embodiment, a program product, such as a machine-readable medium (e.g., a non-transitory machine-readable medium), is provided. The machine-readable medium may have instructions (i.e., the elements implemented in software as described above), which, when executed by a machine, cause the machine to perform the above-described combinations of the various embodiments of this specification. Figures 1-23 The various operations and functions described. Specifically, a system or apparatus equipped with a readable storage medium storing software program code that implements the functions of any of the embodiments described above, and enabling the computer or processor of the system or apparatus to read and execute the instructions stored in the readable storage medium.

[0186] In this case, the program code itself, which can be read from a readable medium, can perform the functions of any of the above embodiments. Therefore, the machine-readable code and the readable storage medium storing the machine-readable code constitute a part of the present invention.

[0187] Examples of readable storage media include floppy disks, hard disks, magneto-optical disks, optical disks (such as CD-ROM, CD-R, CD-RW, DVD-ROM, DVD-RAM, DVD-RW, DVD-RW), magnetic tapes, non-volatile memory cards, and ROMs. Alternatively, program code can be downloaded from a server computer or the cloud via a communication network.

[0188] According to one embodiment, a computer program product is provided, the computer program product including a computer program, which, when executed by a processor, causes the processor to perform the above-described combinations of the various embodiments of this specification. Figures 1-23 The description includes various operations and functions.

[0189] Those skilled in the art will understand that the various embodiments disclosed above can be modified and varied without departing from the spirit of the invention. Therefore, the scope of protection of this invention should be defined by the appended claims.

[0190] It should be noted that not all steps and units in the above process and system structure diagrams are mandatory; some steps or units can be omitted according to actual needs. The execution order of each step is not fixed and can be determined as needed. The device structure described in the above embodiments can be a physical structure or a logical structure; that is, some units may be implemented by the same physical entity, or some units may be implemented by multiple physical entities, or they may be jointly implemented by certain components in multiple independent devices.

[0191] In the above embodiments, the hardware units or modules can be implemented mechanically or electrically. For example, a hardware unit, module, or processor may include permanent dedicated circuitry or logic (such as a dedicated processor, FPGA, or ASIC) to perform the corresponding operation. The hardware unit or processor may also include programmable logic or circuitry (such as a general-purpose processor or other programmable processor), which can be temporarily configured by software to perform the corresponding operation. The specific implementation method (mechanical, dedicated permanent circuitry, or temporarily configured circuitry) can be determined based on cost and time considerations.

[0192] The specific embodiments described above with reference to the accompanying drawings are exemplary embodiments, but do not represent all embodiments that can be implemented or fall within the scope of the claims. The term "exemplary" as used throughout this specification means "serving as an example, instance, or illustration" and does not imply that it is "preferred" or "advantageous" compared to other embodiments. Specific details are included to provide an understanding of the described techniques. However, these techniques can be practiced without these specific details. In some instances, well-known structures and apparatuses are shown in block diagram form to avoid obscuring the concepts of the described embodiments.

[0193] The foregoing description of this disclosure is provided to enable any person skilled in the art to implement or use this disclosure. Various modifications to this disclosure will be apparent to those skilled in the art, and the general principles defined herein can be applied to other variations without departing from the scope of this disclosure. Therefore, this disclosure is not limited to the examples and designs described herein, but is consistent with the widest scope of the principles and novel features disclosed herein.

Claims

1. A method for constructing an intent knowledge graph, comprising: Intent nodes of the intent knowledge graph are mined from user interaction data between users and items on the life service platform. The intent nodes are constructed as functional nodes and product nodes, and compositional relationships are established between the intent nodes and the corresponding functional nodes and product nodes. Determine the semantic primitive nodes corresponding to the functional nodes and the product nodes, and establish relationships between the functional nodes and the product nodes and their corresponding semantic primitive nodes; Determine the hierarchical and sequential relationships between the intent nodes; and Establish hierarchical relationships between intent nodes that have a hierarchical relationship, and establish sequential relationships between intent nodes that have a sequential relationship. The intent nodes have timestamp information, and determining the sequential relationship between the intent nodes includes: Intent nodes are provided to a Bayesian network for relational reasoning to identify multiple candidate intent node pairs with potential sequential relationships; and Based on the node relevance and sequential relationship sensitivity of the intent nodes, intent node pairs with a defined sequential relationship are determined from the candidate intent node pairs.

2. The method of claim 1, further comprising: Intent nodes can be constructed based on functional verbs and product entities in the knowledge base, or manually defined.

3. The method of claim 1, wherein, The user interaction data includes text data, and intent nodes are extracted from user interaction data between users and items on the lifestyle service platform, including: The user interaction data is segmented into words; Perform part-of-speech tagging on the segmented words; and The segmented words are constrained using the intent part-of-speech composition format in order to extract intent nodes.

4. The method of claim 3, wherein, The segmented words are subjected to phrase constraints using the intent part-of-speech composition format in order to extract intent nodes, including: The segmented words are constrained using the intent part-of-speech composition format to obtain candidate intent nodes; The candidate intent nodes are then given phrase scores; and The intent node is determined from the candidate intent nodes based on the phrase score of the candidate intent nodes.

5. The method of claim 1, wherein, Determining the semantic primitive nodes corresponding to the functional node and the product node includes: Using a semantic acquisition model, determine the semantic primitive nodes corresponding to the functional nodes and the product nodes.

6. The method of claim 1, wherein, Determining the hierarchical relationship between the intent nodes includes: Based on the semantic hierarchy of product nodes, the hierarchical relationship between the intent nodes is determined, or Based on the textual semantic information of the product nodes, the hierarchical relationship between the intent nodes is determined.

7. The method of claim 1, wherein, The intent knowledge graph also includes category nodes, and the method further includes: The intent classification model is used to classify the intent nodes to determine the category nodes of the intent knowledge graph. Determine the hierarchical relationship between intent nodes and category nodes, as well as among the category nodes; and Establish hierarchical relationships between the intent node and category nodes with hierarchical relationships, as well as among category nodes with hierarchical relationships.

8. An intent prediction method based on intent knowledge graph, comprising: Based on the user's historical interaction data and intent representation set in the first user interaction data sequence between the user and items on the life service platform, the user intent representation corresponding to the user's historical interaction data is determined. The first user interaction data sequence includes the user's historical interaction data immediately preceding the user's current interaction data. The intent representation set is generated by learning the intent knowledge graph through a graph neural network. The intent knowledge graph is constructed according to any one of the methods described in claims 1 to 7. as well as The user intent representation, interaction location information representation, interaction time information representation, current interaction location information representation, current interaction time information representation, and the intent representation set from the user's historical interaction data in the first user interaction data sequence are provided to the intent prediction model to predict the next user intent.

9. The intention prediction method as described in claim 8, wherein, The intent prediction model adopts a Transformer-based long-order prediction model. Providing the user intent representation, interaction location information representation, interaction time information representation, current interaction location information representation, current interaction time information representation, and the intent representation set from the user's historical interaction data in the first user interaction data sequence to the intent prediction model to predict the next user intent includes: A first user intent sequence and a second user intent sequence are generated based on the first user interaction data sequence and the second user interaction data sequence, respectively. The second user interaction data sequence includes the user's current interaction data and at least one user historical interaction data immediately preceding the user's current interaction data. The first user intent in the first user intent sequence includes a user intent representation, an interaction location information representation, and an interaction time information representation corresponding to the user interaction data. The second user intent in the second user intent sequence includes a user intent representation, an interaction location information representation, and an interaction time information representation corresponding to the user interaction data. The user intent representation in the second user intent corresponding to the user's current interaction data is filled with a specific value. The first user intent sequence is provided to the encoder of the intent prediction model to obtain a feature map of the first user intent sequence; The feature map of the first user intent sequence and the second user intent sequence are provided to the decoder of the intent prediction model to predict the feature representation of the next user intent; and The next user intent is determined based on the feature representation of the next user intent and the intent representation set.

10. The intention prediction method as described in claim 8, wherein, Based on the user's historical interaction data and intent representation set in the first user interaction data sequence between the user and items on the life service platform, the user intent representation of the user's historical interaction data is determined to include: Extract image features of interactive items corresponding to users' historical interaction data; The image feature representation of the extracted image features is concatenated with the data feature representation of the user's historical interaction data and then provided to the item representation model to obtain the item representation of the interactive item; and Based on the item representation of the interactive item and the intent representation set, the item intent representation of the interactive item is determined as the user intent representation corresponding to the user's historical interaction data.

11. A method for recommending items, comprising: Recall candidate items recommended to users from the item pool provided by the lifestyle service platform; Based on the next user intent prediction result, the candidate items are sorted, and the next user intent prediction result is predicted according to the intent prediction method as described in any one of claims 8 to 10. as well as Item recommendations are made based on the ranking of the candidate items.

12. The item recommendation method as described in claim 11, wherein, The ranking of the candidate items based on the next user intent prediction results includes: The candidate items are ranked based on the next user intent prediction result, the intent representation set, and the item intent representation.

13. The item recommendation method as described in claim 12, wherein, The candidate items recalled from the pool of items provided by the lifestyle service platform to recommend to users include: Based on the next user intent prediction result, an intent-based item retrieval method is used to retrieve candidate items recommended to the user from the item pool provided by the life service platform.

14. An apparatus for constructing an intent knowledge graph, comprising: The intent node mining unit mines intent nodes from user interaction data between users and items on the life service platform. The intent nodes are constructed as functional nodes and product nodes. The semantic origin node determination unit determines the semantic origin nodes corresponding to the functional node and the product node; The node relationship determination unit determines the hierarchical and sequential relationships between the intent nodes; as well as The node relationship construction unit constructs compositional relationships between the intent nodes and their corresponding functional and product nodes; constructs relationships between the functional nodes and product nodes and their corresponding semantic nodes; constructs hierarchical relationships between intent nodes with hierarchical relationships; and constructs sequential relationships between intent nodes with sequential relationships. The node relationship determination unit provides the intent nodes to the Bayesian network for relationship reasoning, so as to identify multiple candidate intent node pairs with potential sequential relationships. And based on the node relevance and sequential relationship sensitivity of the intent nodes, intent node pairs with a defined sequential relationship are determined from the candidate intent node pairs.

15. The apparatus of claim 14, wherein, The intent node mining unit includes: The word segmentation module performs word segmentation on the user interaction data; The part-of-speech tagging module performs part-of-speech tagging on the segmented words; and The intent node mining module uses intent part-of-speech composition format to perform phrase constraints on the segmented words in order to mine intent nodes.

16. The apparatus of claim 14, wherein, The user intent knowledge graph also includes category nodes, and the device further includes: The category node determination unit uses an intent classification model to classify each intent node to determine the category nodes of the user intent knowledge graph. The node relationship determination unit determines the hierarchical relationship between the intent node and the category node, as well as among the various category nodes. The node relationship construction unit constructs hierarchical relationships between the intent node and the category nodes with hierarchical relationships, as well as among the category nodes with hierarchical relationships.

17. An intent prediction device based on an intent knowledge graph, comprising: An intent representation set generation unit performs graph neural network learning on an intent knowledge graph to generate an intent representation set, wherein the intent knowledge graph is constructed according to any one of claims 1 to 7. The intent representation determination unit determines the user intent representation corresponding to the user historical interaction data based on the user historical interaction data in the first user interaction data sequence between the user and the items of the life service platform and the intent representation set. The first user interaction data sequence includes the user historical interaction data immediately preceding the user's current interaction data. as well as The intent prediction unit provides the user intent representation, interaction location information representation, interaction time information representation, current interaction location information representation, current interaction time information representation, and the intent representation set from the user's historical interaction data in the first user interaction data sequence to the intent prediction model to predict the next user intent.

18. The intention prediction device as claimed in claim 17, wherein, The intent prediction model adopts a Transformer-based long-order prediction model. The intent prediction unit includes: The intent sequence generation module generates a first user intent sequence and a second user intent sequence based on the first user interaction data sequence and the second user interaction data sequence, respectively. The second user interaction data sequence includes the user's current interaction data and at least one user historical interaction data immediately preceding the user's current interaction data. The first user intent in the first user intent sequence includes a user intent representation, an interaction location information representation, and an interaction time information representation corresponding to the user interaction data. The second user intent in the second user intent sequence includes a user intent representation, an interaction location information representation, and an interaction time information representation corresponding to the user interaction data. The user intent representation in the second user intent corresponding to the user's current interaction data is filled with a specific value. The intent sequence feature map determination module provides the first user intent sequence to the encoder of the intent prediction model to obtain the feature map of the first user intent sequence. The intent feature representation prediction module provides the feature map of the first user intent sequence and the second user intent sequence to the decoder of the intent prediction model to predict the feature representation of the next user intent; and The intent determination module determines the next user intent based on the feature representation of the next user intent and the intent representation set.

19. The intention prediction device as claimed in claim 17, wherein, The intent representation determination unit includes: The item image feature extraction module extracts the image features of interactive items corresponding to the user's historical interaction data; The item representation generation module concatenates the image feature representations of the extracted image features with the data feature representations of the user's historical interaction data and provides them to the item representation model to obtain the item representation of the interactive item; and The intent representation determination module determines the item intent representation of the interactive item based on the item representation and the intent representation set, and uses it as the user intent representation corresponding to the user's historical interaction data.

20. An intent prediction system based on an intent knowledge graph, comprising: The apparatus for constructing an intent knowledge graph as described in any one of claims 14 to 16; as well as The intent prediction device as described in any one of claims 17 to 19.

21. An item recommendation device, comprising: The candidate item recall unit recalls candidate items recommended to users from the item pool provided by the life service platform; The candidate item sorting unit sorts the candidate items based on the next user intent prediction result, wherein the next user intent prediction result is predicted according to any one of the intent prediction methods as described in claims 8 to 10. as well as The item recommendation unit recommends items based on the ranking results of the candidate items.

22. An item recommendation system, comprising: The apparatus for constructing an intent knowledge graph as described in any one of claims 14 to 16; Intent prediction device as described in any one of claims 17 to 19; as well as The item recommendation device as described in claim 21.

23. An apparatus for constructing an intent knowledge graph, comprising: At least one processor, Memory coupled to the at least one processor, and A computer program stored in the memory, which is executed by the at least one processor to implement the method for constructing an intent knowledge graph as described in any one of claims 1 to 7.

24. An intent prediction device based on an intent knowledge graph, comprising: At least one processor, Memory coupled to the at least one processor, and A computer program stored in the memory, which is executed by the at least one processor to implement the intent prediction method based on the intent knowledge graph as described in any one of claims 8 to 10.

25. An item recommendation device, comprising: At least one processor, Memory coupled to the at least one processor, and A computer program stored in the memory, which is executed by the at least one processor to implement the item recommendation method as described in any one of claims 11 to 13.

Images