In e-commerce, ranking algorithms based on relevance and engagement signals have often shown improvement in sales and gross merchandise value (GMV). Designing such algorithms becomes particularly challenging when serving customers across diverse regional markets, as shopping preferences and cultural traditions vary significantly. We propose the SEQ+MD framework, which combines sequential learning for multi-task learning (MTL) with a region-based feature mask for handling multi-distribution data. This approach utilizes the sequential order within tasks and accounts for regional heterogeneity, enhancing performance on multi-source data. Unlike traditional sequential models that rely on tracking user interaction histories, SEQ operates on user-item feature pairs and generates task-specific predictions in sequence. Moreover, SEQ supports eficient parameter sharing across tasks and allows new tasks to be added easily. Notably, SEQ trained on data from only two tasks outperforms the baseline model trained on data from all three tasks when evaluated on the full three-task setting. Experiments on in-house data showed significant gains in high-value engagements, including add-to-cart and purchase actions. Furthermore, our multi-regional learning module can be flexibly applied to enhance other MTL applications.
In e-commerce, the design of item display algorithms is crucial for enhancing the customer shopping
experience [
Designing efective machine learning algorithms for global e-commerce involves two major challenges.
First, models often need to handle multiple tasks with unevenly distributed data. For example, click data
is much more abundant than purchase data [
Task 1 (b) Learning multi-task as a sequence (ours)
Task 2 … e e g g … e … g … g e …
Input g Gate e Expert g e e
R R x1 …
R R x2 … … … Seq Processor
Input xk Input for taskk R
R Unrolled RNN
R R xk
Existing methods usually address these two challenges separately. To the best of our knowledge,
no single model currently solves both challenges efectively. Regarding multi-task learning (MTL),
many approaches treat tasks independently [
To this end, we propose the learning multi-task as a SEQuence + Multi-Distribution (SEQ+MD) framework, which can tackle the two challenges simultaneously. For the multi-task component, we observe that many user actions follow a natural sequence, such as clicking before purchasing, which can be modeled efectively as a sequential learning problem. Rather than treating each task independently, our SEQ architecture generates task predictions as a sequence, as shown in Fig.1-(b). The input pair of user and item features is first encoded into a sequence, and the model then outputs a probability token for each task in order. The most closely related work, HTLNet[18], also uses the output of earlier tasks as input for later ones. However, their approach relies on separate task towers, while our SEQ model uses a recurrent neural network (RNN) [19] that shares the same weights across tasks. This design supports eficient expansion to new tasks and maintains strong performance without the need for additional training. For handling mixed input distributions, we separate input features into region-invariant and region-dependent groups. The region-dependent features are processed with a country embedding in our multi-distribution (MD) learning module, meaning these features are transformed according to their region, and then concatenated with the region-invariant features. An advantage of this approach is that the MD module is easy to plug in and can enhance the performance of any multi-task learning model on multi-source data.
We evaluated our framework on our in-house data ofline and observed a 1.8% performance increase in the critical purchase task while keeping the click task performance positive compared to baseline models. In summary, our contributions are: • We introduced a new framework SEQ for multi-task learning with an improvised RNN architecture, specifically designed to handle tasks with sequential order. SEQ not only extracts and utilizes the sequence relation between tasks, reduces redundant computations among related tasks but also demonstrates excellent transferability when adding new tasks. By decomposing a complex task (a) Search “birthday” on CA and GB sites (b) Feature distribution shift between CA and GB
Listing views per query count
User gift purchase count
into simpler, sequential tasks, SEQ efectively enhances the multi-task learning process. • We developed a module MD for learning regional data with diferent distributions. The MD module enables the model to capture region-specific features while sharing region-invariant features, allowing for efective training with a more extensive and diverse dataset.
• Our in-house data experiments demonstrate improvements with this new framework.
Multi-task learning (MTL) trains models on multiple tasks simultaneously. By sharing information
across tasks, the model can learn more robust features, leading to improved performance on each
individual task. MTL can be categorized into two types: hard parameter sharing and soft parameter
sharing. Hard parameter sharing involves an architecture where certain layers are shared among all
tasks in the base model, while other layers remain specific to individual tasks in separate task "towers."
The "Shared-bottom" approach [
Multi-distribution learning trains models using data from various sources, each with distinct feature distributions. Multi-regional data is an example of multi-distribution input, with prior work largely focusing on language-agnostic approaches to create a unified, unbiased embedding space [ 24] or on learning consistent similarities across diferent markets [ 25, 26]. In contrast, our approach utilizes regionally distinct signals to enhance model diversification . Bonab et al. [27] propose learning in an MTL setting where each market is treated as a task. However, this approach faces challenges when market data is imbalanced, especially for smaller markets with limited data. Model-agnostic meta learning (MAML) [28] tackles this through a dual-loop training process: an inner loop optimizes each market individually, while an outer loop optimizes across markets, but the need for separate parameter ifne-tuning for market adaptation makes MAML ineficient in this context. More recently, Market-Aware (MA) models [29] have used market-specific embeddings to create market-adapted item embeddings. Our MD module is similar to MA, yet we observed that not all features are region-specific [ 30], making it more efective to distinguish between shared and region-specific features.
In this section, we introduce our SEQ+MD framework, which includes two model components: a multi-task learning architecture SEQ and a multi-distribution learning module MD. We provide formal definitions for the problem followed by detailed explanations for our framework in the subsections.
Consider an online shopping dataset that records users’ queries and interactions (e.g., click, purchase) with the returned listings. Let = {(, )}=1 be the dataset with samples, where = (, ), refers to the -dimensional features about the user and query, refers to the -dimensional features about the target listing, and = {}=1 is the score set for tasks. The score for each task is calculated based on the user interaction sequences. A complete sequence would be ["click", "add to cart", "purchase"]. The last action in this sequence represents the final step. For example, if the sequence is ["click", "add to cart"], it means the user clicked on the listing and added it to the cart but did not Country purchase it. If none of these actions occurred, the sequence is ["no interaction"]. We assign specific scores to each action ("no interaction", "click", "add to cart", "purchase"), and the final task score is a combination of these action scores.
The multi-task learning architecture SEQ focuses on making predictions for the tasks simultaneously given a single input . Meanwhile, the multi-distribution learning module MD is designed for unified learning across the entire input set {()}=1, where the distribution of for certain regions shows significant diferences compared to other regions. (See Fig. 2-(b) for examples.) The multi-task learning architecture and multi-distribution learning module can be applied separately. We combine these two parts in our final framework and Fig. 3 shows the overall structure.
Some tasks naturally form a sequence, e.g., click, add to cart, purchase, where each action occurs in a sequential order, conditional on the previous ones. However, most multi-task learning architectures do not account for the sequential nature of the problem, making the output tasks order-agnostic and interchangeable.
Introducing "order" into multi-task learning ofers several benefits. First, sequential ordering allows the model to prioritize more complex tasks later in the sequence. In e-commerce, those later tasks (e.g. purchase) are often more critical than earlier (e.g. click) tasks because of their higher monetization values. At the same time, the data sparsity of the purchase task makes it more dificult to optimize. By establishing a sequence, knowledge from earlier (and typically easier) tasks can be used to address later (and often harder) tasks. Second, sequential ordering facilitates the transfer or addition of new tasks. Since the model learns tasks in a "continuous" manner, adding new tasks in the sequence requires minimal training cost. Journey Ranker [31] recognized the importance of task order having each task model predict the conditional probability based on the previous task. However, the MLP components in their model are isolated, not fully utilizing the knowledge exchange of the sequential tasks.
To address this, we connect RNNs [19] with multi-sequential-task learning. In RNN [19], the prediction of later tokens is based on previous tokens; similarly, our predictions for later user actions are conditioned on previous actions. In RNN [19], each token position shares the same set of weights (e.g. ℎℎ, ℎ and ℎ in Eq. 2, 3) with the only diference being the input token and the hidden input from previous tokens. In our approach, as shown in Eq. 1, we process the single input feature through an MLP for each token, transforming the input feature specifically for each task (see Fig. 3-(a)). The hidden input can be seen as the knowledge passed down from previous actions. As shown in Eq. 2, the knowledge for the current task (ℎ ) is from both the input for task ( ()) and knowledge from the previous task − 1 (ℎ− 1). The score for task ( ) depends on the knowledge (ℎ ). Gated Recurrent Unit (GRU) [32] is applied in our SEQ architecture.
[1, ..., ] = [ 1(), ..., ()] ℎ = ℎ(ℎℎℎ− 1 + ℎ ()) = ℎℎ (1) (2) (3) (4)
Fig. 3 shows our sequential task learning together with MD module. Given a single input feature, the first step is passing it through − 1 MLPs to create a length- sequence, where is the number of tasks. After passing through multiple layers of RNN, the output scores are in sequence form, with each score token corresponding to a task.
To further strengthen the learning with sequence, we add the Descending Probability Regularizer [31]. Based on the prior knowledge that the probability of a sequence of actions decreases from the beginning to the end (i.e., the probability of a user "clicking" the listing is greater than or equal to the probability of "purchasing"), we add a sigmoid multiplication at the end of the output. Each output score is activated with a sigmoid function and then multiplied by the previous sigmoid scores. As shown in Eq. 4, the score for task , ˜ is the product of the sigmoid activations of the logits from all previous tasks. This ensures that the output probabilities of later actions are always smaller than those of previous actions, aligning with the prior knowledge.
= ˜ = ∏︁ ()
=1
Looking at the distribution of each raw input feature, we noticed that there are multi-distributions for certain features (e.g. average number of purchases, see examples in Fig.2-(b)). If the goal of training a machine learning model is to learn the transition from a input distribution to the output distribution, then this multi-distribution will pose significant challenges to the model, ultimately leading to a failure in learning [33].
Fig. 4 shows the overall structure of the multi-distribution adaptor module. We first break the input features into three parts: country features (which is the deciding factor of the distribution shift), dependent features (with distribution shifts across countries), and invariant features (which are countryagnostic features). The feature split is done in a heuristic way: country features are manually selected, and the dependent features and invariant features are separated with a distribution distance threshold. i.e., when the average of the distribution distance among all countries is greater than a certain threshold, the feature is categorized as a dependent feature.
After splitting the input features, diferent operations are applied to these three groups of features. Country features are used to generate country mask weights for the dependent features. Country mask weights have the same dimension as the dependent features, and elementwise-multiplication is performed between the mask and dependent features. The multiplied input is fed into an MLP, which transforms the output into invariant features. These are then concatenated with the invariant features from the original input, resulting in a transformed input with consistent distributions.
This multi-distribution adaptor module MD can be easily plugged in for all MTL frameworks. Adding this module directly after the input and then sending the transformed input to the model is clean and simple. We also explore other options for combining this adaptor module with our sequential task learning framework, as shown in Fig. 3. Instead of concatenating the transformed dependent features with the input feature directly, we can concatenate them with the invariant feature model output from the previous layers. Block (b) in Fig. 3 shows how the multi-distribution module works in our sequential learning architecture. Each task has its own country mask. For a single input (country Algorithm 1 SEQ+MD
Input: Feature (, ), heuristic feature selector , network for generating country mask MLPcountry mask, networks for each task input transformation {MLPtask }, sequential learning network RNNseq task.
Output: Scores {}=1 for tasks. 2: (country, dependent, invariant) ← 3: // Generate country mask 1: // Separate country features, dependent features, and invariant features for the input (, ) 4: ← MLPcountry mask(country) 5: // Transformed dependent features ◁ element-wise multiplication ◁ denotes the sigmoid function features, dependent features) transformed with -task country masks, the output is also a length- input sequence. Concatenated with the invariant feature output, the new input features can be processed with the following sequential learning layers to finally get the task scores.
To evaluate our methods, we conducted experiments on our ofline in-house datasets. Four baseline
methods were selected for comparison. The Shared-Bottom model [
We used 14 days of ofline in-house data for training and three days of data for evaluation, and we report the relative increase in the average Normalized Discounted Cumulative Gains ( ) [34] in the result tables (see Sec. 4.2 for more details). Due to the varying nature of diferent trafic sources, the results are divided into two sections: Webpage search trafic (Web), and Mobile App search trafic (App). We track multi-tasks across all trafic.
The results focus on two main areas: the efectiveness of the sequential learning architecture for MTL and the "plug-in" multi-distribution learning module for SOTA MTL methods. Ablation studies and alternative designs are discussed in Sec. 5.
We select a few state-of-the-art multi-task learning methods without any multi-distribution adjustments as the baselines. For multi-distribution learning challenges, most related work [35, 36] focuses on learning invariant features, whereas our goal is to better capture regional preferences. Thus, we use training with single or multi-distribution data as the baselines for multi-distribution learning comparisons.
Shared-bottom [
PLE [
Adatt-sp [
We exclusively use our in-house data for experiments because public search datasets [37] often omit feature details for data security reasons. This omission makes it dificult to isolate country features and generate accurate country mask weights. Our ofline in-house dataset contains over 20 million <user, query, listing> interaction sequences from 10 regions and 2 platforms. Unless otherwise specified, we train the models with data from all regions and platforms. Results are evaluated separately for each platform. Normalized Discounted Cumulative Gain ( ) [34] is our evaluation metric, commonly used for measuring the efectiveness of search engines by summing the gain of the results, discounted by their ranked positions. The rankings of the search listings are ordered by the output scores from the model, and is calculated based on the user interaction sequences. As discussed in Sec. 3.2, e-commerce prioritizes the purchase task over click, making purchase-ndcg our prioritized metric for model evaluation.
SEQ. Table 1 presents the multi-task learning performance on click and purchase tasks across diferent platforms. State-of-the-art MTL baseline methods demonstrate various improvements in the purchase task but show a slight decline in the click task. In contrast, our SEQ model shows improvement across all tasks, adding MD module (SEQ+MD) achieves the best on the critical purchase task. We observed a performance drop in the click task after adding the MD module to SEQ, making the final click performance only slightly positive compared to the share-bottom baseline. This may be due to the click data being noisier and having higher variance. Another possible explanation is that the region-dependent features isolated by the MD module are more closely related to user/listing purchase history, which may have a greater impact on the purchase task.
MD: Multi-Distribution Learning Module. Table 2 illustrates the efectiveness of our
multidistribution learning module as a "plug-in" component for state-of-the-art MTL methods. The adapted
models demonstrate overall improvements, with PLE [
A significant advantage of learning multi-task sequences lies in the inherent properties of RNNs, where weights are shared across all tokens in the sequence. This has two main benefits. First, it reduces redundant calculations among related tasks. For instance, tasks like click and purchase share many commonalities in the buyer’s decision process, i.e. a listing that a user clicks on is also likely to be purchased. Second, by reinforcing the connections between tasks, later tasks in the sequence can be learned more efectively by decomposing them and beginning with easier tasks. As the sequence progresses, task dificulty can be seen as increasing, with earlier tasks acting as processors for the later ones. This recurrent learning process, from easier to harder tasks, is advantageous. For example, predicting which listing is likely to be purchased is challenging, but if the model starts by learning click behavior, it can learn better. We hypothesize that the sequential learning model will benefit from more tasks. In our experiment, we add an add to cart task between the click and purchase sequence to better reflect the buyer’s shopping journey. The results in Table 3 support this hypothesis.
An important consideration for multi-task models is how easily they can adapt to additional tasks, the SEQ+MD model demonstrates a significant advantage. Adding new tasks requires almost no increase in parameters compared to the state-of-the-art models which increase parameter size by 30% on average. Moreover, reusing weights trained on previous tasks can also lead to improved performance in new task evaluations. Figure 5 illustrates the performance comparison of evaluating a three-task setup using weights from a two-task model. The RNN in SEQ+MD uses consistent weights across sequence positions, allowing a new task to be added by simply appending a token to the input sequence. This setup enables predictions for the new task without fine-tuning or additional data. In our three-task evaluation, we averaged the MLP weights from the click and purchase tasks to initialize the MLP weights for the add to cart task. After transforming the inputs separately with three MLPs as a sequence, we applied the RNN using weights trained on only two tasks. Notably, without exposure to add to cart data during training, the model still outperforms the baseline trained on three tasks in both click and purchase tasks. These results support our hypothesis that utilizing the sequential order of tasks can improve multi-task learning efectiveness.
Learning multi-task as a sequence not only enhances knowledge sharing among tasks but also simplifies the integration of output regularization. In our SEQ design, we incorporate a descending probability regularizer that enforces the model to output task scores in a non-increasing order. This regularization is based on the observation that the probability of a user purchasing a listing cannot exceed the probability of them clicking on it, as a click typically precedes a purchase. The results in Fig. 6 demonstrate the efectiveness of this regularizer.
Our SEQ+MD model demonstrates a superior ability to align with regional preferences compared
to other baselines. Figure 7 illustrates the changes in the percentage of domestic listings relative to
the shared-bottom [
GB
CA Better preferences. In contrast, our SEQ+MD model efectively captures these regional trends, providing more accurate rankings that better align with the buyers’ preferences.
In this paper, we introduce the SEQ+MD framework, which integrates sequential learning for multitask problems with multi-distribution data. While SEQ and MD can be applied independently, their combination yields stronger results, particularly on complex tasks. The motivation behind learning multi-task as a sequence stems from the natural sequential order of tasks. Our experiments and analyses highlight two primary benefits: First, SEQ reduces redundant computation across tasks and enhances transferability between diferent task sets, requiring minimal additional parameters while efectively utilizing weights from previous models. Second, by breaking down a complex task into simpler subtasks that serve as processors in the sequence, the model demonstrates improved performance on more challenging tasks. Additionally, our MD module efectively handles multi-distribution data, it can also enhance the performance of state-of-the-art multi-task learning models.
Future work. 1. Improve robustness against noisy data. Even though the primary goal of our approach is to improve performance on complex tasks such as add to cart and purchase, we see opportunities in making SEQ+MD have a neutral impact on click compared to SEQ only. One hypothesis is that click data tends to be noisier than other tasks, with a significant amount of "false clicks" present, particularly on mobile platforms. For example, users may accidentally click on a listing due to the touch screen’s sensitivity. Learning with task-specific noise within a multi-task learning framework could be a valuable direction for future research. 2. Generalize multi-distribution data from region-wise to other scenarios. While this paper focuses on regional diferences as an example of multi-distribution, other multi-distribution exists in e-commerce search data. For instance, diferent platforms (web, app) may show distinct shopping patterns. Extending our MD module to address these scenarios could be a promising research direction.
During the preparation of this work, the author(s) used GPT-4 in order to: Grammar and spelling check. After using these tool(s)/service(s), the author(s) reviewed and edited the content as needed and take(s) full responsibility for the publication’s content. [15] E. Yuan, W. Guo, Z. He, H. Guo, C. Liu, R. Tang, Multi-behavior sequential transformer recommender, in: Proceedings of the 45th international ACM SIGIR conference on research and development in information retrieval, 2022, pp. 1642–1652. [16] X. Ma, L. Zhao, G. Huang, Z. Wang, Z. Hu, X. Zhu, K. Gai, Entire space multi-task model: An efective approach for estimating post-click conversion rate, in: The 41st International ACM SIGIR Conference on Research & Development in Information Retrieval, 2018, pp. 1137–1140. [17] X. Tao, M. Ha, Q. Ma, H. Cheng, W. Lin, X. Guo, L. Cheng, B. Han, Task aware feature extraction framework for sequential dependence multi-task learning, in: Proceedings of the 17th ACM Conference on Recommender Systems, 2023, pp. 151–160. [18] X. Tang, Y. Qiao, F. Lyu, D. Liu, X. He, Touch the core: Exploring task dependence among hybrid targets for recommendation, in: Proceedings of the 18th ACM Conference on Recommender Systems, 2024, pp. 329–339. [19] K. Cho, B. Van Merriënboer, C. Gulcehre, D. Bahdanau, F. Bougares, H. Schwenk, Y. Bengio, Learning phrase representations using rnn encoder-decoder for statistical machine translation, arXiv preprint arXiv:1406.1078 (2014). [20] D. Xi, Z. Chen, P. Yan, Y. Zhang, Y. Zhu, F. Zhuang, Y. Chen, Modeling the sequential dependence among audience multi-step conversions with multi-task learning in targeted display advertising, in: Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining, 2021, pp. 3745–3755. [21] H. Zhang, J. Pan, D. Liu, J. Jiang, X. Li, Deep pattern network for click-through rate prediction, in: Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, 2024, pp. 1189–1199. [22] B. Hidasi, A. Karatzoglou, Recurrent neural networks with top-k gains for session-based recommendations, in: Proceedings of the 27th ACM international conference on information and knowledge management, 2018, pp. 843–852. [23] B. Hidasi, A. Karatzoglou, L. Baltrunas, D. Tikk, Session-based recommendations with recurrent neural networks, arXiv preprint arXiv:1511.06939 (2015). [24] A. Ahuja, N. Rao, S. Katariya, K. Subbian, C. K. Reddy, Language-agnostic representation learning for product search on e-commerce platforms, in: Proceedings of the 13th International Conference on Web Search and Data Mining, 2020, pp. 7–15. [25] J. Cao, X. Cong, T. Liu, B. Wang, Item similarity mining for multi-market recommendation, in: Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval, 2022, pp. 2249–2254. [26] X. Li, Z. Qiu, J. Jiang, Y. Zhang, C. Xing, X. Wu, Conditional cross-platform user engagement prediction, ACM Transactions on Information Systems 42 (2023) 1–28. [27] H. Bonab, M. Aliannejadi, A. Vardasbi, E. Kanoulas, J. Allan, Cross-market product recommendation, in: Proceedings of the 30th ACM International Conference on Information & Knowledge Management, 2021, pp. 110–119. [28] C. Finn, P. Abbeel, S. Levine, Model-agnostic meta-learning for fast adaptation of deep networks, in: International conference on machine learning, PMLR, 2017, pp. 1126–1135. [29] S. Bhargav, M. Aliannejadi, E. Kanoulas, Market-aware models for eficient cross-market recommendation, in: European Conference on Information Retrieval, Springer, 2023, pp. 134–149. [30] J. Cao, S. Li, B. Yu, X. Guo, T. Liu, B. Wang, Towards universal cross-domain recommendation, in: Proceedings of the Sixteenth ACM International Conference on Web Search and Data Mining, 2023, pp. 78–86. [31] C. H. Tan, A. Chan, M. Haldar, J. Tang, X. Liu, M. Abdool, H. Gao, L. He, S. Katariya, Optimizing airbnb search journey with multi-task learning, in: Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 2023, pp. 4872–4881. [32] J. Chung, C. Gulcehre, K. Cho, Y. Bengio, Empirical evaluation of gated recurrent neural networks on sequence modeling, arXiv preprint arXiv:1412.3555 (2014). [33] B. Peng, The sample complexity of multi-distribution learning, in: The Thirty Seventh Annual
Conference on Learning Theory, PMLR, 2024, pp. 4185–4204. [34] H. Valizadegan, R. Jin, R. Zhang, J. Mao, Learning to rank by optimizing ndcg measure, Advances in neural information processing systems 22 (2009). [35] J. Cha, K. Lee, S. Park, S. Chun, Domain generalization by mutual-information regularization with pre-trained models, in: European conference on computer vision, Springer, 2022, pp. 440–457. [36] J. Cha, S. Chun, K. Lee, H.-C. Cho, S. Park, Y. Lee, S. Park, Swad: Domain generalization by seeking lfat minima, Advances in Neural Information Processing Systems 34 (2021) 22405–22418. [37] P. Li, R. Li, Q. Da, A.-X. Zeng, L. Zhang, Improving multi-scenario learning to rank in e-commerce by exploiting task relationships in the label space, in: Proceedings of the 29th ACM International Conference on Information & Knowledge Management, 2020, pp. 2605–2612.