Neural recommender models for sparse and skewed behavioral data

Krishnan, Adit

Content Files

KRISHNAN-DISSERTATION-2021.pdf

Download Options

Download All Files Download Counts…

Loading…

Download Files

Loading…

Download Counts (All Files)

Loading…

Edit File

Loading…

Permalink

https://hdl.handle.net/2142/110473 Copy

Description

Title

Neural recommender models for sparse and skewed behavioral data

Author(s)

Krishnan, Adit

Issue Date

2021-04-15

Director of Research (if dissertation) or Advisor (if thesis)

Sundaram, Hari

Doctoral Committee Chair(s)

Sundaram, Hari

Committee Member(s)

• Han, Jiawei
• Zhai, ChengXiang
• Das, Mahashweta

Department of Study

Computer Science

Discipline

Computer Science

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

Ph.D.

Degree Level

Dissertation

Date of Ingest

2021-09-17T01:10:54Z

Keyword(s)

• Recommender Systems
• Machine Learning
• Artificial Intelligence
• Data Sparsity
• Data Skew

Abstract

Modern online platforms offer recommendations and personalized search and services to a large and diverse user base while still aiming to acquaint users with the broader community on the platform. Prior work backed by large volumes of user data has shown that user retention is reliant on catering to their specific eccentric tastes, in addition to providing them popular services or content on the platform. Long-tailed distributions are a fundamental characteristic of human activity, owing to the bursty nature of human attention. As a result, we often observe skew in data facets that involve human interaction. While there are superficial similarities to Zipf's law in textual data and other domains, the challenges with user data extend further. Individual words may have skewed frequencies in the corpus, but the long-tail words by themselves do not significantly impact downstream text-mining tasks. On the contrary, while sparse users (a majority on most online platforms) contribute little to the training data, they are equally crucial at inference time. Perhaps more so, since they are likely to churn. In this thesis, we study platforms and applications that elicit user participation in rich social settings incorporating user-generated content, user-user interaction, and other modalities of user participation and data generation. For instance, users on the Yelp review platform participate in a follower-followee network and also create and interact with review text (two modalities of user data). Similarly, community question-answer (CQA) platforms incorporate user interaction and collaboratively authored content over diverse domains and discussion threads. Since user participation is multimodal, we develop generalizable abstractions beyond any single data modality. Specifically, we aim to address the distributional mismatch that occurs with user data independent of dataset specifics; While a minority of the users generates most training samples, it is insufficient only to learn the preferences of this subset of users. As a result, the data's overall skew and individual users' sparsity are closely interlinked: sparse users with uncommon preferences are under-represented. Thus, we propose to treat these problems jointly with a skew-aware grouping mechanism that iteratively sharpens the identification of preference groups within the user population. As a result, we improve user characterization; content recommendation and activity prediction (+6-22% AUC, +6-43% AUC, +12-25% RMSE over state-of-the-art baselines), primarily for users with sparse activity. The size of the item or content inventories compounds the skew problem. Recommendation models can achieve very high aggregate performance while recommending only a tiny proportion of the inventory (as little as 5%) to users. We propose a data-driven solution guided by the aggregate co-occurrence information across items in the dataset. We specifically note that different co-occurrences are not equally significant; For example, some co-occurring items are easily substituted while others are not. We develop a self-supervised learning framework where the aggregate co-occurrences guide the recommendation problem while providing room to learn these variations among the item associations. As a result, we improve coverage to ~100% (up from 5%) of the inventory and increase long-tail item recall up to 25%. We also note that the skew and sparsity problems repeat across data modalities. For instance, social interactions and review content both exhibit aggregate skew, although individual users who actively generate reviews may not participate socially and vice-versa. It is necessary to differentially weight and merge different data sources for each user towards inference tasks in such cases. We show that the problem is inherently adversarial since the user participation modalities compete to describe a user accurately. We develop a framework to unify these representations while algorithmically tackling mode collapse, a well-known pitfall with adversarial models. A more challenging but important instantiation of sparsity is the few-shot setting or cross-domain setting. We may only have a single or a few interactions for users or items in the sparse domains or partitions. We show that contextualizing user-item interactions helps us infer behavioral invariants in the dense domain, allowing us to correlate sparse participants to their active counterparts (resulting in 3x faster training, ~19% recall gains in multi-domain settings). Finally, we consider the multi-task setting, where the platform incorporates multiple distinct recommendations and prediction tasks for each user. A single-user representation is insufficient for users who exhibit different preferences along each dimension. At the same time, it is counter-productive to handle correlated prediction or inference tasks in isolation. We develop a multi-faceted representation approach grounded on residual learning with heterogeneous knowledge graph representations, which provides us an expressive data representation for specialized domains and applications with multimodal user data. We achieve knowledge sharing by unifying task-independent and task-specific representations of each entity with a unified knowledge graph framework. In each chapter, we also discuss and demonstrate how the proposed frameworks directly incorporate a wide range of gradient-optimizable recommendation and behavior models, maximizing their applicability and pertinence to user-centered inference tasks and platforms.

Graduation Semester

2021-05

Type of Resource

Thesis

Permalink

http://hdl.handle.net/2142/110473

Copyright and License Information

Copyright 2021 Adit Krishnan

Owning Collections