Scenario-based application and exploration of graph technology in Meituan takeout

In the CTR prediction modeling of food delivery advertisements, we have carried out a series of explorations on the scene based on graph technology. This paper first introduces the starting point of using graph network technology, and then from the abstract graph relationship at the feature level to subgraph expansion and scene-aware subgraphs, gradually introduces how to use graph technology to model business problems, and focuses on the details of online services for joint training. The effect is explained and analyzed, hoping to bring some help or inspiration to the students engaged in related work.

1 Introduction

Food delivery has become a very important part of public life, and everyone is gradually feeling the convenience brought by food delivery. With the advent of the era of big data and deep learning, the click-through rate (CTR)/conversion rate (CVR) estimation technology has been greatly developed, and deep learning technology has become the mainstream method in the industry. Meituan Waimai has also achieved significant online gains by applying the deep model. What the prediction model does is to model the correlation between users and commodities in a specific scenario (ie "people-goods-field") contained in the data. Taking click-through rate estimation as an example, portrait features, contextual features, behavioral features, etc. can be modeled, and the model can perceive the relationship between users and products in this scenario.

Meituan Food Delivery is a scenario-based business: the user's current decision-making is the result of a combination of factors in different scenarios, including but not limited to LBS geographic location, business operations, and meal times. For example, in the bustling business district/small city (LBS), on working days/non-working days/dinner/afternoon tea (time meal period), circle the business according to the business conditions of the business. Compared with the traditional e-commerce business, the restrictions on LBS and time period are added, and the scene factors are more abundant. At the same time, food delivery has a strong immediate demand nature, and the user’s decision-making link will be very short, and it is rare to “go to” the food delivery app for a long time. Therefore, a single user decision is short-term, which further further the takeaway scenario. ization adds more modeling factors.

Therefore, how to model the user's food delivery needs in scenarios, so as to improve the user's ordering experience when using food delivery, has become a key problem to be solved by the food delivery prediction model.

1.1 Problems and challenges

Compared with traditional e-commerce, user interests in the takeaway business show more obvious scene-based characteristics, with the link of [user-scene-interest-decision]: that is, users combine their own needs and personal dietary interests in a specific scene. generate decisions.

In essence, scene modeling is to match users with the best products based on user interests under given scene conditions, such as geographic location, meal time, weather, etc. Focusing on the goal of scenario-based modeling, the industry has carried out a series of technical explorations from different perspectives:

• Feature modeling: Construct statistical features/cross-features of user-product-scene interaction, such as: user's category preference during lunch time, statistics on the number of merchants clicked by users during supper time, etc.
• Sequence modeling: Sequence of behaviors in different scenarios, and detailed description of user interests in different scenarios, such as session behaviors of users in different cells and session behaviors in different time periods.

The above modeling methods can model the influence of scene factors when users make decisions about products, but there are some problems:

• Feature modeling, especially in the process of feature intersection, is easy to introduce noise, which will have a negative impact on model learning.
• Sequence modeling, depending on the richness of user behaviors, is not friendly to low-frequency users’ interest characterization under the division of scene-based behaviors, while the interest characterization of high-frequency users is easy to fall into the closed circle of personal interests.
• Intersection and sequence paradigms do not fully describe the scene. Intersection paradigms have an upper limit of dimension, and sequence paradigms are limited to users' existing behavioral preferences.

Therefore, scenario modeling has the following challenges:

• How to extract valid scene feature intersections.
• How to break the closed interest under sequence modeling.
• How to effectively describe the user decision-making scene completely.

In view of the characteristics of the above problems, after layer-by-layer decomposition, we found that we need a more complete and efficient information expression method, which can have: relationship prediction ability, global information dissemination ability, high-level expression ability, and we found in the field of graph technology. Corresponding solutions will be developed in the following sections for these problems and solutions of graph technology. I hope that these thoughts and practical experience can be helpful or enlightening to everyone.

1.2 Introduction to graph technology

In recent years, with the rapid development of Graph Neural Networks ^[1] , more and more people have begun to pay attention to graph data. The industry has also seen the successive implementation of graph technology, and many application scenarios can be abstracted into graph tasks such as node vectorized representation, classification, clustering, and link prediction.

• For recall scenarios, based on the naturally existing relational interaction graphs between various entities, the embedding expressions learned by constructing a deep matching model can be directly used to measure user-product preferences and product-product associations.
• For the prediction model, the most naive perspective is to perform link prediction on the user-product bipartite graph.

In the internal business of Meituan, Graph Embedding/GNN technology has been implemented in multiple technical teams; such as the related work of Meituan platform ^{[2], [4], [10]} and the Represent-Learning project of food delivery technology, all of which have achieved good results positive earnings.

Compared with the traditional European-style spatial structure type data, the graph structure data has the following characteristics:

• Structural : The graph network has a topological structure, and this structural characteristic often represents some laws. Such as node importance, community structure, etc.
• Association : Graph networks provide a measure of complex relationships and interactions. For example, associations and dependencies can be measured through graph representation learning.
• Globalization : Graph network construction uses global data. Compared with private domain sequence data, graph structure data can better reflect group preference information.
• Strong generalization : Using the message dissemination mechanism of the graph network, the rich information of the nodes on the graph is easier to spread to the unpopular nodes, and the expression ability of the unpopular nodes is improved.

Judging from the development trend of information representation in the industry, information representation is increasing in dimension: from manual features, to sequence modeling, and then to graph modeling, behind it is the demand for more complete information to effectively describe.

• Manual feature era : Mining user/merchant/commodity portraits based on behavior log statistics. The disadvantage is that the characterization is coarse, inaccurate, and limited in expression ability.
• Neural Network (NN) Sequence Model Era : Designing a user interest extraction module based on original behavior sequences to learn user interests end-to-end. The advantage is that the accuracy of user interest expression is improved to a certain extent from the aspect of multimodal and dynamic evolution. The disadvantage is that it is difficult to learn structured information, and there is a strong dependence on behavior data, the behavior of users in the head is dense, and the behavior of users in the middle and long tails is sparse.
• The era of Graph Neural Network (GNN) : everything is a graph. A sequence can be regarded as a subgraph. Compared with a sequence, the graph-structured data is more structured, complete and rich in the expression of information.

In daily business optimization, we also found that if we want to find a formal modeling language that can accurately and completely translate our business scenarios, then the modeling language presented is "graph".

Therefore, we look forward to realizing the scene modeling in the takeaway scene by means of graph technology. In the following, we will describe our attempts and experience in graph technology scene modeling from two aspects: graph algorithm exploration and specific engineering practice.

2. Scenario-based exploration of graph technology

Scenario-based takeaway refers to the common patterns mined based on the complete interaction information (<User, POI, Time, Location>) of users-merchants/products. We characterize and extract this pattern by constructing a user-merchant/commodity interaction scene graph, and introduce the prior knowledge of the scene into the prediction model to assist decision-making. There have been cutting-edge explorations in the industry to apply GNN to LBS scene modeling. For example, the STGCN ^[2] of the Meituan platform describes the application of GNN in the LBS scene from the perspective of space-time integration, and the "gated hypergraph GNN" of the takeaway data set ^[3] The application of hypergraph in scene-based modeling of food delivery LBS is described; compared with ordinary GNN methods, the SOTA effect has been achieved.

In view of the scene-based modeling characteristics of Meituan Takeaway, we have also conducted a series of explorations on graph algorithms, focusing on users in different scenarios in three aspects: scene feature intersection, subgraph expansion perception, and meta-path scene graph. -The goal of POI modeling has been explored in many aspects, and good results have been achieved in offline evaluation and online business.

2.1 Scene feature cross modeling based on feature map

2.1.1 Scene Feature Intersection

Features are the driving force of machine learning models, and it is often said in the industry that "the upper limit of the feature determines the upper limit of the model". Before the NN era, a large part of the improvement of the model effect came from feature engineering. As the model enters the NN era, the fitting ability of the NN model and the blessing of the data dividend have greatly reduced the energy expenditure of algorithm engineers in feature engineering, and the focus of work has begun to focus on the model structure. Although NN can theoretically fit all functions, it cannot quickly approach the upper limit of the effect within limited data and training cycles. In this context, explicit feature intersection has regained everyone's attention. Focusing on automatic intersection features, the industry has successively iterated models such as FM/xDeepFM/DCN/CAN, and achieved very good results.

In the Meituan takeaway scene, it also experienced the manual crossover feature in the first stage and the automatic crossover feature in the second stage. However, in scene modeling, we found that cross features bring information gain, but often also bring "noise" problems; for example, in terms of sample granularity, the effective cross features required by different types of samples are not completely Consistent, there are differences. In the past two years, some work in the industry, such as Fi-GNN, L0-SIGN, Ali FIVES, etc., have also found the problem of noise introduced by the cross of full information in the corresponding business.

Therefore, from the perspective of iterative thinking, it is hoped that more cross features can be introduced, and the introduction of noise information can also be reduced, so as to achieve "personalized" cross features at the sample granularity.

2.1.2 Feature Intersection from Graph Perspective

Feature intersection can be abstracted as "selecting K groups of features in pairs from the full feature set to achieve the purpose of bringing efficient nonlinear expression capabilities to the model". In essence, it can be regarded as the relationship prediction of the bipartite graph between features and features: the feature is regarded as a node, and the relationship between features is regarded as an edge, and the relationship between all features can be regarded as a graph network. The weight of the edge between two nodes can be regarded as the strength of the cross relationship between the corresponding features. By embedding this relation graph into the model training process, the weights of different edges in the relation graph reflect the importance of the intersection of different features.

The N features in each sample form a fully connected graph with each other, denoted as M, each vertex in the graph represents the feature F, and the edge between the vertices represents the degree of association between any two features Fi and Fj, denoted as M< i, j>. By jointly training the relationship graph and the prediction model, the parameter matrix M is updated to keep the semantic information of the relationship graph and the prediction model relevant and consistent. The main process is shown in Figure 3 below:

The main steps can be described as follows:

• Step1. Establish a parameter matrix (obtained by random initialization), do a dot product on the vector representation corresponding to the feature, and use the result as a correlation coefficient.
• Step2. Apply L0 penalty to the matrix to limit the parameters of matrix M to be as close to 0 as possible to reduce useless scene intersections.
• Step3. The parameter matrix is 0,1 to determine the nodes that need to participate in the aggregation.
• Step4. Graph aggregation. For each feature, perform aggregation operations on multiple features that interact with it. The aggregation method uses Attention.
• Step5. Use the aggregated feature vector representation as a new vector representation of the feature and use it in the joint training of downstream CTR estimation.

The feature intersection coefficient is abstracted as the edge weight importance evaluation problem of the graph, so that the model has the ability to predict the strength of the correlation between scene features, so that the intersection features can be introduced more efficiently, providing users with more scene-based modeling. information input.

2.2 Behavioral graph modeling based on subgraph extension

2.2.1 Pain points and diagram solutions for scene sequence modeling

In the iteration of the food delivery algorithm model, the sequence modeling has also undergone a long period of iteration. From the single-scene short-term session modeling to the multi-scene user long-term behavioral interest modeling, detailed explorations have been made in multiple directions. Sequence modeling can fully explore users' interests and preferences in the user's historical behavior, but because the user behavior sequence itself is a "biased sampling" feedback: different users' geographic locations, business supply, usage frequency, etc. Moreover, the distribution of click behaviors of high- and low-frequency users is obviously different, showing the characteristics of abundant aggregation of high-frequency user behaviors and sparse behavior of low-frequency users.

For high-frequency users, the circle of interest may be closed and the model modeling cannot escape the existing circle of interest; for low-frequency users, the lack of information leads to incomplete characterization of their interests. Therefore, we need to have the information expansion ability to expand the user's interest boundary and the ability to expand single-point information; that is, to find a new data structure, break the two-dimensional linear limit, and achieve three-dimensional expansion. Based on this idea, we start from the figure From the perspective of rethinking user behavior modeling: use the linear behavior sequence in the private domain as the basis for interest characterization, and use the global interconnection graph in the public domain as the supplement of interest to establish the connection between individual differences and group commonality.

2.2.2 Behavior POI subgraph design

User behavior information refers to the user's click/order activity record on the platform, which is the most primitive and direct depiction of user interest. Especially for users with sparse behavior, any node in the behavior POI sequence is worthy of attention. However, for this part of users, it is difficult to model interest only by relying on personal behavior POI. If you can expand user behavior through graphs, you can have the opportunity to escape the limitation of personal and private domain behavior and capture potential interests in different scenarios through global behavior graphs. .

We build a POI network based on user session behavior: in the same session, the POIs clicked by the user are related, then a connected graph can be constructed for the POIs in each session; since the POIs of different user behaviors overlap, Through these overlapping POI links, a POI network can be formed between different POI clusters in the entire takeaway scenario. For any POI, we can find the first-hop and second-hop neighbors of the POI from the POI behavior network, and these neighbors can be used as relevant supplements for the POI information. In this way, the characterization of the user's behavioral interest is not limited to the user's existing sequence, but can be further expanded through subgraphs.

Compared with traditional sequence modeling methods, graph network modeling can utilize the high-order network structure interconnected by global user behaviors, and expand user interests with the help of POI sequences:

• For behaviorally sparse users, the user interest modeling cues are supplemented through the global behavioral interconnection graph.
• Alleviate the Matthew effect based on intensive user behavior modeling, jump out of historical behavior to explore potential interests, and improve the generalization of recommendation results.

Specifically, for each POI in the user behavior sequence, it can be expanded through a subgraph, and the expanded subgraph forms a vector representation of the POI by convolution, as shown in the following figure (left). Through the expansion of the behavior sequence, the user behavior is supplemented, so that the user's personal interest can be jumped out of the limit, and the information expression of the user and POI can be enriched.

On the basis of POI subgraphs, we further consider how to effectively aggregate different POIs to achieve a more complete expression of subgraph information.

• Model the relationship between different POIs with different strengths and weaknesses, and use the Attention structure to dynamically allocate to determine the contribution of POIs in the subgraphs to which they belong.
• Considering that the POI sub-graph is constructed by Session, there are differences in the user's behavior sequence, and correspondingly, there are differences in the expression of POI information in different user sequences, and the POI sub-graph information should be adaptively expressed in different behavior context sequences.
• In order to capture this difference, in the process of subgraph convolution, we aggregate the central node with other nodes in the current behavior sequence to model the contextual relevance of behavior.

The central node is not only affected by sequences and subgraphs, but also by the current candidate POI. Therefore, in the process of joint downstream training, we aggregate the central node with the sample Target POI.

During offline training, based on the user behavior sequence, each POI in the sequence is expanded with subgraphs of interest; while the subgraphs are generated using the method of pre-sampling (Message Passing) + joint training aggregation; through the above three aggregation methods, get The vector representation corresponding to the POI in the action sequence. Since this process not only expands POIs, but also takes sequence information and candidate POI information into account in the representation of each node, we call this subgraph-aware convolution. Through sub-graph-aware convolution, the vector representation of POI is associated with contextual information during the convolution process, so that the embedded representation of POI incorporates more accurate interest information.

2.3 Meta-path-based scene graph modeling

2.3.1 Starting from business characteristics - the original intention of meta-path modeling

We abstract the user decision-making process, define an interaction between a user and a merchant POI in a given context as an event, and define the result of multiple users and POI interactions as an event chain. For multiple strongly related event chains (different event chains are connected through common nodes), a scene is formed, and there is a connection between the User and POI subjects between the scenes. In this way, the extension actually constitutes a "scene" Topological network diagram, as shown in the following figure:

It can be seen that the "scenario" composed of the event chain is actually a heterogeneous graph: for example, a user U with certain attributes (consumption frequency, dining preferences, etc.), under a certain context C (time, location, etc.), Click/purchase merchant P with certain attributes (category, main dishes, etc.), this decision-making process is actually the simplest UCP meta-path. The event chain is extended on the basis of this meta-path to get a longer meta-path instance (U-C1-P2-C2-U'). Through the event chain, the relationship link of scene elements can be established, and the complete description of the scene is the extraction of scene element representation and element relationship.

So far, we have abstracted scene-based modeling as the problem of meta-path modeling on heterogeneous graphs. Next, we will introduce how to model user decision-making scenarios on this scene graph.

2.3.2 To the business - the realization of meta-path modeling

The semantics expressed by the meta-path UCPCU are: different users click on the same merchant in different time and space scenarios. Of course, different meta-paths can be defined for different scenarios. In order to further integrate the rich semantic expressions in meta-paths, if only relying on a single meta-path, the expressed semantics will be limited. Starting from user U, a series of meta-path instances can be extended through this meta-path, and these instances describe the cross-temporal relationship between users and POIs in different scenarios. The entire modeling process is divided into the following steps:

• Step1. User and merchant nodes have more attributes, which contain more semantic information than node topology. We regard attribute information as nodes, which are aggregated into user and merchant representations by means of GraphSAGE.
• Step2. Starting from the user U, based on the meta-path, extend the instances (event chains) of multiple meta-paths. The following figure shows multiple instances, including: U-C1-P1, U-C2-P2-C3-U', U-C2-P3-C4-U''-C5-P4; by extension, user U and The association of merchant P4.
• Step3. Meta-path instance query vector representation is then spliced, and interacts with the user (Target User) in the sample. The design of multiple candidate meta-paths can break through the information loss caused by the single meta-path dependent information clipping. The interaction method adopts Attention, which calculates the association between the current user and all candidate meta-paths, and finally acts on the downstream estimation.

For online requests, retrieve Emie in the same scene.

Through the graph network technology, we expand the potential interests of users in different scenarios, capture the current decision-making with the help of attention mechanism, and the relevance of the scenarios corresponding to different meta-paths, so as to achieve scenario-based modeling.

3. Implementation of engineering practice

3.1 Main structure of scene graph CTR model

In the process of model structure design, we follow the principle of "low coupling, high cohesion, pluggability, and no dependency", and we combine the "behavior subgraph extension module", "meta-path scene subgraph module", "feature graph cross-section" The three sub-direction iterations of "module" are inserted into the model as independent modules.

3.2 Behavior graph joint training based on subgraph expansion

In "Behavior Graph Modeling Based on Subgraph Extension", for each request online module needs to perform "behavior sequence length * number of neighbors" vector queries and calculations, when the behavior sequence is long, the online CTR estimation service will It is said that there will be larger RT challenges.

Considering that the "global behavioral interconnection graph" between POIs and the long-term interests of users are relatively stable, for the long-term behavioral subgraph expansion, the aggregation based on behavioral POIs and sampling subgraphs is adopted without relying on candidate Target POIs, and the short-term behavioral subgraph expansion scheme is still adopted. In 2.2.2, the scheme is dynamically calculated according to the sequence and candidate Target POIs, and the long and short-term POIs use different Embedding spaces. Based on this online solution, the idea of offline calculation and online query of long-term behavior subgraphs is adopted, the subgraph embeddings of users' long-term behavior sequences are calculated offline and filled into the database, and Redis is queried online to obtain subgraph embeddings to participate in subsequent calculations.

The long-term behavior module based on subgraph expansion is shown in the following figure:

We divide the whole process into three steps. The core of ensuring that the online time consumption does not increase lies in the sub-graph splitting and reconstruction of Step2. Because the behavior sub-graph Embedding representation will not change during the online Serving process, the results are consistent using sub-graph query or aggregated table lookup.

• Step1. End-to-end training

On the basis of the original CTR model, subgraph query expansion is performed for POI behavior sequences: each POI applies for neighborhood convolution sampling operation from the graph engine, that is, from second-order neighbors to first-order neighbors, and then aggregated to POI itself .

• Step2. Split the subgraph query module, aggregate Embedding, and fill the library

After the training is completed, the original sub-graph query module is split, and a sub-graph POI-Embedding aggregation operation is performed on the full POI in the long-term behavior to obtain the sub-graph Embedding of the behavior POI. In the filling stage, the Embedding of the long-term behavior subgraph POI and the Embedding of other features are written into the Redis storage.

• Step3. Refactor and export the jointly trained model to generate a new online Serving model

The long-term behavior "subgraph extended behavior graph module" neutron subgraph aggregation calculation is refactored to directly query the Embedding representation, and the Embedding is retrieved online from the results of Step2 filling the library.

This avoids the time-consuming of a large number of online convolution operations. Online experiments have verified that the time-consuming of new modules of TP99 and TP999 during the peak period is basically the same.

3.3 Scene Subgraph Module

The meta-path-based scene graph modeling uses meta-path Metapath to represent the user subgraph and merchant subgraph in the CTR task sample. Meal information) heterogeneous behavior scene graph.

3.3.1 Offline Heterogeneous Graph Construction

Since Context will connect User and POI at the same time, the Context node will become a super node on the heterogeneous graph (which can connect almost all Users and POIs in a cell), and POI type nodes may also become super nodes (connecting all Context nodes in the area) ; When a common super node appears, it will lead to difficulty in sampling upstream of the graph and increase the noise. We prevent this problem by designing and building Context nodes and heterogeneous edges:

• Step1. The Context node is used as a space-time context, running through users and stores; if the Context nodes are refined (such as including honeycombs, meals, and categories), the connection edges between User and Context, and the connection edges between Context and POI will be greatly reduced.
• Step2. Different users may jump to different POIs through the Context node. In order to prevent the Path from the User node from jumping to an irrelevant POI during sampling, the Context node should better reflect the user's interest (category information is actually one kind) .
• Step3. There are restrictions on edge rights, which can prevent Context type nodes from becoming super nodes, and the problem of POI will also be solved.

3.3.2 Meta-path sampling modeling

User interests and behaviors are relatively scattered. Starting from the User node, you can jump to different Contexts along the edges to obtain a relatively wide range of instances, and the instances sampled by Metapath can be regarded as a set of User decision-making scene representations. The specific process is as follows :

• Step1. Take a path such as UCPCU in the user subgraph as an example, sample N instances, and store them according to the first Context spread by user nodes. If M groups of instances are obtained, the formula is as follows. Where Ci is the number of instances per group, Uuid: <Context1: [Metapath1, Metapath2]; Context2: [Metapath3, Metapath4]; ; ContextM: [Metapath5, Metapath6]>.

• Step2. For a user request, query the User subgraph according to the Uuid and the current request context, and obtain the subgraph representation of the user decision scene that can match the decision information of the current scene.
• Step3. Use the attention mechanism to capture the current decision, calculate the correlation between different Metapath instances and the corresponding scene, so as to realize the scene modeling.

3.4 Effect Benefit

Subgraph expansion joint training offline AUC+2 thousand points; feature intersection subgraph module offline AUC+1 thousand points; scene subgraph interaction offline AUC+2.5 thousand points.

3.4.1 High and low frequency scene perception

Through graph modeling design, our model can perceive high and low frequency scenes, thereby improving the effect of corresponding users in the scene.

Specifically, in the CTR estimation business of takeaway display advertisements, analyze the user performance of the scene graph algorithm at different frequencies: count high frequency (the number of POI clicks on Meituan Takeaway in the past month > 150 times) and low frequency (the past month For users who clicked POI times in takeout <15 times), compare the online indicators (click-through rate, CTR/Gross Merchandise Volume, GMV) of the experimental group compared to the control group (with the Base model AB experiment without scene graph). At the same time, we also counted the per capita exposures of the third-level categories according to high and low frequency users, and compared the per capita exposures of the baseline.

	CTR	GMV	Number of categories exposed per capita
low frequency users	+1.58%	+1.08%	+0.02 category
high frequency users	+2.68%	+1.94%	+0.3 category

From the results of the above table, it can be seen that:

• The CTR and GMV of low-frequency and high-frequency users are both improved, which proves that the perceptual subgraph convolution can effectively capture high-frequency and low-frequency scenes, and realize the characterization of user interest in the scene.
• The per capita exposure categories of both low-frequency and high-frequency users have increased, and the per capita categories of high-frequency users have increased more, indicating that they have better interest mining and exploration capabilities, which can help high-frequency scene users escape the existing circle of interest.

3.4.2 Specific spatiotemporal scene perception

In order to verify the ability of the scene graph model to identify and describe different scenes, we further compared and introduced the scene graph joint modeling to compare the perception changes of no scene graph under the time category and space requirements (the following sets of statistical results are accumulated over multiple days/same period). result).

3.4.2.1 Time Category Scenario

The milk tea category is a hot-selling category in the afternoon. From the perspective of the exposure conversion ratio, the afternoon tea period is higher and the time period effect is obvious. We have counted the milk tea category of a business for multiple days at the same time period, and the number of exposures accounted for the total number of exposures in this time period. From the perspective of time period, the exposure ratio of milk tea in the experimental group increased during the afternoon tea time (14-16:00), while the exposure ratio during the evening dinner decreased, indicating the category of the scene graph model. The ability to perceive the time period has been strengthened and the flow selection tends to be high-quality flow. After the model is launched, the overall posterior CTR index of the milk tea category shows a positive performance.

3.4.2.2 Space requirement scenarios

User needs and behaviors on takeaways have obvious cyclical effects: that is, users' behaviors are quite different on working days and non-working days. For example, on non-working days, most users are at home (community) rather than office buildings, and there will be a greater tendency to purchase dishes and daily necessities on Meituan Takeaway (higher conversion exposure ratio).

Let's take the convenience category exposure of a supermarket as an example. We counted the proportion of supermarket convenience exposure in the total exposure of the day for 7 consecutive days from Monday to Sunday, and compared the exposure proportion according to the experimental group and the control group. As can be seen from the figure, the exposure of the experimental group decreased on Mondays and Tuesdays, and increased on weekends, indicating that the model captured the difference between convenient shopping scenes in supermarkets on working days and non-working days. After the model was launched, the overall posterior CTR of convenience categories over time Indicators are also positive.

4. Summary and Outlook

Different from traditional e-commerce recommendation, food delivery recommendation presents scene-based features: supply is strongly constrained by LBS, user decision-making links are short, and easily affected by the environment. Graph network technology has been deeply explored in academia and industry. For the challenges encountered in the scene-based modeling of Meituan food delivery, we have also carried out corresponding optimization solutions through graph technology, through the intersection of scene feature graphs, Expansion of scene sequence subgraphs and meta-path scene graphs have been explored in terms of cross-feature denoising, breaking through the closed circle of user interest, and complete scene graph characterization.

In the implementation of the algorithm, considering the time-consuming problem of online, we split and reconstruct the long-term behavior subgraph in the Serving phase, which reduces the computational complexity without changing the model results and meets the online performance requirements. Although the scene-based implementation of graph technology has achieved certain benefits, there are still some problems that need to be solved. For example, under the constraint of computing power, feature graph intersection can only satisfy part of feature intersection; for scene meta-path characterization, it still relies on prior manual work. The definition has not yet embarked on the road of "autonomous driving", and we will continue to explore in the future.

5. About the author

Ren Jian, Zhang Wei, Yufeng, Huang Kun, Hui Nan, Pengye, and Zhang Bo are all from the Meituan Takeaway Advertising Technology Team.

6. References

• [1] Li F, Chen Z, Wang P, et al. Graph intention network for click-through rate prediction in sponsored search[C]//Proceedings of the 42nd international ACM SIGIR conference on research and development in information retrieval. 2019: 961-964.
• [2] Han H, Zhang M, Hou M, et al. STGCN: a spatial-temporal aware graph learning method for POI recommendation[C]//2020 IEEE International Conference on Data Mining (ICDM). IEEE, 2020: 1052- 1057.
• [3] Yang T, Zhang L, Shi C, et al. Gated Hypergraph Neural Network for Scene-Aware Recommendation[C]//International Conference on Database Systems for Advanced Applications. Springer, Cham, 2022: 199-215.
• [4] Wu L, Li Z, Zhao H, et al. Learning the implicit semantic representation on graph-structured data[C]//International Conference on Database Systems for Advanced Applications. Springer, Cham, 2021: 3-19.
• [5] Xie Y, Wang Z, Li Y, et al. Fives: Feature interaction via edge search for large-scale tabular data[C]//Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining. 2021: 3795 -3805.
• [6] Chang J, Gao C, Zheng Y, et al. Sequential recommendation with graph neural networks[C]//Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval. 2021: 378-387.
• [7] Shao Z, Xu Y, Wei W, et al. Heterogeneous Graph Neural Network with Multi-view Representation Learning[J]. arXiv preprint arXiv:2108.13650, 2021.
• [8] Li Z, Cui Z, Wu S, et al. Fi-gnn: Modeling feature interactions via graph networks for ctr prediction[C]//Proceedings of the 28th ACM International Conference on Information and Knowledge Management. 2019: 539 -548.
• [9] Fu X, Zhang J, Meng Z, et al. Magnn: Metapath aggregated graph neural network for heterogeneous graph embedding[C]//Proceedings of The Web Conference 2020. 2020: 2331-2341.
• [10] Wang Y, Xu H, Yu Y, et al. Ensemble Multi-Relational Graph Neural Networks[J]. arXiv preprint arXiv:2205.12076, 2022.

Read more collections of technical articles from the Meituan technical team

Frontend | Algorithm | Backend | Data | Security | O&M | iOS | Android | Testing

| Reply keywords such as [2021 stock], [2020 stock], [2019 stock], [2018 stock], [2017 stock] in the public account menu bar dialog box, you can view the collection of technical articles by the Meituan technical team over the years.

| This article is produced by Meituan technical team, and the copyright belongs to Meituan. Welcome to reprint or use the content of this article for non-commercial purposes such as sharing and communication, please indicate "The content is reproduced from the Meituan technical team". This article may not be reproduced or used commercially without permission. For any commercial activities, please send an email to [e-mail] to apply for authorization.