A group purchase recommendation model based on a multi-task learning framework

Abstract

The present application belongs to the field of artificial intelligence and information retrieval technology, and specifically relates to a group purchase recommendation model based on a multi-task learning framework. The group purchase recommendation model comprises a graph representation learning module, a multi-task learning module, a representation fine-tuning and prediction module; the present application firstly defines the group purchase recommendation task in a formula, and splits the group purchase task into two related sub-tasks; the graph representation learning module divides multiple views through nodes and relation types, learns the feature representation of users and items on the multiple views respectively, and aggregates to obtain the vector representation of each user and node; the multi-task learning module simultaneously learns the two tasks through an expert network and a gating unit, promotes the learning of the two tasks, and adjusts the gating unit to improve the information interaction ability among multiple learning tasks; the representation fine-tuning and prediction module fine-tunes the feature representation through an auxiliary loss function to improve the generalization of the feature representation. The present application can realize efficient and accurate group purchase recommendation.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and information retrieval technology, specifically relating to a group-buying recommendation model based on multi-tasks. Background Technology

[0002] In today's rapidly developing mobile internet era, people's lives are flooded with information. Platforms offering services via the internet are providing an ever-increasing variety and quantity of products and services, from music and videos to news and shopping. How to enable users to quickly find the products they need, how to present suitable products and services to users, and how to attract and retain users are all challenges facing all platforms. Because different people have different preferences, and these preferences change in different scenarios, how to help users discover their preferred products and services more quickly and effectively is a crucial issue that deserves attention. Personalized recommendation systems serve this purpose. They aim to create user profiles by analyzing historical interaction records, capturing user characteristics, modeling user preferences, and recommending matching personalized results based on each user's unique preferences. Personalized recommendation systems can be divided into two categories based on the recommendation object: item recommendation and user recommendation. Item recommendation aims to predict the probability of a user interacting with candidate items, while user recommendation aims to recommend other users to a given user based on user interaction records and similarities.

[0003] Traditional recommender system research often focuses on capturing the preferences of each individual user and recommending products and services that match those preferences. Group recommendation, as a branch of personalized recommender systems, expands the scope of recommendations beyond just capturing the preferences of a single user. It focuses on capturing and aggregating the shared preferences of a group of users when they jointly purchase products or services. For example, when a family is preparing to go to a restaurant or tourist attraction, finding each individual's preferences and identifying their common preferences is crucial. Current research on group recommendation emphasizes how to aggregate the preferences of each user within a group to form a group preference. Common methods include attention-based mechanisms and knowledge graph-based approaches.

[0004] In recent years, group buying has gained more attention, but researchers have conducted relatively little research on group buying recommendations. Many existing studies are based on group recommendations, which, while able to identify and aggregate the preferences of a group of users, do not effectively solve the group buying problem. This is due to the inherent differences between group buying recommendations and general group recommendation problems. First, users in group buying recommendations are distinguished by their identities. The initiator and participants of a group purchase have different identities, so preferences cannot be generalized. Second, the number of users in group buying recommendations is unknown at the time of recommendation. New users continuously join during the group buying process, meaning that the number and type of users in the group are unknown at the time of recommendation, whereas in group recommendations, these details are known at the time of recommendation. Third, group buying recommendations naturally consist of two stages. The first stage is when users discover their preferred items and initiate a group purchase; the second stage is when other users join the group until the purchase is successful. The recommendation objectives in these two stages also differ. Because of these three main differences between group buying recommendations and general group recommendations, group recommendation methods are difficult to apply directly. This invention proposes, for the first time, a formalized definition of the group-buying recommendation problem, breaks it down into two related sub-tasks, and proposes a novel group-buying recommendation model MGBR based on a multi-task learning framework. Summary of the Invention

[0005] The purpose of this invention is to provide a group-buying recommendation model based on a multi-task learning framework, so as to better solve group recommendation tasks in scenarios similar to e-commerce group buying.

[0006] In this invention, the defined group-buying recommendation task is specifically described as follows: For a group-buying scenario involving a group-buying initiator (also known as a user) u, a group-buying participant p, and a group-buying item i, the ultimate goal is to select suitable i and p to maximize the success rate of the group purchase. This yields the group-buying group triplet data.<U,I,G> Let U represent the set of users u, I represent the set of items i, and G = {p1, p2, ..., p...} |G| Let} represent the set of all participants in a group purchase, and the prediction target is the success rate of the group purchase, i.e., the joint probability P(u,i,p1,p2,…,p). |G| ).

[0007] This invention decomposes the aforementioned group-buying recommendation task into two related sub-tasks. The first sub-task, denoted as Task A, calculates the score function s(i|u) for initiating a group purchase, that is, recommending candidate item i given a user u, enabling the user to initiate a group purchase. The second sub-task, denoted as Task B, calculates the score function s(p|u,i) for completing a group purchase, recommending other participating users p given user u and candidate item i, enabling the group purchase to be completed.

[0008] Specifically, the group-buying recommendation model based on a multi-task learning framework provided by this invention, abbreviated as MGBR, has the following architecture diagram: Figure 1 As shown, it includes the following three modules: a graph representation learning module, a multi-task learning module, and a representation fine-tuning and prediction module. First, the group-buying recommendation task (including two sub-tasks) learns vector representations through the graph representation learning module. Then, these vector representations are input into the multi-task learning module for information interaction and extraction. Finally, they are input into the representation fine-tuning and prediction module to obtain the predicted recommendation results. The model is then updated using the loss functions of the two tasks and the auxiliary loss function of representation fine-tuning. Wherein:

[0009] The graph representation learning module constructs three undirected graphs using a multi-view approach. The node types are (u,i), (p,i), and (u,p), representing the initiator view, participant view, and social view, respectively. GCN (Graph Convolutional Neural Network) is used to learn representations on these three views to obtain the corresponding vector representation for each view. This ensures that each of u,i, and p can obtain vector representations from two different views. Finally, the vector representations from different views are aggregated using vector concatenation to obtain the vector representations corresponding to u,i, and p.

[0010] The multi-task learning module consists of multi-layer expert networks and adjustment gating units. Each layer of the multi-task learning module contains a set of expert networks and adjustment gating units. This set of expert networks is divided into three parts, used to learn task A information, task B information, and shared information, respectively, and are denoted as: expert network group A, expert network group B, and expert network group S of the l-th layer. Each expert network is a multi-layer perceptron (a tower structure composed of three fully connected neural networks), which accepts the input of the gating unit of the previous layer or the original input, and outputs a corresponding vector or matrix, which is the result of information extraction. There are three adjustment gating units in each layer, corresponding to the three expert network parts, and are denoted as gating unit A, gating unit B, and gating unit S of the l-th layer.

[0011] Gating A consists of two parts. The first part weights the outputs of expert network groups A and S based on the inputs of expert network groups A and S. The second part weights the outputs of expert network groups A, B, and S based on a combination of two of the three: initiator, participant, and item (u,i,p). The weighted sum of the two parts is the output of gating A.

[0012] Gating B also consists of two parts. The first part weights the outputs of expert network groups A and S based on the inputs of expert network groups B and S. The second part weights the outputs of expert network groups A, B, and S based on a combination of two of the three: initiator, participant, and item (u,i,p). The weighted sum of the two parts is the output of gating B.

[0013] The gated control S has only a single part, and its output is a weighted sum of the outputs of the expert network groups A, B, and S based on the inputs of the expert network groups A, B, and S.

[0014] The fine-tuning and prediction module includes a prediction layer and auxiliary loss functions. The prediction layer consists of two independent multilayer perceptrons (MLPs) in a dual-tower structure. The prediction layer takes the output of the multi-task learning module as input and calculates the score functions on the two sub-tasks as output. Simultaneously, Bayesian Personalized Ranking (BPR) loss functions are designed for tasks A and B. The fine-tuning part constructs two auxiliary loss functions based on the score functions output by the prediction layer: the auxiliary loss function for task A is a listwise loss, calculated by comparing positive samples with negative samples where i and p are replaced. Positive samples refer to the (u, i, p) combinations observed in the group-buying records. The auxiliary loss for task B is the BPR loss, calculated by comparing positive samples with negative samples where i is replaced. The final optimization objective of the model is a weighted sum of the BPR loss for task A, the listwise auxiliary loss for task A, the BPR loss for task B, and the BPR auxiliary loss for task B.

[0015] In this module, the representation processing in the graph representation learning module and the multi-task learning module is fine-tuned through an auxiliary loss function, making the vector representations of users and items more generalizable.

[0016] The following further elaborates on the task definition and specific details of each part of the model of this invention. Lowercase letters represent vectors, and uppercase letters represent matrices.

[0017] (I) Definition of Group Buying Recommendation Task

[0018] In this invention, u, i, and p represent the group-buying initiator, candidate items, and group-buying participants, respectively. The group-buying recommendation process differs inherently from the general group recommendation process.

[0019] In group-buying recommendation scenarios, existing group recommendation models are difficult to apply directly for two main reasons. First, group-buying recommendation can naturally be divided into two processes: the initiation of the group purchase and the completion of the group purchase, while existing group recommendation methods are basically only applicable to the first process. Second, users in group-buying recommendations have dual identities; they can act as both initiators and participants in the group purchase, making the relationship between users and items more complex. This invention decomposes group-buying recommendation into two related sub-tasks and enhances the information interaction between the two sub-tasks through a multi-task learning framework, thereby successfully modeling group-buying recommendation and achieving better group-buying recommendation results.

[0020] Specifically, this invention breaks down group-buying recommendations into two related sub-tasks, defined in a formula as follows:

[0021] Task A: Calculate the score function s(i|u) to recommend an item i for a given initiator u.

[0022] Task B: Calculate the score function s(p|u,i) to recommend participants p for a given initiator u and item i.

[0023] The rationale for splitting the above two related subtasks will be explained below. This invention uses triples.<u,i,G> This represents the observed group-buying group data, where G = {p1, p2, ..., p...} |G| Let} represent the set of all participants in a group purchase. For a recommendation system, it is necessary to predict the success rate of the group purchase after the recommendation, i.e., the joint probability P(u,i,p1,p2,…,p). |G| Since participants in most group-buying scenarios are not very familiar with each other, we can make the assumption that they are independent. Therefore, the joint probability can be decomposed as follows:

[0024] P(u,i,p1,p2,…,p |G| )=P(p1,p2,…,p |G| |u,i)P(i|u)P(u), (1)

[0025] P(u,i,p1,p2,…,p |G| )∝P(p1,p2,…,p |G| |u,i)P(i|u),(2)

[0026] P(p1,p2,…,p |G| |u,i)=P(p1|u,i)P(p2|u,i)…P(p |G| |u,i). (3)

[0027] As shown above, the scoring functions s(i|u) and s(p|u,i) in this invention correspond to P(i|u) and P(p|u,i) respectively, which shows the rationality of splitting the group buying task into two related sub-tasks in this invention.

[0028] (II) Graph Representation Learning Module

[0029] As described above, the objective score functions of this invention are s(i|u) and s(p|u,i), corresponding to three objectives: u, i, and p. This invention employs graph representation learning to obtain their vector representations in a given dimensional space. This invention constructs three undirected graphs, each containing two types of nodes and their corresponding relationships. The node types in these three graphs are (u,i), (p,i), and (u,p), representing the initiator's view, participant's view, and social view, respectively emphasizing three user characteristics: initiator's preference, participant's preference, and shared user preferences. In this module, a GCN (with H layers) is used to perform representation learning from the graphs corresponding to the three views to obtain the vector representations of u, i, and p. UI G PI G UP This represents three images. The uppercase letter combinations of the subscripts indicate the different relationships between the corresponding lowercase letters u, i, and p (this is how it will be represented in the following text), starting from G. UI G PI It is possible to extract user preferences for items, the only difference being the user's identity, while from G... UP The similarity between users can be extracted.

[0030] Specifically, using Representing three views G respectively UI G PI G UP The node representation learned at the l-th layer of GCN is then computed to the vector representation as follows:

[0031]

[0032]

[0033]

[0034] in, To regularize and add self-loops to the adjacency matrix, This is the parameter matrix, where σ represents the Sigmoid function. d represents the vector dimension, and |U| and |I| represent the sizes of sets U and I, respectively.

[0035] After obtaining vector representations through graph representation learning, each of u, i, p can obtain vector representations from two different views. The two representations are then aggregated by vector concatenation to obtain the vector representations corresponding to u, i, p.

[0036] (III) Multi-task learning module

[0037] As mentioned above, the two subtasks are related, so it's natural to consider using a multi-task learning framework to enable them to mutually reinforce each other. In Task A, when user u selects item i to initiate a group purchase, it's not only based on their personal preference but also on whether the item will be favored by a wide range of potential participants. Therefore, the learning and training of Task A requires the intervention and interference of participant information from Task B. Similarly, the learning and training process of Task B also requires information about the initiator and the item from Task A to provide Task B with more user preference information. Therefore, the information in the vector representation should be fully exchanged and shared between Task A and Task B to achieve better learning results. This invention uses a multi-task learning framework based on expert networks. Expert network learning is divided into three parts: learning information from Task A, learning information from Task B, and learning shared information. Each expert network is associated with a corresponding gating unit to complete aggregation and output.

[0038] Specifically, in this invention, A, B, and S are used to represent different expert network sub-modules, such as... This represents the output vector of the i-th A expert network at layer l. This represents the output vector of the i-th B-expert network at layer l. Let represent the output vector representation of the i-th S expert network at layer l; each expert network learns a new vector representation based on the output of the previous layer, and the calculation process is as follows.

[0039]

[0040]

[0041]

[0042] Taking equation (7) as an example, where, This represents the output of gate A at level l-1. This is the learnable parameter matrix; the other symbols are similar.

[0043] After each expert network learns a new vector representation, the multi-task learning framework selects and aggregates the vector representations of all expert networks through gating units to generate a new output vector. The calculation process is as follows:

[0044]

[0045]

[0046]

[0047]

[0048]

[0049]

[0050] Where W is a learnable parameter matrix, For the output of each expert network in the layer above layer l, This represents the input representation of expert networks A, B, and S at layer l. The expert network obtains a new representation at layer l by utilizing information processed and extracted from the previous layer.

[0051] Furthermore, This represents the second output of gates A and B, which are generated through a deeper consideration of the triple (u,i,p). For example, for task A, the information of (u,i) comes more from task A itself, while the information of (u,p) and (i,p) related to p needs to be obtained more from the shared expert S. A similar approach is taken for task B, thus yielding... Therefore, the outputs of the gating units for both Task A and Task B are derived from a weighted sum of two parts, α. A ,α B ∈[0,1] represents the control parameters of the gating units for tasks A and B, ultimately yielding the output. K represents the number of expert networks in each layer.

[0052] (iv) Represents the fine-tuning and prediction module

[0053] The prediction layer in the prediction module of this invention utilizes the outputs obtained from gate A and gate B in the last layer. The prediction score functions s(i|u) and s(p|u,i) are used. During the prediction phase, the two tasks no longer exchange information but predict independently. The calculation process is as follows:

[0054]

[0055] Among them, MLP A\B () represents a multilayer perceptron network (MLP). The items with the highest s(i|u) are used as the recommendation results for task A, while the participants with the highest s(p|u,i) are used as the recommendation results for task B.

[0056] During model training, this invention selects Bayesian Personalized Ranking (BPR) loss as the loss function for both sub-tasks, and uses... Let these represent the positive and negative training samples for task A(B). The loss function is as follows.

[0057]

[0058]

[0059] Where i′ and p′ represent items and participants that do not exist in positive triplets (the same applies below);

[0060] In addition, this invention further constructs two auxiliary loss functions to help the MGBR model fine-tune the representation, so that the learned representation has higher quality and better generalization.

[0061] For Task A, although the information provided by Task B is helpful as previously mentioned, Task A should still focus more on the preference matching between u and i. This means that in a positive triplet (u, i, p) that actually exists in historical data, replacing i should significantly change the predicted score, but replacing p should not. Therefore, the model should strive to ensure that:

[0062] s(u,i,p)>s(u,i′,p),s(u,i,p′)>s(u,i′,p) (19)

[0063] Based on the above ideas, the auxiliary loss proposed in this invention for task A is:

[0064]

[0065] in, This means replacing the item in the triple with i′. This means replacing the participants in the triple with p′, y (u,i,p) =1, if the triplet Otherwise, y (u,i,p) =0. |T| represents the number of negative samples in the auxiliary loss function;

[0066] Similarly, this invention also designs an auxiliary loss function for Task B. In Task B, it is considered that changes to either item i or user p should have a significant impact on the predicted score. This is because the changed item i′ no longer matches participant p, thus failing to satisfy preferences and attract participants. Therefore, the model should, as far as possible, ensure that:

[0067] s(p|u,i)>s(p|u,i′),s(p|u,i)>s(p′|u,i) (21)

[0068] Based on the above ideas, this invention still uses the auxiliary loss function of BPR loss design task B:

[0069]

[0070] Therefore, in this invention, the overall optimization objective loss function is as follows:

[0071] L = L A +βL B +β A L A′ +β B L B′ (twenty three)

[0072] Where, β A ,β B For the auxiliary loss function L A′ L B′ The control parameters, β, are the loss function L. B The control parameters.

[0073] Features and advantages of the present invention

[0074] 1. It is the first deep recommendation model proposed for real group-buying recommendation tasks in domestic e-commerce scenarios;

[0075] 2. The model architecture is based on a multi-task learning framework, and a novel gating adjustment unit is designed to enable full interaction of information between the two sub-tasks in group buying recommendation, thereby significantly improving the recommendation effect. Attached Figure Description

[0076] Figure 1 This is a schematic diagram of the group-buying recommendation model MGBR of the present invention.

[0077] Figure 2 This is a diagram of the adjustment gate control unit of the present invention. Detailed Implementation

[0078] The advantages of this invention are illustrated below through specific implementation examples and comparative experiments.

[0079] 1. Evaluation settings.

[0080] This invention uses MRR@N (Mean Reciprocal Ranking) and NDCG@N (Normalized Decay Cumulative Gain) as evaluation metrics for the model's recommendation performance. While both metrics characterize the ranking superiority of the recommendations, NDCG is more sensitive to the location information of positive examples than MRR. In the selection of the positive-to-negative example ratio during testing, this invention uses a 1:9 ratio to calculate MRR / NDCG@10 and a 1:99 ratio to calculate MRR / NDCG@100.

[0081] 2. Hyperparameter settings.

[0082] The experiment was conducted on a GeForce RTX 3090 workstation equipped with 24GB of RAM, running Ubuntu 18.04.5 and Torch 1.8.0. Regarding hyperparameter settings, the hyperparameters of all other benchmark models were determined based on the fine-tuning experiments according to this invention; the hyperparameters of MGBR are shown in the table below.

[0083] Table 1 Hyperparameter settings in this invention

[0084] Hyperparameters value explain d 12 Vector represents dimension H 2 Number of GCN layers K 6 Number of expert networks at each layer |T| 99 Number of negative samples in the auxiliary loss function <![CDATA[α A ]]> 0.1 Control coefficients of Task A gating unit <![CDATA[α B ]]> 0.1 Control coefficients of the gating unit in Task B β 1 Control coefficients of the loss function for task B <![CDATA[β A ]]> 0.3 Control coefficients of the auxiliary loss function for Task A <![CDATA[β B ]]> 0.3 Control coefficients of the auxiliary loss function for Task B ρ 0.0002 Learning rate |B| 64 Batch size .

[0085] 3. Comparative Experiment Results and Analysis

[0086] The evaluation results of the MGBR model compared with all baseline models in this invention are shown in Table 2. The results show the performance of MRR and NDCG in two tasks, with positive-to-negative sample ratios of 1:9 and 1:99 respectively. Table 2 shows that MGBR performs optimally in both sub-tasks, especially in Task B. This indicates that MGBR can extract rich information from the relationships between initiators, participants, and items, and promotes synergy between the two tasks through multi-task learning.

[0087] Table 2 Comparative test evaluation results of MGBR in this invention

[0088]

[0089]

[0090] Table 3 shows the ablation experiment results of MGBR in this invention. MGBR-M represents a variant that removes the shared parts of the multi-task learning framework, MGBR-R represents a variant that removes the auxiliary loss function, MGBR-MR represents a variant that removes both the shared information and the auxiliary loss function in the multi-task learning framework, MGBR-G represents a variant that replaces the optimized gating unit with a general gating unit, and MGBR-D represents a variant that replaces three views with a heterogeneous graph. The evaluation results shown in Table 3 reveal that MGBR-M shows the largest performance reduction compared to other variants, indicating that the shared parts of the multi-task learning framework play a crucial role in the model's performance. Furthermore, the auxiliary loss function has a positive effect regardless of whether it is used in conjunction with the multi-task learning framework, indicating that the auxiliary loss function of this invention does indeed help the model's performance through representation fine-tuning. The MGBR-G variant shows a more significant performance degradation on task B, indicating that the optimization of the gating unit in this invention is more suitable for handling the more complex task B. The MGBR-D variant demonstrates that using multiple views to leverage the different relationships between users and items can better facilitate representation learning.

[0091] Table 3. Evaluation results of ablation tests on MGBR in this invention.

[0092]

[0093]

Claims

1. A group-buying recommendation model based on a multi-task learning framework, characterized in that, The system includes a graph representation learning module, a multi-task learning module, and a representation fine-tuning and prediction module. First, for the group-buying recommendation task, the graph representation learning module learns vector representations, which are then input into the multi-task learning module for information interaction and extraction. Finally, these vector representations are input into the representation fine-tuning and prediction module to obtain the predicted recommendation results. The model is then updated using the loss functions of the two tasks and an auxiliary loss function for representation fine-tuning to better address the group recommendation task. The group-buying recommendation task is specifically described as follows: In a group-buying scenario, this includes the group-buying initiator, also known as the user. Group buying participants Group buying items The ultimate goal is to select the appropriate and This maximizes the success rate of group buying; thus obtaining the triplet data of the group buying group. , Indicates user gather, Represents items gather, This represents the set of all participants in a group purchase, and the prediction target is the success rate of the group purchase, i.e., the joint probability. ; The above group-buying recommendation task is broken down into two related sub-tasks. The first sub-task, denoted as Task A, is to calculate the score function that enables initiating a group purchase. That is, given a certain user Recommended candidate items The first subtask enables users to initiate group buying; the second subtask, denoted as Task B, calculates the score function required to complete the group buying. In a given user and candidate items In this case, other participating users are recommended. This enables group buying to be completed; The graph representation learning module constructs three undirected graphs using a multi-view approach, with node types as follows: , and This represents three views: the initiator's view, the participant's view, and the social view. GCN is used to learn representations on these three views to obtain the corresponding vector representation for each view, thus achieving... Each element in the process receives a vector representation from two different views. Finally, the vector representations from the different views are aggregated using vector concatenation to obtain the final result. The corresponding vector representation; The multi-task learning module consists of multi-layer expert networks and adjustment gating units. Each layer of the multi-task learning module contains a set of expert networks and adjustment gating units. This set of expert networks is divided into three parts, used to learn task A information, task B information, and shared information, respectively, and are denoted as: expert network group A, expert network group B, and expert network group S of the l-th layer. Each expert network is a multilayer perceptron, which is a tower structure composed of three fully connected neural networks. It accepts the input or original input from the gating unit of the previous layer and outputs a corresponding vector or matrix, which is the result of information extraction. Each layer has three adjustment gating units, corresponding to the three expert network parts, and are denoted as gating unit A, gating unit B, and gating unit S of the l-th layer. Gating A consists of two parts. The first part weights the outputs of expert network group A and expert network group S based on the inputs of expert network group A and expert network group S. The second part weights the outputs of expert network group A and expert network group S based on the initiator, participants, and items. Two of the three are combined to weight the outputs of expert network groups A, B, and S, and the weighted sum of the two parts is used as the output of gate A. Gating B also consists of two parts. The first part weights the outputs of expert network groups A and S based on the inputs of expert network group B and expert network group S. The second part is based on the initiator, participants, and items. Two of the three are combined to weight the outputs of expert network groups A, B, and S, and the weighted sum of the two parts is used as the output of gated B. The gated control S has only a single part, and the output of the gated control S is a weighted sum of the outputs of the expert network groups A, B, and S based on the inputs of the expert network groups A, B, and S. The fine-tuning and prediction module includes a prediction head and auxiliary loss functions. The prediction layer consists of two independent multilayer perceptrons with a dual-tower structure. The prediction layer takes the output of the multi-task learning module as input and calculates the score functions on the two sub-tasks as output. Simultaneously, a Bayesian personalized ranking loss function (BPR) is designed for tasks A and B, indicating that the fine-tuning part is based on the score functions output by the prediction layer, and two auxiliary loss functions are constructed: the auxiliary loss function for task A is a listwise loss, which compares positive samples with... and The replaced negative samples are used to calculate the auxiliary loss function value. Positive samples refer to those observed in the group-buying records. Combination; the auxiliary loss on task B is the BPR loss, which is achieved by comparing positive samples with... The loss function value is calculated for the replaced negative samples. The final optimization objective of the model is composed of the weighted sum of the BPR loss of task A, the listwise auxiliary loss of task A, the BPR loss of task B, and the BPR auxiliary loss of task B.

2. The group-buying recommendation model based on a multi-task learning framework according to claim 1, characterized in that, The process of breaking down the group-buying recommendation task into two related sub-tasks is formally defined as follows: Task A: Calculate the scoring function , for the given initiator I recommend an item ; Task B: Calculate the scoring function , for the given initiator and items Recommended participants ; Predicting the success rate of group buying after recommendation, i.e., the joint probability. Since the participants in a group purchase are not very familiar with each other, we can make the assumption that they are independent of each other. Therefore, the joint probability can be decomposed as follows: （1） , （2） ， （3） Therefore, the scoring function and Corresponding to and .

3. The group-buying recommendation model based on a multi-task learning framework according to claim 2, characterized in that, In the graph representation learning module, the target score function is... and Corresponding to Three objectives are identified, and their vector representations in a given dimensional space are obtained using graph representation learning. The graph representation learning module constructs three undirected graphs, each containing two types of nodes and their corresponding relationships. The node types in these three graphs are... , and This represents three views: the initiator's view, the participant's view, and the social view, emphasizing three user characteristics: initiator's preference, participant's preference, and shared user preferences, respectively. In this module, GCN is used to learn representations from the graphs corresponding to these three views to obtain... Vector representation; using This represents three images; the uppercase letter combination in the subscript represents the corresponding lowercase letter. The different relationships between them, from This allows us to extract user preferences for items, and from... Extract the similarity between users; Specifically, using , , Representing three views respectively In GCN's The process of calculating the vector representation of the nodes learned by the layers is as follows: ， （4） ， （5） ， （6） in, , , To regularize and add self-loops to the adjacency matrix, , , This is the parameter matrix. This represents the Sigmoid function. Represents the vector dimension. Representing sets The size of I; After learning the vector representation through graph representation, Each of these can obtain vector representations from two different views. These two representations are then aggregated using vector concatenation to obtain... The corresponding vector representation.

4. The group-buying recommendation model based on a multi-task learning framework according to claim 3, characterized in that, In the multi-task learning module, an expert network-based multi-task learning framework is used. Expert network learning is divided into three parts: learning task A information, learning task B information, and learning shared information. Each expert network is associated with a corresponding gating unit to complete aggregation and output. Specifically, using This represents different expert network submodules. , represents the output vector representation of the i-th A expert network at layer l. , indicating the i-th The output vector representation of the expert network at layer l. , indicating the i-th The expert network's output vector representation at layer l; each expert network learns a new vector representation based on the output of the previous layer, and the calculation process is as follows. （7） （8） （9） in, This indicates that gate A is in the first... The output of the layer, This is the learnable parameter matrix; other notations are similar. After each expert network learns a new vector representation, the multi-task learning framework selects and aggregates the vector representations of all expert networks through gating units to generate a new output vector. The calculation process is as follows: （10） （11） （12） , （13） （14） （15） in, All of these are learnable parameter matrices. For the output of each expert network in the layer above layer l, This represents the input representation of expert networks A, B, and S at layer l. The expert network obtains a new representation in layer l by utilizing information processed and extracted from the previous layer. This represents the second output section of gate A and gate B, which are obtained through... This triplet is derived from a deeper consideration; the outputs of the gating units for both Task A and Task B are obtained by weighted summation of two parts. , which are the control parameters of the gating units for Task A and Task B, and finally the output is obtained. , The number of expert networks at each layer.

5. The group-buying recommendation model based on a multi-task learning framework according to claim 4, characterized in that, In the aforementioned fine-tuning and prediction module, the prediction layer utilizes the outputs obtained from gates A and B in the last layer. Predictive score function and During the prediction phase, the two tasks no longer exchange information but predict independently. The calculation process is as follows: （16） in, () represents a multilayer perceptron network; The top-ranked items are used as the recommended results for Task A, while also... The top-performing participants were used as the recommendation results for Task B; During model training, Bayesian personalized ranking loss was chosen as the loss function for both sub-tasks. Let these represent the positive and negative training samples for task A(B). The loss function is as follows. （17） （18） in, This indicates that there are no negative items or participants in the positive triplet; In addition, two auxiliary loss functions are constructed to fine-tune the vector representation, so that the learned representation has higher quality and better generalization. For task A, task A focuses on The preference matching between them represents positive triples that actually exist in historical data. Replace This causes the predicted score to change, replacing To prevent changes in the predicted score, the model should guarantee: （19） Therefore, the auxiliary loss on task A is: （20） in, This means replacing the items in the triple with , This means replacing the participants in the triplet with , If triples ,otherwise, , This indicates the number of negative samples in the auxiliary loss function; For Task B, an auxiliary loss function was designed, which considers both items... Or users Any changes should affect the predicted score, because of the changed items. With the participants If the model no longer matches preferences and fails to attract participants, it should guarantee that: （21） Therefore, the auxiliary loss function for Task B will still be designed using the Bayesian personalized ranking loss: （22） Therefore, the overall optimization objective loss function is as follows: （23） in, Auxiliary loss function , The control parameters, loss function The control parameters.

Images