An item debiasing recommendation method and system
By fusing text and graph embedding representations in the recommendation system and using Transformer and graph convolution to optimize node embeddings, the popularity bias problem is solved, the accuracy and performance of the recommendation system are improved, and items that meet the user's real interests are recommended.
Patent Information
- Application Number
- CN202411271153.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-11
- Publication Date
- 2025-09-05
- Estimated Expiration
- 2044-09-11
AI Technical Summary
Existing recommendation systems are difficult to accurately recommend items that match users' real interests due to the influence of popularity bias, and methods to remove popularity bias usually impair recommendation performance.
By obtaining the user interaction history sequence and user-item interaction graph, using the Transformer encoder and decoder to extract text embedding representation, combining graph convolution and contrastive learning to optimize the node embedding representation, building an item debiased recommendation system, fusing text and graph embedding representations, and generating a recommendation list that meets the user's interests.
It effectively removes popularity bias, improves recommendation performance, increases the recommendation frequency of long-tail nodes, alleviates the center effect of graph convolutional networks, and achieves a win-win situation in recommendation performance and debiasing performance.
Smart Images
Figure CN119377392B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field related to information recommendation, and more specifically, relates to a method and system for debiasing item recommendation. Background Art
[0002] Recommender systems, as a key branch of computer information technology, have become an indispensable component of internet technology amidst the explosive growth of information on the internet. The ever-increasing amount of information on various social media platforms is causing users to face the challenge of information overload. This overload makes it difficult for users to filter useful content from this vast amount of information, but recommendation systems can effectively address this challenge.
[0003] With the widespread adoption of recommendation systems and their increasing user adoption, many issues have gradually emerged. One such issue is that current recommendation systems, influenced by popularity bias, often fail to accurately recommend items that align with users' true interests. Recommendation systems designed to eliminate popularity bias in recommendation results aim to reduce or eliminate the bias caused by item popularity, ensuring fairer and more diverse recommendation lists that meet users' true interests and needs.
[0004] Currently, the removal of popularity bias in recommendation systems generally involves modifying the loss function or post-processing the recommendation results. This method does not consider the balance between debiasing performance and recommendation performance, resulting in deterioration of recommendation performance while removing bias. Summary of the Invention
[0005] In response to the above-mentioned defects or improvement needs of the prior art, the present invention provides an item debiasing recommendation method and system thereof, which aims to remove popularity bias and recommend items that meet the user's real interests.
[0006] To achieve the above objectives, the present invention provides a method for debiasing item recommendation, which comprises:
[0007] Obtain a user interaction history sequence and a user-item interaction graph, wherein the user interaction history sequence includes the user's historical interaction items, and the user-item interaction graph is a user-item adjacency matrix; concatenate the attribute text representations of all items in the same user interaction history sequence, perform feature extraction and fusion, and obtain a text embedding representation of the corresponding user; perform feature extraction and fusion on the attribute text representations of the same item to obtain a text embedding representation of the corresponding item; the attribute text of the item includes popularity-related attributes and non-popularity-related attributes;
[0008] The user's text embedding representation and the item's text embedding representation are used as the initial node embedding representation of the corresponding node in the user-item interaction graph; all nodes in the user-item interaction graph are traversed, the directly connected nodes of the current node are divided into two sets, the nodes in each set are aggregated respectively, one of the aggregated results is used as the embedding representation of the current node, the other aggregated result is used as the positive sample of the current node, and the positive samples of multiple nodes are randomly selected from nodes other than the directly connected nodes as negative samples of the current node. The embedding representation of each node is optimized through comparative learning;
[0009] Perform graph convolution feature extraction on the user-item interaction graph after contrastive learning to obtain the final graph embedding representation of the nodes;
[0010] The text embedding representation of the target user and the final graph embedding representation are fused to obtain the final embedding representation of the target user;
[0011] Calculate the similarity between the target user's final embedding representation and the item's text embedding representation, and determine the item's recommendation priority based on the similarity. The higher the similarity, the higher the recommendation priority.
[0012] In some embodiments, obtaining a user interaction history sequence and a user-item interaction graph includes:
[0013] After removing items and users whose number of interactions is lower than a preset requirement, the user interaction history sequence and the user-item interaction graph are generated based on the interaction records.
[0014] In some embodiments, the feature extraction and fusion includes:
[0015] First, use the Transformer encoder to extract features from the input text;
[0016] The extracted features are then decoded using the Transformer decoder to achieve feature fusion.
[0017] In some embodiments, the concatenation of attribute text representations of all items in the same user interaction history sequence, followed by feature extraction and fusion, to obtain a text embedding representation of the corresponding user includes:
[0018] Concatenate the attribute text representations of all items in the order of recent to distant interaction time to form a long text sentence;
[0019] Splitting the long text sentence into a series of sub-paragraphs;
[0020] Use the Transformer encoder to extract features from each sub-paragraph one by one to obtain corresponding sub-paragraph features;
[0021] Splicing the sub-paragraph features in the order of the sub-paragraphs to obtain splicing features of the long text sentence;
[0022] The concatenated features of the long text sentence are decoded using a Transformer decoder to obtain a text embedding representation of the corresponding user.
[0023] In some embodiments, performing graph convolution feature extraction on the user-item interaction graph after contrastive learning to obtain a final graph embedding representation of the nodes includes:
[0024] The node aggregation calculation of graph convolution is implemented according to the following aggregation formula:
[0025]
[0026] Where u is the node to be updated, is the set of neighbor nodes of node u, d u is the degree of node u, k is the number of aggregations, is the embedding representation of node u after k+1 aggregations, d i is the degree of node i, is the embedding representation of node i after k aggregations;
[0027] After performing a preset number of aggregations, the results of each aggregation are weighted summed to obtain the final graph embedding representation of the node.
[0028] In some embodiments, the step of fusing the text embedding representation of the target user and the final graph embedding representation to obtain the final embedding representation of the target user includes:
[0029] The target user's text embedding representation and the final graph embedding representation are concatenated and sent to a multi-layer perceptron for linear transformation and nonlinear activation to obtain the final embedding representation of the target user.
[0030] In some embodiments, calculating the similarity between the final embedding representation of the target user and the text embedding representation of the item includes:
[0031] The similarity between the final embedding representation of the target user and the text embedding representation of the item is calculated using the vector inner product formula.
[0032] The present invention also provides an item debiasing recommendation system, which includes:
[0033] The text feature extraction module is used to obtain a user interaction history sequence and a user-item interaction graph. The user interaction history sequence includes the user's historical interaction items, and the user-item interaction graph is a user-item adjacency matrix. The module concatenates the attribute text representations of all items in the same user interaction history sequence, performs feature extraction and fusion, and obtains a text embedding representation of the corresponding user. The module also performs feature extraction and fusion on the attribute text representations of the same item to obtain a text embedding representation of the corresponding item. The attribute text of the item includes popularity-related attributes and non-popularity-related attributes.
[0034] A graph contrast learning module is configured to use the user's text embedding representation and the item's text embedding representation as the initial node embedding representation of the corresponding node in the user-item interaction graph; traverse all nodes in the user-item interaction graph, divide the directly connected nodes of the current node into two sets, aggregate the nodes in each set respectively, use one of the aggregated results as the embedding representation of the current node, use the other aggregated result as the positive sample of the current node, randomly select multiple positive samples of nodes from nodes other than the directly connected nodes as negative samples of the current node, and optimize the embedding representation of each node through contrast learning;
[0035] The graph convolution module is used to perform graph convolution feature extraction on the user-item interaction graph after contrastive learning to obtain the final graph embedding representation of the nodes;
[0036] The feature fusion module is used to fuse the text embedding representation of the target user and the final graph embedding representation to obtain the final embedding representation of the target user;
[0037] The output module is used to calculate the similarity between the final embedding representation of the target user and the text embedding representation of the item, and determine the recommendation priority of the item based on the similarity. The higher the similarity, the higher the recommendation priority.
[0038] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program implements the steps of any of the above methods when executed by a processor.
[0039] The present invention also provides a computer program product, comprising a computer program or instructions, wherein when the computer program or instructions are executed by a processor, the steps of any of the above methods are implemented.
[0040] In general, the above technical solutions conceived by the present invention have the following beneficial effects compared with the prior art:
[0041] 1. The present invention integrates text embedding representation and graph embedding representation, and uses the generalization of natural language to solve the popularity bias problem existing in traditional models, and accurately captures user interests through text matching to improve the model recommendation performance. Among them, when obtaining the text embedding representation, by splicing the attributes of all items in the user's historical interaction sequence and extracting information, the text attribute information is fully utilized, which includes both popularity-related attributes and non-popularity-related attributes, and is not limited to the specific attributes of the items, and has strong generalization. When obtaining the graph embedding representation, graph-enhanced contrastive learning is adopted, and node positive and negative samples are constructed through node partitioning, so that more long-tail nodes can participate in feature extraction, thereby improving the recommendation frequency of long-tail nodes and solving the popularity bias problem caused by the long-tail distribution of data.
[0042] 2. Furthermore, the embodiments of the present invention also optimize the aggregation weight of graph convolution, which can suppress the information propagation of popular nodes, reduce popularity deviation, narrow the influence of popular nodes and non-popular nodes on other nodes, alleviate the center effect of graph convolutional networks, and achieve a win-win situation in debiasing performance and recommendation performance.
[0043] 3. Furthermore, in an embodiment of the present invention, the user's long text sentences are segmented and then features are extracted, and then the features are concatenated and fused to generate the user's text embedding representation. This can effectively reduce the time complexity of attention calculation while taking into account the semantic features at different levels in the historical interaction sequence. BRIEF DESCRIPTION OF THE DRAWINGS
[0044] Figure 1 is a flowchart of the steps of a method for debiasing item recommendation in one embodiment of the present invention;
[0045] Figure 2 This is a schematic diagram of some expanded attributes of a certain product in one embodiment of the present invention;
[0046] Figure 3 is a schematic diagram of comparative learning of a certain block diagram in one embodiment of the present invention;
[0047] Figure 4 2 is a schematic diagram of the structure of an item debiasing recommendation system in one embodiment of the present invention. DETAILED DESCRIPTION
[0048] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely for the purpose of explaining the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.
[0049] Example 1
[0050] like Figure 1 The figure shows a flowchart of the steps of the item debiasing recommendation method in one embodiment of the present invention, which mainly includes steps S1 to S5. Each step is introduced in detail below.
[0051] Step S1: Obtain a user interaction history sequence and a user-item interaction graph. The user interaction history sequence includes the user's historical interaction items, and the user-item interaction graph is the intersection adjacency matrix of users and items. Concatenate the attribute text representations of all items in the same user interaction history sequence, perform feature extraction and fusion, and obtain the text embedding representation of the corresponding user. Perform feature extraction and fusion on the attribute text representations of the same item to obtain the text embedding representation of the corresponding item.
[0052] This step can be specifically performed as follows.
[0053] Step S11: Obtain user interaction history sequence.
[0054] Specifically, items and users with interaction numbers lower than preset requirements can be removed. For example, items and users with too few interactions can be removed using the ten-degree screening method, and then a user interaction history sequence is generated based on the interaction records.
[0055] Step S12: Traverse the user interaction history sequence and generate a user-item interaction adjacency matrix.
[0056] Specifically, both items and users can be used as nodes. For example, users can be used as column nodes, and items can be used as column nodes. The element in the Mth row and Nth column indicates whether the Mth user and the Nth item have an interaction relationship. The interaction adjacency matrix is converted into a compression coefficient row format for storage.
[0057] Step S13: Generate text embedding representations for each user and each item.
[0058] For user nodes, a textual representation of each user is obtained by concatenating the attribute textual representations of all items in their historical interaction sequence. Each item has multiple attributes, including popularity-related attributes such as rating score and number of reviews, as well as attributes unrelated to popularity such as item name and address. In this invention, the selected attributes include both popularity-related and non-popularity-related attributes. Before concatenating the attribute textual representations of different items, the multiple attribute key-value pairs for the same item are first expanded.
[0059] For item nodes, expand their multiple attributes to obtain the text representation of each item.
[0060] like Figure 2The figure shows a schematic diagram of expanding some attributes of a commodity in one embodiment of the present invention. For any item v, its attributes can be expanded into a sentence F(v), which is expressed as follows:
[0061] F(v)=Attr1:v(Attr1),…,Attr k :v(Attr k )
[0062] In the formula, Attr is the key of a specific attribute pair, and v(Attr) represents the value of the attribute pair.
[0063] In one embodiment, all attribute information of an item can be selected to implement full-text information extraction and obtain the full-text representation embedding of the node. The advantage of using this full-text representation embedding is that it has no obvious bias. The embedding of long-tail items and popular items will not be affected by popularity bias, and it can fully utilize the generalization advantages of natural language and the advantages of pre-trained models. Ultimately, through text matching between users and items, the real interests of users can be better captured, achieving a win-win situation in recommendation performance and debiasing performance.
[0064] Specifically, we can first use the Transformer encoder to extract features from the text of each node, and then use the Transformer decoder to decode the extracted features to achieve feature fusion and obtain the text embedding representation of the corresponding node.
[0065] Considering that the user's text representation is long, it is difficult to extract fine-grained features and multi-layer semantics from long text sequences. In order to improve the effect of feature extraction and fusion, in one embodiment, the following steps S131 to S135 are used to generate a text embedding representation for each user.
[0066] Step S131: Concatenate the attribute text representations of all items in the order of the interaction time of the items from recent to distant to form a long text sentence.
[0067] In order to understand the user's interaction history To encode, we first need to reverse the user's interaction history to In the form of v t is the item that the current user interacted with at time t. Intuitively, the more recent the interaction record, the higher the weight. After reversing the interaction history, the text representation of the item can be used to replace the interaction history. Finally, add an identifier to the front of each interaction record to indicate the beginning, as shown below:
[0068]
[0069] The final text representation for each user contains the attributes of all items they've interacted with, equivalent to a long sentence composed of attribute key-value pairs from all items in their historical interaction sequence. Among the various item attributes, some are related to popularity, such as item ratings and the number of item reviews. This can also introduce popularity bias, but given that this tendency isn't always harmful, some popularity-related attributes are retained in the final text representation.
[0070] Step S132: Divide the long text sentence into a series of sub-paragraphs.
[0071] Specifically, split the long text sequence into a series of sub-paragraphs as follows:
[0072]
[0073] Here S n Represents the nth subsequence split into, each subsequence has m tokens.
[0074] Step S133: Use the Transformer encoder to extract features from each sub-paragraph one by one to obtain corresponding sub-paragraph features.
[0075] Specifically, the Transformer encoder is used for feature extraction, and the extracted features represent the user's interest preferences over a short period of time.
[0076] For example, for any subparagraph S i , perform feature extraction to obtain sub-paragraph feature h i , which can be expressed as:
[0077] h i =TEncoder(S i )
[0078] Step S134: splicing the sub-paragraph features in the order of the sub-paragraphs to obtain the splicing features of the long text sentence.
[0079] Splice the sub-paragraph features of the same user to represent the splicing features of the long text sentence It is expressed as follows:
[0080]
[0081] Step S135: Use the Transformer decoder to decode the concatenated features of the long text sentence to obtain the text embedding representation of the corresponding user.
[0082] Join features By sending it to the Transformer decoder for decoding, we can get the user's complete interest preferences, as shown below:
[0083]
[0084] Generating the user's text embedding representation through the above steps can not only effectively reduce the time complexity of attention calculation, but also take into account the semantic features of different levels in the historical interaction sequence.
[0085] The method for generating text embedding representations for items is similar to that for user text embedding representations. However, the attribute text representation of an item only has the attributes of a single item. The text is shorter, there is no concatenation operation, and it does not need to be split into sub-paragraphs. The Transformer encoder and Transformer decoder are directly used for feature extraction and fusion to obtain the text embedding representation of the corresponding item.
[0086] Step S2: Use the user's text embedding representation and the item's text embedding representation as the initial node embedding representation of the corresponding node in the user-item interaction graph; traverse all nodes in the user-item interaction graph, divide the directly connected nodes of the current node into two sets, aggregate the nodes in each set respectively, use one of the aggregation results as the embedding representation of the current node, and use the other aggregation result as the positive sample of the current node. Randomly select multiple positive samples of nodes from nodes other than the directly connected nodes as negative samples of the current node, and optimize the embedding representation of each node through comparative learning.
[0087] Specifically, the user-item interaction graph can be divided into multiple block graphs, such as Figure 3 The figure shows a block diagram comparative learning diagram in one embodiment of the present invention. For any node i, its directly connected nodes can be divided into a set and collection Respectively in the set and located in the set The directly connected nodes in the cluster are aggregated to obtain the embedding representation e of node i i and the positive sample e' of node i i , randomly select multiple nodes from nodes other than directly connected nodes and use them as negative samples of node i, and optimize the embedding representation e of each node i through contrastive learning i .
[0088] The above process can be decomposed into the following operation steps S21 to S23.
[0089] Step S21: For a target node i, its directly connected nodes are divided into two sets and Then, the neighbor nodes of the two sets are aggregated to obtain the embedded representation e of node i i and positive sample e' i .
[0090] Specifically, for each node, its embedding representation and positive samples are obtained through aggregation.
[0091] For a target node i, its neighbor nodes are divided into two sets and The block graph to which it belongs is divided into two subgraphs, and then the embedding representation e of node i is obtained by aggregating the neighbor nodes of the two subgraphs i and e' i Here i and e' i is obtained from two completely different sets of neighbor sets, but the two neighbor sets are actually neighbors of node i in a block graph, so e i and e' i should be similar, so e' i As e i A positive sample of .
[0092] Step S22: Randomly select multiple nodes from the remaining non-directly connected nodes and use their positive samples as negative samples of the target node i. Here, not every negative sample in each batch is involved in the calculation, but a part of the nodes are randomly selected from the negative sample set to reduce the computational complexity.
[0093] Step S23: Perform contrastive learning using the InfoNCE evaluator.
[0094] The loss function of contrastive learning is:
[0095]
[0096] Where B is the node set in the block graph divided by the user-item interaction graph, and B i is the set of nodes selected as negative samples of node i in the block graph, sim(e i ,e' i ) is the calculation of the embedding representation e i and positive sample e' i The similarity function, sim(e i ,e' j ) is the calculation of the embedding representation e i and negative samples e' j Here, cosine similarity can be used as a measure of vector similarity to avoid the scale effect and zero vector problems caused by direct dot product.
[0097] The optimization method of contrastive learning is to regard the node representation as a random variable and maximize the mutual information between variables. Each node is aggregated to obtain e i and e' i Two expressions, where e' i is the positive sample of the node, and the node represents the positive sample that should be closer to itself, while the nodes other than the neighboring nodes are the targets that should be kept away from, so e' i As positive samples, the positive samples e' of nodes other than neighbor nodes are taken j As its negative sample, and optimization is performed based on the comparison between positive and negative samples. In the present invention, through comparative learning, long-tail nodes can be involved in training to reduce popularity deviation.
[0098] Step S3: Perform graph convolution feature extraction on the user-item interaction graph after contrastive learning to obtain the final graph embedding representation of the node.
[0099] For each node, the features of its neighboring nodes are aggregated through graph convolution to learn the relationship between nodes, thereby achieving feature extraction and classification.
[0100] Furthermore, in order to suppress the overflow effect of popularity nodes on other nodes, in one embodiment, the aggregation weight of graph convolution is designed. For any target node u, when aggregating its neighbor nodes to obtain the representation of the target node u, the weight is used. As the weight of neighbor node i, where d u is the degree of the target node u (that is, the number of neighboring nodes), and the specific aggregation formula is as follows:
[0101]
[0102] Where u is the node to be updated, is the set of neighbor nodes of node u, d u is the degree of node u, k is the number of aggregations, is the embedding representation of node u after k+1 aggregations, d i is the degree of node i, is the embedding representation of node i after k aggregations.
[0103] After performing a preset number of aggregations, the results of each aggregation are weighted summed to obtain the final graph embedding representation of the node.
[0104] From the perspective of information propagation, if a node i can reach node j by passing through several other nodes, that is, there is a path from i to j in the graph, then the representation of node i will have an impact on node j after propagation. If node i can reach node j, then mathematically speaking, the number of paths should be related to the degree of node i. The higher the degree, the greater the expected number of paths. Here, we use the derivative of the BPR loss function to evaluate the influence of one node on another in the graph, and the results are shown below:
[0105]
[0106]
[0107] Here Φ u represents the set of paths of length L-1 from node u to node j, and Φ is the set of paths of length L from node i to node j, and from Φ to Φ u The change will be a degree d of i i , in terms of expectations Therefore, we can extract a d from the formula i , and then divided by the normalized Here L is node i, and we can get the second line of the formula Finally, use δ i Replacing the expected value in the formula can be more concise.
[0108] It can be found that the influence of node i on node j is related to its degree, and the greater the degree, the greater its influence. The larger the value, the larger the value. Therefore, based on this conclusion, the weight of node aggregation can be dynamically adjusted. When the target node u is aggregated with neighboring nodes to obtain the representation, for each neighbor node i, use As its weight, this can effectively suppress the spillover effect of popularity nodes on other nodes. Through the improved aggregator, the gap in the influence of popular nodes and non-popular nodes on other nodes is narrowed, and the trend of other nodes approaching popular nodes is slowed down, which can largely solve the problem of central effect and effectively eliminate the popularity bias in the system.
[0109] Step S4: Fuse the target user’s text embedding representation and the final graph embedding representation to obtain the target user’s final embedding representation.
[0110] Among them, the target user is the user to whom items need to be recommended.
[0111] Specifically, the text embedding representation and the final graph embedding representation of the target user can be concatenated and sent to a multi-layer perceptron for linear transformation and nonlinear activation to obtain the final embedding representation of the target user.
[0112] The main function of feature fusion is to fuse different embedding representations to obtain the final user representation, which is then used to recommend candidate item sets for the target user. For a given user historical interaction sequence, text information is extracted to obtain its text-based text embedding representation T u , and then the graph embedding representation I is obtained by extracting the interaction information from the user-item interaction graph through the graph convolutional network u , concatenate the two vectors:
[0113] X u =Concat(T u ,I u )
[0114] The concatenated vector is input into the multi-layer perceptron, and after linear transformation and nonlinear activation, the final embedding representation of the target user can be obtained: u =MLP(X u ).
[0115] Step S5: Calculate the similarity between the final embedding representation of the target user and the text embedding representation of the item, and determine the recommendation priority of the item based on the similarity. The higher the similarity, the higher the recommendation priority.
[0116] Specifically, the similarity between the target user’s final embedding representation and the item’s text embedding representation can be calculated using the vector inner product formula. The calculation formula is as follows:
[0117]
[0118] Where, P w is the text embedding representation of item w in the candidate set, It is the similarity score between the final embedding representation of the target user and the text embedding representation of the item w, and is also the ranking score in the final recommendation candidate set. The top-k items with the highest scores are selected as the final output and recommended to the user.
[0119] In general, the technical solution of the embodiment of the present invention, by integrating text embedding representation and graph structure embedding representation and introducing the advantages of comparative learning models, comprehensively analyzes the user's interest data, generates a recommendation list of points of interest that meets the user's preferences, removes the popularity bias in the recommendation system, reduces the proportion of popular items in the recommendation results, recommends items that truly meet the user's interests, and optimizes the user's usage experience.
[0120] Example 2
[0121] The present invention also relates to an item debiasing recommendation system, such as Figure 4FIG. 1 is a schematic diagram of the structure of an item debiasing recommendation system according to an embodiment of the present invention, which includes:
[0122] The text feature extraction module is used to obtain the user interaction history sequence and the user-item interaction graph. The user interaction history sequence contains the user's historical interaction items, and the user-item interaction graph is the user-item adjacency matrix. The attribute text representations of all items in the same user interaction history sequence are concatenated and then subjected to feature extraction and fusion to obtain the corresponding user's text embedding representation. The attribute text representations of the same item are subjected to feature extraction and fusion to obtain the corresponding item's text embedding representation. The item attribute text includes popularity-related attributes and non-popularity-related attributes.
[0123] The graph contrastive learning module uses the user's text embedding representation and the item's text embedding representation as the initial node embedding representation of the corresponding node in the user-item interaction graph. It traverses all nodes in the user-item interaction graph, divides the directly connected nodes of the current node into two sets, aggregates the nodes in each set, and uses one of the aggregated results as the embedding representation of the current node, and the other aggregated result as the positive sample of the current node. It randomly selects multiple positive samples of nodes from nodes other than the directly connected nodes as negative samples of the current node, and optimizes the embedding representation of each node through contrastive learning.
[0124] The graph convolution module is used to perform graph convolution feature extraction on the user-item interaction graph after contrastive learning to obtain the final graph embedding representation of the nodes;
[0125] The feature fusion module is used to fuse the text embedding representation of the target user and the final graph embedding representation to obtain the final embedding representation of the target user;
[0126] The output module is used to calculate the similarity between the final embedding representation of the target user and the text embedding representation of the item, and determine the recommendation priority of the item based on the similarity. The higher the similarity, the higher the recommendation priority.
[0127] It can be understood that the system in Example 2 can be used to execute the method in Example 1. For specific details, please refer to the introduction of Example 1 and will not be repeated here.
[0128] Example 3
[0129] The present invention also relates to a computer-readable storage medium having a computer program stored thereon, which implements the steps of the above method when the computer program is executed by a processor.
[0130] Specifically, the memory may include a high-speed random access memory, and may also include a non-volatile memory, such as a hard disk, a memory, a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), at least one disk storage device, a flash memory device, or other volatile solid-state storage device.
[0131] Example 4
[0132] An embodiment of the present invention provides a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the steps of the method of the above embodiment of the present invention.
[0133] The technical features of the above embodiments can be combined in any manner. To simplify the description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification. It should be noted that the phrases "in one embodiment", "for example", "and another example", etc. of the present invention are intended to illustrate the present invention and are not intended to limit the present invention.
[0134] The above embodiments merely illustrate several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the patent application. It should be noted that a person skilled in the art would be able to make numerous variations and improvements without departing from the spirit of the present invention, and all such variations and improvements fall within the scope of protection of the present invention.
Claims
1. A method for debiasing item recommendation, characterized in that: include: Obtain a user interaction history sequence and a user-item interaction graph, wherein the user interaction history sequence includes the user's historical interaction items, and the user-item interaction graph is a user-item adjacency matrix; concatenate the attribute text representations of all items in the same user interaction history sequence, perform feature extraction and fusion, and obtain a text embedding representation of the corresponding user; perform feature extraction and fusion on the attribute text representations of the same item to obtain a text embedding representation of the corresponding item; the attribute text of the item includes popularity-related attributes and non-popularity-related attributes; Using the user's text embedding representation and the item's text embedding representation as initial node embedding representations of corresponding nodes in the user-item interaction graph; Traverse all nodes in the user-item interaction graph, divide the directly connected nodes of the current node into two sets, aggregate the embedding representations of the nodes in each set, use one of the aggregation results as the embedding representation of the current node, and use the other aggregation result as the positive sample of the current node. Randomly select multiple positive samples of nodes from nodes other than the directly connected nodes as negative samples of the current node, and optimize the embedding representation of each node through comparative learning; Perform graph convolution feature extraction on the user-item interaction graph after contrastive learning to obtain the final graph embedding representation of the nodes; The text embedding representation of the target user and the final graph embedding representation are fused to obtain the final embedding representation of the target user; Calculate the similarity between the target user's final embedding representation and the item's text embedding representation, and determine the item's recommendation priority based on the similarity. The higher the similarity, the higher the recommendation priority.
2. The method for debiasing item recommendation according to claim 1, wherein: The obtaining of the user interaction history sequence and the user-item interaction graph includes: After removing items and users whose number of interactions is lower than a preset requirement, the user interaction history sequence and the user-item interaction graph are generated based on the interaction records.
3. The method for debiasing item recommendation according to claim 1, wherein: The feature extraction and fusion process includes: First, use the Transformer encoder to extract features from the input text; The extracted features are then decoded using the Transformer decoder to achieve feature fusion.
4. The method for debiasing item recommendation according to claim 3, wherein: The method of concatenating the attribute text representations of all items in the same user interaction history sequence and then performing feature extraction and fusion to obtain the text embedding representation of the corresponding user includes: Concatenate the attribute text representations of all items in the order of recent to distant item interaction time to form a long text sentence; Splitting the long text sentence into a series of sub-paragraphs; Use the Transformer encoder to extract features from each sub-paragraph one by one to obtain corresponding sub-paragraph features; Splicing the sub-paragraph features in the order of the sub-paragraphs to obtain splicing features of the long text sentence; The concatenated features of the long text sentence are decoded using a Transformer decoder to obtain a text embedding representation of the corresponding user.
5. The method for debiasing item recommendation according to claim 1, wherein: The method of performing graph convolution feature extraction on the user-item interaction graph after contrastive learning to obtain the final graph embedding representation of the nodes includes: The node aggregation calculation of graph convolution is implemented according to the following aggregation formula: Where, is the node to be updated, For nodes The set of neighbor nodes of For nodes The degree, is the number of aggregations, For passing The nodes obtained after the second aggregation The embedding representation of For nodes The degree, For passing The nodes obtained after the second aggregation Embedded representation of After performing a preset number of aggregations, the results of each aggregation are weighted summed to obtain the final graph embedding representation of the node.
6. The method for debiasing item recommendation according to claim 1, wherein: The feature fusion of the target user's text embedding representation and the final graph embedding representation to obtain the target user's final embedding representation includes: The target user's text embedding representation and the final graph embedding representation are concatenated and sent to a multi-layer perceptron for linear transformation and nonlinear activation to obtain the final embedding representation of the target user.
7. The method for debiasing item recommendation according to claim 1, wherein: Calculating the similarity between the final embedding representation of the target user and the text embedding representation of the item includes: The similarity between the final embedding representation of the target user and the text embedding representation of the item is calculated using the vector inner product formula.
8. An item debiasing recommendation system, characterized in that: include: The text feature extraction module is used to obtain a user interaction history sequence and a user-item interaction graph. The user interaction history sequence includes the user's historical interaction items, and the user-item interaction graph is a user-item adjacency matrix. The module concatenates the attribute text representations of all items in the same user interaction history sequence, performs feature extraction and fusion, and obtains a text embedding representation of the corresponding user. The module also performs feature extraction and fusion on the attribute text representations of the same item to obtain a text embedding representation of the corresponding item. The attribute text of the item includes popularity-related attributes and non-popularity-related attributes. A graph comparison learning module is configured to use the user's text embedding representation and the item's text embedding representation as initial node embedding representations of corresponding nodes in the user-item interaction graph; Traverse all nodes in the user-item interaction graph, divide the directly connected nodes of the current node into two sets, aggregate the embedding representations of the nodes in each set, use one of the aggregation results as the embedding representation of the current node, and use the other aggregation result as the positive sample of the current node. Randomly select multiple positive samples of nodes from nodes other than the directly connected nodes as negative samples of the current node, and optimize the embedding representation of each node through comparative learning; The graph convolution module is used to perform graph convolution feature extraction on the user-item interaction graph after contrastive learning to obtain the final graph embedding representation of the nodes; The feature fusion module is used to fuse the text embedding representation of the target user and the final graph embedding representation to obtain the final embedding representation of the target user; The output module is used to calculate the similarity between the final embedding representation of the target user and the text embedding representation of the item, and determine the recommendation priority of the item based on the similarity. The higher the similarity, the higher the recommendation priority.
9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.
10. A computer program product comprising a computer program or instructions, characterized in that When the computer program or instruction is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.