Comprehensive demand map of local living (GENE: Lifestyle General Needs Net (Lifestyle General Needs Net), from the perspective of user NEeds, deeply explores the diversified NEeds of users in local life scenarios, and forms a knowledge map associated with multi-industry and multi-type supply, aiming to improve the efficiency of supply-demand matching of the platform and boost business growth. This paper introduces the background, system design and algorithm practice of local life comprehensive demand graph, and shows the application in more than Meituan business lines, hoping to bring some help or inspiration to everyone.

The background,

1.1 Business Status

With the mission of “helping people eat better and live a better life”, Meituan covers hundreds of industries such as takeout, catering, hotel, home stay, tourism, ticket, film/performance, leisure/entertainment, beauty, medical care, parents and children, education, marriage and life services, and meets 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 users and merchants in terms of quantity and quality, it is also important to match users’ demand and merchants’ supply more efficiently.

In order to improve the efficiency of matching, we need to have a fuller and deeper understanding of user demand and merchant supply, and try to organize and manage supply from the perspective of users. At present, “industry – category – merchant – commodity” is a relatively common way of supply organization and management. However, with the rapid development of business and industry, the pain points brought by this way of organization become increasingly prominent. For example:

  • Part of the directivity is not clear user needs, it is difficult to get suitable matching results. For example, “Where do you go with your baby this weekend?” Due to the unsatisfactory matching results of the platform, users often can only complete the category decision offline, 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 are you going to relax with your friends this weekend?” After the user completes the offline category decision, the optional categories include KTV, bar, secret room, board game, etc. However, the bearer pages for each purpose are independent of each other, and the user needs to switch back and forth between the bearer pages.
  • In partially well-directed categories, users still struggle to find supplies that meet their needs. For example, under the category of medical aesthetics, due to the lack of relevant knowledge, users often do not know the respective functions, suitable parts and materials of various service items provided by merchants, so they cannot efficiently find the service supply suitable for them.

The essential reason for the above problems is that the supply is organized from an industry perspective and does not fully consider the user perspective. In the current market environment where the primary goal is to meet the needs of users, we need to iterate and perfect the existing supply organization.

1.2 Problem Analysis

In order to solve the above problem, we try to analyze it from an external perspective, combined with first principles. In the whole human society, according to Maslow’s hierarchy of needs [1], human needs can be summarized and stratified. If we think of human society as a system, humans satisfy their needs through transactions in one of the subsystems called “markets”.

Starting from the hierarchy 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 broken down into “soul-searching -> consideration -> selection evaluation -> transaction purchase -> performance/service”. And, through the first three stages, users’ requirements gradually evolve from coarse-grained to fine-grained, from abstract to concrete. The following will be combined with specific examples to interpret:

  • Hierarchy of needs: In Maslow’s Hierarchy of needs, there is an emotional level, which corresponds to a variety of human emotional needs, including family, friendship, love and so on. As a mother, there is a strong mother-child bond with her baby and she wants to reinforce it.
  • Echoing thoughts: To this end, mothers often try to spend more time with their babies. By increasing the dimension of the way of accompanying, it becomes “playing with the baby”; By adding time, space and other dimensions, it becomes “Where do you go to play with your baby on weekends?”
  • Consideration: Mother will come up with many solutions to the above problems, such as outdoor barbecues, farm picking, theme parks, etc. When she decides on an outdoor barbecue option, it translates into a specific product requirement, such as a grill.
  • Evaluating options: The mother then chooses from a range of accessible supplies. As people often say, shopping around is based on different factors, such as price, quality, word of mouth and so on.
  • Transaction Purchase: When the selection is made, the mother makes a transaction in exchange for goods or services.
  • Performance services: namely the arrival of goods, completion of services, etc.

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 a kind of online subsystem, which provides search, recommendation and other forms of supply and retrieval capabilities. The current situation is that users usually complete the process from “thinking” to “considering” offline, which is transformed into specific demand for goods/services, and then enter the e-commerce platform to complete the follow-up process from “selection evaluation” to “performance service” (as shown in Figure 1 below). E-commerce platforms tend to focus on the latter three stages of capacity improvement, and tend to ignore the first two stages.

Therefore, it is difficult for users to form the mind to complete “thinking” and “considering” on the platform, while most e-commerce platforms organize and manage the supply in the way of “industry, category, merchant, commodity”. In the end, users and e-commerce platforms form a relationship between the constraints.

In fact, compared with the specific commodity demand of “grill”, users still have a lot of abstract, vague, unclear demand, which still stays in the first two stages. For example, where do you go with your baby on the weekend? Where do you go to relax with your friends at the weekend? How can you make yourself more beautiful before marriage? How to cultivate children’s hands-on ability in summer vacation? At the same time, such requirements often span multiple categories, or there are multiple choices within the same category.

Only by breaking the existing constraints and providing users with the ability of the first two stages (soul-stirring and consideration), can the e-commerce platform further meet the needs of users. The decision cost of users is further reduced, the decision process is more coherent, the user experience can also be improved correspondingly, and the transaction process of users 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 needs of users in the first two stages and establish a new correlation between them and supply, supplemented by search, recommendation and other retrieval capabilities, it is possible for users to complete the first two stages online. As a semantic network that reveals the relationships between entities, knowledge graph is particularly suitable to solve the above problems.

Second, solutions

2.1 Solutions

Continuing the example from 1.2, the mother will ask “Where do you go with your baby on the weekend?” The demand of “barbecue” is transformed into specific “outdoor barbecue” demand, which is extended to more specific “grill” and “farm music group purchase” demand. At this time, my mother will go to various physical e-commerce platforms and life service e-commerce platforms represented by Meituan respectively for “selection evaluation”. The two e-commerce platforms respectively use search and recommendation to locate specific goods/services in the physical supply pool or service supply pool, and then feedback to the mother.

As for the purpose stated in 1.2, the technical team expects to achieve, a representative reference case at present is AliCoco [2], Alibaba’s e-commerce cognition map. The basic construction idea of it is from the user’s perspective. It first carries out various types of atomic words mining, and then further combines atomic words and mines relevant candidate phrases, and then identifies the real user needs from them, and finally relates to the corresponding supply. Its hierarchy is shown in Figure 2:

  • Classification layer: a complete classification system is constructed, including all kinds of classifications in the vast world, including generic categories such as space and time, as well as colors, functions and the most important categories involved in e-commerce.
  • Atomic Concept Layer: Extended on the basis of many categories in the classification layer, including atomic concepts under various categories (such as Space -> Outdoor, Event -> BBQ, Time -> Christmas, Color -> Red, Function -> Warm, Category -> Dress) and relationships between atomic concepts.
  • E-business concept layer: above the atomic concept layer, it contains users’ shopping needs composed of atomic concepts or directly excavated phrase granularity, namely, e-business concept (such as outdoor barbecue), so as to explicitly express users’ shopping needs with a phrase in line with natural language.
  • Commodity layer: contains the relationship between the commodity and various atomic concepts, e-commerce concepts (such as outdoor > grill, butter, tin foil).

Based on the graph above, the mother could express her need for “barbecue outside” directly on Tmall, rather than the more specific “grill”; Tmall also feeds back to the mother other important items related to the barbecue outside of the grill. From the perspective of correspondence, the e-commerce concept layer of Alicoco corresponds to the “consideration” stage, and the commodity layer corresponds to the “selection evaluation” stage. Obviously, thanks to Alicoco, Tmall is able to intervene in the user’s transaction process from the “consider” stage.

Corollary to this, we should be able to build a more complete atlas to cover the “soul-stirring” phase. At this stage, human needs are gradually materialized by adding one or more dimensional constraints to Maslow’s hierarchy of needs. For such dimensional constraints, we collectively call them “scenario constraints”. Therefore, the requirements corresponding to the “consider” phase are called “concrete requirements”; The requirements corresponding to the “soul-stirring” stage are called “scene requirements”. To this end, we hope to build a Gene: Lifestyle General Needs Net, as shown in Figure 3 below. For the hundreds of comprehensive industries involved in the local living scenarios, we believe that the new supply organization can be more close to the needs of users, and also solve the problem of supply and demand matching from the perspective of users.

2.2 Specific Plan

Continue in 2.1, we try to build a multi-level graph structure, and will be “representational demand” and “scene demand” split into separate layers, both to avoid the two types of demand caused by the confusion, and at the same level of and can carry out more detailed requirements of the user scenarios apart, more rich. Gene of local life is mainly composed of six parts, including scene demand layer, scene element layer, concrete demand layer, demand object layer, industry system layer and supply layer, as shown in Figure 4 below:

In the scene demand layer, we use human-readable short sentences to represent the scene user needs, such as “Where to play with the 3-year-old baby on National Day”, “Make yourself more beautiful before getting married”, “Elementary school students improve their thinking ability”, etc. In the expression of a scene demand, it usually contains characters, purpose, time, space, mode and other elements. Take “Where to play with a 3-year-old baby on National Day” as an example. “3-year-old baby” is the character, “play with the baby” is the purpose, and “National Day” is the time.

At the level of scene elements, in order to better express the scene requirements, we disassemble these short sentences and refine them into a number of fine-grained words. These words are used to complete the coverage and systematic organization of the characters, purpose, time, space, mode and other elements in the scene requirements, which are called “scene elements”.

In the representational demand layer, since the description of scenario demand often does not explicitly point to a specific service/supply, but implies a group of potential services/supplies suitable for this scenario. For example, in the example of “Where to play with a 3-year-old baby on National Day”, outdoor barbecues, feeding alpacas, playing on slides, riding ponies and other specific services are all suitable for the requirements of this scene. Therefore, we need to explicitly show all these specific services in the form of phrases, which directly reflect the specific service needs of users, known as “concrete needs”.

In the requirement object layer, in order to further understand the concrete requirement, we divide the concrete requirement into the object corresponding to the specific service requirement, which is called “requirement object”, and the interaction behavior between the user and the object in the service. For example, the feeding of alpaca with 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, a variety of concrete demands such as touching alpaca, riding alpaca and watching alpaca performance can also be generated. In addition to containing the requirement object nodes, this layer also covers the attribute information of the requirement object for a more detailed description of the requirement object. For example, outdoor barbecue for representational demand can be divided into barbecue (demand object), outdoor (demand object attribute) and experience (implied service interaction).

In the industry system layer, because the user’s scene demand and concrete demand often span multiple traditional service categories, in order to determine a specific business scope for the user demand, we also need to build the category system involved in each industry, as the business foundation for the construction of the above layers.

In the supply layer, it contains virtual supply such as content and physical supply 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 demand. For example, a supply that provides outdoor barbeque will be associated with the figurative demand “outdoor barbeque”, and further with the scene demand “where to play with a 3-year-old baby on National Day”.

To sum up, in the comprehensive demand graph of local life, users’ scenario-based demands and specific service demands are respectively expressed as scenario-based demands at the phrase level and figurative-based demands at the phrase level. These two requirements are expressed through scenario elements and requirement objects respectively. Finally, different types of supply will be associated with scene demand and representational demand, so as to improve the matching efficiency between supply and user with user demand as a link.

Third, the implementation method

At present, the comprehensive demand map of local life has initially covered the diverse needs of users in the three industries related to local life: entertainment, medical beauty and education. In the construction process of the graph, we build it layer by layer in order of industry system layer, demand object layer, concrete demand layer, scene element layer and scene demand layer from bottom to top, and establish association relationship between nodes of each layer and various types of supply. The following sections describe the details of how each layer of the graph is built and the algorithms involved, taking the play industry as an example.

3.1 Industry system layer

3.1.1 Construction of industry category tree

In the play industry, the industry system layer contains the categories that can provide play services, and the category information is represented by a tree structure. Since the construction of the system of the play industry requires a high level of expert knowledge, and the design of this part is crucial for the subsequent knowledge mining at all levels, we did not directly define it manually, but based on our current mature industry category tree, we built it by pruning and splitting.

First, the first level category nodes related to play are selected from the category tree, including “leisure and entertainment”, “parent-child”, “tourism”, “catering” and so on. For each category, we further screened the next category related to play until the leaf category, and cut out the categories unrelated to play. In addition, we also split the leaf categories related to play that can be subdivided in the pruned category tree. For example, we subdivided “bath” into “private soup” and “bath center”, etc., and finally obtained a complete category tree of play industry.

3.1.2 Supply association for class purposes

After determining the category tree, we also need to obtain the subordinate relationship between physical supply (merchants and commodities) and virtual supply (content, such as UGC) and category purpose, so as to provide data support for a series of subsequent mining. Since both the goods and content can be linked to the merchant, we only need to obtain the merchant’s affiliation with the class object. The play category tree is obtained by pruning and splitting our existing categories. Except for the new categories that are split, the relationship between other categories and merchants can directly inherit the original results. For the new split category, we need to rebuild the relationship between merchants and their affiliation.

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

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

  • Merchant name, commodity name and commodity details: they are all text data, which are directly extracted by Bert [3] and output after text features.
  • Merchant UGC: Merchant UGC usually has a large number of UGC. In order to make effective use of its information, first of all, the features of UGC are obtained through Encode in the way of Doc2Vec[4], and then a self-attention [5] module is used to process the features and then output them.
  • Merchant portrait: converted into one-hot feature, output after nonlinear mapping through the full connection layer.

The above three features are connected and fused, and the final category discrimination is realized through the full connection layer and the Softmax layer. Based on multi-source data fusion modeling, merchant information is fully utilized. Taking the bathing subcategory as an example, only the merchant name, commodity name and commodity detail data were used, and the accuracy was 92% based on BERT discrimination, while the accuracy was improved to 98% based on the multi-source fusion model discrimination.

3.2 Requirements Object Layer

In the demand object layer, we hope to be able to dig out the play object words related to various items in the play industry system as the nodes of this layer. These words can describe the interaction objects of users in the actual play process, which is the basis for forming the representational play requirements. To ensure the comprehensiveness of play object mining, we adopt a multi-source, multi-method approach.

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

  • The first is unsupervised expansion. Before starting mining, the operator will first provide some play object words as seed input based on experience. We use the corpus in advance to construct the Word2Vec model with unsupervised skip-gram structure, extract word vectors from the seed words of business input, and quickly expand relevant object words in combination with cosine similarity.
  • The second is supervised tagging, which defines the sequence tagging problem and uses the Bert +CRF model to automatically identify new object words in the corpus.

In practice, in order to excavate more efficiently, we carry out text matching for the object words after the expansion of unsupervised links and quality inspection in the corpus, and transform the matching results into training samples with supervised labeling links. At the same time, the results marked with supervision will also be taken as the input of unsupervised expansion after operation quality inspection. Through the combination of the two links, we complete the mining of play objects. The complete process is shown in Figure 6 below. In addition, the operating manual audit to play in the process of the object word, for some of the core of the business is known to play object, will also enter the business side is associated with precipitation characteristics as its properties, further perfect play object information, for example, “kill” the play object, to increase the corresponding “live” and “desktop” and other types of properties.

After obtaining the object words, we also need to know which category the object words belong to, so as to mine the representational demand and supply association in the next step. Therefore, we construct the relationship between the object words and the category. It is the most intuitive and accurate method to measure the relationship between the two words by referring to the number of object words in various corpus texts. Therefore, we directly use object words to carry out text matching in the corpus under each category, and determine the relationship through the word frequency. At the same time, we further construct the hypophysis and synonyms between object words. At present, there are common supervision methods such as relation discrimination through projection and classification (such as the relationship model between Bert sentences). In the actual process, we use the rule to assist the manual, and guide the manual to complete the construction quickly based on the statistical characteristics of the object words and the co-occurrence results of the Pattern.

3.3 The figurative demand layer

3.3.1 Mining of figurative needs

Representational demand layer may be regarded as the user to play in the industry specific set of service requirements, each representational play demand is the layer of a node, the play object superposition pluralistic interaction between the user and the object and object description information, it is in the form of a phrase expresses the essence of the user to play service supply demands. The specific play demand mining process can be divided into two steps:

  1. Candidate phrase generation: A large number of phrases containing play objects are generated around the play object words as a candidate set of representational play requirements.
  2. Phrase quality discrimination: A semantic discrimination model is established to extract the real representational play requirements from the candidate set.

Candidate phrase generation

In Step 1, we first take demand object words as the core and use the same corpus as play object mining to generate candidate phrases. Commonly used phrase mining algorithms, such as Autophrase [6], use Ngram to carry out phrase combination, but this form is too redundant for phrases with demand objects, so we consider phrase mining based on syntactic structure.

In order to make the generated phrases conform to the requirements of syntax, we use the preset syntactic relations as the template for mining. In order to excavate syntactic relationships more efficiently in a large corpus, we obtained the Embedding of each component of a sentence based on a more convenient Electra [7] pre-training model, and then used Biaffine [8] to predict the syntactic relationships. Through dependency parsing, we excavate syntactically related phrases containing the corresponding play objects from the corpus of each category. In addition, the properties of objects in the requirements object layer are also used as object descriptions for phrase generation. Finally, all the phrases mined will be used as candidate sets of representational play requirements after being roughly screened by statistical features such as word frequency. The mining example is shown in Figure 7(a).

Phrase quality judgment

In Step 2, although the phrases in the candidate set obtained through Step 1 conform to the preset syntactic relationship, there are still a large number of semantic expressions inconsistent with the actual needs of users. Through sampling analysis, we find that less than 10% of the phrases meet the requirements. How to select the phrases that reflect the real users’ demand for figurative play from the vast number of candidate phrases becomes an urgent problem to be solved.

Autophrase uses the discriminant model based on phrase statistical features to score phrases. However, it is difficult to identify phrases with low semantic quality only through statistical features. Therefore, we further build a discriminant model of Wide&Deep[9] structure based on the joint modeling of statistical and semantic features. It is hoped that the discriminant model can filter out a large number of low-quality phrases to determine whether the phrases in the candidate set are figurative play needs, so as to save a large amount of human cost for the operation. The overall structure of the discrimination model is shown in Fig. 7(b), where:

  • The Wide part extracts the global and contextual statistical features of candidate phrases and outputs them after nonlinear mapping through the full connection layer.
  • In the Deep part, the Deep semantic features of candidate phrases are extracted, and the corresponding features are extracted by Bert and then output.

The above output features of the Wide part and the Deep part are connected and fused to complement each other. The final phrase discrimination is realized through the full connection layer and the softmax layer. Accumulated in the process of practice, in addition to the direct use of the phrase tags as is sample, we still have some common sense through the preset Pattern constructed from a candidate set is samples, such as ornamental [plants], touch/animal, and sample for sampling structure of candidates, complete) model training, after combining with the active learning, after several rounds of iteration, The model achieved 92 percent recall rate and 85 percent accuracy. The phrases retained after passing the quality judgment will be reviewed and refined by the operation personnel to become the final figurative play requirements.

3.3.2 Supply correlation of representational demand

In the layer of representational needs, since representational play needs are obtained by play objects, there is a natural correspondence between them. As for the hypophysis and synonymy between the requirements of representational play, we can assist the construction of them in the link of manual review based on the relationship between objects and its syntactic relationship. Beyond that, it is more important to relate the need for figurative play to physical offerings (merchants and goods) and virtual offerings (content, such as UGC).

We abstract this problem into a semantic matching problem, which is realized by matching the figurative play demand with the text information provided by the corresponding class object, where the merchant uses the text information of the merchant name, the commodity uses the text information of the commodity name and commodity details, and UGC uses its own text information. Since UGC and goods are part of the merchant, the relationship between figurative play needs and UGC/ goods will also be incorporated into the construction of the relationship with the merchant. The overall matching process is shown in Figure 8 below. We first matched the figurative play demand with UGC/ goods, and then combined with the matching results of merchant name text to associate them with merchants through rule aggregation.

Due to the large number of demands for figurative play and the text information provided at the same time usually contains multiple clauses, we divided the matching process into two stages, recall and sorting, for the balance between efficiency and effect.

In the recall phase, we sifted out clauses that might be potentially relevant to the need for representational play. For the representational play requirements, we extend the synonymous label of representational play requirements based on the constructed synonymy, and carry out coarse-grained Pattern matching with the clause text. For the clauses in the matching, we proceed to the sorting stage to carry out fine correlation calculation.

In the sorting stage, we built a semantic matching model based on the relationship classification between Bert sentences, and realized the classification by adding a full connection layer and a softmax layer after Bert. The model predicts the coarse sieve samples obtained in the recall phase and identifies the semantic matching relationship (correlation/discorrelation) between the two. The average recall rate and accuracy of the final supply association were 90% and 95%, respectively.

3.4 Scene element layer

3.4.1 Disassembly of scene elements

The scenario element layer contains the scenario elements that make up the user’s scenario requirements. As mentioned at the beginning, to describe a scene, specific characters, time, space, purpose and other elements need to be explained. For example, to the question “Where do you go with your 3-year-old on National Day?” For this scenario demand, we can break it down as follows: time – National Day, character – 3-year-old baby, purpose – family companionship (playing with baby). Therefore, we disassembled the scene elements in accordance with the above methods, in order to excavate and comb the scene elements as comprehensively and systematically as possible.

3.4.2 Scene element mining

After the dismantling of scene elements is completed, the next step is to mine scene elements in each dismantled category. As the contextual information of the representational demand, scene elements often come from the intuitive feelings of users, so for the mined corpus, we choose the contextual corpus of UGC associated with representational play demand. Similar to the method of demand object mining, we take the extracted and summarized scene elements of each category as seed words, and complete the mining of scene elements by means of relevant 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 figurative play needs, that is, for each scene element, find out its suitable figurative play needs, such as the suitable for viewing cherry blossoms in spring and the suitable for children to get close to animals. After analyzing the UGC text, we found that when users in UGC talk about a specific need for figurative play, they often give some relevant information of scene elements. Therefore, we continue to choose the contextual corpus of UGC associated with the need for figurative play as the data source for relationship construction.

At first, we adopted the method based on Pattern. By inducting patterns that can be used to judge the relationship between scene elements and representational play needs, we extracted text containing both elements directly from the corpus. However, due to the diversity of user expressions, not only the accuracy could not be guaranteed, but the limited Pattern also affected the recall. Therefore, we further tried to use the method based on model discrimination to improve the generalization and perfect the construction of the relationship.

Since the representational play needs in the corpus we use are known, if the scene elements are regarded as the attributes of representational play needs, then the problem can be regarded as an aspect-based classification problem. Referred to the practice of attribute-level sentiment classification [10], we constructed auxiliary sentences by presetting sentence templates, combining scene elements and representational play requirements, and transformed attribute-level classification into a QA-like sentence pair classification problem. For example, for a corpus already associated with the need to feed alpaca: “This Saturday we went to the farmhouse to feed alpaca”, one of the auxiliary sentences is “This weekend is suitable for feeding alpaca”.

We adopted the Bert inter-sentence relationship classification model to realize sentence pair classification, as shown in Figure 9. The auxiliary sentences and corpus text are connected by [SEP] and input to the model for discrimination, and the model outputs the discrimination result (fit/not fit). Finally, we extracted the results according to the relations on all the corpus, voted the relationship between each scene element and the representational demand, calculated the score, and then determined the relationship between them.

3.5 Scene requirements layer

3.5.1 Assembly of scene requirements

In the scene requirements layer, we will assemble the information of the scene element layer and the representational requirements layer, so as to generate a large number of scene requirements. The assembled scene requirements may only contain scene elements, such as “Where to play with a 3-year-old baby on National Day?” It does not contain any figurative demand, but can contain scene elements and figurative demand at the same time. For example, in “picking strawberries in the suburbs at the weekend”, the weekend and the countryside are scene elements, while picking strawberries is figurative demand.

3.5.2 Scene requirement discrimination

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 “bestie” and “parent-child” are contradictory play scenes. To do this, we first need to calculate the relationship score between the scene elements to guide the assembly of the scene requirements. Scene elements are meaningful only when they rely on concrete requirements and match appropriate gameplay. Therefore, for the construction of a reasonable relationship between scene elements, we try to evaluate the correlation between two scene elements through relationship transmission based on the score of the relationship between scene elements and representational requirements.

In Section 3.4.2, we have quantified the score of the relationship between scene elements and representational requirements. The most intuitive idea is to calculate the relationship between scene elements through the transfer of the relationship between scene elements, representational requirements and scene elements. As shown in Fig. 10(a), the relationship scores of two scene elements, “parent-child” and “bestie”, can be obtained by taking the figurative demand “feed the alpaca” as the link.

We first construct the scene elements and representational demand relationship score matrix, considering the style number to meet the long tail distribution situation, demand representational dimensions of matrix column normalized processing, at the same time in order to keep the scene elements – scene autocorrelation coefficients of the matrix is 1, the normalized after scene elements – representational demand matrix line L2 norm normalized processing, Therefore, the new matrix obtained by multiplying the normalized matrix with its transpose matrix can be used as the scoring matrix of the relationship between scene elements and scene elements.

The relationship score among scene elements can be obtained quickly through the above method. However, this method only calculates the direct common reference strength of scene elements on the representational demand in the relationship transfer mode of scene elements – representational demand – scene elements, resulting in insufficient coverage of scene element relations. For this reason, we extend the chain transfer mode of longer node relationship, and the transfer relationship between such nodes is subject to Markov nature, as shown in Figure 10(b). But as the delivery path increases, the cost of computing increases exponentially. Therefore, the one-step time-series difference method in reinforcement learning [11] was adopted to solve the problem. The concept of “maximizing the expected cumulative return” was taken as the value of the node, the node set of scene elements was taken as the state space in reinforcement learning concept, and the node set of concrete demand was taken as the action space.

For example, when we are in the “parent-child” scene element state, we can jump to the next state “Bestie” or “Outdoor” by selecting “Feed the Alpaca” or “Role Play”. The decision function of the state jump process randomly selects a specific demand node as the decision behavior according to all the specific demand associated with the state of the current scene element, and the probability of extraction is positively correlated with the score. The state transition probability is a random jump to the associated scene elements under the decision of the representational demand node, and the jump probability is positively correlated with the score.

At the same time, for specific mutually-exclusive relationships, we formulated a reward matrix according to the actual business application requirements to achieve a diversified scoring model of scene element relationships. In this way, the node relation transfer model was transformed into Markov decision model, and the value iteration expression and node pair relation score prediction formula were derived by combining with Behrman’s optimal principle, as shown in Fig. 10(c). Formula according to the pie chart, on the premise of keep unchanged strategy by adopting the idea of bootstrap iterative calculating the value of the node, and further calculate the relationship between the scene elements, which can ensure to make better use of existing network information, improve the relationship between coverage, but also lower the constraint relation through reward matrix mutex relationship, the influence of the flexibility to adapt to different business needs.

Finally, based on the relationship score among scene elements, we selected the scene requirements with high score from the assembled scene requirements set, and generated the final expression of scene requirements according to the preset template, such as “relaxing with friends on weekends”, “playing with girlfriends”, “good place for outdoor barbecue with children on National Day”. Through the scene elements/representational play requirements contained in them, these scene requirements can be linked to the corresponding representational play requirements, and then related supplies can be associated, so as to provide users with scenario-based play solutions.

Fourth, application practice

Local life needs comprehensive atlas, covers the user scene demand and representational demand, on the one hand, more lead to participate in the user decision, in “who have decided to”, “consideration”, “choosing assessment” and so on the multiple stages influence users, reduce its decision-making cost, on the other hand, provide a more diversified supply, supply and demand matching efficiently. In terms of application mode, it is applied to various business forms such as search and recommendation.

After nearly a year of construction, the current comprehensive demand spectrum contains hundreds of thousands of core concrete requirements and scene node, and the relationship between tens of millions of, and in Meituan parent-child, leisure entertainment, medical beauty, education and training, and other business carried on the preliminary practice of application of an example by introducing the following specific application mode and application effect.

4.1 parents

Parent-child original user channel page matching efficiency between supply and demand, the ICON according to the traditional parent-child category, unable to meet the needs of users of different types (figure 11 (a) left), and guess or supply at the bottom of the form of a single, and reflect the high quality supply of user requirements, lack of decision-making information (figure 11 (b) left), thus to redesign the parent-child channel page. In order to fit the business characteristics of parent-child, we applied the demand nodes and relationships related to parent-child play to multiple traffic bits after channel revision to provide label and supply data support.

Among them, for ICONS, demand ICONS are generated across categories based on high-frequency scenes and representational demands, such as “being close to animals”, “taking baby to bubble up”, etc. (in Fig. 11(a)) and corresponding second-level pages (on the right of Fig. 11(a)). These ICONS contain similar demands in multiple categories originally and provide decision information for users in the “considering” stage.

Recommended for the bottom and we supply needs of parent-child play representational, optimization, as the supply of high quality for the contents of its associated recommendations, and for each supply extract contains corresponding text representational demand, recommended as a reason for the leakage, these sentences from the perspective of the user’s actual demand presents information, greatly attracted users (FIG. 11 (b) to the right). In addition, based on browsing and transaction behavior, the association relationship between representational demand and users is further established with supply as the medium, which is applied to the optimization of recall and ordering of personalized recommendation. The revised parent-child channel page can meet the diversified recommendation needs of users and greatly improve the user experience.

4.2 Leisure and Entertainment

In the leisure and entertainment channel page, we carried out a series of applications around the scene demand and the concrete demand. On the one hand, based on organization play scene demand new scene ICON, such as meet user outdoor play “walk in the garden”, meet the users in the tide of indoor play “play indoor tide”, meet like the nighttime users “nightlife” have fun, meet friends and colleagues party party “group”, the ICON from the user scene of the play, It breaks the restriction of traditional categories and makes the matching between users and supplies more smooth. The secondary page of each ICON will show the concrete gameplay requirements of each scene, as well as the associated merchants and contents.

On the other hand, in the scene navigation module of the channel page, it tries to use the scene requirements to further display the scene-based play information, including more than ten play scene themes such as “one person is happy”, “family warmth” and “birthday”, and makes recommendations for merchants associated with the concrete needs of these scenes. These scenario-based applications (Figure 12(a)) act on users in the “soul-stirring” stage and improve their decision-making efficiency.

In addition, some representational needs through rewriting can be used directly to the corresponding class purpose merchant list page quick sieve, live-action script to kill/desktop killed, for example, change/hanfu experience/lu pet will go/flight simulator (figure 12) in (b) left and, while we in the industry system layer class purpose segmentation result, also can be quick screening of merchants, For example, the subdivision of the bathing category (Figure 12(b), right), the application of these quick screens makes it more convenient for users to choose stores.

Fifth, summary and outlook

In the local life service, how to continuously improve the matching efficiency between the supply and the user is a difficult problem. We try to connect the supply and user by digging into the user’s needs and using them as a link. In order to comprehensively explore and understand user needs, we try to build a comprehensive demand map of local life, which is constructed layer by layer in order of industry system layer, demand object layer, concrete demand layer, scene element layer and scene demand layer, and establishes correlation relationship for various types of supply.

At present, the results of the comprehensive requirements graph can be applied to various business forms such as search and recommendation, and practical results have been achieved in more than Meituan business scenarios. However, we are still in the preliminary stage of exploration, and there is still a long way to go for iteration. Hereby, we propose some subsequent thoughts and prospects:

  • Wider industry coverage: on the one hand, deepen the construction of the existing entertainment, medical beauty and education industries, dig more nodes and relationships, and better understand user needs; On the one hand to beauty, marriage and other more industries for horizontal coverage; In addition, it will further expand to the full link of user decision-making, build a service experience map, cover the performance service link, analyze the user needs and feedback, and better enable businesses to improve user experience.
  • More data introduction: the current construction of the graph is mainly based on the text corpus of the users and merchants of the platform. In the next step, more modal data such as images will be used, and external knowledge will be introduced to improve and supplement the current nodes and relationships.
  • Deeper mapping applications: current map in search and recommended practices are mainly concentrated on the label and its associated supply direct application of the follow-up to consider further deepen the map application, make full use of the scene and scene elements of information demand, provide more accurate recommend side user intention recognition with support, so as to increase supply and users’ matching efficiency, make knowledge map more value.

reference

  • [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: 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. (in Chinese)
  • Neural Dependency Parsing [J]. Neural Dependency Parsing: Neural Dependency parsing[J]. Neural Dependency Parsing: Neural Dependency parsing[J]. Neural Dependency Parsing: Neural Dependency parsing. 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.

Author’s brief introduction

Li Xiang, Chen Huan, Zhiwei, Xiaoyang, Yanting, Xue, Cao Zhen, etc., all come from the platform technology department of Meituan to the integrated business data team.

Recruitment information

Meituan to the store platform technology department – to integrated business data team, long-term recruitment algorithm (natural language processing/recommendation algorithm), data warehouse, data science, system development and other positions of students, coordinate Shanghai. Interested students are welcome to send resumes to: mailto:[email protected].

Read more collections of technical articles from the Meituan technical team

Front end | | algorithm back-end | | | data security operations | iOS | Android | test

| in the public bar menu dialog reply goodies for [2020], [2019] special purchases, goodies for [2018], [2017] special purchases such as keywords, to view Meituan technology team calendar year essay collection.

| this paper Meituan produced by the technical team, the copyright ownership Meituan. You are welcome to reprint or use the content of this article for non-commercial purposes such as sharing and communication. Please note that the content is reprinted by Meituan Technical Team. This article shall not be reproduced or used commercially without permission. For any commercial activity, please send an email to mailto:[email protected] for authorization.