An item recommendation method based on logical preference fusion reasoning

Abstract

The present application relates to the technical field of machine learning, and in particular to an item recommendation method based on logical preference fusion reasoning. The recommendation method includes: based on the bipartite graph of user-item interaction and the item knowledge graph, using neural symbols to determine the query embedding representation including the feature part and the symbol part; using neural symbolic operators to infer the user's interest preferences and logical needs, and determine the interest preference embedding representation, the logical need embedding representation and the real desire embedding representation; using a co-guided fusion gate to mine the relationship between the interest preference embedding representation, the logical need embedding representation and the real desire embedding representation, and determine the real desire representation; based on the real desire representation, determine the predicted probability and loss function of the target item as a query answer set; based on the loss function, perform iterative training to generate an item recommendation model based on logical preference fusion reasoning. The item recommendation method of the present application has high item recommendation accuracy and good robustness.

Description

Technical Field

[0001] The present application relates to the field of machine learning technology, and in particular to an item recommendation method based on logical preference fusion reasoning. Background Art

[0002] Recommendation systems are currently widely used in online networks to solve the problem of data overload. The task of recommendation systems is to extract user preferences from a large amount of data and recommend items of potential interest to users.

[0003] Collaborative Filtering (CF) is a recommendation algorithm based on user-item intersection. Its principle is to filter massive amounts of information by collaborating with many users’ feedback, evaluations and opinions, and select information that users may be interested in for recommendation. Therefore, it is widely used in recommendation systems. In the prior art, when using collaborative filtering, in order to utilize rich information, a method has emerged that integrates knowledge graphs (KGs) as auxiliary information of projects into CF for recommendation, namely, knowledge graph recommendation (KGRec).

[0004] However, the KGRec method uses KGs as auxiliary information to simulate the neighborhood relationship of items. It only focuses on mining user preferences, but ignores the user logical needs embedded in the logical relationship between items and entities, resulting in low recommendation accuracy and robustness. Summary of the invention

[0005] In order to address the deficiencies in the prior art, the purpose of this application is to provide an item recommendation method based on logical preference fusion reasoning to improve the recommendation accuracy of the recommendation system.

[0006] To achieve the above objectives, the present application provides an item recommendation method based on logical preference fusion reasoning, including:

[0007] Based on the bipartite graph of user-item interaction and the item knowledge graph, a query embedding representation including a feature part and a symbol part is determined using neural symbols, wherein the query embedding representation includes an item embedding representation, an entity embedding representation, a user embedding representation, a logical relationship embedding representation, and a preference relationship embedding representation;

[0008] Based on the query embedding representation, using a neural symbolic operator to infer the user's interest preferences and logical requirements, and determining an interest preference embedding representation, a logical requirement embedding representation, and a true desire embedding representation;

[0009] Using a co-guided fusion gate to mine the relationship between the interest preference embedding representation, the logical demand embedding representation and the real desire embedding representation, and determine the real desire representation;

[0010] Determining a predicted probability and a loss function of the target item as a query answer set based on the true desire representation;

[0011] Based on the loss function, iterative training is performed to generate an item recommendation model based on logical preference fusion reasoning.

[0012] Furthermore, the specific step of using neural symbols to determine the query embedding representation including the feature part and the symbol part based on the bipartite graph of user-item interaction and the item knowledge graph adopts the following formula:

[0013] ;

[0014] ;

[0015] ;

[0016] ;

[0017] in, Embedding representation for the query, Indicates the characteristic part, Indicates the symbol part, is the embedding dimension, It is a boundary; is the item embedding representation, is the entity embedding representation, Embedding representation for users, is the logical relationship embedding representation, is the preference relation embedding representation, is the relation embedding representation, The characteristic part of the relationship, The symbolic part that represents a relationship.

[0018] Furthermore, the neural symbolic operator includes a neural symbolic projection operator, using the following formula:

[0019] ;

[0020] in, is the neural symbolic projection operator, , is a connection symbol;

[0021] in, is a restriction as well as The activation function is as follows:

[0022] ;

[0023] in, and is a vector slices, is the hyperbolic tangent function, yes function.

[0024] Furthermore, the neural symbol operator also includes a neural symbol and an operator, using the following formula:

[0025] ;

[0026] ;

[0027] ;

[0028] ;

[0029] ;

[0030] ;

[0031] in, is the query embedding representation of the i-th entity, Embedding representation for query The feature representation of entity i in Embedding representation for query The symbolic representation of entity i in , are neural symbols and operators, It is a self-attention neural network with the following formula:

[0032] ;

[0033] ;

[0034] in, is the attention weight used to detect changes in the symbol part, .

[0035] in, is a fuzzy AND operator, and the formula is as follows:

[0036] .

[0037] Furthermore, the neural symbol operator also includes a neural symbol or operator, using the following formula:

[0038] ;

[0039] in, Represents the maximum value operation on the last dimension.

[0040] Furthermore, the specific steps of using the co-guided fusion gate to mine the relationship between the interest preference embedding representation, the logical demand embedding representation and the real desire embedding representation and determining the real desire representation include:

[0041] Based on the interest preference embedding representation, the logical demand embedding representation and the real desire embedding representation, the relationship between the interest preference, the logical demand and the real desire is determined respectively, and the formula is as follows:

[0042] ;

[0043] ;

[0044] in, , is a learnable parameter, Respectively represent the relationship between interest preference, logical needs and real desires, Embedding representation for interest preferences, Embed preferences for logical requirements, To embed representations for real needs;

[0045] The weights during training are adjusted through the memory gate. The formula is as follows:

[0046] ;

[0047] ;

[0048] ;

[0049] ;

[0050] in, is a learnable parameter, A memory gate that models the relative relationship between user interest preferences. To model the memory gates of the relative relationship between user logic requirements, A memory gate that models the specific relationship between user interest preferences and real wishes. A memory gate that models the specific relationship between users' logical needs and real wishes.

[0051] The mutual relationship between embeddings is modeled by the fusion gate and jointly guides the learning, as follows:

[0052] ;

[0053] ;

[0054] ;

[0055] ;

[0056] ;

[0057] in, is the user interest preference fusion gate, It is the user logic demand fusion gate. is a learnable parameter;

[0058] in, , , They are interest preference representation, logical demand representation and real demand representation.

[0059] Furthermore, the specific step of determining the predicted probability and loss function of the target item as the query answer set based on the true desire representation adopts the following formula:

[0060] ;

[0061] ;

[0062] in, Target item As a query answer set The predicted probability of the answer, represents different types of query embeddings, where Indicate interest preferences, Indicates logical requirements, Express true desire, is a hyperparameter distance boundary, represents L1 norm regularization;

[0063] is the loss function, Represents the true value of item i.

[0064] To achieve the above purpose, the electronic device provided by the present application includes:

[0065] processor;

[0066] a memory having stored thereon one or more computer program instructions executed on the processor;

[0067] Wherein, when the processor runs the computer program instructions, it executes the item recommendation method based on logical preference fusion reasoning as described above.

[0068] To achieve the above objectives, the present application provides a computer-readable storage medium on which computer instructions are stored. When the computer instructions are executed by a processor, the steps of the item recommendation method based on logical preference fusion reasoning as described above are executed.

[0069] The present invention discloses an item recommendation method based on logical preference fusion reasoning, which uses a neural symbolic method to explicitly model the user's interest preferences and logical needs for reasoning. It predicts the user's real desire from the perspective of perception and cognition. Through the co-guided fusion gate, it further mines and fuses the unique information and mutual relationship of the user's interest preferences, logical needs and real desires. It greatly improves the accuracy and robustness of the recommendation model.

[0070] Other features and advantages of the present application will be set forth in the following description, and in part will become apparent from the description, or may be understood by practicing the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] The accompanying drawings are used to provide a further understanding of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation of the present application. In the accompanying drawings:

[0072] Figure 1 A flowchart of the item recommendation method based on logical preference fusion reasoning of this application;

[0073] Figure 2 This is a schematic diagram of the structure of the item recommendation model based on logical preference fusion reasoning of this application;

[0074] Figure 3 This is a schematic diagram of the structure of the data set;

[0075] Figure 4 This is a schematic diagram of comparative test results of the item recommendation model based on logical preference fusion reasoning of this application;

[0076] Figure 5 This is a schematic diagram of the comparative test results of the item recommendation model based on logical preference fusion reasoning in this application under the H@20 indicator. DETAILED DESCRIPTION

[0077] The embodiments of the present application will be described in more detail below with reference to the accompanying drawings. Although certain embodiments of the present application are shown in the accompanying drawings, it should be understood that the present application can be implemented in various forms and should not be construed as being limited to the embodiments described herein. Instead, these embodiments are provided to provide a more thorough and complete understanding of the present application. It should be understood that the drawings and embodiments of the present application are only for exemplary purposes and are not intended to limit the scope of protection of the present application.

[0078] It should be understood that the various steps described in the method implementation of the present application can be performed in different orders and / or performed in parallel. In addition, the method implementation may include additional steps and / or omit the steps shown. The scope of the present application is not limited in this respect.

[0079] The term "including" and its variations used herein are open inclusions, i.e., "including but not limited to". The term "based on" means "based at least in part on". The term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; the term "some embodiments" means "at least some embodiments". The relevant definitions of other terms will be given in the following description.

[0080] It should be noted that the modifications of "one" and "plurality" mentioned in this application are illustrative rather than restrictive, and those skilled in the art should understand that unless otherwise clearly indicated in the context, it should be understood as "one or more". "Plurality" should be understood as two or more.

[0081] To address the problems in the existing technology, a method called Logic Preference Fusion Reasoning on Recommendation (LPFR) was proposed.

[0082] The interactions between users and items can be defined in a bipartite graph: ,in Represents users, items, and users respectively For items interest preference relationship, and Respectively The set of users and items in . express and The positive interaction between .

[0083] In addition, the recommendation task can be transformed into a complex logical query task, which can be represented by the knowledge graph of items: ,in Represents triples The head entity, relationship and tail entity in and Respectively The entity sets and relationship sets in .

[0084] Based on the above definition, we formulate the recommendation task to be solved as a logical query task. The input includes a bipartite graph of user-item interactions and knowledge graph , output recommended items that represent the user’s true wishes by considering the user’s interest preferences and logical needs.

[0085] It should be noted that the process of using the user-item interaction bipartite graph and the knowledge graph to recommend item information matching the prediction results to users includes:

[0086] Collect historical behavior data of users, such as purchase records, browsing records, click records, etc.; collect knowledge graph data of items, including attributes, classifications, brands, etc. of items; implicitly process the collected behavior data, which includes converting behavior into binary implicit feedback, such as converting clicks or purchases into 1, and converting no clicks or no purchases into 0; filter out user-item pairs that meet the 5core data standard based on the number of users-items, and only retain data with at least 5 interactions between users and items, and build a user-item interaction bipartite graph based on the filtered 5core data; map items to Freebase entities through title matching, consider triples directly related to entities aligned with items, or extract item information from local business information networks as knowledge graph data, filter out entities with less than 10 in the dataset, and retain relations that appear in at least 50 triples; randomly select 80% of each user's interaction history as a training set, and the rest as a test set; given the user-item interaction bipartite graph and knowledge graph, predict the likelihood of interacting with all items through the recommendation model, and use the most likely ones as recommendation results.

[0087] Furthermore, first, we introduce the neural symbolic operator and view the query embedding as the concatenation of features and symbols. For the symbolic part, we use vector logic, a variant of fuzzy logic, to perform explicit real logic reasoning.

[0088] In the feature part, we define the neural network and combine the results from the symbolic part. This approach allows us to combine the computational advantages of embeddings with the reasoning advantages of logic.

[0089] Subsequently, we conduct logic and preference reasoning to explicitly extract users’ interest preferences and logical needs as users’ real wishes from the perspective of perception and cognition. Thus, three answer sets are generated, including interest preferences, logical needs, and real wishes.

[0090] However, using only the true wish query answer set for prediction will ignore information specific to the other two different answer sets. This may be detrimental to recommendation.

[0091] To better utilize each answer set, we design a fusion gate to model the mutual relations among the three answer sets and promote collaborative learning. The fused true wish set is then used to predict user behavior.

[0092] In short, we propose a neural-symbolic approach that explicitly models users' preferences and logical requirements for reasoning on recommendations.

[0093] This approach enables us to predict users' true desires from the perspective of perception and cognition, and combines the computational advantages of embedding with the advantages of logical reasoning; We design a jointly guided fusion gate layer to model the interrelationships between interest preferences, logical requirements, and true desire queries, which enables us to effectively explore and utilize information unique to these three types of answer queries; To verify the effectiveness of our proposed LPFR model, we conduct comprehensive experiments on three real-world datasets. The experimental results demonstrate the effectiveness of the proposed method.

[0094] Furthermore, as mentioned in the background section, user preferences and logical requirements jointly represent the user's true wishes. However, how to effectively represent and fuse user preferences and logical requirements is still worth discussing. Therefore, in this paper, we: (1) design a neural symbolic embedding method for recommender systems to represent user preferences and logical requirements in the neural symbolic space, respectively, (2) define neural symbolic logic operators on these neural symbolic embeddings, (3) use logical preference reasoning to mine user's true wishes, and (4) use co-guided fusion gates to further mine and fuse the unique information and mutual relations of answer sets.

[0095] Hereinafter, embodiments of the present application will be described in detail with reference to the accompanying drawings.

[0096] Example 1

[0097] An embodiment of the present application provides an item recommendation method based on logical preference fusion reasoning, which will be referred to below. Figure 1 The item recommendation method based on logical preference fusion reasoning of the present application is described in detail.

[0098] Step S101: Based on the bipartite graph of user-item interaction and the item knowledge graph, a query embedding representation including a feature part and a symbol part is determined using neural symbols;

[0099] The query embedding representation includes item embedding representation, entity embedding representation, user embedding representation, logical relationship embedding representation and preference relationship embedding representation;

[0100] It should be noted that the interaction between users and items is defined in a bipartite graph:

[0101] ;

[0102] in Represents users, items, and users respectively For items interest preference relationship, and Respectively The set of users and items in . express and The positive interaction between .

[0103] In this implementation, the recommendation task can be transformed into a complex logic query task, which is represented by the knowledge graph of items: ,in Represents triples The head entity, relationship and tail entity in and Represents knowledge graphs The entity sets and relationship sets in .

[0104] It should be noted that the process of using the user-item interaction bipartite graph and the knowledge graph to recommend item information matching the prediction results to users includes:

[0105] Collect historical behavior data of users, such as purchase records, browsing records, click records, etc.; collect knowledge graph data of items, including attributes, classifications, brands, etc. of items; implicitly process the collected behavior data, which includes converting behavior into binary implicit feedback, such as converting clicks or purchases into 1, and converting no clicks or no purchases into 0; filter out user-item pairs that meet the 5core data standard based on the number of users-items, and only retain data with at least 5 interactions between users and items, and build a user-item interaction bipartite graph based on the filtered 5core data; map items to Freebase entities through title matching, consider triples directly related to entities aligned with items, or extract item information from local business information networks as knowledge graph data, filter out entities with less than 10 in the dataset, and retain relations that appear in at least 50 triples; randomly select 80% of each user's interaction history as a training set, and the rest as a test set; given the user-item interaction bipartite graph and knowledge graph, predict the likelihood of interacting with all items through the recommendation model, and use the most likely ones as recommendation results.

[0106] In the implementation of the present application, a neural symbolic embedding representation is formulated for query q, entity e, item i, logical relation l, user u, and preference relation p. We consider dividing the query embedding into two different parts: features and symbols, to encode the user's true desire from the perspective of perception and cognition respectively, and the two parts are learned independently. Among them, the feature part is calculated by a neural network. The symbol part is calculated independently and strictly follows vector logic, which is a form of fuzzy logic.

[0107] Specifically, the specific steps of using neural symbols to determine the query embedding representation including the feature part and the symbol part are as follows:

[0108] ;

[0109] ;

[0110] ;

[0111] ;

[0112] in, Embedding representation for the query, Indicates the characteristic part, Indicates the symbol part, is the embedding dimension, It is a boundary; is the item embedding representation, is the entity embedding representation, Embedding representation for users, is the logical relationship embedding representation, is the preference relation embedding representation, is the relation embedding representation, The characteristic part of the relationship, The symbolic part that represents a relationship.

[0113] It should be noted that, since the item embedding representation will not change, the entity embedding representation, user embedding representation and relationship embedding representation will change with training, so the above entity embedding representation, user embedding representation and relationship embedding representation are all initialized.

[0114] It is understandable that logical relationships are embedded From the knowledge graph Initialize the relationship between items and entities, and embed the preference relationship User-Item Bipartite Graph Initialize the relationship between users and items.

[0115] In summary, unlike previous neural symbolic work that uses the same ID embedding as both feature representation and symbolic variable, the implementation method of the present application divides the embedding representation into two parts: a feature part and a symbolic part to describe features and perform logical reasoning respectively.

[0116] Step S102: Based on the query embedding representation, use a neural symbolic operator to infer the user's interest preferences and logical requirements to determine the interest preference embedding representation, the logical requirement embedding representation and the real desire embedding representation;

[0117] In this embodiment, three operators are used, including:

[0118] The neural symbolic projection operator uses the following formula:

[0119] ;

[0120] in, is the neural symbolic projection operator, , It is a connection symbol;

[0121] in, is a restriction as well as The activation function is as follows:

[0122] ;

[0123] in, and is a vector slices, is the hyperbolic tangent function, yes function.

[0124] From the above formula, we can see that the neural symbolic projection operator It maps query embedding to item embedding. , in relational embedding Perform projection operation.

[0125] Neural Symbols and Operators,Neural Symbols and Operators, which applies query embedding to both the symbolic part and the feature part. The symbolic part is embedded by vector logic, which ensures that at least all logical operators can be modeled. The feature part adopts the symbolic information and calculates the feature information through the self-attention mechanism.

[0126] Neural symbols and operators use the following formula:

[0127] ;

[0128] ;

[0129] ;

[0130] ;

[0131] ;

[0132] ;

[0133] in, is the query embedding representation of the i-th entity, Embedding representation for query The feature representation of entity i in Embedding representation for query The symbolic representation of entity i in , are neural symbols and operators, It is a self-attention neural network that learns hidden information from the symbol part and applies it to the feature part. The formula is as follows:

[0134] ;

[0135] ;

[0136] in, is the attention weight used to detect changes in the symbol part, .

[0137] in, is a fuzzy AND operator, and the formula is as follows:

[0138] .

[0139] It should be noted that the logical AND operation of the symbol part only depends on the previous symbol part.

[0140] Neural Symbols OR Operators,It should be noted that the queries in the training phase only involve,neural symbolic projection and neural symbolic AND operations, and the,queries of neural symbols OR operations are only introduced in the testing phase.

[0141] Use the following formula:

[0142] ;

[0143] in, Represents the maximum value operation on the last dimension.

[0144] In the implementation mode of the present application, a multi-hop relationship design in a knowledge graph is adopted, and a neural symbolic operator is used to perform logical reasoning on queries consisting of user interest preferences (user embedding, preference relationship embedding) and logical needs, and ultimately obtain three types of answer sets: interest preference embedding representation, logical need embedding representation, and real desire embedding representation.

[0145] The true desire embedding representation is defined as follows:

[0146]

[0147] in, It is a neural symbol and operator. Represents the user's true desire embedding representation. Indicates the user's preference interests. Represents the user's logical requirements.

[0148] in, and It is calculated based on the above neural symbolic operator.

[0149] Step S103: using a co-guided fusion gate to mine the relationship between the interest preference embedding representation, the logical demand embedding representation and the real desire embedding representation, and determining the real desire representation;

[0150] Since the user's interest preferences and logical needs influence each other and jointly determine the user's true desire, there are limitations in using only neural symbols and operators to determine the final predicted true desire. This approach may ignore specific information related to the user's interest preferences and logical needs, which may be beneficial to the recommendation. At the same time, this specific information may influence each other and determine the user's choice.

[0151] Therefore, in order to obtain the real desire more accurately, in this implementation:

[0152] First, based on the interest preference embedding representation, the logical demand embedding representation and the real desire embedding representation, the relationship between interest preference, logical demand and real desire is determined respectively, and the formula is as follows:

[0153] ;

[0154] ;

[0155] in, , is a learnable parameter, Respectively represent the relationship between interest preference, logical needs and real desires, Embedding representation for interest preferences, Embed preferences for logical requirements, To embed representations for real needs;

[0156] Then, the weights during training are adjusted through the memory gate, and the formula is as follows:

[0157] ;

[0158] ;

[0159] ;

[0160] ;

[0161] in, is a learnable parameter, A memory gate that models the relative relationship between user interest preferences. A memory gate that models the relative relationship between user logic requirements. A memory gate that models the specific relationship between user interest preferences and true desires. A memory gate that models the specific relationship between users' logical needs and real wishes.

[0162] Finally, the mutual relationship between embeddings is modeled through the fusion gate and jointly guides the learning, as follows:

[0163] ;

[0164] ;

[0165] ;

[0166] ;

[0167] ;

[0168] in, is the user interest preference fusion gate, It is the user logic demand fusion gate. is a learnable parameter;

[0169] in, , , They are interest preference representation, logical demand representation and real demand representation.

[0170] Step S104: determining the predicted probability and loss function of the target item as a query answer set based on the true desire representation;

[0171] In the embodiment of the present application, in order to expect the distance between the target item and the recommendation query to be as small as possible, a distance function is defined to measure the embedding distance of the target item. With query embedding The distance between items As a query answer set The probability of the answer is calculated as follows:

[0172] ;

[0173] in Target item As a query answer set The predicted probability of the answer, represents different types of query embeddings, where Indicate interest preferences, Indicates logical requirements, Express true desire, is a hyperparameter distance boundary, represents L1 norm regularization.

[0174] In the embodiment of the present application, the true value is approximated by binary cross entropy loss , which is a one-hot encoded vector representing Is it in the answer set? express In the answer set, otherwise .

[0175] The final loss function is expressed as:

[0176]

[0177] in, is the loss function, Represents the true value of item i.

[0178] Step S105: Based on the loss function, iterative training is performed to generate an item recommendation model based on logical preference fusion reasoning.

[0179] In the implementation of the present application, in order to evaluate the trained item recommendation model based on logical preference fusion reasoning (hereinafter referred to as LPFR), experiments are conducted on three widely used datasets for recommendation. The datasets are Amazon-book, LFM-1b and Yelp. We use the Amazon-book and LFM-1b datasets to compare with the recommendation model based on the knowledge graph, and use the Amazon-book and Yelp datasets to compare with the query embedding method for complex logical reasoning. The detailed data of the datasets are as follows: Figure 3 shown.

[0180] In the implementation of this application, two of the most commonly used performance evaluation indicators for recommendation tasks, namely hit rate (HIT) and normalized discounted cumulative gain (NDCG), are used. HIT is the proportion of ground-truth items that appear in the top N recommendations. NDCG is a ranking loss that measures the position of real items in the top N recommendations. The higher the results of these two indicators, the more effective the model. We report the results of HIT@{10,20} and NDCG@{10,20}.

[0181] We compared with common baseline models. The baselines used to compare with our model can be divided into two groups. Recommendation models based on knowledge graphs: For this group, we compared methods based on knowledge graph embeddings (MKR), methods combining embeddings and paths (RippleNet), methods based on user-item knowledge graphs (CFKG), methods based on GNNs (KGCN), and multi-task methods that encode logical requirements (LogicRec). Query embedding models: For this group, we compared our framework with some query embedding methods for complex logical reasoning. They are geometric operation embedding methods (GQE, Q2B), probabilistic embedding methods (BetaE, GammaE), fuzzy logic embedding methods (FuzzQE), and neural symbolic embedding methods (FLEX).

[0182] See also Figure 4-Figure 5 , the comparative test results of the recommended performance are as follows Figure 4-Figure 5 As shown, where “H” is the abbreviation of HIT, “N” is the abbreviation of NDCG, underlined numbers represent suboptimal results, and bold numbers represent the best results.

[0183] from Figure 4 The experimental results in show that, compared with the baseline model of the recommendation method based on knowledge graph, LPFR achieved the best performance in both HIT and NDCG indicators on both datasets. The recommendation model based on knowledge graph only uses knowledge graphs (KGs) as auxiliary information to obtain better entity representation, without considering the logical relationship between entities. Our method takes into account the logical relationship between items and entities to simulate the logical needs of users. Such experimental results show that it is beneficial to consider the logical relationship between entities in the project recommendation task. Compared with the existing method LogicRec that considers user logical needs, our method performs better due to the advantages of neural symbolic computing.

[0184] from Figure 5The experimental results in show that LPFR achieves the best performance in the HIT@20 metric on both datasets compared with the baseline model of the query embedding method. This result demonstrates the effectiveness of LPFR in the logical part. This proves the effectiveness of our proposed method in embedding entities in terms of both features and symbols. Because the implementation of the baseline model only utilizes the intersection part, it does not consider the inherent information in user preferences and project relationships, which is helpful for recommendation. Therefore, this also reflects the necessity of exploring and integrating the inherent information in user preferences and project relationships.

[0185] Example 2

[0186] In this embodiment, an electronic device is also provided, including a processor and a memory. The memory is used to store non-transitory computer-readable instructions. The processor is used to run the non-transitory computer-readable instructions, and when the non-transitory computer-readable instructions are run by the processor, one or more steps of the item recommendation method based on logical preference fusion reasoning described above can be executed. The memory and the processor can be interconnected through a bus system and / or other forms of connection mechanisms.

[0187] For example, the processor can be a central processing unit (CPU), a digital signal processor (DSP), or other forms of processing units with data processing capabilities and / or program execution capabilities, such as a field programmable gate array (FPGA); for example, the central processing unit (CPU) can be an X86 or ARM architecture, etc.

[0188] For example, the memory may include any combination of one or more computer program products, and the computer program product may include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. Volatile memory may include, for example, random access memory (RAM) and / or cache memory (cache), etc. Non-volatile memory may include, for example, read-only memory (ROM), hard disk, erasable programmable read-only memory (EPROM), portable compact disk read-only memory (CD-ROM), USB memory, flash memory, etc. One or more computer program modules may be stored on the computer-readable storage medium, and the processor may run one or more computer program modules to implement various functions of the electronic device. Various applications and various data, as well as various data used and / or generated by the application, etc. may also be stored in the computer-readable storage medium.

[0189] It should be noted that in the embodiments of the present application, the specific functions and technical effects of the electronic device can refer to the above description of the item recommendation method based on logical preference fusion reasoning, which will not be repeated here.

[0190] Example 3

[0191] In this embodiment, a computer-readable storage medium is also provided, and the storage medium is used to store non-transitory computer-readable instructions. For example, when the non-transitory computer-readable instructions are executed by a computer, one or more steps in the item recommendation method based on logical preference fusion reasoning described above can be executed.

[0192] For example, the storage medium can be applied to the above-mentioned electronic device. For example, the storage medium can be the memory in the electronic device of Example 2. For example, the relevant description of the storage medium can refer to the corresponding description of the memory in the electronic device of Example 2, which will not be repeated here.

[0193] It should be noted that the storage medium (computer-readable medium) mentioned above in the present application may be a computer-readable signal medium or a non-transitory computer-readable storage medium or any combination of the above two. The non-transitory computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or device, or any combination of the above. More specific examples of non-transitory computer-readable storage media may include, but are not limited to: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above.

[0194] In the present application, a non-transitory computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in combination with an instruction execution system, apparatus, or device. In the present application, a computer-readable signal medium may include a data signal propagated in a baseband or as part of a carrier wave, which carries a computer-readable program code. Such propagated data signals may take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. A computer-readable signal medium may also be any computer-readable medium other than a non-transitory computer-readable storage medium, which may send, propagate, or transmit a program for use by or in combination with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (radio frequency), etc., or any suitable combination of the above.

[0195] The computer-readable medium may be included in the electronic device, or may exist independently without being installed in the electronic device.

[0196] Computer program code for performing the operations of the present application may be written in one or more programming languages, or a combination thereof, including, but not limited to, object-oriented programming languages, such as Java, Smalltalk, C++, and conventional procedural programming languages, such as "C" or similar programming languages. The program code may be executed entirely on the user's computer, partially on the user's computer, as a stand-alone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server.

[0197] The flow chart and block diagram in the accompanying drawings illustrate the possible architecture, function and operation of the system, method and computer program product according to various embodiments of the present application. In this regard, each square box in the flow chart or block diagram can represent a module, a program segment or a part of a code, and the module, the program segment or a part of the code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some implementations as replacements, the functions marked in the square box can also occur in a sequence different from that marked in the accompanying drawings. For example, two square boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each square box in the block diagram and / or flow chart, and the combination of the square boxes in the block diagram and / or flow chart can be implemented with a dedicated hardware-based system that performs a specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

[0198] The units involved in the embodiments described in this application may be implemented by software or hardware, wherein the name of a unit does not, in some cases, constitute a limitation on the unit itself.

[0199] The functions described above herein may be performed at least in part by one or more hardware logic components. For example, without limitation, exemplary types of hardware logic components that may be used include: field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), systems on chips (SOCs), complex programmable logic devices (CPLDs), etc.

[0200] The above description is only a partial embodiment of the present application and an explanation of the technical principles used. Those skilled in the art should understand that the scope of disclosure involved in the present application is not limited to the technical solution formed by a specific combination of the above technical features, but should also cover other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the above disclosed concept. For example, the above features are replaced with the technical features with similar functions disclosed in this application (but not limited to) by each other to form a technical solution.

[0201] In addition, although each operation is described in a specific order, this should not be construed as requiring these operations to be performed in the specific order shown or in a sequential order. Under certain circumstances, multitasking and parallel processing may be advantageous. Similarly, although some specific implementation details are included in the above discussion, these should not be construed as limiting the scope of the application. Some features described in the context of a separate embodiment can also be implemented in a single embodiment in combination. On the contrary, the various features described in the context of a single embodiment can also be implemented in multiple embodiments individually or in any suitable sub-combination mode.

[0202] Although the subject matter has been described in language specific to structural features and / or methodological logical actions, it should be understood that the subject matter defined in the appended claims is not necessarily limited to the specific features or actions described above. On the contrary, the specific features and actions described above are merely example forms of implementing the claims.

Claims

1. An item recommendation method based on logical preference fusion reasoning, characterized in that: include: Based on the bipartite graph of user-item interaction and the item knowledge graph, neural symbols are used to determine the query embedding representation including the feature part and the symbol part, and the query embedding representation includes the item embedding representation, the entity embedding representation, the user embedding representation, the logical relationship embedding representation and the preference relationship embedding representation; wherein, based on the bipartite graph of user-item interaction and the item knowledge graph, the specific step of using neural symbols to determine the query embedding representation including the feature part and the symbol part adopts the following formula: ； ； ； ； in, Embedding representation for the query, Indicates the characteristic part, Indicates the symbol part, is the embedding dimension, It is a boundary; is the item embedding representation, is the entity embedding representation, Embedding representation for users, is the logical relationship embedding representation, is the preference relation embedding representation, is the relation embedding representation, The characteristic part of the relationship, The symbolic part that represents the relationship; Based on the query embedding representation, the user's interest preference and logical demand are inferred using a neural symbolic operator to determine the interest preference embedding representation, the logical demand embedding representation and the real desire embedding representation, wherein the neural symbolic operator includes a neural symbolic projection operator, and the following formula is used: ； in, is the neural symbolic projection operator, , It is a connection symbol; in, is a restriction as well as The activation function is as follows: ； in, and is a vector slices, is the hyperbolic tangent function, yes function; Using a co-guided fusion gate to mine the relationship between the interest preference embedding representation, the logical demand embedding representation and the real desire embedding representation, and determine the real desire representation; Determining a predicted probability and a loss function of the target item as a query answer set based on the true desire representation; Based on the loss function, iterative training is performed to generate an item recommendation model based on logical preference fusion reasoning.

2. The item recommendation method based on logical preference fusion reasoning according to claim 1 is characterized in that: The neural symbol operator also includes a neural symbol and an operator, using the following formula: ； ； ； ； ； ； in, is the query embedding representation of the i-th entity, Embedding representation for query The feature representation of entity i in Embedding representation for query The symbolic representation of entity i in , are neural symbols and operators, is a self-attention neural network with the following formula: ； ； in, is the attention weight used to detect changes in the symbol part, ; in, is a fuzzy AND operator, and the formula is as follows: 。 3. The item recommendation method based on logical preference fusion reasoning according to claim 2 is characterized in that: The neural symbolic operator also includes a neural symbol or operator, using the following formula: ； in, Represents the maximum value operation on the last dimension.

4. The item recommendation method based on logical preference fusion reasoning according to claim 3 is characterized in that: The specific steps of using the co-guided fusion gate to mine the relationship between the interest preference embedding representation, the logical demand embedding representation and the real desire embedding representation and determining the real desire representation include: Based on the interest preference embedding representation, the logical demand embedding representation and the real desire embedding representation, the relationship between the interest preference, the logical demand and the real desire is determined respectively, and the formula is as follows: ； ； in, , is a learnable parameter, Respectively represent the relationship between interest preference, logical needs and real desires, Embedding representation for interest preferences, Embed preferences for logical requirements, To embed representations for real needs; The weights during training are adjusted through the memory gate. The formula is as follows: ； ； ； ； in, is a learnable parameter, A memory gate that models the relative relationship between user interest preferences. To model the memory gates of the relative relationship between user logic requirements, A memory gate that models the specific relationship between user interest preferences and real wishes. Memory gates to model the specific relationship between user logical needs and real desires; The mutual relationship between embeddings is modeled by the fusion gate and jointly guides the learning, as follows: ； ； ； ； ； in, is the user interest preference fusion gate, It is the user logic demand fusion gate. is a learnable parameter; in, , , They are interest preference representation, logical demand representation and real demand representation.

5. The item recommendation method based on logical preference fusion reasoning according to claim 4 is characterized in that: The specific step of determining the predicted probability and loss function of the target item as the query answer set based on the true desire representation adopts the following formula: ； ； in, Target item As a query answer set The predicted probability of the answer, represents different types of query embeddings, where Indicate interest preferences, Indicates logical requirements, Express true desire, is a hyperparameter distance bound, represents L1 norm regularization; is the loss function, Represents the true value of item i.

6. An electronic device, characterized in that: include: processor; a memory having stored thereon one or more computer program instructions executed on the processor; Wherein, when the processor runs the computer program instructions, it executes the item recommendation method based on logical preference fusion reasoning as described in any one of claims 1-5.

7. A computer-readable storage medium, characterized in that: Computer instructions are stored thereon, and when the computer instructions are executed by the processor, the steps of the item recommendation method based on logical preference fusion reasoning as described in any one of claims 1 to 5 are executed.

Images