Multi-behavior recommendation method and system based on noise reduction and prompt adjustment
The multi-behavior recommendation framework based on noise reduction and cue adjustment, constructed using a three-stage learning method, solves the problems of noisy interaction of auxiliary behaviors and semantic gap, and achieves efficient improvement in the accuracy of multi-behavior recommendation.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- HARBIN ENG UNIV
- Filing Date
- 2023-05-25
- Publication Date
- 2026-07-07
AI Technical Summary
In existing multi-behavior recommendation methods, the noisy interactions of auxiliary behaviors and the semantic gap between multiple behaviors lead to a decrease in recommendation accuracy, and the lack of supervision signals labeled with noisy interactions makes it difficult to solve the problem.
We adopt a multi-behavior recommendation method based on noise reduction and cue adjustment. We construct a recommendation framework through a three-stage learning approach, including a pattern-enhanced graph encoder, a behavior-aware noise reduction module, and a prediction layer. We use a data-driven approach to learn complex patterns and reduce the impact of noise, thereby alleviating the semantic gap and improving recommendation performance.
It requires no additional supervision signals, effectively reduces noise and mitigates behavioral semantic gaps, and significantly improves the accuracy and performance of multi-behavior recommendation.
Smart Images
Figure CN116644282B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of multi-behavior recommendation technology, and in particular to a multi-behavior recommendation method, system, device, and medium based on noise reduction and cue adjustment. Background Technology
[0002] In multi-behavior recommendation, user interactions with items can take many forms (e.g., clicks, adding to cart, and purchases). Typically, auxiliary behaviors (i.e., clicks and adding to cart) have more interaction records, while the target behavior (i.e., purchases) has relatively fewer. Therefore, a large number of auxiliary behaviors can introduce irrelevant information into the recommendation model, reducing the model's accuracy in recommending users based on their target behavior. Specifically, a large number of auxiliary behaviors presents two key challenges:
[0003] 1. Noisy Interactions in Auxiliary Behaviors: In real-world recommendation scenarios, some user interactions with items in auxiliary behaviors are noisy, such as accidental interactions. These noisy interactions cannot accurately reflect user interests, thus negatively impacting the multi-behavioral knowledge learned by the model. Therefore, when transferring learned knowledge to target behavior recommendations, the greater the proportion of auxiliary behaviors in the dataset, the greater the negative impact of noise, thereby affecting recommendation accuracy. However, in real-world scenarios, there is often a lack of supervised signals annotating noisy interactions, making this challenge difficult to solve.
[0004] 2. Semantic Gap Between Multiple Behaviors: Although user interactions overlap across multiple behaviors, the semantic features of the target behavior still differ from those of auxiliary behaviors. For example, in e-commerce scenarios, a large number of clicks do not ultimately translate into purchases. Therefore, knowledge learned from numerous auxiliary behaviors will be excessively compressed into the semantic space of those auxiliary behaviors, leading to the inherent challenge of fitting the semantic gap between multiple behaviors. Specifically, how can we effectively extract the target behavior-specific semantics from such knowledge and transfer this semantics to target behavior recommendations without compromising the informative content of the knowledge? Summary of the Invention
[0005] This invention provides a multi-behavior recommendation method and system based on noise reduction and cue adjustment. Existing multi-behavior recommendation methods may introduce irrelevant information into the recommendation model, reducing the accuracy of the model in recommending user target behaviors.
[0006] One embodiment of the present invention provides a multi-behavior recommendation method based on noise reduction and cue adjustment, including:
[0007] Step S1: Obtain multi-behavioral interaction data between users and projects to construct a dataset, and preprocess the dataset to divide it into a training set and a test set;
[0008] Step S2: Construct a multi-behavior recommendation framework based on noise reduction and cue adjustment, and train the multi-behavior recommendation framework based on noise reduction and cue adjustment using the training set. The multi-behavior recommendation framework based on noise reduction and cue adjustment includes an embedding layer, a pattern enhancement graph encoder, a behavior-aware noise reduction module, and a prediction layer.
[0009] Step S3: Input the target user into the trained multi-behavior recommendation framework based on noise reduction and prompt adjustment, calculate the recommendation score of the item to be recommended relative to the target user based on its prediction layer, and recommend the item to the target user based on the recommendation score.
[0010] Another embodiment of the present invention provides a multi-behavior recommendation system based on noise reduction and cue adjustment, comprising:
[0011] The preprocessing module is used to acquire multi-behavioral interaction data between users and projects to construct a dataset, and to preprocess the dataset into a training set and a test set.
[0012] The training module is used to construct a multi-behavior recommendation framework based on noise reduction and cue adjustment, and to train the multi-behavior recommendation framework based on noise reduction and cue adjustment using the training set. The multi-behavior recommendation framework based on noise reduction and cue adjustment includes an embedding layer, a pattern enhancement graph encoder, a behavior-aware noise reduction module, and a prediction layer.
[0013] The recommendation module is used to input the target user into the trained multi-behavior recommendation framework based on noise reduction and prompt adjustment, calculate the recommendation score of the item to be recommended relative to the target user based on its prediction layer, and recommend the item to the target user based on the recommendation score.
[0014] The technical solution of the present invention achieves at least the following beneficial technical effects:
[0015] (1) A three-stage learning method is used. In the first stage, a pattern-enhanced graph encoder is used to learn complex patterns in a data-driven manner to guide the behavior-aware denoising module to identify inherent noise in auxiliary behaviors and generate denoised user item multi-behavior interaction data for subsequent stages. In the second and third stages, a lightweight fine-tuning method and a continuous deep cue adjustment method are used to effectively reduce the influence of noise in auxiliary behaviors and alleviate the semantic gap between behaviors. This method does not require any additional supervision signals (labeling noise data) and adjusts or adds a small number of learnable parameters to effectively denoise auxiliary behaviors and alleviate the semantic gap between behaviors to improve the performance of multi-behavior recommendation.
[0016] (2) Compared with the prior art, the performance of multi-behavior recommendation is improved by denoising the auxiliary behavior and bridging the semantic gap between the auxiliary behavior and the target behavior.
[0017] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description
[0018] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein:
[0019] Figure 1 This is a flowchart of a multi-behavior recommendation method based on noise reduction and cue adjustment according to an embodiment of the present invention;
[0020] Figure 2 This is a multi-behavior recommendation framework diagram based on noise reduction and cue adjustment according to an embodiment of the present invention. (a) is the constructed relationship diagram, (b) is the first stage of the multi-behavior recommendation framework diagram based on noise reduction and cue adjustment, (c) is the second stage of the multi-behavior recommendation framework diagram based on noise reduction and cue adjustment, and (d) is the third stage of the multi-behavior recommendation framework diagram based on noise reduction and cue adjustment.
[0021] Figure 3 This is a schematic diagram of the structure of a multi-behavior recommendation system based on noise reduction and prompt adjustment according to an embodiment of the present invention. Detailed Implementation
[0022] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.
[0023] The following describes, with reference to the accompanying drawings, a multi-behavior recommendation method and system based on noise reduction and cue adjustment proposed according to an embodiment of the present invention. First, the multi-behavior recommendation method based on noise reduction and cue adjustment proposed according to an embodiment of the present invention will be described with reference to the accompanying drawings.
[0024] Figure 1 This is a flowchart of a multi-behavior recommendation method based on noise reduction and cue adjustment according to an embodiment of the present invention.
[0025] like Figure 1 As shown, this multi-behavior recommendation method based on noise reduction and cue adjustment includes the following steps:
[0026] In step S1, multi-behavioral interaction data between users and projects is acquired to construct a dataset, and the dataset is preprocessed and divided into a training set and a test set.
[0027] Furthermore, in one embodiment of the present invention, step S1 specifically includes:
[0028] Step S101: Obtain multi-behavioral interaction data between users and projects to extract user ID, user multi-behavioral interaction records and project ID information, and construct a dataset;
[0029] Step S102: Filter the data to include users with low engagement and less popular items;
[0030] Step S103: Divide the filtered dataset into training set and test set according to time.
[0031] Specifically, user IDs, user multi-behavior interaction records, and project IDs are first extracted from user project multi-behavior interactions to construct a dataset. Users with too few interactions and overly unpopular projects are filtered from the dataset, such as users with fewer than 2 interactive projects and projects with fewer than 5 interactions. The filtered user multi-behavior interaction data is then divided into training and test sets according to time. For example, the most recent 20% of user interactions under the target behavior are divided into the test set, and the remaining historical data are divided into the training set.
[0032] In step S2, a multi-behavior recommendation framework based on noise reduction and cue adjustment is constructed, and the multi-behavior recommendation framework based on noise reduction and cue adjustment is trained using a training set. The multi-behavior recommendation framework based on noise reduction and cue adjustment includes an embedding layer, a pattern enhancement graph encoder, a behavior-aware noise reduction module, and a prediction layer.
[0033] Furthermore, step S2 in this embodiment of the invention specifically includes:
[0034] Step S201: Construct a user-item multi-behavior graph, an item relationship graph, and a user relationship graph based on the training set.
[0035] In other words, based on the multi-behavioral interaction data of all users and items in the training set, a user-item multi-behavioral graph, an item relationship graph, and a user relationship graph are constructed.
[0036] like Figure 2 As shown in (a), the multi-behavioral interaction data between the user and the project is denoted as... in U = {u1, u2, ..., u} |U| User set, I = {i1, i2, ..., i} |I|} represents a set of items, b a b indicates auxiliary behavior t Indicates the target behavior. A * Any element in represents user u. no Interact with item i under behavior type *, where *∈{a,t}.
[0037] User project multi-behavior diagram notation is as follows: For any behavior graph edge set ε * There exists a weighted undirected edge between a user and a project if and only if in And *∈{a,t}.
[0038] Project relationship diagram is denoted as There exists a weighted directed edge of behavior type * that points from item i to item j. If and only if Cnt(i→j,*)>0, where cnt(,) is a counter. weight Calculations based on the project's order of precedence, such as the Jaccard similarity coefficient. Indicates the strength of the sequential relationship between items.
[0039] The user relationship diagram is denoted as... There exists a weighted undirected edge between user u and user v under behavior type *. If and only if and side weight Calculated based on user similarity, such as the Jaccard similarity coefficient. This indicates the strength of the similarity relationship between users.
[0040] In step S202, the embedding layer, pattern enhancement graph encoder, behavior-aware denoising module, and prediction layer are constructed sequentially to obtain a multi-behavior recommendation framework based on denoising and cue adjustment.
[0041] Specifically, the embedding layer is constructed first: based on the initial embedding table, the one-hot encoding of the ID is mapped to a dense vector using learnable parameters.
[0042]
[0043] in, x is a learnable parameter u ,x i , User u, project i, and auxiliary behavior b are respectively. a Target behavior b t The one-hot encoded vector.
[0044] Reconstruct the pattern enhancement graph encoder, which includes a user relationship aggregation layer, a user-item interaction aggregation layer, and an item relationship aggregation layer.
[0045] The user relationship aggregation layer is used to learn the behavioral patterns of similar users and generate specific encoding vectors for the layer:
[0046]
[0047] in, The aggregation function encodes user relationships, where l∈R is the level. For the image middle u * Neighbors This is a set of learnable parameters. For example, it can be implemented through convolution operations.
[0048]
[0049] Among them, Conv U,(l) (·) represents a convolution kernel with a stride of 1 and a size of 2×1. for Normalized weights, || for join operation, This is the set of learnable parameters for this convolution operation.
[0050] The project relationship aggregation layer is used to learn the sequential patterns of projects and generate layer-specific encoding vectors:
[0051]
[0052] in, An aggregate function for encoding project relationships. For the image in i * The neighboring country on the border, For the image in i * The neighbor beyond the border, This is a set of learnable parameters. For example, it can be implemented using attention mechanisms and convolution operations.
[0053]
[0054]
[0055]
[0056] Among them, Conv I,(l) (·) is related to Conv U,(l) (·) Convolutional kernels with the same shape but different learnable parameters for Normalized weights of the middle edge, It is a set of learnable parameters, including the attention matrix. With convolutional filters
[0057] The user-item interaction aggregation layer is used to learn multiple behavioral patterns between users and items, and generate specific encoding vectors for the layer:
[0058]
[0059] in An aggregate function that links data from projects to users. This refers to aggregate functions that route data from users to projects. For example, this can be implemented using LightGCN. and
[0060]
[0061] Next, a behavior-aware denoising module is constructed. This module uses a behavior-aware graph decoder as a discriminator to perform the information restoration task. The behavior-aware graph decoder takes behavior embeddings, behavior-aware user representations, and item representations as inputs and outputs a parameterized user-item multi-behavior graph.
[0062]
[0063] in, For the parameterized graph under behavior*, For the user's embedded collection,
[0064] The project's embedded collection.
[0065] Finally, a prediction layer is constructed, and the embedding layer, pattern enhancement graph encoder, behavior-aware denoising module, and prediction layer are combined to build a multi-behavior recommendation framework based on denoising and cue adjustment.
[0066] Step S203: Construct the first stage of a multi-behavior recommendation framework based on noise reduction and cue adjustment, and obtain a noise-reduced user item multi-behavior graph.
[0067] Furthermore, in one embodiment of the present invention, step S203 specifically includes:
[0068] In the embedding layer, obtain the embeddings of the first user, the first item, the first assistive behavior, and the first target behavior;
[0069] The embedding of the first user, the embedding of the first item, the user-item multi-behavior graph, the item relationship graph, and the user relationship graph are respectively input into the pattern enhancement graph encoder to obtain the first behavior-aware user representation and item representation, and the first multi-behavior user representation and item representation.
[0070] The first behavior-aware user representation and item representation are input into the behavior-aware noise reduction module to obtain the first parameterized user-item multi-behavior graph;
[0071] Based on the first parameterized user item multi-behavior graph, the edge weights of the auxiliary behaviors in the user item multi-behavior graph are binarized, and the edge weights of the target behaviors in the user item multi-behavior graph are reset to obtain a denoised user item multi-behavior graph.
[0072] Specifically, such as Figure 2 As shown in (b), the user and item embeddings and the three types of graphs described in step S201 are respectively input into the three aggregation layers of the pattern enhancement graph encoder to obtain the first behavior-aware user representation and item representation, as well as the first multi-behavior user representation and item representation.
[0073] It should be noted that the behavior-aware user representation, item representation, and multi-behavior user representation, as well as the item representation, are fused from different encodings obtained from each of the three aggregation layers of the pattern-enhanced graph encoder. For example, using a gating network:
[0074]
[0075]
[0076] Where σ(·) is the activation function, such as the Sigmoid activation function. As weight, These are learnable parameters. Based on this, the multi-behavior embeddings of the (l+1)th layer are obtained:
[0077]
[0078] By concatenating the user embeddings and item embeddings at each layer, we obtain the final behavior-aware user embeddings and item embeddings, and multi-behavior user embeddings and item embeddings:
[0079]
[0080]
[0081] Among them, f U and f I Neural networks with different learnable parameters, such as feedforward neural networks activated by the ReLU function, have different learnable parameters.
[0082] The behavior-aware user representation and the item representation are input into the behavior-aware noise reduction module to obtain the first parameterized user-item multi-behavior graph.
[0083] Based on the first parameterized user item multi-behavior graph, the edge weights of the auxiliary behaviors in the user item multi-behavior graph are binarized, and the edge weights of the target behaviors in the graph are reset to obtain a denoised user item multi-behavior graph.
[0084] Specifically, binarizing the edge weights of auxiliary behaviors in the user project multi-behavior graph involves converting edge weights between 0 and 1 to either 0 or 1. For example, using a threshold interception method: when... season Otherwise, let Where δ is a hyperparameter, for example, δ = 0.2.
[0085] Among them, resetting the edge weights of the target behavior in the user project multi-behavior graph is: resetting the edge weights of the target behavior in the parameterized graph to the edge weights of the target behavior in the user project multi-behavior graph.
[0086] Step S204: Train the first stage using the training set.
[0087] Furthermore, in one embodiment of the present invention, step S204 specifically includes:
[0088] The training set is input into the first stage to obtain the first behavior-aware user representation and item representation, and the first multi-behavior user representation and item representation.
[0089] The second behavior-aware user representation and item representation are input into the behavior-aware noise reduction module to obtain the second parameterized user-item multi-behavior graph.
[0090] The first parameterized user item multi-behavior interaction graph and the user item multi-behavior interaction graph are input into the cross-entropy loss function;
[0091] The first multi-behavior user representation and item representation are input into the Bayesian personalized ranking loss function;
[0092] Minimize the cross-entropy loss function and the Bayesian personalized ranking loss function to jointly train the first stage.
[0093] Specifically, the data in the training set is input into the first stage of the multi-behavior recommendation framework based on noise reduction and cue adjustment to obtain the second behavior-aware user representation and item representation, and the second multi-behavior user representation and item representation.
[0094] The second behavior-aware user representation and item representation are input into the behavior-aware noise reduction module to obtain the second parameterized user-item multi-behavior graph.
[0095] The second parameterized user item multi-behavior interaction graph and the user item multi-behavior interaction graph are input into the cross-entropy loss function;
[0096] The cross-entropy loss function after input is:
[0097]
[0098]
[0099] The second multi-behavioral user representation and item representation are input into the Bayesian personalized ranking loss function;
[0100] The Bayesian personalized ranking loss function after input is:
[0101]
[0102]
[0103] Where sim(·) is the similarity calculation function, such as inner product similarity or neural network, u∈U,i * ∈I, j * For items randomly sampled from I, where
[0104] The first stage of the multi-behavior recommendation framework, based on noise reduction and cue adjustment, involves minimizing the cross-entropy loss function after input and the Bayesian personalized ranking loss function after input.
[0105] Step S205: Construct the second stage of a multi-behavior recommendation framework based on noise reduction and cue adjustment based on the first stage after training.
[0106] Furthermore, in one embodiment of the present invention, step S205 specifically includes:
[0107] Load and freeze all learnable parameters from the first phase after training;
[0108] In the embedding layer, the embeddings of the second user, the second item, the second auxiliary behavior, and the second target behavior are obtained;
[0109] The pruning-mode enhanced graph encoder retains only the user item interaction aggregation layer and reinitializes and updates the learnable parameters of the user item interaction aggregation layer to build the graph encoder;
[0110] The user-item multi-behavior graph, which includes the embedding and denoising of the second user, second item, second auxiliary behavior, and second target behavior, is input into the graph encoder to obtain the second multi-behavior user representation and item representation.
[0111] Specifically, such as Figure 2 As shown in (c), all learnable parameters in the first stage of the trained multi-behavior recommendation framework based on denoising and cue adjustment are loaded and frozen.
[0112] The embeddings of users, items, assistive behaviors, and target behaviors are obtained from the embedding layer;
[0113] Build a graph encoder, and enhance the graph encoder with pruning mode, retaining only the user item interaction aggregation layer, and reinitialize and update the learnable parameters of the user item interaction aggregation layer;
[0114] The user-item multi-behavior graph, which embeds and denoises the user, item, auxiliary behavior, and target behavior, is input into the graph encoder to obtain the third multi-behavior user representation and item representation.
[0115] Step S206: Train the second stage using the training set.
[0116] Furthermore, in one embodiment of the present invention, step S206 specifically includes:
[0117] The training set is input into the second stage to obtain the second multi-behavioral user representation and item representation;
[0118] The second multi-behavior user representation and item representation are input into the Bayesian personalized ranking loss function, and the input Bayesian personalized ranking loss function is minimized to train the second stage.
[0119] Specifically, the data in the training set is input into the second stage of the multi-behavior recommendation framework based on noise reduction and cue adjustment to obtain the third multi-behavior user representation and item representation.
[0120] The third multi-behavior user representation and item representation are input into the Bayesian personalized ranking loss function, and the input Bayesian personalized ranking loss function is minimized to train the second stage of the multi-behavior recommendation framework based on noise reduction and cue adjustment.
[0121] The Bayesian personalized ranking loss function after input is:
[0122]
[0123] Where u∈U,i * ,j * ∈I,
[0124] Step S207: Construct the third stage of a multi-behavior recommendation framework based on noise reduction and cue adjustment based on the second stage after training.
[0125] Furthermore, in one embodiment of the present invention, step S207 specifically includes:
[0126] Load and freeze all learnable parameters in the second phase;
[0127] The embeddings of the second user, the second item, the second auxiliary behavior, and the second target behavior are obtained based on the embedding layer;
[0128] The embeddings of the second auxiliary behavior and the second target behavior are fused into a cue embedding, and the learnable parameters of the embedding of the second target behavior are updated.
[0129] The cues are embedded layer by layer and added to the graph encoder built in the second stage;
[0130] The denoised user-item multi-behavior graph, along with the embeddings of the second user and the items, are input into the graph encoder to obtain the third multi-behavior user representation and item representation.
[0131] Specifically, such as Figure 2 As shown in (d), all learnable parameters in the second stage of the trained multi-behavior recommendation framework based on denoising and cueing adjustment are loaded and frozen;
[0132] Based on the embedding layer, obtain the embeddings of users, projects, assistive behaviors, and target behaviors;
[0133] The embeddings of auxiliary behaviors and target behaviors are fused into a cue embedding, and the learnable parameters of the target behavior embedding are updated.
[0134] The fusion of auxiliary behavior embeddings and target behavior embeddings into cue embedding methods is a fusion method without additional learnable parameters, such as pooling operations.
[0135]
[0136] The cue embeddings are added layer by layer to the graph encoder. The process of adding cue embeddings layer by layer to the graph encoder is as follows:
[0137]
[0138]
[0139] Methods for embedding prompts layer by layer into a graph encoder include direct addition, layer-by-layer stitching, and projection blending. For example, direct addition:
[0140]
[0141] The denoised user-item multi-behavior graph and user embedding and item embedding are input into the graph encoder to obtain the fourth multi-behavior user representation and item representation.
[0142] Step S208: Train the third stage using the training set.
[0143] Furthermore, in one embodiment of the present invention, step S208 specifically includes:
[0144] The training set is input into the third stage to obtain the third multi-behavioral user representation and item representation;
[0145] The third multi-behavioral user representation and item representation are input into the Bayesian personalized ranking loss function of the user under the target behavior, and the loss function is minimized to train the third stage.
[0146] Specifically, the data in the training set is input into the third stage of the multi-behavior recommendation framework based on noise reduction and cue adjustment to obtain the sixth multi-behavior user representation and item representation.
[0147] The third stage of training a multi-behavior recommendation framework based on denoising and cue adjustment is to minimize the Bayesian personalized ranking loss function for users under target behaviors.
[0148] The current Bayesian personalized ranking loss function is:
[0149]
[0150] In step S3, the target user is input into the trained multi-behavior recommendation framework based on noise reduction and cue adjustment. The recommendation score of the item to be recommended relative to the target user is calculated based on its prediction layer. The item is then recommended to the target user based on the recommendation score.
[0151] Specifically, the target user is input into the trained multi-behavior recommendation framework based on noise reduction and cue adjustment. The recommendation score of the item to be recommended relative to the target user is calculated according to its prediction layer. The recommendation scores of the items are sorted, and the top K items with the highest scores are recommended to the user.
[0152] The multi-behavior recommendation method based on denoising and cue adjustment proposed in this invention constructs a denoising and cue adjustment framework through a three-stage learning method. In the first stage, user-item multi-behavior graphs, item relationship graphs, and user relationship graphs are constructed using multi-behavior interaction data of users and items in the training set. A pattern-enhanced graph encoder is then used to learn behavior-aware user and item representations. Finally, the user and item representations are input into a denoising module, which outputs a denoised user-item multi-behavior graph for subsequent stages of the denoising and cue adjustment framework. In the second and third stages, two lightweight fine-tuning methods based on pruning and cue adjustment are used respectively to further denoise and bridge the semantic gap between behaviors, thereby greatly improving the recommendation performance under the target behavior. At the same time, no additional supervision signals (labeling noisy data) are required, and a small number of learnable parameters are adjusted or added to efficiently denoise auxiliary behaviors and alleviate the semantic gap between behaviors to improve multi-behavior recommendation performance.
[0153] Next, referring to the accompanying drawings, a multi-behavior recommendation system based on noise reduction and prompt adjustment according to an embodiment of the present invention is described.
[0154] Figure 3 This is a schematic diagram of the structure of a multi-behavior recommendation system based on noise reduction and prompt adjustment according to an embodiment of the present invention.
[0155] like Figure 3 As shown, the system 10 includes: a preprocessing module 100, a training module 200, and a recommendation module 300.
[0156] The preprocessing module 100 is used to acquire multi-behavioral interaction data between users and projects to construct a dataset, and to preprocess the dataset into a training set and a test set.
[0157] In one embodiment of the present invention, the preprocessing module 100 has the following functions:
[0158] Acquire multi-behavioral interaction data between users and projects to extract user IDs, user multi-behavioral interaction records, and project ID information, and construct a dataset;
[0159] Filter the dataset to include users with low engagement and less popular items;
[0160] The filtered dataset is divided into training and test sets according to time.
[0161] The training module 200 is used to construct a multi-behavior recommendation framework based on denoising and cue adjustment, and to train the multi-behavior recommendation framework based on denoising and cue adjustment using the training set. The multi-behavior recommendation framework based on denoising and cue adjustment includes an embedding layer, a pattern enhancement graph encoder, a behavior-aware denoising module, and a prediction layer.
[0162] In one embodiment of the present invention, the training module 200 is specifically used for:
[0163] Construct a user-item multi-behavior graph, an item relationship graph, and a user relationship graph based on the training set;
[0164] An embedding layer, a pattern enhancement graph encoder, a behavior-aware denoising module, and a prediction layer are constructed sequentially to obtain a multi-behavior recommendation framework based on denoising and cue adjustment. The pattern enhancement graph encoder includes a user relationship aggregation layer, a user item interaction aggregation layer, and an item relationship aggregation layer.
[0165] The first stage of constructing a multi-behavior recommendation framework based on noise reduction and cue adjustment, and obtaining a noise-reduced user item multi-behavior graph;
[0166] The first stage is trained using the training set;
[0167] The second phase builds a multi-behavior recommendation framework based on noise reduction and cue adjustment based on the first phase after training.
[0168] The second phase is trained using the training set;
[0169] The third stage builds a multi-behavior recommendation framework based on noise reduction and cue adjustment based on the second stage after training;
[0170] The third stage is trained using the training set.
[0171] The recommendation module 300 is used to input the target user into the trained multi-behavior recommendation framework based on noise reduction and prompt adjustment, calculate the recommendation score of the item to be recommended relative to the target user based on its prediction layer, and recommend the item to the target user based on the recommendation score.
[0172] It should be noted that the foregoing explanation of the multi-behavior recommendation method based on noise reduction and cue adjustment also applies to the multi-behavior recommendation system based on noise reduction and cue adjustment in this embodiment, and will not be repeated here.
[0173] The multi-behavior recommendation system based on denoising and cue adjustment proposed in this embodiment of the invention constructs a denoising and cue adjustment framework through a three-stage learning method. In the first stage, user-item multi-behavior graphs, item relationship graphs, and user relationship graphs are constructed using multi-behavior interaction data of users and items in the training set. Then, a pattern-enhanced graph encoder is used to learn behavior-aware user and item representations. Finally, the user and item representations are input into a denoising module and the denoised user-item multi-behavior graph is output for subsequent stages of the denoising and cue adjustment framework. In the second and third stages, two lightweight fine-tuning methods based on pruning and cue adjustment are used respectively to further denoise and bridge the semantic gap between behaviors, thereby greatly improving the recommendation performance under the target behavior. At the same time, no additional supervision signals (labeling noisy data) are required, and a small number of learnable parameters are adjusted or added to efficiently denoise auxiliary behaviors and alleviate the semantic gap between behaviors to improve multi-behavior recommendation performance.
[0174] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.
[0175] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.
[0176] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.
Claims
1. A multi-behavior recommendation method based on noise reduction and cue adjustment, characterized in that, Includes the following steps: Step S1: Obtain multi-behavioral interaction data between users and projects to construct a dataset, and preprocess the dataset to divide it into a training set and a test set; Step S2: Construct a multi-behavior recommendation framework based on noise reduction and cue adjustment, and train the multi-behavior recommendation framework based on noise reduction and cue adjustment using the training set. The multi-behavior recommendation framework based on noise reduction and cue adjustment includes an embedding layer, a pattern enhancement graph encoder, a behavior-aware noise reduction module, and a prediction layer. Step S201: Construct a user item multi-behavior graph, an item relationship graph, and a user relationship graph based on the training set; Step S202: Construct the embedding layer, the pattern enhancement graph encoder, the behavior-aware denoising module, and the prediction layer in sequence to obtain the multi-behavior recommendation framework based on denoising and prompt adjustment. The pattern enhancement graph encoder includes a user relationship aggregation layer, a user item interaction aggregation layer, and an item relationship aggregation layer. Step S203: Construct the first stage of the multi-behavior recommendation framework based on noise reduction and cue adjustment, and obtain a noise-reduced user item multi-behavior graph; The embeddings of the first user, the first item, the first auxiliary behavior, and the first target behavior are obtained in the embedding layer; The embedding of the first user, the embedding of the first item, the user-item multi-behavior graph, the item relationship graph, and the user relationship graph are respectively input into the pattern enhancement graph encoder to obtain the first behavior-aware user representation and item representation, and the first multi-behavior user representation and item representation. The first behavior-aware user representation and item representation are input into the behavior-aware noise reduction module to obtain a first parameterized user-item multi-behavior graph; Based on the first parameterized user item multi-behavior graph, the edge weights of the auxiliary behaviors in the user item multi-behavior graph are binarized, and the edge weights of the target behaviors in the user item multi-behavior graph are reset to obtain the denoised user item multi-behavior graph. Step S204: Train the first stage using the training set; Step S205: Construct the second stage of the multi-behavior recommendation framework based on noise reduction and cue adjustment according to the first stage after training; Load and freeze all learnable parameters from the first phase after training; The embedding of the second user, the second item, the second auxiliary behavior, and the second target behavior is obtained in the embedding layer; The pattern-enhanced graph encoder is pruned to retain only the user item interaction aggregation layer, and the learnable parameters of the user item interaction aggregation layer are reinitialized and updated to construct the graph encoder. The embeddings of the second user, the second item, the second auxiliary behavior, and the second target behavior, along with the denoised user-item multi-behavior graph, are input into the graph encoder to obtain the second multi-behavior user representation and item representation. Step S206: Train the second stage using the training set; Step S207: Construct the third stage of the multi-behavior recommendation framework based on noise reduction and cue adjustment according to the second stage after training; Load and freeze all learnable parameters from the second phase; The embeddings of the second user, the second item, the second auxiliary behavior, and the second target behavior are obtained based on the embedding layer; The embeddings of the second auxiliary behavior and the second target behavior are fused into a cue embedding, and the learnable parameters of the embedding of the second target behavior are updated. The cues are embedded layer by layer and added to the graph encoder built in the second stage; The denoised user item multi-behavior graph, along with the embedding of the second user and the embedding of the item, are input into the graph encoder to obtain a third multi-behavior user representation and item representation. Step S208: Train the third stage using the training set; Step S3: Input the target user into the trained multi-behavior recommendation framework based on noise reduction and prompt adjustment, calculate the recommendation score of the item to be recommended relative to the target user based on its prediction layer, and recommend the item to the target user based on the recommendation score.
2. The multi-behavior recommendation method based on noise reduction and cue adjustment according to claim 1, characterized in that, Step S1 specifically includes: Step S101: Obtain the multi-behavior interaction data between the user and the project to extract the user ID, the user's multi-behavior interaction records and the project ID information, and construct the dataset; Step S102: Filter the dataset to include users with low interaction and unpopular items; Step S103: Divide the filtered dataset into the training set and the test set according to time.
3. The multi-behavior recommendation method based on noise reduction and cue adjustment according to claim 1, characterized in that, Step S204 specifically includes: The training set is input into the first stage to obtain the first behavior-aware user representation and item representation, and the first multi-behavior user representation and item representation. The first behavior-aware user representation and item representation are input into the behavior-aware noise reduction module to obtain a first parameterized user-item multi-behavior graph; The first parameterized user item multi-behavior interaction graph and the user item multi-behavior interaction graph are input into the cross-entropy loss function; The first multi-behavior user representation and item representation are input into the Bayesian personalized ranking loss function; The first stage is jointly trained by minimizing the cross-entropy loss function after input and the Bayesian personalized ranking loss function after input.
4. The multi-behavior recommendation method based on noise reduction and cue adjustment according to claim 1, characterized in that, Step S206 specifically includes: The training set is input into the second stage to obtain the second multi-behavior user representation and item representation; The second multi-behavior user representation and item representation are input into the Bayesian personalized ranking loss function, and the input Bayesian personalized ranking loss function is minimized to train the second stage.
5. The multi-behavior recommendation method based on noise reduction and cue adjustment according to claim 1, characterized in that, Step S208 specifically includes: The training set is input into the third stage to obtain the third multi-behavioral user representation and item representation; The third multi-behavioral user representation and item representation are input into the Bayesian personalized ranking loss function of the user under the target behavior, and the loss function is minimized to train the third stage.
6. A multi-behavior recommendation system based on noise reduction and cue adjustment, characterized in that, The system is used to execute the multi-behavior recommendation method based on noise reduction and cue adjustment as described in any one of claims 1 to 5, including: The preprocessing module is used to acquire multi-behavioral interaction data between users and projects to construct a dataset, and to preprocess the dataset into a training set and a test set. The training module is used to construct a multi-behavior recommendation framework based on noise reduction and cue adjustment, and to train the multi-behavior recommendation framework based on noise reduction and cue adjustment using the training set. The multi-behavior recommendation framework based on noise reduction and cue adjustment includes an embedding layer, a pattern enhancement graph encoder, a behavior-aware noise reduction module, and a prediction layer. Construct a user-item multi-behavior graph, an item relationship graph, and a user relationship graph based on the training set; The embedding layer, the pattern enhancement graph encoder, the behavior-aware denoising module, and the prediction layer are constructed sequentially to obtain the multi-behavior recommendation framework based on denoising and cue adjustment. The pattern enhancement graph encoder includes a user relationship aggregation layer, a user item interaction aggregation layer, and an item relationship aggregation layer. The first stage of constructing the multi-behavior recommendation framework based on noise reduction and cue adjustment is carried out, and a noise-reduced user item multi-behavior graph is obtained. The embeddings of the first user, the first item, the first auxiliary behavior, and the first target behavior are obtained in the embedding layer; The embedding of the first user, the embedding of the first item, the user-item multi-behavior graph, the item relationship graph, and the user relationship graph are respectively input into the pattern enhancement graph encoder to obtain the first behavior-aware user representation and item representation, and the first multi-behavior user representation and item representation. The first behavior-aware user representation and item representation are input into the behavior-aware noise reduction module to obtain a first parameterized user-item multi-behavior graph; Based on the first parameterized user item multi-behavior graph, the edge weights of the auxiliary behaviors in the user item multi-behavior graph are binarized, and the edge weights of the target behaviors in the user item multi-behavior graph are reset to obtain the denoised user item multi-behavior graph. The first stage is trained using the training set; The second stage of constructing the multi-behavior recommendation framework based on noise reduction and cue adjustment is based on the first stage after training. Load and freeze all learnable parameters from the first phase after training; The embedding of the second user, the second item, the second auxiliary behavior, and the second target behavior is obtained in the embedding layer; The pattern-enhanced graph encoder is pruned to retain only the user item interaction aggregation layer, and the learnable parameters of the user item interaction aggregation layer are reinitialized and updated to construct the graph encoder. The embeddings of the second user, the second item, the second auxiliary behavior, and the second target behavior, along with the denoised user-item multi-behavior graph, are input into the graph encoder to obtain the second multi-behavior user representation and item representation. The second stage is trained using the training set; The third stage of constructing the multi-behavior recommendation framework based on noise reduction and cue adjustment is based on the second stage after training; Load and freeze all learnable parameters from the second phase; The embeddings of the second user, the second item, the second auxiliary behavior, and the second target behavior are obtained based on the embedding layer; The embeddings of the second auxiliary behavior and the second target behavior are fused into a cue embedding, and the learnable parameters of the embedding of the second target behavior are updated. The cues are embedded layer by layer and added to the graph encoder built in the second stage; The denoised user item multi-behavior graph, along with the embedding of the second user and the embedding of the item, are input into the graph encoder to obtain a third multi-behavior user representation and item representation. The third stage is trained using the training set; The recommendation module is used to input the target user into the trained multi-behavior recommendation framework based on noise reduction and prompt adjustment, calculate the recommendation score of the item to be recommended relative to the target user based on its prediction layer, and recommend the item to the target user based on the recommendation score.