The stay time of users in the takeaway scene is lower than that of traditional e-commerce, and there are higher requirements for the understanding and feedback of users' real-time needs. In response to business problems, the food delivery recommendation team has continued to invest since 2021, and finally explored a set of interactive recommendation architecture and strategies suitable for food delivery scenarios, and achieved good returns. The following will introduce in detail the challenges and solutions encountered in building interactive recommendations for the takeaway homepage feed.
-
1. Background
-
1.1 What is interactive recommendation?
-
1.2 Why do we need interactive recommendation?
-
-
2. Issues and Challenges
-
3. Main tasks
-
3.1 Interactive Recommendation Framework
-
3.2 Evaluation methods and evaluation indicators
-
3.3 Understanding User Intent
-
3.4 Recommended sorting strategy
-
-
4. Summary and Outlook
-
5. Author of this article
-
6. References
-
7. Recruitment information
1. Background
1.1 What is interactive recommendation?
Interactive recommendation is an interactive real-time recommendation product module, which mainly makes recommendations in an interactive way by understanding user needs. Interactive recommendation was proposed by Youtube in 2018 [1], which is mainly used to solve the delay of the recommendation system [2] and the problem of weak interaction with users.
Starting from the second half of 2021, the Meituan food delivery recommendation technical team will continue to explore on the food delivery homepage feed, and the full amount will be completed in the first half of 2022. The specific process is shown in Video 1: After the user enters the merchant details page from the homepage feed and exits, new recommended content is dynamically inserted into the user recommendation list. Its main advantage is to dynamically insert cards for feedback according to the real-time needs of users, thereby enhancing the user experience.
Video 1 Interactive recommendation form in the feed of the takeaway homepage
1.2 Why do we need interactive recommendation?
We found that the takeaway homepage feed has pain points in the capture and feedback of users' instant interests , and the shopping efficiency and experience of "comparative" users are not good. The takeaway homepage feed is one of the main shopping scenarios for pan-intent users. Users usually need to make some comparisons in the process of browsing to order, so as to gradually converge their intentions, and then make a final decision.
However, limited by the page-turning mode of the long list, the homepage feed is not capable of real-time adjustment of the recommendation results according to user needs. The specific performance is that the browsing depth of some users is less than one page, and the recommendation system has no additional opportunity to adjust the recommendation results according to the user's interests. Although the other part of users has a deep browsing depth, the recommendation system needs to wait until the page is turned to re-understand the user's intentions, and the real-time performance is insufficient.
The main product forms for optimizing such problems in the industry include interactive recommendation, dynamic page turning, and on-device rearrangement. Since the interactive recommendation is inserted within the user's visible range, the user's perception is strong; the latter two mainstream forms update the recommendation in the user's invisible area, and the user's perception is relatively weak. In fact, these three forms have been tried in Meituan Waimai. This article focuses on the introduction of interactive recommendations.
2. Issues and Challenges
When we build interactive recommendations in the delivery scene, we mainly face the following difficulties and challenges:
-
Different from traditional recommendation systems, interactive recommendations are those triggered by users. In the food delivery scenario, how to better match the real-time needs of users and build a recommendation system based on the terminal intelligent framework suitable for food delivery is our primary solution. The problem.
-
As a personalized module inside the homepage feed, it is not enough to optimize a single module for interactive recommendation, but also consider the overall shopping efficiency of the homepage feed. Then, how to choose the optimization target, and how to measure the effect and benefits are the second problems before us.
-
The mainstream feed form is a double-column waterfall stream of products, but the takeaway homepage feed is a single-column list dominated by merchants. How to avoid the "interference" caused by interaction on the user's selection path and trigger interactive recommendations at the right time, is the third problem we face.
-
The interactive recommendation has a dynamic insertion effect, and the user's feeling of good or bad recommendation results will be more obvious. How to better understand the user's immediate intentions and how to use the recommendation results of the homepage feed list to optimize the interactively recommended single-merchant card is the fourth problem we face.
From the above four aspects, this article will introduce in detail the whole process of building the interactive recommendation of the takeaway homepage feed from 0 to 1, as well as the solutions to the above problems.
3. Main tasks
3.1 Interactive Recommendation Framework
3.1.1 Overall link
As mentioned above, in order to realize interactive recommendation, it is very important to build a recommendation system suitable for takeaway and based on the terminal intelligent framework. The construction idea can be summarized by "4W1H":
-
Where/How : Where is the interactive recommendation card displayed? What is the display form of the interactive recommendation card? related to product form.
-
Who/When : What kind of users should interactive recommendations trigger? When is it triggered? Involves user intent understanding.
-
What : What does an interactive recommendation card need to show? Referral strategies are involved.
Based on the thinking and discussion of the above issues, we finally built a set of interactive recommendation links suitable for takeaway homepage feeds together with multiple related teams such as product, terminal intelligence, client, application service, and recommendation engineering.
Figure 1 Interactive recommendation overall link
Figure 1 above shows the whole process from "the user clicks on the homepage feed merchant card" to the display of the interactive recommendation card. After the user enters the ordering page, the client invokes the intelligent intent understanding engine of the client; After the conditions are triggered, feature processing, calculation and storage are performed, and the calculated features are passed to the client to assemble the recommendation request; the recommendation request is transparently transmitted to the shuffling service by the application service layer, and then the shuffling calls the merchant recommendation module. Recall, sorting, mechanism, revealing stages, and finally return the results to the client for display.
3.1.2 Product form
Video 1 at the beginning of the article is our final form online (inserting a single merchant card under the user’s click on the merchant), but before that, we made several rounds of attempts on the interactive recommendation card form and interaction logic.
-
In terms of card forms, we have successively explored and launched various forms such as search term cards, multi-merchant aggregation cards (as shown in video 2), and single-merchant cards (as shown in video 1), to test the impact of different card types on user purchases. Impact.
-
In terms of interaction logic, in order to avoid the "interference" of inserting animations to users' purchases, the two interactions of "covering on the clicked card" and "inserting under the clicked card" were also compared to test the impact on users' purchases.
Video 2 Interactive Recommendation Dual Merchant Card Display Style
When observing the difference in the effect of different product forms, we focused on the impact of the inserted interactive card on the turnover of 1,000 people on the homepage feed. The experimental data is shown in the table below:
Among them, UV_CXR = number of trading users/number of exposures.
During the exploration process, we also iterated the following three cognitions:
-
First of all, in the single-column list, the more original cards (higher similarity to mainstream cards), the less interference to users, and easier to be accepted by users, resulting in behaviors such as clicks and orders.
-
Secondly, whether it is search word recommendation or multi-merchant aggregation recommendation, although it seems that there is more exposed supply, the landing page link is added to the conversion link, and the actual loss will be higher (as shown in Figure 2 below) ;At the same time, because the interactive recommendation must ensure a certain degree of relevance, the form of the landing page has higher requirements for the richness of the supply, but the supply in the LBS (location-based) recommendation is relatively less, so it is more difficult.
-
In addition, "multi-store comparison" is a very common scenario during the user's purchase process. Therefore, although the click-through merchant is covered to save a hole, the negative impact it brings is greater than its positive benefit.
3.2 Evaluation methods and evaluation indicators
The goal of the interactive recommendation is to improve the overall shopping efficiency of the homepage feed, thereby enhancing the user experience. The core indicator for evaluating its revenue is based on the overall conversion efficiency of the homepage feed. However, interactive recommendations have trigger policy constraints (see Section 3.3 for details), and the proportion of traffic is relatively low. At the same time, there is a "crowding" effect on user orders with the home page feed. The overall efficiency of the home page feed can only be leveraged when the shopping efficiency of interactive cards is greatly improved. Therefore, only observing the overall efficiency of the home page feed cannot guide the iteration and effect analysis of the daily strategy of interactive recommendation. More direct and reliable indicators are needed to measure "what is a good interactive recommendation algorithm".
We evaluate interactive recommendation algorithms, mainly considering two dimensions:
-
Coverage for inserting cards
-
Matching degree of inserted card
In order to cope with the evaluation of the above two dimensions, we respectively introduce the "proportion of exposed pages" and "increment of orders at the same location" to measure the impact of interactive recommendations on the exposure of homepage feeds.
To evaluate coverage, common indicators are exposure and exposure ratio. However, after the interactive recommendation card is inserted, it will change the exposure of the homepage feed, and it is unreasonable to directly calculate the proportion of its exposure in the homepage feed. Therefore, we changed the dimension of statistical exposure from "quantity" to "page", and evaluated the coverage of cards by calculating the proportion of interactive card exposure Pages to home page exposure Pages.
Observing the proportion of Pages (hereinafter referred to as "the proportion of exposed pages") is convenient for evaluating the difference between the coverage of interactive recommendations and its theoretical upper limit. The proportion is the same as the Page CTR (click-through rate) of the homepage feed. Therefore, combined with the Page CTR of the homepage feed, we can observe the gap between the interactive recommendation coverage and its theoretical upper limit, which is convenient for continuous optimization to approach the upper limit. The page exposure ratio is not sensitive to multiple triggers on the same page. Therefore, we introduced the interactive card exposure PV ratio (the ratio of the interactive card exposure to the home page feed exposure), exposure UV (the user who exposed the interactive card The ratio of users who are exposed to the homepage feed) to assist in observing the impact on the homepage feed.
To evaluate the matching degree, the common indicator is the conversion rate from exposure to order. We believe that merchants inserted by interactive recommendations should be more relevant to users' current interests than other merchants in context. Therefore, the most intuitive indicator is to compare the conversion rate of interactive cards and other cards in the homepage feed. However, when comparing these two types of cards, there will be three deviations:
-
Crowd bias : The people exposed to interactive recommendations are users who have clicked, and the conversion rate of this part of the user group is naturally higher than that of the "big market".
-
Position deviation : The interactive recommendation card is triggered by user clicks. Since the merchants ranked in the front are often more likely to be clicked, the exposure position of the interactive card is relatively higher.
-
Resource type deviation : The feed list on the homepage involves many heterogeneous traffic such as topics and advertisements. Most of the heterogeneous traffic is not sorted based on the conversion rate.
The above three deviations make the conversion rate of interactive cards naturally higher than that of other cards in the homepage feed. Therefore, simply comparing the difference between interactive cards and normal cards in the homepage feed cannot correctly evaluate the value of the interactive recommendation itself. The product feature of the interactive recommendation is to squeeze the cards that were originally exposed on the homepage feed one by one, so only when the conversion rate of the interactive recommendation card is higher than that of the original card Only when the card next to the user’s , interactive recommendation can have a positive effect.
Based on this, we use "comparing the relative difference between the estimated conversion rate of the next natural merchant in the same request" (hereinafter referred to as the relative next difference) to measure the matching degree of the recommended card. "Same request" solves the problem of crowd The problem of deviation, "next person" alleviates the problem of location deviation, and "natural merchant" solves the problem of resource type deviation.
In addition, expanding coverage usually leads to a decline in matching. In order to balance these two indicators, we introduce "relative next-digit difference multiplied by interactive card exposure" as an auxiliary observation indicator for strategy iteration. Its physical meaning is to insert interactive After pressing down the original card, the increment of the expected number of orders generated at this position (hereinafter referred to as "order increment at the same position").
3.3 Understanding User Intent
Interactive recommendation is triggered by the user's "interaction" perceived by the recommendation system. Its process of understanding user intent mainly includes two stages: 1) Which behaviors of the user on the recommendation system can trigger interactive recommendations; 2) What is the user's immediate intent when interactive recommendations are triggered. The following will focus on these two parts.
3.3.1 First trigger strategy
In order to explore the impact of different triggering timings on indicators such as "relative next difference" and "ratio of exposure page Page" of interactive cards, we tried adding shopping carts, clicking on dishes, and staying time (10s/ 5s/2s/immediately when entering the store) and other timings. Experiments have shown that relaxing the trigger conditions will increase the risk of interactive recommendation unbelief triggering, resulting in a decrease in the efficiency of interactive cards, but it will increase the exposure of interactive cards, cover more users, and facilitate subsequent strategy iterations. In the end, we adopted the first trigger strategy of "as soon as the user enters the store from the homepage feed".
3.3.2 Continuous trigger strategy
Since the duration of the user’s stay on the merchant’s details page cannot be determined, and it is impossible in engineering to request the recommendation service and display the results at the moment the user returns to the list, the system needs to request recommendations from the server multiple times during the user’s browsing of the merchant’s details page As a result, the longer the user stays on the business details page and the clearer the needs, the more accurate the recommendation results of the server will be. Therefore, we adopt the "continuous triggering strategy", that is, as the user's stay in the store increases, or after the user generates new dish clicks or additional purchase features, the client will initiate multiple consecutive requests to the back-end service to update the recommendation results .
3.3.3 Understanding of real-time user needs
How to better understand user intent through end-to-end intelligence [3-4] is the focus of our attention. Compared with the server, the characteristics of users on the end mainly have the following two characteristics:
-
Better real-time performance : from "quasi real-time" to "super real-time" interaction.
-
Finer dimension : Evolved from "interaction item" to "micro-granularity of item interaction".
Therefore, with the help of terminal intelligence capabilities, we are no longer limited by the paging request update mechanism of the home page feed. We can better understand user needs based on user behavior, make real-time intelligent decisions to update recommendation results, and alleviate the problem of feedback signal perception delays.
The main behavior of the user in the store after clicking the merchant card can help us better understand the real-time needs of the user. Figure 4(a) shows some in-store behaviors, and Figure 4(b) analyzes some of the different behaviors compared to viewing the merchant's introduction behavior. , and there is a difference in takeaway orders on that natural day), indicating that there are obvious differences in the needs of users under different behaviors.
Figure 4(a) Main behaviors of users in the store
Figure 4(b) Differences in the day-to-day order rate difference between the main behaviors of users in the store and "view comments"
3.4 Recommended sorting strategy
The homepage feed shows the entire listing at a time, while the interactive recommendations only show 1 business card at a time. In order to recommend more accurate results, interactive recommendations are needed to better understand users' real-time food delivery needs. Therefore, on the basis of the recommended link on the homepage feed, we optimized the "recall->sort->mechanism->display" link to cover more users while continuously improving the matching degree of interactive recommendations and user interests .
3.4.1 Recall
The recall phase is divided into two steps (as shown in Figure 5 below):
-
Using multiple recall algorithm strategies to recall hundreds of candidate POI merchants from the user's vicinity.
-
Use the similar category filtering scheme to filter merchants that are obviously different from the current user intention, and pass the generated candidate results to the sorting stage. In order to better understand the user's immediate intention, we propose the Item2Item Multi-Trigger recall and similar category filtering scheme.
Figure 5 Flow chart of the recall phase
First of all, on the one hand, we directly reused the recall link recommended by the homepage feed, and integrated multiple recall algorithms such as Twin Towers recall [5], User2Item recall [6], and hot sale recall. On the other hand, in order to strengthen the understanding and attention to the user's immediate intention, we added a new item2item Multi-Trigger bypass recall.
The specific method is: we use the POI that the user clicks and purchases on the homepage feed as the trigger for the Item2Item recall, and recall more merchants that meet the user's immediate intention. The number of Triggers for each user is different, and the number of merchants recalled by each Trigger is also different, and the number satisfies N/M (N is the total number of POIs recalled by I2I Multi-Trigger, and M is the number of Triggers).
Secondly, the merchants that the user has recently clicked on can help us better understand the user's immediate intentions. Considering the majority of "contrasting" users in the takeaway scene, in order to bring users a better experience, we propose the "same leaf category" strategy: interactive card merchants that restrict exposure must be the same as the leaf category that triggers the merchant ( It reflects the taste of the merchant, which is related to the main dishes, such as skewers, chicken rolls). But this solution will bring 2 problems:
-
Under the constraints of LBS, there are fewer merchants of the same leaf category, resulting in less exposure of interactive cards.
-
In the takeaway scenario, the category definition of merchants has different granularities (each merchant includes multiple levels of categories from coarse to fine). Interactive recommendation requires a unified category definition method. It is necessary to ensure that the recommended merchant category and user interest are high. Correlation requires a certain degree of diversity in the recommendation results.
Therefore, we refer to the definition of existing merchant categories, consider the dimensions of merchant taste, consumer similarity, and the distribution of commodity categories included in the merchant, and redefine similar categories for interactive recommendation through clustering. Specifically, we define about 200 fine-grained categories as about 70 coarse-grained categories, which not only meet the needs of "comparative" users, but also bring novelty and diversity to more users. experience.
The strategy we proposed, while significantly increasing the Page ratio of the interactive card exposure page, compared with the next one, the difference is significantly improved. The specific effect can be seen in the table below:
3.4.2 Sorting
In the sorting stage, the main task of the model is to predict CTR and CXR (exposure conversion rate), and pass the estimated results to the mechanism stage.
To optimize the ranking model for interactive recommendation, we mainly face the problems of sample distribution differences and few training samples. There is a natural difference between the single-merchant card form of interactive recommendation and the list form of the homepage feed, which leads to obvious inconsistencies in the sample distribution (such as click-through rate, conversion rate, and crowd distribution). The recommendation model that directly uses the homepage feed lacks the support for interactive recommendation. The effect of scene personalized attention will be significantly attenuated. The simple way is to directly use the samples of interactive recommendation to train the model, but the number of single-scene samples of interactive recommendation is small, which will lead to insufficient robustness of the model.
Therefore, we chose the fine-tune method commonly used in the industry. Based on the homepage feed sorting model, we used the interactively recommended sample Fine-tune sorting model. At the same time, we make full use of the user real-time demand understanding module built in Section 3.3.3 to optimize the model effect. Of course, we have also explored how different network structures can improve the model effect, but due to constraints such as computing power resources, we have not launched a more complex interactive recommendation ranking model. The specific model structure is shown in Figure 6 below.
The Embdding data of the input model is passed through the MMoE[7] layer and the 3-layer MLP network to obtain the predicted pCTR and pCXR results. Among them, the input features of the model are divided into five types: 1) user features; 2) merchant features; 3) contextual features; 4) sequence features; 5) merchant features that trigger interactive recommendations. Sequence features, including real-time exposure, click and other sequences, and use Micro-Behavior[8] details.
The characteristics of merchants that trigger interactive recommendations include embedding representations of merchants, delivery information, discount information, etc. The offline/online effects of the sorting model are shown in the table below. It can be seen that compared with the homepage feed sorting model, the optimized interactive recommendation sorting is significantly better than the effect of directly using the homepage feed sorting model in terms of card efficiency.
3.4.3 Mechanism
In order to carry business goals more flexibly, we have introduced a mechanism module, the goal is to adjust the order of the candidate merchants passed in the sorting stage according to different business goals (such as CTR, CXR, novelty, etc.). Finally, with the goal of novelty in mind, interactive recommendations are sorted by estimated CXR to maximize the conversion rate of cards and lists. At the same time, we also optimized the experience from two aspects: solving negative feedback and optimizing the experience:
-
Resolve negative feedback through business rule constraints, mainly including: repeated exposure merchant filtering, pre-order merchant filtering, same-brand merchant filtering, user dislike & blacklist merchant filtering, and avoid inserting high delivery fees and delivery distances merchant.
-
In terms of interactive experience: 1) Provide users with a richer recommendation experience by exploring novelty and other goals; 2) Explain the reasons for users through the optimization of recommendation reasons.
3.4.4 Leaking out
In the revealing stage, it is mainly to judge whether the top 1 merchants transmitted in the mechanism stage are displayed to the user. Theoretically, every time a user "triggers" an interactive recommendation, the system may recommend a new merchant for display. However, the recommendation strategy that does not consider the quality of merchants will greatly damage the user experience and the effect of the home page feed. Therefore, we explored the revealing strategy of the card, that is, whether the Top 1 card transparently transmitted in the mechanism stage is displayed.
As shown in Figure 7 below, the merchant display area is divided into four areas: ABCD: interactive recommended card insertion position (A), trigger merchant (B), trigger merchant above (C), and trigger merchant below (D). After the interactive card is inserted, the first merchant in area D will slide down, and the animation effect will attract the user's attention to the interactive card A. However, whether the user places an order in the interactively recommended merchant A is not only related to whether it meets user preferences, but also inseparable from the comparison effect with the contextual merchants C, B, and D—at least, A should be better than C, B, and Merchants in area D are more in line with the user's current intention.
We pay more attention to the conversion rate of the home page feed. Therefore, when the pCXR of an interactively recommended merchant is higher than that of other merchants in the same context, this merchant should be exposed. The format is:
Interactively recommend merchants to the next merchant
Then, there are a few questions: Which merchants are compared in areas B, C, and D? How much higher should the pCXR of the interactively recommended merchant A be compared with the comparison merchant (controlled by the proportional coefficient α in formula 1) before it should be displayed? For the latter, we obtained through experiments; for the former, we analyze as follows:
-
Area C: The user has already browsed, and the possibility of clicking/ordering is lower. Obviously, it doesn't make much sense to compare businesses in this area.
-
Area B: the merchants of the user's "last click" && "triggered interactive recommendation", the user is very interested in this merchant. The comparison with it seems to have a clear meaning, but the merchant can be exposed, which means that it is a leader among similar/similar merchants. It is difficult to find a merchant with better pCXR without changing the ranking model/features.
-
Area D: The user has not browsed it. Once the interactive merchant card A is displayed, the dynamic effect of inserting the card will make the user pay more attention to this area. Therefore, it is more intuitive to compare with these merchants.
Due to the limitation of the revealing conditions of the card, the exposure of the interactive card is significantly reduced. Experiments have shown that when the pCXR of the interactive recommendation card is higher than that of the next merchant's card, the increase in orders at the same location is the highest, the loss of the exposure page Page ratio is the least, and the strategy is optimal. We adopt this scheme. It can be seen from the experimental data that when comparing the average value of N-bit pCXR, as the value of N changes, the exposure and efficiency of the interactive card will be affected, and the effect is equivalent to directly adjusting the filtering threshold α of pCXR. In the actual production environment, it is enough to select the parameter α when the "interactively recommended same-position order increment" is high, and here we take 1.
4. Summary and Outlook
This article introduces our attempts to explore interactive recommendations on the homepage feed, which mainly includes:
-
Relying on the ability of terminal intelligence, combined with the user's "multi-store comparison" shopping characteristics in the takeaway scene, an interactive recommendation system of "dynamically inserting single-card merchants" was built.
-
Fully considering the impact of inserting cards on the feed context of the home page, indicators such as "increment of orders at the same location" are proposed, and an evaluation method of "what is a good interactive recommendation system" is constructed from two aspects of matching degree and coverage.
-
Starting from business understanding, user demand modeling, etc., by optimizing the "trigger->recall->sort->mechanism->disclosure" link, the accuracy of the system's understanding of user intentions is improved and user experience is optimized.
At present, the interactive recommendation has been fully loaded on the homepage feed, and we have also reaped the following business benefits:
-
The turnover of 1,000 people on the homepage feed +0.43%, and the novelty of exposure on the homepage feed +1.16%.
-
Interactive cards have a +132% conversion rate to the next organic merchant.
In the future, we will explore and optimize from the following directions:
-
Optimize product form : continue to optimize the product function of interactive recommendation from multiple perspectives such as customized recommendation reasons and trigger timing, and expand the dynamic recommendation capability to other scenarios of takeaway.
-
Carrying more business goals : On the premise of satisfying the precise recommendation of users, it integrates multiple differentiated business goals such as novelty and diversity, and performs modeling optimization.
-
Expand the intelligent advantages of the end : the existing interactive recommendation system puts feature processing, recall, sorting, mechanism and other processes on the server to complete, but the performance of the server <-> cloud limits the processing and utilization of more information, which can be used in the future Put it on the terminal to complete training and estimation, and effectively protect the privacy of users while realizing the ultimate personalized experience of "thousands of people and thousands of models". At the same time, we can use the advantages of terminal intelligence to explore solutions for reordering on the terminal.
5. Author of this article
| Ji Chen, Yacheng, Wang Wei, Jackie Chan, Jiang Fei, Wang Cong, Bei Hai, etc., from Daojia Business Group/Daojia R&D Platform/Search and Recommendation Technology Department.
| Shu Yang, Zhang Jing, etc., from Daojia Business Group/Takeaway Business Department/Product Department.
6. References
-
[1] Christakopoulou K, Beutel A, Li R, et al. Q&R: A two-stage approach toward interactive recommendation[C]//Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. 2018: 139-148.
-
[2] Xinran He, Junfeng Pan, et al. Practical lessons from predicting clicks on ads at facebook. In Proceedings of the Eighth International Workshop on Data Mining for Online Advertising. ACM, 1–9, 2014.
-
[3] Gong Y, Jiang Z, Feng Y, et al. EdgeRec: recommender system on edge in Mobile Taobao[C]//Proceedings of the 29th ACM International Conference on Information & Knowledge Management. 2020: 2477-2484.
-
[4] The application practice of terminal intelligence in the search reordering of Dianping .
-
[5] Po-Sen Huang, Xiaodong He, Jianfeng Gao, Li Deng, Alex Acero, and Larry Heck. 2013. Learning Deep Structured Semantic Models for Web Search Using Clickthrough Data (CIKM ’13). Association for Computing Machinery, New York, NY, USA, 2333–2338.
-
[6] Yue Shi, Martha Larson, and Alan Hanjalic. 2014. Collaborative Filtering beyond the User-Item Matrix: A Survey of the State of the Art and Future Challenges. ACM Comput. Surv. 47, 1, Article 3 (may 2014), 45 pages.
-
[7] Jiaqi Ma, Zhe Zhao, Xinyang Yi, Jilin Chen, Lichan Hong, and Ed H Chi. 2018. Modeling task relationships in multi-task learning with multi-gate mixture-of-experts. In Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining. 1930–1939.
-
[9] Zhou G, Zhu X, Song C, et al. Deep interest network for click-through rate prediction[C]//Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining. 2018: 1059-1068.
---------- END ----------
Job Offers
Search Recommendation Algorithm Engineer
-
Develop large-scale deep learning, graph learning and other technologies, use modules such as attention mechanism, memory network, and relationship network to understand user needs and discover user interests from massive data across multiple time and space scenarios, and optimize click-through rate and conversion rate models. Show users more suitable and interesting delicacies and products.
-
Develop reinforcement learning, explainable deep learning, multi-modal learning, multi-objective optimization and other technologies, optimize rearrangement and mixing models, intelligently control traffic distribution, optimize platform ecology, and achieve a win-win situation for consumers and businesses.
-
Using technologies such as knowledge graphs, computational vision, and natural language generation, it helps merchants automatically and intelligently generate display content and copywriting based on user interests, improving promotion efficiency.
-
Track and study the cutting-edge technology of artificial intelligence, and explore the application of technology in retail and medical e-commerce scenarios.
Interested students can send their resumes to: [email protected]. Looking forward to working with you to create a better future.
recommended reading
| The Practice and Exploration of Meituan Waimai Recommendation Intelligent Traffic Distribution
| Quality Model and Practice of Meituan Comprehensive Business Recommendation System
| TensorFlow GPU training optimization practice in Meituan food delivery recommendation scene