Providing personalized recommendations based on the dynamic sequential behaviors of users plays an important role in e-commerce platforms since it can considerably improve a user's shopping experience. Previous works apply a unified model pipeline to build recommender systems, without considering the diferentiated behavior patterns and intrinsic shopping tendencies on diferent pages of an e-commerce website. In this paper, we focus on generating a personalized recommender system optimized to both the View Item Page and Homepage by elaborately designing strategies for data formulation and model structure. Our proposed model (PWPRec) consists of a causal transformer encoder together with a fusion module designed for diferent pages, built on the basis of the classical two-tower structure. This provides the capability to capture a balanced long-short interest or diverse multiple interests of a user during their shopping journey across multiple types of pages. We have conducted experiments both on in-house datasets as well as public datasets to validate the efectiveness of our model, all showing significant improvements on Recall@k metrics compared to the commonly applied sequential models of recent years. Additionally, we built a state-of-the-art deep learning based retrieval system utilizing real-time KNN search as well as near real-time (NRT) user embedding updates to reduce the recommendation delay to a few seconds. Our online A/B test results show a big advantage compared to the previous GRU-based sequential model in production, with a 38.5% increase in purchased items due to model improvements and 107% increase in purchased items due to the engineering innovations.
Recommender systems play a fundamental role in ecommerce marketplaces, ofering personalized recommendation products based on a user’s specific interests which will largely improve a user’s shopping experience.
In this work, we focus on ”user context based” recommender systems that generate recommendations using a user’s historical interactions as the main context. There are several diferent landing pages which display recommendations to the user on an e-commerce platform and in this work we focus on two pages: View Item Page (VIP) and Homepage (HP). For the VIP, users usually have a specific shopping mission when they navigate a detailed item page, thus they tend to spend more time comparing similar products and trying to find the most appropriate one. Figure 1 depicts an example of a VIP with a user context based recommendations module based on users recent views. For the HP, usually at the beginning of a user’s shopping session, users tend to wander through the whole page without a specific shopping mission. They could be attracted by discounted or hot-sale products, or diversified categories they have been consistently internEvelop-O Systems and User Modeling, jointly with the 16th ACM Conference on 2. We develop a novel model architecture by com- Most prior work generates a single embedding to reprebining a causal transformer encoder with a long- sent a user, this is reasonable for recommendation pages short or multi-interest fusion module in order to or placements with specified target items. But in some generate user embedding(s). occasions such as Homepage recommendations we may 3. We deploy our recommender system to our pro- like to provide users with more diversified set of recomduction industrial setting building a state-of-the- mendations reflecting the multiple interests of a user. It art deep learning based retrieval system in the can be observed that the problem of how to capture the process. multiple interests of a user a popular topic in recent years.
Weston et al. [9] introduced a highly scalable method for
The paper is organized in the following sections. Sec- learning a non-linear latent factorization to model the tion 2 covers related approaches in literature to our multiple interests of a user. Li et al. [10] proposed a multimethod. The main model architecture is discussed in interest extractor layer based on capsule network with Section 3. The datasets and sampling strategies used for the dynamic routing mechanism. Cen et al. [11] explored our ofline experiments are then discussed in Section 4. a self-attentive method for multi-interest extraction, and An overview of our production engineering architecture utilized an aggregation module to balance accuracy and as well as A/B tests is presented in Section 5. We conclude diversity. our work in Section 6. In terms of engineering system architecture, there are several works which describe large scale embedding 2. Related Works based retrieval systems. Pal et al. [12] describes an industrial embeddings based retrieval system which uses Adding personalization in recommender systems is a well HNSW model [13] for the approximate nearest neighbor studied problem both in academia and in industrial ap- (ANN) component. There are several production systems plications. Recently, deep neural networks have been that utilize a two tower model for search and retrieval, adopted in personalized recommendations, with the abil- including in the social media space [14] as well as in eity to build a more generalized model by capturing com- commerce space [15, 16]. We will now discuss the details plex content-based features, which can also serve well in of our model architecture. cold-start situations or volatile situations. To generate personalized recommendations, the sequential behaviors 3. The PWPRec Model of users are efectively exploited by applying diferent sequential encoder networks. Many works apply Recurrent In our application scenario, we find that a user’s distribuNeural Networks (RNN) for sequential recommendation tion of recently viewed items difers for diferent pages. and obtain promising results, among those Hidasi et al. For VIP, users usually have a definite shopping purpose [1] proposed an GRU-based network to model the se- and are thus more likely to click on items related to their quential behaviors of users, and adopts the last output most recently viewed items. However, for Homepage, as the user embedding (known as GRU4Rec), Hidasi and users have a less focused shopping purpose and may click Karatzoglou [2] proposed a top-k gain ranking loss func- on diferent categories of items. Thus we build our model tion used in RNNs for session-based recommendations, in consideration of diferent pages and placements, which Li et al. [3] was also based on RNN network, but pro- can better capture and understand the diferent behavior posed a way to balance a user’s local interest and global intentions of users. interest (known as NARM). Besides RNN network, the recently well-known self-attention mechanism [4] for sequential modeling has also been commonly applied 3.1. Page-Wise Sequential Behavior in recommendations, Kang and McAuley [5] proposed Analysis a self-attention based network to capture the sequential Before introducing our detailed model, we first present behaviors of users, and the encoded value of last item an analysis of a user’s shopping behavior as a function in the sequence is regarded as the ultimate user vector of time. In our sequential modeling approach, every (known as SASRec), Sun et al. [6] adopted the Bidirec- training example is composed of a positive target clicked tional Encoder Representations from Transformers which item, several negative items a user did not click on in the trained a bidirectional model to predict the masked items impression, and a series of user historical items. in the sequence. Also, there are some methods based We build up a histogram of the overlap between the on graph neural networks proposed for sequential rec- category of the target item and historical items for all ommendation, Wu et al. [7] modeled session sequences users in the dataset. Figure 2 demonstrates the diference as graph-structured data to take item transitions into between the VIP and HP distributions. The horizontal account, Xu et al. [8] proposed a graph contextualized axis represents the number of hours between the target self-attention model for session-based recommendation. clicked item and a historical item, while the vertical axis represents the category overlap between the target and historical items. It can be seen from the graph that for the View Item Page (orange in Figure 2), about 80% of users are also viewing the same category in the first hour before, 5% are viewing same category in the second hour before, indicating that users are focused on the same category of most recent items. While for Homepage, the curve is more gradual with only 30% overlap in the first hour before, indicating that on homepage users show interest in categories they interacted with in a longer period, thus target item category may correlate to more diverse historical categories. at the same time. The overall architecture of PWPRec is shown in Figure 3. Following our previous work [17] we keep using the same structure for the item tower, and focus on optimizing the user tower. The original user tower adopted the recurrent neural network as the base encoder of user’s historical events and an average fusion strategy to generate the final user embedding. Here we optimize the user encoder network by two architectural modules: 1) a sequential encoder to better capture the ordered historical events, and 2) a fusion network to better adapt to pages with diferent historical item distributions. In the next sections, we will delve deep into these modules in detail.
The transformer network and self-attention mechanism described in [4] are widely applied in NLP related tasks, and achieved state-of-the-art performance. Here we adopt the idea of transformer and self-attention to function as the sequential encoder, and we have made some modifications in order to capture the order information which is of vital importance to recommendation scenarios.
(1) (2) (3)
Based on the above analysis, we decide to adopt different data formulation strategies and diferent model structures for diferent pages. • For the View Item Page, considering users usually have specific shopping missions and less interests in other categories, we organize training data in a ”session-based way” with most recent past behaviors in a shorter period. The ultimate output will be a singular item set showing the user’s most recent interest. • For the Homepage, users may show interest in a diverse set of categories that they interacted with, even several days before. In this case, we organize the training data in a ”user-based way” incorporating more days and more past behaviors. The ultimate output will be multiple item sets showing the user’s multiple interests through the long shopping journey. (
) = − where is the position of the target item and is the position of items prior to the target item. The relative position is then encoded into embedding P , and the ifnal input to the transformer encoder is calculated as: IN = ITEM + P where ITEM is the original item embedding, and the ifnal input vector IN is a vector addition of the item embedding with the positional embedding P .
The vanilla transformer encoder attends to any positions in a sequence by self-attention and multi-head mechanism, with each head output being formulated as:
Our proposed approach for personalized recommendations is based on a two-tower deep learning model structure to generate user embedding(s) and item embeddings (
head = ( Q, K, V) = (
QWiQ, KWiK, VWiV) QKT √ )V long-short fusion generating a comprehensive user embedding or multi-interest fusion generating multiple user embeddings. where Q, K, V are the packed matrices of queries, keys interest applied for the VIP, named as Long-Short Fuand values, is the dimension of queries and keys and sion; the other for generating multiple interests for the in self-attention mechanism [4].
Q Wi , WiK
, WiV are the parameter matrices, as described
For the sequential recommendation scenario, a causal mask [18] needs to be performed to guarantee that postclicked items cannot be seen when predicting previous items, otherwise this may lead to data leakage. Therefore, we apply a lower triangle attention mask matrix (as shown in the left part of Figure
3) to guarantee the causality between items, and in this way the self-attention can be formulated as : HP, named as Multi-Interest Fusion.
For generating one single interest, we would like to adopt a network architecture which combines a user’s shortterm and long-term interests. The short-term takes the last position output of the transformer encoder, indicating the most recent preferences; while the long-term takes outputs of all the positions into consideration, indicating their global preferences. We involve the attention mechanism to calculate a weighted average of all the outputs to form a long-term interest. which can be inter
In our industrial recommendation scenario, we design two fusion networks to handle the diferent recommendation targets, one for generating a comprehensive single represents item embeddings of all the positions, A1 transinto a latent space, A2 plays the same role , and is the sigmoid function.
After learning a long-term and a short-term embedding, the last important step is to integrate them appropriately. Here we chose the gated way to learn contribution coeficients of long-term and short-term embeddings, which is illustrated in Figure 4, and can be calculated as: U = ( = (1 − ) ×
A = ((ℎ(
In this section, we describe the dataset we utilized to train and validate our PWPRec model. We have adopted diferent data formulation strategies for the View Item Page and Homepage respectively. We also conducted experiments on both our eBay dataset and public dataset to validate the efectiveness of our model. retailer. In this dataset, only item id and category Statistics is provided to generate an item embedding.
In order to better validate the efectiveness of sequential encoders, we filter out very short sessions with sequence length of less than 4. The data statistics of the
For the Homepage, we organize the data in a user-based way within a longer time window and thus much longer user’s sequential length is obtained. Here we choose two user-based datasets in our experiments, one is collected from our eBay in-house data, the other is the public Taobao dataset [22].
• eBay (user-based) dataset
This dataset is also derived from our real world eBay production trafic containing clicked items on Homepage as the target label, and all the items that a user have viewed within 30 days before the clicked item as the sequential historical events. • Taobao dataset
This dataset contains the sequential behavior of users collected from Taobao, which consists of 1 million users shopping behaviors within 10 days. We follow the same training/validation data splitting methods as in [11].
The statistics of the above datasets are shown in Table 2. Also, we build up a histogram of the overlap between the category of the target item and historical items for the Taobao dataset in Figure 6, which shows a gradual curve more similar to the eBay HP than VIP. So we adopted the Taobao dataset to validate the efectiveness of the multiple interests model targeted for Homepage.
For model training and validation, diferent negative sampling strategies and loss calculations are adopted for View Item Page and Homepage. where v ∈ ℝ is a -dimensional vector for the embedding of user , v
∈ ℝ is a -dimensional vector for the embedding of positive item, is the afinity function between user and item (we adopt the inner product result as the afinity score), and
∪ target positive and sampled negative items. is the union set of the
As for the Homepage, data samples are grouped in the user-based way with longer sequential behaviors within a 30 days time window, batch negative sampling is adopted to select 1000 samples both in training and validation entropy with inverse temperature, 2) adopt a attentionphase. Here the loss calculation logic for training and val- based weighted sum mechanism to generate the ultiidation process is tackled diferently for accelerating the mate embedding. We call our baseline model GRU4Recconvergence of multiple user embeddings model struc- Enhanced. For our model PWPRec proposed in this ture.
• Training phase. As we have the positive item for the target label information, we can use the positive item embedding to choose one final user embedding from multiple embeddings as the one to calculate the training loss.
v = V [ (
V v )] where v ∈ ℝ is the final user embedding we select to calculate the loss in equation (8), V is the multiple embeddings genenerated for the user, and v
∈ ℝ is the positive item embedding. • Validation phase. Diferent from the training phase, label information like positive item cannot be used in metrics calculation, otherwise this would result in label leakage. Here we applied a simplified trick to fastern the procedure, which paper, we add the sufix (LS) to represent Long-Short fusion strategy.
We see from Table 3 that on both of the datasets we have depicted in Section 4.1, our model PWPRec(LS) achieved the best performance. Our model gains 10+% increase on Recall@1 compared to the baseline model GRU4Rec-Enhanced. We notice that on the YooChoose dataset, the recall values are lower, possibly because of (9) the smaller size of training set as well as the lack of item features, like titles or aspects. However, our model has a bigger advantage on this dataset even for recall with larger Ks, which implies that even in a situation where less features are available, our model can perform better and have better generalization capabilities.
The experimental results can be found in Table 4. For our model PWPRec proposed in this paper, we add the sufix is selecting one user embedding having the maxi- (MI) to represent Multi-Interest fusion strategy. Here mum summarized afinity score with the candidate item set as the final user embedding for loss and metrics calculation.
v = V [ ( ∑ ∈
V v )] where v ∈ ℝ is the final user embedding we select to calculate model metrics, V is the multiple
embeddings genenerated for the user, and v ∈ ℝ is the item embedding contained in the candidate item set for validation.
we chose the model ComiRec described in [11] as the baseline, and also the well-known multi-interest model MIND [10] for comparison. We see from Table 4 that our model PWPRec(MI), with the transformer encoder (10) and multi-interest fusion layer, outperforms the other two models by 20+% on Recall@1 metrics, and a high 10+% increase for recall with larger K. Similar to previous experiments, our model gains a more significant improvement on the public datasets which lacks item feature information.
several = 1, 5, 10, 20 . For P impressions, the metric is defined as: 1 =1 # total relevant items (11) We then explain the experimental results conducted on diferent pages with diferent datasets.
The experimental results can be found in Table 3. We select three other models for comparison: GRU4Rec [1], NARM [3], SASRec [5]. Our baseline model , described in [17], is very similar to GRU4Rec but has the following enhancements: 1) changes the loss function to cross
Details of the continuous improvements we have made to the engineering architecture of this system can be seen in our eBay Tech Blog post [23]. Most of the modeling innovations described in Sections 3.2 and 4 were A/B tested against the baseline version of the system described in our previous work [17]. In our previous approach, most of the model calculations were performed ofline with daily batch jobs to generate user/item embeddings and perform KNN for every user embedding searching over the space of item embeddings. There is a clear disadvantage, namely the delay between ofline calculation of predictions (performed daily) and displaying the recommendations to the user could lead to stale outdated recommendations and a degraded user experience.
To overcome this issue and reduce this delay to a few seconds, we built a state-of-the-art deep learning based retrieval system utilizing real-time KNN search as well as near real-time (NRT) user embedding updates, details displayed in Figure 7.
To enable fast real-time KNN search for vector em- As a result of this system architecture, the delay bebeddings, we have built an in-house KNN microservice tween generating personalized recommendations based based on HNSW [13] method, where a user embedding on the user’s session data and displaying them is reduced is sent as an input, an ANN search is performed in the to a few seconds. This system is in production serving item embedding space, and then item recommendations high volume trafic to a diverse set of users. Next we will discuss our online A/B testing results which support our ofline model evaluations.
In order to understand how our models perform online, we deployed them to serve real world users and production trafic. We compare results respectively for View Item Page and Homepage, and the NRT architecture as well.
We performed A/B testing on the View Item Page on the desktop platform comparing our PWPRec(LS) model to our previous baseline model [17] named as GRU4RecEnhanced. Our model outperformed the previous baseline with a 38.53% increase on purchases. This implies our model with transformer encoder better captures the sequential behavior of a user, and the Long-Short fusion mechanism is also a good choice to automatically balance the weights captured from long interests and short interests, better than the previous weighted sum fusion way.
In this paper, we presented an approach for generating 5.1.2. NRT A/B Test personalized recommendations by considering diferent It was interesting to see the impact of the reduced delay user behavior patterns on diferent pages in an indusbetween recommendation generation and serving on the trial e-commerce setting. Diferent strategies on data operational metrics of the system as we deployed the formulation and fusion layer adoption have been elaboNRT engineering architecture to production. The pur- rately designed to capture a user’s sequential behavior chases were improved by 107% compared to the previous on the View Item Page and the Homepage. The overofline system. This makes sense from a user experience all structure is based on a two-tower model aiming to perspective, as the user shopping journey evolves in real learn embeddings of items and users in a shared vector time, the model embedding is updated in real time, the space. To model the user’s sequential behavior, we adopt recommendation relevance quality is improved, and op- a causal transformer encoder together with Long-Short erations metrics are better. [17] fusion or Multi-Interest fusion determined by page settings on the user tower side. This approach captures the 5.1.3. Homepage Multi-Interest User Scrapes user’s long-short interests and multiple interests well. In order to verify the efectiveness of our model, we We are in the process of serving our multi-interest model have conducted experiments on our in-house datasets online for A/B test. However, we wanted to share some and commonly adopted public datasets as well. All exmulti-interest user recommendations from production en- periments showed significant improvements over the vironment to demonstrate the performance of the model. baseline approaches in comparison. Furthermore, a perWe can see the generated recommendations in Figure 8 sonalized recommender system with NRT engineering which depicts 3 distinct sets of recommended items based architecture has been launched to production and is now on the multiple interests of a user from related browsing serving recommendations at scale to eBay buyers. This history. Based on the user’s past viewed items shown on system reacts quickly based on instant user interactions the first line, our model captures three interests for this and generates large improvements in the buyer shopping user, which accurately reveals the intrinsically diverse set experience. Online A/B tests have also been conducted of interests of a user throughout their shopping journey. for our proposed model as well as the NRT architecture, which also show increases on downstream business metrics, such as purchases.
We are actively working to enhance the performance and extend the application scenario of our model as well
We wanted to thank the generous support of Sathish Veeraraghavan, Bing Zhou, Arman Uygur, Marshall Wu, Sriganesh Madhvanath, Santosh Shahane, Leonard Dahlmann, Menghan Wang, and Jef Kahn for the help with the production system as well as review comments during the manuscript preparation. as engineering system. One direction of future work is to incorporate more rich user features (e.g. demographic features like buyer age and behavioral features like purchase quantity) as well as item features (e.g. item price, popularity). Another direction is to add a deep learning ranking model after the multiple recommendation sets have been retrieved, in order to further optimize for operational metrics, like engagement or conversion. Last but not least, besides the current View Item Page and Homepage we are serving, we plan to extend our personalized recommneder system to more scenarios like infinite feed as well as checkout success placements, in order to give users more personalized and diverse choice with NRT experiences in their shopping journey. [21] Yoochoose, Recsys challenge 2015, 2015. URL: https: //recsys.acm.org/recsys15/challenge/. [22] Alimama, User behavior data from taobao for recommendation, 2018. URL: https://tianchi.aliyun.c om/dataset/dataDetail?dataId=649&userId=1&lan g=en-us. [23] Y. M. Brovman, Building a deep learning based retrieval system for personalized recommendations, 2022. URL: https://tech.ebayinc.com/engineering/b uilding-a-deep-learning-based-retrieval-system-f or-personalized-recommendations/.