COUPA: An Industrial Recommender System for Online to Offline Service Platforms

Recommender systems, which aim at matching user interests, are playing vital role in various online websites and mobile applications, including Taobao (Gong et al., 2020), Youtube (Covington et al., 2016) and Google Play (Cheng et al., 2016). To fully mine user preference in online services, numerous methods are proposed to learn the effective matching function between target users and items with subtle feature engineer (Xiao et al., 2017; Lian et al., 2018; Guo et al., 2017) or abundant historical records (Zhou et al., 2018, 2019; Hidasi et al., 2016; Kang and McAuley, 2018; Hou et al., 2022). Although these methods achieve superior performance to some extent, we argue that they are unsuitable to recommendation scenarios for Online to Offline (O2O) service platforms (e.g., Meituan ¹¹1https://about.meituan.com/en, Grubhub ²²2https://www.grubhub.com/ and Uber Eats ³³3https://www.ubereats.com/), where recurrence based temporal patterns are ubiquitous.

In the case of Alipay ⁴⁴4https://www.alipay.com/, whose goal is to guide user for locally discovering retail services including entertainment, travelling, delivery and other services, there is a clear and strong need to capture dynamic preference over time for recommendation. Concretely, we illustrate an example in Figure 1. There are numbers of channels (e.g., “Food Delivery” and “Travel” channel) in our scenarios, each of which represents a certain business field. Besides, several items are displayed under each channel, which are called contents of channel. Taking the contents of “Food Delivery” channel ⁵⁵5Obviously, similar findings can also be revealed for other channels (e.g., “Travel” channel). as an example (as shown in the left part of Figure 1), users may be interested in daily necessities (e.g., “Coffee”) or weekly intents (e.g., “HotPot”), leading to different click demands at different times. On the other hand, as shown in the right part of Figure 1, since the feeds-like styles are widely employed in the recommendation scenarios of O2O service platforms, users typically scan the screen of mobile phone from top to bottom, while their attention distributions are indeed distinct. That is, users may prefer clicking items with certain positions regardless of the relevance, which brings so-called personalized position biases to hurt the recommendation effectiveness.

In light of the above observations, we believe it is of crucial importance to design an industrial recommender system for O2O service platforms, centering on characterizing user preference from time and position bias aware perspectives simultaneously. The idea is appealing, while the solution is challenging in real recommendation scenarios, which is summarized as follows: (C1) Time aware preference modelling: how to effectively capture recurrence based preferences for users over continuous time? Recurrence based temporal patterns are ubiquitous in recommendation scenarios for local services. For example, a certain user may periodically visits specific malls and enjoy specific delicacies (e.g., Hot Pot) in a reasonable period (e.g., weekend). Therefore, explicitly capturing such temporal patterns over continuous time is potential to benefit for the preference characterization in real recommendation scenarios. (C2) Position aware preference modelling: how to deal with user’s preferences for positions and build a debiasing model in a personalized way? On account of the feeds-like exhibition in our scenarios and users’ browsing habits, position biases widely exist in the behavior data. Moreover, users have different preferences towards each position. For example, most users prefer clicking items ranked first whereas some users have preferences for other certain positions (e.g., the second and even last position). Blindly fitting the observational behavior data without considering the inherent position biases may seriously deteriorate the recommendation effectiveness. (C3) System design: how to devise the base system to support the complicated temporal interaction scenarios? Our temporal interaction scenario put forward the urgent request of low delay for system design since our approach is expected to have awareness of user real-time behaviors for online serving. Moreover, building the complete chain to balance the trade off between data storage and timely feedback as well as guaranteeing low delays for online serving also remain challenging.

To this end, we propose COUPA, a novel COntinuoUs time and Position bias Aware recommender system targeted for O2O service platforms (e.g., Alipay). Specifically, we borrow the idea of functional time encoding, a promising way to embed timestamp into vector space, and propose a novel continuous time aware point process with an attention mechanism to explore the excites of inhabits from previous interaction records in continuous time (C1). On the other hand, inspired by the idea of knowledge transfer, a position selector component, cooperated with a position personalization module is elaborately designed to perform personalized position debiasing in our scenarios (C2). To satisfy the requirements of online serving, the implementation and deployment are carefully designed to not only collect user behaviors as well as corresponding positions in a real-time manner, but also perform efficient online inference in a two-stage mode (C3).

In sum, we make the following contributions:

• •

  We analyze and highlight the importance of capturing time and position aware preferences in Alipay, which is potentially generalized to the universal recommendation scenarios for O2O service platforms (e.g., Meituan, Grubhub and Uber Eats).

• •

  We propose COUPA, which is not only capable of modeling continuous time aware preference for recommendation, but also well deal with position bias in a personalized manner.

• •

  System deployment and corresponding implementation details are uncovered, where edge, streaming and batch computing are jointly employed for capturing the real-time user behaviors, and a two-stage online serving mode is also carefully designed.

• •

  We conduct extensive offline and online experiments, demonstrating that COUPA consistently achieves superior performance and provides intuitive evidences for recommendation.

In this section, we give an intuitive analysis for user preference w.r.t. temporal pattern and position existed in recommendation scenarios for O2O platform, with Alipay as an example. All of the real data are collected within one month, consisting of user interactions with channels and contents under each channel.

Temporal Pattern. As mentioned above, recurrence based temporal patterns are widely existed in recommended scenarios for O2O services. Since users may be interested in daily necessities or weekly intents, which implies that the reasons of users accessing O2O service platform (e.g., Alipay) are periodical. With the real data, we aim at digging out the willingness of users to purchase items under the same categories w.r.t. time interval.

We respectively select the categories “Hot Pot” and “Coffee” and plot the corresponding intensities (i.e., the number of purchase) against time intervals in Figure 2 (a) and (b). Here, $t=0$ refers to the timestamp of first purchase. From results in Figure 2 (a), we observe that the intensity reaches local peak when $t=24\cdot n(n=1,2,...)$. Moreover, among these peaks, the intensity is more large when $t=7\cdot 24\cdot n(n=1,2)$. These observations show that the intention of purchasing “Hot Pot” has strong daily and weekly periodicity. On the contrary, the intensity under the “Coffee” achieve a peak with 24 hours after the first purchase, whereas it sharply declines as time passes. Both cases reveal that user behaviors have strong temporal patterns w.r.t. different categories of items in our recommendation scenarios, which is vital information that should be fully considered in our approach.

Position bias. Click distributions are always distinct over positions for different users in feeds-like scenarios of O2O platform, which brings in personalized bias in modelling caused by that user feedback data is observational rather than experimental.

We conduct a micro-view analysis based on 3 types of users with different click distributions over positions in Figure 2 (c), (d), (e). Specifically, “User type 1”, whose behavior is the most common case, prefers clicking items with higher positions (i.e., Position “0” and “1”), whereas “User type 2” is more likely to click position “2” and “3”. As for “User type 3”, who treat each position almost equally, their preferences are more stable. In sum, we conclude that position bias is not only dependent, but also personalized, which has not been sufficiently explored in previous studies.

Functional time encoding aims to find a mapping $\Phi:\mathcal{T}\rightarrow\mathbb{R}^{d}$ from time domain $\mathcal{T}$ to $d$-dimensional vector space, targeting for preserving evolving nature of user interest/intent. Intuitively, the temporal pattern related to the timespan between any two timestamps $t_{1},t_{2}\in\mathcal{T}$ can be denoted as inner product between their functional encodings, i.e., $\langle\Phi(t_{1}),\Phi(t_{2})\rangle$. Therefore, we formulate above temporal patterns with a translation-invariant kernel $\mathcal{K}$ with $\Phi$ as the mapping function associated with $\mathcal{K}$.

Suggested by the Mercer’s Theorem (Minh et al., 2006; Xu et al., 2019a; Hu et al., 2022), we formulate the mapping function $\Phi$ with frequency parameter $\omega$ as follows:

With the help of the nice truncation properties provided by such a Fourier series-like form, we truncate above mapping function $\Phi_{\omega}^{\mathcal{M}}(t)$ as $\Phi_{\omega,d}^{\mathcal{M}}(t)$. Subsequently, by concatenate multiple truncated periodic mapping function with the frequency set $\{\omega_{1},...,\omega_{k}\}$, we represent time with functional encoding as:

Temporal point process has been commonly adopted to model dynamic in sequences, which is a kind of stochastic process that generates a list of discrete events at different times, denoted as $\mathcal{S}=\{v_{i},t_{i}\}_{i=1}^{n}$ where $v_{i}\in\mathcal{V}$ is the type (i.e., item in our study) of $i$-the event and $t_{i}\in\mathcal{T}$ is the timestamp of the $i$-th event. The process is well characterized via the conditional intensity function $\lambda^{*}(t):=\lambda(v,t|\mathcal{H}_{t})$, which makes prediction for next arrival time $t$ based on the history $\mathcal{H}_{t}=\{(v_{i},t_{i})\in\mathcal{S}|t_{i}<t\}$. Here, the $*$ symbol reminds us of dependence on $\mathcal{H}_{t}$. The core of the temporal point process is to design the conditional intensity function $\lambda^{*}(t)$ for capturing various interests. As consequence, a series of typical point processes equipped with well-designed intensity functions are proposed, including Possion process, Hawkes process and recently proposed fully neural network based proint process (FullyNN).

Maximum likelihood estimation (MLE) is commonly adopted to learn the parameters of temporal point process, and the corresponding likelihood over a time interval $[0,T]$ is given by :

where the first term models the sum of log-intensity functions of past events and the second term is the log-likelihood of infinitely many non-events, where the negative sampling strategy is always adopted.

In this section, we present COUPA, a novel COntinUous time and Position Aware recommender system for O2O platforms, whose model part mainly focuses on user preferences towards temporal patterns (i.e., time aware preference modelling) and positions (i.e., position bias aware modelling). The overall architecture of COUPA is illustrated in Figure 3.

Before the elaboration of the model design for COUPA, we give a brief introduction for the inputs, which clearly consist of position related features (e.g., the sequence of user click positions and the corresponding item ids), user profile, context feature and user behaviors (including item ids and category ids) from left to right in Figure 3. Following common strategies adopted in (Guo et al., 2017; Lian et al., 2018), we transform the involved features into low-dimensional representations (called embedding) by look-up operation, concatenation and multi-layer perceptron successively. At last, given a user $u$ and an item $v$, we represent the embedding for position related features as $\mathbf{p}_{u,v}$, user profile as $\bm{e}_{u}$, context features as $\mathbf{c}_{u,v}$ and item profile (i.e., item id and category id) as $\bm{e}_{v}$, respectively.

In this section, we introduce the online deployment of the proposed COUPA in Alipay ⁶⁶6It is noteworthy that COUPA is also adapted for arbitrary recommendation scenarios in O2O service platforms. to serve ranking tasks for channels and contents of each channel simultaneously, where the system design is still challenging. Most importantly, our temporal sequential model has urgent requirements of low delay for collecting real-time sequential data. Compared with traditional methods (e.g., DeepFM), COUPA is relatively more complicated. It drives us to avoid amounts of delays for online inference, which needs to support high QPS (queries per second). To tackle these challenges, we carefully design the industrial system, roughly comprised of three main modules: data processing, two-stage online serving and offline training.

Data processing. User feedback data (i.e., exposure and click) are necessary to train the recommender system, where the trade off between data storage and timely feedback need to be balanced. Hence, our principle is to store 90-day click data in huge volume via cheap storage that only guarantee latency in days, and timely feedback data in tiny volume via edges that guarantee latency in seconds. As such, we design the following batch, streaming and edge computing in a cooperative manner, followed by a sequence fusion procedure.

Batch computing. Based on the powerful MaxCompute ⁷⁷7https://www.alibabacloud.com/product/maxcompute, an offline and low-cost computing platform developed by Alicloud ⁸⁸8https://www.alibabacloud.com/, we perform batch computing for generating user click sequences in the last 90 days, involving hundreds of billions of interaction records, with a latency of days.

Streaming computing. Streaming computing is mainly based on the frigate system (similar to Kafka ⁹⁹9https://kafka.apache.org/ and Storm ¹⁰¹⁰10https://storm.apache.org/ system) in the Ant Group, which directs at generating user click sequences in the last 2 days. Due to the affects of log back and real-time reporting, this part of the data has a latency of tens of seconds.

Edge computing. Due to the urgent real-time demand of COUPA, we employ edge computing for collecting click sequences on the client in completely real time with a latency of only few seconds. Specifically, user click data will be captured by the client at once after reported, and recorded into local database. As soon as a user proposes a request, the client timely processes his/her click data in the last 3 hours and sends them to online servers.

Sequence fusion. Sequence fusion is applied to fuse user click sequences in different time spans. It stitches the sequence data, generated by batch, streaming and edge computing, in chronological order and eliminates duplicates with timestamps. Subsequently, a complete sequence in the last 90 days is passed to COUPA as input.

Two-stage online serving. This module is designed to relax delays of the online system for COUPA with two main stages. Stage I: Channel Ranking Worker performs the coarse-grained ranking on channels with dozens of candidates based on COUPA, ignoring features of contents under channels. Then only top $M$ (i.e., $M=5$) channels are selected to request their corresponding contents (i.e., “Food Delive” and “Trave” channels) recommender systems with tens of millions of candidates in Content Recommendation Worker, whose ranking model is COUPA. Stage II: With contents under channels are determined in Stage I as well as collected content features, Channel Ranking Worker performs the fine-grained ranking considering content features on top $M$ channels to decide their displayed orders. In this way, system latency is greatly relaxed and ensures the stability of COUPA for online inference.

Offline training. This module aims at the training of the COUPA based on the user historical click records and online feature logs. With help of the MaxCompute and PAI ¹¹¹¹11https://www.alibabacloud.com/product/machine-learning platform, we are capable of training our model on tens of billions of data, whose details are stated in the Section 4.

In this section, we conduct a series of offline and online experiments to demonstrate the effectiveness of COUPA. In short, we aim at answering the following three research questions:

• •

  RQ1: Does our proposed model outperforms other state-of-the-art methods on both temporal recommendation and position debiasing task.

• •

  RQ2: Does our proposed model achieve superior performance in real-world scenarios.

• •

  RQ3: Does our proposed model intuitively provide convincing evidences for recommendation.

Recent years have witnessed the great success of sequential recommendation methods, which devote to predicting user interests based on his/her historical behavior sequences. Early works mainly focus on modelling behavior sequences with Markov chains (Rendle et al., 2010; He and McAuley, 2016; Zimdars et al., 2001). Due to the ability of capturing complex feature interaction, the deep neural network (DNN) has yielded excellent performance in various applications. To marry up the advantage of DNN and sequential recommendation, a series of methods attempt to extract user preferences from his/her previous behaviors through recursive neural network (RNN) (Hidasi et al., 2016; Tan et al., 2016), convolutional neural network (CNN) (Xu et al., 2019b; Tang and Wang, 2018; Yan et al., 2019) and memory network (Chen et al., 2018) based structures. Meanwhile, recently emerging neural attention mechanism (Vaswani et al., 2017) is commonly adopted in characterizing long- and short-term preferences in a more fine-grained principle (Zhou et al., 2018, 2019; Kang and McAuley, 2018; Luo et al., 2020). Furthermore, many recent studies have proved that temporal information is of crucial importance for recommendation, which implies users’ underlying reasons to click/purchase a target item. Therefore, a few works are proposed to incorporate such temporal information into sequential recommendation with extended LSTM structure (Zhu et al., 2017), temporal point process (Bai et al., 2019) and time aware neural network (Ye et al., 2020). Here, we please refer the reader to recent surveys (Fang et al., 2020; Wu et al., 2021) for a thorough review,. Unfortunately, due to the time sensitivity of user preferences, how to fully exploit influences derived from historical interaction records still remains challenging. Moreover, in the industrial settings, automatically collecting real-time sequential data and guaranteeing low delays for on online serving also need to be carefully considered.

On the other hand, recent years have seen a surge of efforts on exploring the impacts of the biases and performing necessary debiasing in recommender systems. In our paper, we center on the widely studied position bias, which denotes that users are more likely to click the higher ranked items regardless of the corresponding relevance. To solve the position bias, a number of works have been proposed, which roughly fall into two lines. The first line makes some basic hypotheses about the user browsing behaviors. Specifically, these methods maximize the likelihood of the observed interaction data for the true relevance feedback with self-defined examination hypothesises (Craswell et al., 2008; Dupret and Piwowarski, 2008; Chapelle and Zhang, 2009), personalized transition probability based cascade models (Guo et al., 2009; Zhu et al., 2010) and state-of-the-art deep recurrent survival model (Jin et al., 2020). The other line is based on the recently emerging counterfactual learning (Joachims et al., 2017; Wang et al., 2018), which weights each sample with a position aware value and collects the user feedback with inverse propensity score (Agarwal et al., 2019). Following this line, some recent works (Qin et al., 2020; Ai et al., 2018) further adopt the dual leaning for jointly learning a propensity model as well as a recommendation model with well-designed EM algorithms. To learn more works about debising in recommender systems, please refer to the elaborate survey (Chen et al., 2020). Nevertheless, these methods mainly assume the independence of positions, while the personalization of position bias is still unexplored.

In this paper, we proposed a novel continuous time and position aware recommender system for O2O service platforms (e.g., Alipay), called COUPA, which comprehensively takes user preferences towards temporal patterns and position biases into consideration. To better learn temporal patterns in real-world applications, we develop a continuous time aware point process equipped with continuous time aware attention mechanism to chronologically summarize influences derived from historical behaviors for recommendation. Moreover, a position selector module cooperated with a Multi-gate Mixture-of-Experts (MMoE) block and a Gated Linear Units (GLU) black is introduced for mitigating position bias in a personalized manner. At last, we devoted to the design and implementation of our whole system, which jointly employs edge, streaming and batch computing for capturing the real-time user behaviors and adopt a two-stage mode for efficient online serving. Extensive offline and online experiments demonstrated the superiority of our proposed COUPA.

Currently, our approach is able to effectively extract user preferences from historical interaction records for recommendation. While, there exist a number of users/items with sparse interactions in real-world scenarios, for which current approach fails to learn high-quality representations. As future work, we will consider how to incorporate graph structure data (e.g., social network and knowledge graph) for compensating these cold-start users/items. In addition, we will also consider adapting our approach to well deal with geographical information, which plays an important role in O2O service platforms.