Construction and Application of Comprehensive Demand Atlas of Local Life

The comprehensive lifestyle GEneral NEeds net (GENE: lifestyle GEneral NEeds net) is a knowledge map that deeply explores the diverse needs of users in local life scenarios from the perspective of user needs, and forms a knowledge map related to multi-industry and multi-type supply , Which aims to improve the efficiency of platform supply and demand matching and help business growth. This article introduces the background, system design, and algorithm practices involved in the comprehensive demand map of local life, and shows the application of multiple business lines in Meituan, hoping to bring you some help or inspiration.

1. Background

1.1 Business status

With the mission of "helping everyone eat better and live better", Meituan covers takeaways, catering, hotels, homestays, travel, tickets, movies/performances, leisure/play, beauty, medical care, parenting, education, marriage, Hundreds of industries including life services meet the diversified life service needs of hundreds of millions of users. In order to continuously increase the value of the platform, in addition to promoting the continuous improvement of the number and quality of users and merchants, more efficiently matching user needs and merchant supply is also an important starting point.

In order to improve the efficiency of matching, we need to fully and deeply understand user needs and merchant supply, and try to organize and manage supply from the user's perspective. At present, "industry-category-merchant-commodity" is a relatively common method of supply organization and management. However, with the rapid development of business and industry, the pain points brought by this organization method have become increasingly prominent. For example:

• Some user needs with unclear directivity make it difficult to obtain suitable matching results. For example, "Where to accompany my baby to play on weekends?" Due to the unsatisfactory matching results of the platform, users often have to complete category decisions offline and decide to take the baby to the farmhouse barbecue, and then search for the corresponding farmhouse group purchase on the platform.
• Some requirements span multiple categories, and the matching process is not smooth enough. For example, "Where to relax with friends on weekends?" After the user completes the category decision offline, the selectable categories include KTV, bar, secret room, board game, etc., but the hosting pages for various purposes are independent of each other, and the user needs to be between the hosting pages Switch back and forth.
• In some specific categories, users still find it difficult to find a supply that meets their needs. For example, in the medical aesthetics category, because of the lack of relevant knowledge, users often do not understand the functions of the various service items provided by the merchants, which parts are suitable for them, and what materials should be used, and they cannot efficiently find a service supply that suits them.

The essential reason for the above problems is that the organization of supply is mainly based on the perspective of the industry, and does not fully consider the perspective of users. In the current market environment where satisfying user needs is the primary goal, we need to iteratively improve the existing supply organization.

1.2 Problem analysis

In order to solve the above problems, we try to analyze it from an external perspective, combined with first principles. In the entire human society, according to Maslow ^[1] , human needs can be summarized and stratified. If human society is regarded as a system, human beings meet their needs through transactions in one of the subsystems called "markets".

Starting from the level of needs, human beings complete transactions in the market, and the final needs are met. Then the process of trading in the market can be disassembled as "inspiration->consideration->choice evaluation->transaction purchase->performance/service" . And, after the first three stages, user needs gradually evolved from coarse-grained to fine-grained, from abstract to concrete. The following will be combined with specific examples for interpretation:

• Level of : There is an emotional level in Maslow’s level of needs, which corresponds to the rich and diverse emotional needs of human beings, including family affection, friendship, love, and so on. As a mother, there is often a strong mother-child relationship with her baby, and she hopes to continue to strengthen this emotion.
• from : For this reason, mothers often try to spend more time with their babies. By increasing the dimension of companionship, it becomes "play with the baby"; by increasing the dimensions of time and space, it becomes "where to play with the baby on weekends?"
• Consider : For the above problems, the mother will find many options, such as outdoor barbecue, farm picking, theme parks and so on. When she decides to choose an outdoor barbecue, it will translate into specific commodity needs, such as buying grills.
• evaluates and chooses : The mother then chooses among the available supply ranges. As people often say, shop around, the basis for selection will vary, such as price, quality, reputation and so on.
• transaction purchase : After completing the selection, the mother will conduct a transaction in exchange for goods or services.
• performance service : that is, the arrival of the goods, the completion of the service, and so on.

The market is a subsystem of human society, and the e-commerce platform is a subsystem of the market. At the same time, the e-commerce platform is an online subsystem that provides search, recommendation and other forms of supply retrieval capabilities. The current status quo is that users often complete the process from "thinking" to "consideration" offline, transforming them into specific product/service requirements, and then entering the e-commerce platform to complete the follow-up from "choice evaluation" to "contract performance service" Process (shown in Figure 1 below). E-commerce platforms tend to focus on improving the capabilities of the latter three stages, and it is easy to overlook the first two stages.

Therefore, it is more difficult for users to form the mentality to complete the "inspiration" and "consideration" on the platform, and most e-commerce platforms organize and manage the supply in the manner of "industry-category-merchant-commodity". In the end, a mutually restrictive relationship has formed between users and e-commerce platforms.

In fact, compared with the clear commodity demand of "grill", users still have many abstract, vague, and ambiguous demands, which are still stuck in the first two stages. For example, where do you go with your baby on weekends? Where can I relax with friends on the weekend? How can I make myself more beautiful before getting married? How to cultivate children's hands-on ability during summer vacation? At the same time, such requirements often span multiple categories, or there are multiple choices under similar categories.

Only by breaking the existing constraints and providing users with the capabilities of the first two stages (thinking and considering) can the e-commerce platform further meet the needs of users. The user's decision-making cost is further reduced, the decision-making process is more consistent, and the user experience can be improved accordingly. At the same time, the user's transaction process in the market can also be further realized online.

Taking "industry-category-merchant-commodity" as a reference, if the e-commerce platform can identify the user's needs in the first two stages, and establish a new relationship between it and the supply, supplemented by search, recommendation, etc. With search capabilities, users may complete the first two stages . As a semantic network that reveals the relationships between entities, knowledge graphs are particularly suitable to solve the above problems.

Two, the solution

2.1 Solutions

Continuing the example in 1.2, this mother transformed the need of "Where to accompany the baby to play on weekends?" into specific "outdoor barbecue" requirements, and extended to more specific "grills" and "farmhouse music group purchases" requirements. At this time, the mother will go to various physical e-commerce platforms and life service e-commerce platforms represented by Meituan to conduct "selection evaluation". The two e-commerce platforms respectively use search and recommendation techniques in the physical supply pool or service supply pool to locate specific goods/services and give feedback to the mother.

Regarding the purpose described in 1.2, the technical team expects to achieve the goal, the current representative reference case is the Alibaba e-commerce cognitive map AliCoCo ^[2] . Its basic construction idea is to start from the user's perspective, first carry out various types of atomic word mining, and then further combine the atomic words and mine related candidate phrases, then identify the real user needs from them, and finally relate to the corresponding supply. Its hierarchical structure is shown in Figure 2:

• classification layer : Build a complete classification system, including various classifications of the world, including general space, time and other categories, as well as color, function and the most important categories involved in e-commerce.
• atomic concept layer : expand on the many categories of the classification layer, including atomic concepts under various categories (such as space -> outdoor, event -> barbecue, time -> Christmas, color -> red, function -> Warmth, category -> dress) and the relationship between the concept of atom.
• E-commerce concept layer : Above the atomic concept layer, it contains user shopping needs of phrase granularity composed of atomic concepts or directly mined, that is, e-commerce concepts (such as outdoor barbecue), so as to explicitly use user shopping needs A phrase that conforms to natural language.
• Commodity layer : Contains the relationship between commodities and various atomic concepts and e-commerce concepts (such as outdoor barbecue -> grill, butter, tin foil).

Based on the above map, this mother can directly express the need for "outdoor barbecue" on Tmall, rather than a more specific "grill"; Tmall will also feed back other important products related to outdoor barbecue besides the grill to this Mother. From the perspective of the corresponding relationship, AliCoCo's e-commerce concept level corresponds to the "consideration" stage, and the product level corresponds to the "choice evaluation" stage. Obviously, due to the existence of AliCoCo, Tmall can intervene in the user's transaction process from the "consideration" stage.

Inferring from this, we should be able to build a more complete map to cover the stage of "inspiring thoughts". At this stage, human needs, according to Maslow's level of needs, are gradually concretized by adding one or more dimensional constraints. For such dimensional constraints, we collectively refer to them as "scene constraints". Therefore, we call the requirements corresponding to the "consideration" phase "representative requirements"; the requirements corresponding to the "thinking" phase are referred to as "scene requirements". To this end, we hope to construct a comprehensive local life demand map (GENE: lifestyle GE neral NE eds net ) , as shown in Figure 3 below. For the hundreds of comprehensive industries involved in the local life scene, we believe that the new supply organization method can be closer to the needs of users, and it is also to solve the problem of matching supply and demand from the user's perspective.

2.2 Specific plan

Continuing the construction idea in 2.1, we tried to build a set of multi-level graph structure, and split the "representative requirements" and "scenario requirements" into independent levels, which avoided the combination of two types of requirements in the same level. It can cause confusion, and can disassemble and describe the needs of users in a more detailed manner. The local life comprehensive demand map (GENE) is mainly composed of six parts, including the scene demand layer, the scene element layer, the concrete demand layer, the demand object layer, the industry system layer and the supply layer, as shown in Figure 4 below:

At the scene demand level, we use Human-Readable short sentences to characterize scene-oriented user needs, such as "Where to accompany a 3-year-old baby on National Day to play", "Make yourself more beautiful before getting married", and "Primary students improve their thinking ability" "Wait. The expression of a scene requirement usually contains elements such as characters, purpose, time, space, methods, etc. Take “where to play with a 3-year-old baby on National Day” as an example, “3-year-old baby” is a character, and “play with the baby” It is the purpose, and the "National Day" is the time.

At the scene element level, in order to better express the scene requirements, we disassemble these short sentences and refine them into multiple fine-grained vocabulary, and use these words to analyze the characters, purpose, time, space, method and other elements in the scene requirements. For complete coverage and systematic organization, we call it "scene elements".

In the concrete demand layer, because the expression of scene demand often does not explicitly point to a specific service/supply, it implies a batch of potential services/supply suitable for this scene. For example, in the example of "Where to accompany a 3-year-old baby on National Day", outdoor barbecue, feeding alpaca, playing on the slide, riding a pony, etc. are all specific services suitable for the needs of this scene. Therefore, we need to show all these specific services explicitly in the form of phrases, which directly reflect the specific service needs of users, and are called "concrete needs".

At the demand object level, in order to further understand the concrete demands, we divide the concrete demands into objects corresponding to specific service demands, which we call "demand objects", and the interaction between users and objects in the service. For example, feeding alpaca with a concrete demand can be divided into alpaca (demand object) and feeding (service interaction). Due to the diversity of local life services, around the demand object of alpaca, in addition to feeding alpaca, it can also produce There are many concrete needs such as touching alpaca, riding alpaca and watching alpaca show. In addition to the nodes of the demand object, this layer also covers the attribute information of the demand object to describe the demand object in more detail. For example, outdoor barbecue with concrete needs can be divided into barbecue (demand object), outdoor (demand object attribute) and experience (implicit service interaction).

At the industry system level, since the user’s scene requirements and concrete requirements often span multiple traditional service categories, in order to determine a specific business scope for user needs, we also need to construct a category system related to each industry, as the above The business foundation built by each layer.

At the supply layer, it includes virtual supplies such as content and physical supplies such as merchants and commodities. These supplies will be associated with nodes such as concrete demand and scene demand, so as to provide corresponding supply support for user needs. For example, a supply that provides outdoor barbecue will be associated with the concrete demand "outdoor barbecue", and further associated with the scene demand "where to accompany a 3-year-old baby to play on National Day".

To sum up, in the comprehensive demand map of local life, user-scenario-based needs and specific service needs are expressed as sentence-level scenario needs and phrase-level concrete needs, respectively. These two kinds of needs are expressed through scene elements and demand objects respectively. Finally, different types of supply will be related to the demand of the scene and the demand of the concrete, so as to use the user's demand as the link to improve the matching efficiency of the supply and the user.

Three, realization method

At present, the comprehensive demand map of local life has initially covered the diversified needs of users in the three local life-related industries of play, medical beauty, and education. In the process of constructing the graph, we build from the bottom up in the order of the industry system layer, the demand object layer, the concrete demand layer, the scene element layer, and the scene demand layer. The nodes of each layer are combined with various types of supply. Establish an association relationship. The following will take the play industry as an example to introduce the construction details of each layer of the map and the algorithms involved.

3.1 Industry system layer

3.1.1 Construction of industry category tree

In the play industry, the industry system layer contains categories that can provide play services, and category information is represented by a tree structure. Due to the construction of the play industry system, the requirements for expert knowledge are very high, and the design of this part is very important for the subsequent knowledge mining at various levels, so we did not directly carry out manual definitions, but based on our current mature industry category tree As the foundation, it is constructed by pruning and splitting it.

First, select the first-level category nodes related to play in the category tree, including "leisure and entertainment", "parent-child", "tourism", "catering" and so on. For each first-level category, we further screen the next-level categories related to play to the leaf category, and cut out categories that are not related to play. In addition, we also split the play-related leaf categories that can be subdivided in the pruned category tree, such as subdividing "bath" into "private soup", "bath center", etc., and finally get a complete play Industry category tree.

3.1.2 Supply linkage of category

After determining the category tree, we also need to obtain physical supply (merchants and commodities) and virtual supply (content, such as UGC) and the affiliation of the category, so as to provide data support for the subsequent series of mining. Since products and content can be linked to merchants, we only need to obtain the affiliation between merchants and categories. The play category tree is obtained by pruning and splitting our existing categories. Except for the split new categories, the relationship between other categories and merchants can directly inherit the original results. For the newly split categories, we need to rebuild the affiliation between merchants and them.

To determine which category a merchant belongs to, the most intuitive basis is the merchant name, product name, and product details. However, the merchant name and product details of many merchants often contain less information, which increases the difficulty of category identification. In order to ensure the accuracy of the classification of merchants, we introduced more merchant information, including merchant UGC and merchant portraits, and designed a multi-source heterogeneous data fusion distinguishing model. The overall model structure is shown in Figure 5 below:

Among them, the feature extraction and processing methods of data from different sources are as follows:

• Merchant name, product name and product details: all are text data, which are directly output after extracting text features ^[3]
• Merchant UGC: Because merchants often have a large amount of UGC, in order to effectively use their information, first ^[4] to obtain UGC features, and then use a Self-Attention ^[5] module Output after feature processing.
• Merchant portrait: After being converted to One-Hot feature, it is output after non-linear mapping through the fully connected layer.

After the above three features are connected, they are merged, and the final category discrimination is realized through the fully connected layer and the softmax layer. Based on the fusion modeling of multi-source data, merchant information has been fully utilized. Taking bath classification as an example, using only the merchant name, product name and product details data, the accuracy rate is 92% based on the BERT discrimination, and the accuracy rate is increased to 98% after the discrimination based on the multi-source fusion model.

3.2 Demand object layer

At the demand object level, we hope to dig out the play object words involved in various items in the play industry system as the nodes of this layer. These words can describe the interactive objects of the user in the actual play process, which is used to form concrete play The basis of demand. In order to ensure the comprehensiveness of the play object mining, we adopt the form of multiple sources and multiple methods.

In terms of data, we use relevant texts from merchants and users as mining corpus. In terms of methods, we use two methods to mine the play object words:

• The first is unsupervised expansion. Before mining starts, the operation will first provide some play object words as seed input based on experience. We use the corpus in advance to construct an unsupervised Skip-Gram structured Word2Vec model, and input the seed words for the business. Extract word vectors and combine cosine similarity to quickly expand relevant target words.
• The second type is supervised tagging. We define the sequence tagging problem and use the BERT+CRF-based model to automatically identify new target words in the corpus.

In the process of practice, in order to mine more efficiently, we will match the unsupervised link expansion and quality inspection target words in the corpus, and transform the matching results into training samples of the supervised labeling link; at the same time, for the supervised labeling After the operation quality inspection, it will also be used as the input of unsupervised expansion. Through the combination of the two links, we complete the excavation of the play objects. The complete process is shown in Figure 6 below. In addition, in the process of operating the manual review of the play object words, for some core play objects known by the business, the relevant features that have been deposited on the business side will also be directly input as their attributes to further improve the information of the play object, for example, for "script killing "This play object, add corresponding "real scene" and "desktop" and other types of attributes.

After obtaining the object word, we also need to know which category the object word belongs to, so that the next step of concrete demand mining and supply correlation, for this we build the relationship between the object word and the category. It is the most intuitive and accurate method to measure the relationship between the two by the number of times the target word is mentioned in the corpus texts of various currents. Therefore, we directly use the object words to perform text matching in the corpus under each category, and determine the relationship by the word frequency. At the same time, we further build the upper, lower and synonymous relationships between the object words. Currently, there are supervised methods such as relationship discrimination through projection and classification (such as the BERT inter-sentence relationship model) and other methods. In the actual process, we use rule-assisted manual methods, based on the statistical characteristics of the object words and the results of the pattern co-occurrence to guide the manual to quickly complete the construction.

3.3 The concrete demand layer

3.3.1 Concrete demand mining

The concrete demand layer can be regarded as a collection of users' specific service needs in the play industry. Each concrete play demand is a node of this layer. The play object superimposes the diversified interactive behaviors and object description information between the user and the object. Obtained, it expresses the essential demands of users for the provision of entertainment services in the form of phrases. The process of mining specific play needs can be divided into two steps:

1. candidate phrase generation : Generate a large number of phrases containing the play object around the play object words as a candidate set of concrete play requirements.
2. Phrase quality discrimination : Establish a semantic discrimination model to extract real concrete play needs from the candidate set.

candidate phrase generation

In step 1, firstly, we use the target words as the core and use the same corpus as the object mining to generate candidate phrases. Commonly used phrase mining algorithms, such as AutoPhrase ^[6] , use Ngrams for phrase combination, and this form is too redundant for phrases with demand targets, so we consider phrase mining based on syntactic structure.

In order to make the generated phrases meet the requirements of syntax, we use the preset syntactic relationship as a template for mining. In order to mine syntactic relations more efficiently in large-scale corpus, we use the lighter ELECTRA ^[7] pre-training model to obtain the Embedding of each component of the sentence, and then use BiAffine ^[8] predict its syntactic relations. Through dependency syntax analysis, we dig out phrases that contain the corresponding play objects and conform to the syntactic relationship in the corpus of each category. In addition, the attributes of the objects in the demand object layer will also be used as object descriptions for phrase generation. In the end, all the mined phrases will be selected as the candidate set of concrete play needs after being roughly screened by statistical features such as word frequency. An example of mining is shown in Figure 7(a).

Phrase quality judgment

In step 2, although the candidate set phrases obtained in step 1 conform to the preset syntactic relationship, there are still a large number of semantically inconsistent expressions with the actual needs of users. Through random inspection and analysis, we found that less than 10% of the phrases that meet the requirements are met. How to select a phrase that reflects the actual user's actual play needs from a large number of candidate phrases has become an urgent problem to be solved.

AutoPhrase uses a discriminant model based on phrase statistical features to score phrases. However, it is difficult to identify phrases with low semantic quality only by statistical features. For this reason, we further build a wide&Deep ^[9] structure discrimination based on statistical and semantic features The model determines whether the phrases in the candidate set are concrete and playful needs. We hope that the discriminant model can filter out a large number of low-quality phrases, so as to save a lot of labor costs for operations. The overall structure of the discriminant model is shown in Figure 7(b), where:

• The Wide part extracts the global and contextual statistical features of candidate phrases, and outputs them after nonlinear mapping through the fully connected layer.
• In the Deep part, the deep semantic features of candidate phrases are extracted, and the corresponding features are extracted through BERT and output.

The features output by the above Wide and Deep parts are connected and then fused to complement each other's advantages. The final phrase discrimination is achieved through the fully connected layer and the softmax layer. In practice, in addition to directly using accumulated phrase tags as positive samples, we also construct positive samples from the candidate set by pre-setting some common-sense patterns, such as viewing [plants], touching [animals], and evaluating the candidate set. Sampling and constructing negative samples, completing the training of the first version of the model, combined with active learning, after multiple rounds of iterations, the model finally reached a recall rate of 92% and an accuracy rate of 85%. After passing the quality judgment, the reserved phrases will be handed over to the operation for manual review and refinement to become the final concrete play demand.

3.3.2 Supply linkage of concrete demand

In the concrete demand layer, since the concrete play demand is obtained by the play object, a corresponding relationship is naturally established between the two. As for the subordinate and synonymous relationship between the concrete play needs, it can be based on the relationship between the objects and the syntactic relationship to assist the manual completion of the construction in the link of manual review. In addition, it is more important to link the demand for concrete play with physical supply (merchants and commodities) and virtual supply (content, such as UGC).

We abstract this problem as a semantic matching problem, which is achieved by matching the concrete play needs with the text information provided by the corresponding category. Among them, the merchant uses the text information of the merchant name, the product uses the product name and the text information of the product details, UGC Use its own text information. Since UGC and commodities are part of the merchants, the relationship between the demand for figurative play and UGC/commodities will also be added to the construction of its relationship with merchants. The overall matching process is shown in Figure 8 below. We first match the actual play needs with UGC/commodities, and then combine the matching results of the merchant name text on this basis, and associate them with the merchants after aggregation through rules.

Due to the large number of concrete play needs, the text information provided at the same time usually contains multiple clauses. For the balance of efficiency and effect, we divide the matching process into two stages: recall and sorting.

In the recall phase, we roughly screen out clauses that may be potentially related to the demand for concrete play. For the concrete play needs, we expand the synonymous tags of the concrete needs based on the constructed synonymous relationship, and perform coarse-grained pattern matching with the clause text, and the matching clauses will enter the sorting stage for refinement Association relationship calculation.

In the sorting stage, we build a semantic matching model based on the classification of the relationship between BERT sentences, and realize the classification by adding a fully connected layer and a softmax layer after the BERT. The model predicts the coarse-screened samples obtained in the recall phase, and recognizes the semantic matching relationship (association/non-association) between the two. The average recall rate and accuracy rate of the final supply relationship reached 90% and 95%, respectively.

3.4 Scene element layer

3.4.1 Disassembly of scene elements

The scene element layer contains the scene elements that make up the user's scene-oriented needs. As mentioned at the beginning of the article, to describe a scene, you need to account for specific characters, time, space, purpose and other elements. For example, for "Where to accompany a 3-year-old baby on National Day?" this scene-based requirement, we can disassemble it as follows: time-National Day, characters-3 years old baby, purpose-family companionship (play with the baby). Therefore, we disassemble the scene elements in the above way, in order to explore and sort out the scene elements as comprehensively and systematically as possible.

3.4.2 Scene element mining

After completing the disassembly of the scene elements, the next step is to mine the scene elements in each disassembled category. The scene elements, as the scene information of the concrete needs, often come from the user's intuitive feelings, so we choose the UGC context corpus related to the concrete play needs for the corpus to be excavated. Similar to the method of demand object mining, we use the refined and summarized scene elements of each category as seed words, and complete the mining of scene elements by means of related element expansion and sequence labeling.

After determining the scene elements, the next key is to complete the construction of the relationship between the scene elements and the needs of figurative play, that is, for each element of the scene, find out its suitable needs of figurative play. For example, spring is suitable for viewing cherry blossoms and children are suitable for being close to animals. After analyzing the UGC text, we found that when users in UGC talk about a specific play demand, they often also explain some relevant scene element information, so we continue to choose the UGC context corpus associated with the play demand for the concrete play as a relationship The source of the constructed data.

Initially, we adopted a pattern-based method, by inducing patterns that can be used to determine the relationship between scene elements and concrete play needs, and directly extracting text containing both from the corpus. However, due to the diversity of user expressions, not only the accuracy rate cannot be guaranteed, but the limited pattern also affects the recall. Therefore, we further tried to use the method based on model discrimination to improve generalization and improve the relationship construction.

Since the concrete play needs in the corpus we use are known, if the scene elements are regarded as attributes of the concrete play needs, then the problem can be regarded as an aspect-based classification problem. Referring to the approach of attribute-level sentiment classification ^[10] , we construct auxiliary sentences by pre-setting sentence templates, combining scene elements and concrete play needs, and transforming the attribute-level classification into a QA-like sentence pair classification problem. For example, for a certain corpus that has been associated with the figurative play needs of "feeding alpaca": "This Saturday we went to the farmhouse to feed the alpaca", one of the auxiliary sentences is "suitable for feeding alpaca on weekends".

We use the BERT sentence relationship classification model to realize sentence pair classification, as shown in Figure 9. The auxiliary sentence and the corpus text are connected through [SEP] and then input into the model for discrimination, and the model outputs the discrimination result (suitable/unsuitable). Finally, based on the relationship extraction results on all corpora, we voted on the relationship between each scene element and the concrete demand to calculate the score and then determined the relationship between the two.

3.5 Scenario demand layer

3.5.1 Scenario requirement assembly

In the scene demand layer, we will assemble the information of the scene element layer and the concrete demand layer to generate a large number of scene requirements. The assembled scene requirements may only contain scene elements, such as "Where to accompany a 3-year-old baby to play on National Day?" It does not contain any specific requirements, but can also contain both scene elements and specific requirements, such as "Going to the suburbs to pick strawberries on weekends" In China, weekends and outskirts are scene elements, and strawberry picking is a concrete demand.

3.5.2 Discrimination of Scenario Requirements

For the assembled scene requirements, the most important thing is to ensure its rationality. For example, "weekend" and "parent-child" are reasonable play scenes, while "girlfriends" and "parent-child" are contradictory play scenes. To this end, we first need to calculate the relationship score between the scene elements to guide the assembly of the scene requirements. The scene elements are only meaningful when relying on the concrete requirements and matching the appropriate gameplay to participate in the assembly of the scene. Therefore, for the construction of a reasonable relationship between the scene elements, we try to use the relationship score between the scene elements and the concrete requirements as the basis, and evaluate the correlation between the two scene elements through relationship transfer.

In Section 3.4.2, we have quantified the score of the relationship between the scene elements and the concrete requirements. One of the most intuitive ideas is to calculate the relationship between the scene elements through the transfer of the relationship between the scene elements-the concrete requirements-the scene elements. . As shown in Figure 10(a), with the concrete demand "feeding alpaca" as the link, the relationship scores of the two scene elements of "parent-child" and "girlfriend" can be obtained.

We first construct a scoring matrix of the relationship between the scene elements and the concrete requirements. Considering that the number of gameplays meets the long-tail distribution, the matrix is normalized to the column of the concrete demand dimension, and at the same time, to ensure the autocorrelation coefficient of the scene element-scene element matrix Is 1, the normalized scene element-representative demand matrix is normalized with L2 row norm, so the new matrix obtained by multiplying the normalized matrix and the transposed matrix can be used as the scene element-scene element The relationship score matrix.

Through the above method, the relationship score between scene elements can be quickly obtained. However, this method only uses the transfer mode of scene element-representative demand-scene element relationship to calculate the direct co-referential strength of the scene elements in the concrete demand, resulting in the scene element Insufficient coverage of the relationship. For this reason, we extend to a longer node relationship chain transfer mode, this transfer relationship between nodes obeys Markov properties, as shown in Figure 10(b). But as the transmission path grows, the computational cost will increase exponentially. Therefore, we use ^[11] to solve the problem. The concept of "cumulative return expectation maximization" is used as the value of the node, and the scene element node set is used as the state space in the reinforcement learning concept. , The set of concrete demand nodes is used as the action space.

For example, when we are in the state of "parent-child" scene elements, we can jump to the next state "girlfriends" or "outdoors" by selecting "feeding alpaca" or "role playing". The decision function of the state transition process is to randomly select a concrete demand node from all the concrete demands associated with the current scene element state as a decision-making behavior, and the extraction probability is positively correlated with the score; the state transition probability is under the decision of the concrete demand node, Randomly jump to the scene element associated with it, and the jump probability is positively correlated with the score.

At the same time, for specific mutually exclusive relationships, we formulate a reward matrix according to actual business application requirements to achieve a diversified scenario element relationship scoring model. In this way, we transform the node relationship transfer model into a Markov decision model, combined with the value iteration expression derived from the Bellman optimal principle and the node pair relationship score prediction formula, as shown in Figure 10(c). According to the formula shown in the figure, the idea of bootstrap iteration is used to calculate the value of the node while keeping the strategy unchanged, and the relationship score between the scene elements is further calculated, which can ensure that the existing relationship network information is more fully utilized. Improve relationship coverage, and reduce the influence of mutually exclusive relationships through the constraint relationship of the reward matrix, and flexibly adapt to the needs of different businesses.

Finally, based on the scoring of the relationship between the scene elements, we select high-scoring scene requirements from the assembled scene requirement set, and generate the final scene requirement expression according to the preset template, such as "Relax with friends on weekends" and "With my girlfriends" "Play", "A good place to take your kids to barbecue outdoors on National Day". These scene demands can be linked to corresponding concrete play demands through the included scene elements/figurative play demands, and then related to related supplies, so as to provide users with scene-based play solutions.

Four, application practice

The comprehensive demand map of local life covers the user's scene needs and concrete needs. On the one hand, it participates in user decision-making in advance, and affects users in multiple stages such as "thinking", "consideration", and "choice evaluation". Its decision-making cost, on the other hand, provides more diversified supply options and efficiently matches supply and demand. In terms of application, it is applied to various business forms such as search and recommendation.

After nearly a year of construction, the current comprehensive demand map contains hundreds of thousands of core concrete demand and scene demand nodes, as well as tens of millions of relationships. It is used in the parent-child, leisure and entertainment, medical aesthetics, education and training of Meituan. Preliminary application practice has been carried out. The following examples introduce specific application methods and application effects.

4.1 Parenting

The matching efficiency between user demand and supply on the parent-child original channel page is inefficient. Among them, ICON is divided into traditional parent-child categories, which cannot meet different types of users' needs (Figure 11(a), left). At the bottom, Guaixi’s supply form is single and reflects The high-quality supply of user needs is insufficient, and the decision-making information is insufficient (Figure 11(b), left), so the parent-child channel page has been revised. In order to fit the business characteristics of parent-child, we apply the demand nodes and relationships related to parent-child play to multiple traffic positions after the channel revision to provide support for tags and supply data.

Among them, for ICON, based on high-frequency scenes and concrete requirements, cross-category requirements ICON are generated, such as "close to animals", "bring a baby in a bath", etc. (Figure 11 (a)) and corresponding secondary pages (Figure 11 ( a) Right). These ICONs contain similar requirements in multiple categories and provide decision-making information for users in the "consideration" stage.

For the bottom recommendation, we optimize the supply around the concrete needs of parent-child play, recommend the related content as a high-quality supply, and extract the text containing the corresponding concrete demand for each supply, which is exposed as the reason for the recommendation. These sentences are based on the actual needs of users. Displaying information from the angle of, which greatly attracted users (Figure 11(b), right). In addition, it is further based on browsing and transaction behaviors to use supply as a medium to establish the relationship between the concrete demand and the user, which is applied to the optimization of the recall and ranking of personalized recommendations. The revised parent-child channel page meets the diversified recommendation needs of users and greatly improves the user experience.

4.2 Leisure and entertainment

On the leisure and entertainment channel page, we have carried out a series of applications around the needs of the scene and the needs of the concrete. On the one hand, new scene ICONs are organized based on the needs of play scenes, such as "outing and viewing flowers" to satisfy users' outdoor play, "indoor trendy play" to meet the trend of indoor fun, and "night life" for users who like to have fun at night. Satisfaction "Team building gatherings" for gatherings with friends and colleagues. These ICONs start from user-scenario play, break the traditional category restrictions, and make the match between users and supplies smoother. Each ICON's secondary page will display every The specific gameplay requirements of each scene and the associated merchants and content.

On the other hand, in the channel page scene navigation module, we try to use scene requirements to further display scene-based play information, including more than a dozen play scene themes such as "One-Man Music", "Family Warmth", and "Birthday". The merchants associated with the concrete needs of these scenes are recommended. These scene-based applications (Figure 12(a)) act on users in the "inspirational" stage, which improves the efficiency of users' decision-making.

In addition, some of the actual requirements can be directly used for quick screening of merchants on the list page of the corresponding category after rewriting, such as real-life script kill/desktop script kill, dress-up/hanfu experience/love pet must go/flight simulation, etc. (Figure 12(b) ) Left and Middle), and the results of our category segmentation in the industry system layer can also be a quick screen for merchants. For example, the category of bathing category (Figure 12(b) right). The application of these quick screens, It is more convenient for users to choose stores.

V. Summary and Outlook

In local life services, how to continuously improve the matching efficiency between supply and users is a difficult problem facing us. We try to take the user's attention object as the starting point, by digging into the user's needs and using it as a link to bring related supplies and users. In order to explore and understand user needs in an all-round way, we strive to explore and try to build a comprehensive demand map of local life, which is constructed layer by layer in the order of industry system layer, demand object layer, concrete demand layer, scene element layer, and scene demand layer. Various types of supply establish relationships.

At present, the results of the comprehensive demand map can be applied to various business forms such as search and recommendation, and have achieved practical results in many business scenarios of Meituan. However, we are still in the initial stage of exploration, and there is still a long iterative path to go. Here we propose some follow-up thoughts and prospects:

• Broader industry coverage : On the one hand, deepen the construction of the existing entertainment, medical beauty and education industries, tap more nodes and relationships, and better understand user needs; on the one hand, to more industries such as beauty and marriage Carry out horizontal coverage; in addition, it will further expand to the full chain of user decision-making, build a service experience map, cover the performance service link, analyze user needs and feedback, and better empower merchants to improve user experience.
• More data is introduced into : The current map construction is mainly based on the text corpus of platform users and merchants. The next step will use more modal data such as images, and try to introduce external knowledge to the current node. And the relationship is perfected and supplemented.
• Deeper map application : At this stage, the practice of map search and recommendation is mainly focused on the direct application of tags and their associated supply. Follow-up consideration is to further deepen the application of the map, and make full use of the information of the scene requirements and scene elements for the recommendation side More accurate user intention recognition provides support, thereby improving the matching efficiency of supply and users, and giving play to the greater value of the knowledge map.

references

• [1] Maslow A H. A theory of human motivation[J]. Psychological review, 1943, 50(4): 370.
• [2] Luo X, Liu L, Yang Y, et al. AliCoCo: Alibaba e-commerce cognitive concept net[C]. Proceedings of the 2020 ACM SIGMOD International Conference on Management of Data. 2020: 313-327.
• [3] Devlin J, Chang M W, Lee K, et al. Bert: Pre-training of deep bidirectional transformers for language understanding[J]. arXiv preprint arXiv:1810.04805, 2018.
• [4] Le Q, Mikolov T. Distributed representations of sentences and documents[C]. International conference on machine learning. PMLR, 2014: 1188-1196.
• [5] Vaswani A, Shazeer N, Parmar N, et al. Attention is all you need[J]. arXiv preprint arXiv:1706.03762, 2017.
• [6] Shang J, Liu J, Jiang M, et al. Automated phrase mining from massive text corpora[J]. IEEE Transactions on Knowledge and Data Engineering, 2018, 30(10): 1825-1837.
• [7] Clark K, Luong M T, Le Q V, et al. Electra: Pre-training text encoders as discriminators rather than generators[J]. arXiv preprint arXiv:2003.10555, 2020.
• [8] Dozat T, Manning C D. Deep biaffine attention for neural dependency parsing[J]. arXiv preprint arXiv:1611.01734, 2016.
• [9] Cheng H T, Koc L, Harmsen J, et al. Wide & deep learning for recommender systems[C]. Proceedings of the 1st workshop on deep learning for recommender systems. 2016: 7-10.
• [10] Sun C, Huang L, Qiu X. Utilizing BERT for aspect-based sentiment analysis via constructing auxiliary sentence[J]. arXiv preprint arXiv:1903.09588, 2019.
• [11] Sutton R S, Barto A G. Reinforcement learning: An introduction[J]. 2011.

About the Author

Li Xiang, Chen Huan, Zhiwei, Xiaoyang, Yanting, Xu Le, Cao Zhen, etc., all come from the Meituan to the store platform technology department to the comprehensive business data team.

Job Offers

Meituan to shop platform technology department-to comprehensive business data team, long-term recruitment algorithm (natural language processing/recommendation algorithm), data warehouse, data science, system development and other positions students, coordinates in Shanghai. Interested students are welcome to send their resumes to: [e-mail].

Read more technical articles from the

front-end | algorithm | back-end | data | security |

| in the public account menu bar dialog box, and you can view the collection of technical articles from the Meituan technical team over the years.

| This article is produced by the 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 activity, please send an email to [e-mail] to apply for authorization.