Accepted Submissions

Personalizing E-commerce Search using Embeddings

Arnab Dutta(EBay), Liyang Hao(EBay), Mithcell Donley(EBay)

Abstract: We present a multi pronged personalization strategy for e-commerce search by incorporating high dimensional representations of user and item interactions. We explore and evaluate two major application sub-domains of search science: query expansion and ranking. The former deals with personalizing at the query level by augmenting queries with users’ item condition preferences. Our proposed methodology empirically shows almost 0.7% gain in relevance measures for some class of queries. While the later approach influences the ranking models with personalized factors. This is done by building a multioutput regression model to quantify the purchase price affinity for an user, which is utilized as a factor by the ranking models complementing its standard set of predictors. Offline evaluation for the price affinity model exhibits 35% reduction in price estimates error compared to the baseline method.

DNN approach based Personalized Product Ranking for E-Marketplaces

Linsey Pang (Walmart Labs)

Abstract: E-commerce is becoming the largest retail channel in the world. With more and more consumers are buying retail goods online including essentials, groceries, household supplies, big electronics, etc, more and more sellers now register and sell products through e-marketplaces platforms. With the significantly increased online consumers transactions and larger collection of product assortments, there is a need for e-markerplace sellers to making the right product picks and ordering the right quantities to match products’ demand. In this work, we utilize the signals of product features of seller and competitor obtained from in-network as well as from out-of-network, build multi-target deep learning model to produce the personalized demand ranking of a given product at different granularity for our seller: in-network rank, ROM rank and overall rank. We tackle this problem in these steps: First, we obtain matched items 2 from in-network and ROM by running our owned item matching model. Second, matched items and the correspondent sellers’ listing information are used as training set for our multi-target deep learning model to learn the normalized rank score for a given product listed by a seller. Third, model inferences this given item and generates demand rank, which is used as gauging the popularity of the product and provide personalized recommendation to fill the assortment gap for seller. Extensive experiments are conducted to evaluate our proposed multi-target deep learning approach by applying regression and classification metrics, etc. The experiments show our proposed approach achieves high accuracy, low MSE and high 𝑅2. Additionally, we validate that multi-target approach outperforms single-target approach.

Modeling Heterogenous modality for Personalized Video Recommendation at Scale

Feng Zhang(Meta), Hyun Duk Kim(Meta), Ziliang Zhao(Meta), Jiaying Li(Meta), Bin Kuang(Meta), Chi-Hoon Lee(Meta)

Abstract: As Recommendation systems are adopted for many different domains with success stories, multiple different vertical recommendation systems are often built. Meta Platform Inc. (previously known as Facebook) and Amazon are examples that provides multiple services to their users. Those services within a company are often time using different backend systems to focus on their vertical experiences and to facilitate the development processes. Those paradigm implies that a vertical recommendation system would not take advantages of utilizing signals with respect to cross-products and multi-modality. However, it is not trivial to build a framework to utilize signals from cross-products, due to the scale involved with the user sizes and content volumes as well as dynamics of content updates. To address this, we present a novel framework that collects signals from cross-products by considering low latency, allowing a vertical recommendation system to realize interactions from other products fast. In addition, we present a novel modeling approach in representing signals in heterogenous modality in order to enable to quantify “relatedness” or “similarity” to be mapped for vertical specific entities – that is, signals from cross-products are initially collected in a free text format since they are queried for searching. And those search signals are modeled to represent for quantifying similarity with videos, which is our focused domain for this paper. Our approach using bi-partite graph in modeling heterogenous signals, search queries, for personalized video recommendation system has shown effective outcomes. Online A/B test showed that our approach can statistically significantly improve our top line metrics in terms of watch time, etc.

A Contextual-aware Reranking Framework for Autocomplete in Walmart

Junchao Zheng(Walmart), Vishal Rathi(Walmart), Jun Zhao(Walmart)

Abstract: Autocomplete is a common practice in the search engine for E-commerce. A successful application of Autocomplete requires an understanding of search context. In this work, we propose a novel machine learning based reranking framework with the capacity to consume real-time contextual information but without adding any latency. We also introduce Absolute Mean Reciprocal Rank(AMRR), a variation of Mean Reciprocal Rank(MRR) to evaluate Autocomplete ranking. The reranking framework shows $+17.2\%$ AMRR lift in offline evaluation. It is currently served on Walmart.com as a key component to improve customers search experience.

Call for Papers

Call For Papers

Personalization has been a prevalent concept in the context of recommender systems. And the Web Search engines such as Google and Bing have been incorporating user intents and context to show different ranked lists of results for different users for the same query. Furthermore, in recent years, we are seeing widespread usage of personalization and recommendations in search. Examples of personalized search include e-commerce sites like Etsy.com, Music search on Spotify, restaurants search on UberEats, talent search on LinkedIn etc. showing different users’ different relevant results for the same query. These applications indicate how valuable personalization has been in the context of search. Similarly we find recommendations adding value to search beyond typical retrieval and ranking approaches. In this workshop, we want to bring all such applications and approaches developed to incorporate personalization and recommendation in search algorithms.

This workshop to be held with WSDM 2022 welcomes submissions from academia and industry researchers and practitioners to submit their work related to personalization and recommendations in search. Please note that for the scope of this workshop both personalization and/or recommendations are for improving search experience. Personalization or recommendations as stand-alone topics are out of the scope for this workshop. We also welcome work that can highlight the challenges faced in developing a search algorithm with personalization in real production systems. We invite quality research contributions, including original research, preliminary research results, and proposals for new work, to be submitted as an extended abstract. All submitted papers will be peer reviewed by the program committee and judged for their relevance to the workshop, especially to the topics identified below, and their potential to generate discussion.

Accepted submissions will be presented at the workshop but will not be archived. Hence we encourage submissions of any relevant recent work even if that might have been presented in or submitted to other venues. We also encourage submissions that discuss work in progress in order to facilitate wider discussions.

The topics of interest are (but not limited to) listed below:

Personalized query interpretation and intent disambiguation
Incorporating user context in real time or in-session adaptation for personalization in search
Personalized evidence generation (such as summary, snippet/TAU, topics, among others) and explainability of search results
Fairness and privacy in personalized search
Latency, caching and other infrastructural considerations for real-world personalized search systems
Lessons learnt and challenges for developing large scale personalized search systems
Leveraging different modality of data, including text, image, audio for personalization of search
New learning to rank (LTR) approaches for search personalization as well as beyond traditional LTR approaches, such as reinforcement learning for search personalization
New applications for personalization and recommendations for search (e.g. in healthcare)
Evaluation metrics for personalization and recommendation in search
Optimization of delayed user behavioral rewards for personalization and recommendations in search context
Human-computer interaction considerations in designing user interfaces and interaction for personalized search
Datasets or design of simulator for personalized search/recommendation
Personalizations and recommendations for Conversational Assistants
Personal search (e.g. email search, desktop search, etc.) is not within the scope of the workshop.

Paper submission

Four page extended abstract reporting original results, preliminary results, and proposals for new work, will be considered for presentation/poster session at the workshop. Manuscripts must be self-contained and in English with 4 pages length plus references. Papers must be submitted in PDF according to the new ACM format published in ACM guidelines, selecting the generic “sigconf” sample. The PDF files must have all non-standard fonts embedded. After uploading your submission, please verify the copy stored on the CMT site. Please follow other guidelines for formatting submissions as indicated in the main conference page:

https://www.wsdm-conference.org/2022/

At least one author of each accepted paper must attend the workshop and present the paper. The deadline for paper submission is Dec 18, 2021. Papers can be submitted through CMT:

https://cmt3.research.microsoft.com/PaRiS2022

Important Dates (New extended deadlines)

Please pay attention to the following dates:

Paper Submission Deadline: Jan 14, 2022 (11: 59 P.M. PDT)
Acceptance notification: Jan 17, 2022
Workshop date: Feb 25, 2022

All deadlines are 11:59 pm, anywhere in the world.

Please reach out to paris2022 at googlegroups dot com for any question or clarification regarding the workshop.