Modeling and Analyzing the Influence of Non-Item Pages on Sequential Next-Item Prediction

TL;DR

This work investigates how non-item pages—pages not directly tied to specific items—can augment sequential next-item prediction. It formalizes three representations (UPID, CPID, PE) and uses HypTrails to validate that non-item pages influence subsequent interactions, then adapts eight popular sequential models to incorporate non-item representations. Across a synthetic SynDS and two real-world datasets (Coveo and Fashion), the study demonstrates consistent performance gains in next-item prediction when non-item pages are included, though gains depend on representation quality, model architecture, and data noise. The findings emphasize careful selection of non-item representations and suggest that non-item signals can meaningfully sharpen intent inference in real-world recommender systems.

Abstract

Analyzing sequences of interactions between users and items, sequential recommendation models can learn user intent and make predictions about the next item. Next to item interactions, most systems also have interactions with what we call non-item pages: these pages are not related to specific items but still can provide insights into the user's interests, as, for example, navigation pages. We therefore propose a general way to include these non-item pages in sequential recommendation models to enhance next-item prediction. First, we demonstrate the influence of non-item pages on following interactions using the hypotheses testing framework HypTrails and propose methods for representing non-item pages in sequential recommendation models. Subsequently, we adapt popular sequential recommender models to integrate non-item pages and investigate their performance with different item representation strategies as well as their ability to handle noisy data. To show the general capabilities of the models to integrate non-item pages, we create a synthetic dataset for a controlled setting and then evaluate the improvements from including non-item pages on two real-world datasets. Our results show that non-item pages are a valuable source of information, and incorporating them in sequential recommendation models increases the performance of next-item prediction across all analyzed model architectures.

Figures (18)

• Figure 1: Exemplary click sequence in an online store. The inclusion of non-item pages provides an explanation for the shift in user interest, enabling a more accurate prediction of their preferences.
• Figure 2: The common setup for sequential recommendation models and our approach for including non-item representations.
• Figure 3: Distribution of attributes within SynDS and Coveo-Search dataset.
• Figure 4: The evidence or marginal likelihood for our hypotheses on SynDS, Coveo-Search and Fashion for different hypothesis weighting factor $k$. A higher evidence means a hypothesis is a better explanation for the data, while $k$ can be interpreted as how strongly you believe in the hypothesis.
• Figure 5: HR@10 for models on Rand-SynDS with seed 212 on datasets with a growing percentage of shuffled non-item pages. Genre-ID shows the performance of models including non-item pages based on genres with an ID embedding, while Genre-PE shows those with an Page Embedding. The baseline utilizing items only is marked by "BL Items".