Mining longitudinal user sessions with deep learning to extend the boundary of consumer priming

Highlights

• The proposed framework using NLP and deep learning techniques approximate user heuristics after B&B product shortlisting.

• The proposed framework was an extension of the priming theory into product comparison and shortlisting stages that were traditionally difficult for marketers to tap into.

• This study proposed a recommendation system built based on the learned consumer heuristics to recommend B&B options.

• We collaborated with an online accommodation booking platforms company to analyze their consumer data for this research.

Abstract

Priming is challenging when consumers start shortlisting products before the final purchase. This is because this shortlisting process is performed in multiple user sessions online across time, the shortlist does not stay as a static list, and product comparison in this stage uses the heuristics internal to individual consumers. The goal of this study is two folds: (1) to approximate user heuristics after B&B product shortlisting using NLP and deep learning techniques, and (2) to identify optimized deep learning models for the representation of key elements of consumer heuristics. This offers an extension of the priming theory into product comparison and shortlisting stages that were traditionally difficult for marketers to tap into. By analyzing the B&B product information repeated visited in user sessions, the formation of shortlists is identified and products in the shortlists can then be compared. Subsequent priming and promotions can therefore be performed closer to the actual purchase. Our study also provides marketers keywords and their associated activated words relevant for crafting marketing messages. As these activation words are extracted from the B&B sites and product reviews that the users had visited repeatedly in long-term tracking sessions, they are analogous to effects produced from user participatory design, an approach popular in the IT world. Our work shows that opportunities for marketing decision support, especially into the shortlisting phase, are now possible through machine learning techniques. Both theoretical and practical implications are provided.

Introduction

Message priming refers to an approach to make a message more accessible in order to influence responses of message receivers [1]. For example, the primed message “Only pay for what you need” by Liberty Mutual Insurance activates the semantic importance of price individualization later when a consumer purchases a car insurance. A longer form of message priming includes the primed message either as a standalone message or as an embedded message into product description, ads, and other consumer-facing materials. Priming encourages brand identity, retention [2], customer loyalty [3,4] and purchase [5]. Priming effects weaken as the time goes by [1,2], but stronger if presented closer to the purchase decisions [5], such as during product shortlisting before the final purchase.

Because shortlisted products tend to be similar in product attributes that were used to select products to include in the shortlist, consumers usually compare the shortlisted products using additional supplemental information (e.g., usage experience in product reviews, additional product features, etc.) before making the final purchase [6]. Unlike commodity products where shortlisted items selected into the shopping carts are known to the vendors and certain last-minute marketing strategies (e.g., dynamic pricing in shopping carts [7]) could be applied as a result, such opportunity to tap into consumer shortlists for hotel booking is virtually non-existent for reasons such as consumers may not login when they start shortlisting, shortlisted products may not reside on vendor's platform (such as in the form of shopping carts), consumers may not rely on vendor's platform to retain their shortlists, and hotel comparison is done through micro-increments across different time [8]. Coupled with the restriction of the HTTP protocol being stateless where web servers treat every user visit as an independent, different user visit even if it was made by the same user at different times. As a result, linking consumer visits at different times even when they had not logged in is important to the success of shortlist identification. Without being able to identify the formation of shortlists in hotel bookings, the effects of marketing interventions (such as priming and other promotion strategies), identification of competition during shortlisting, last-minute promotion with the right primed message and primed keywords to promote product differentiation, are either limited, difficult or impossible. For example, shortlisted products are quite homogeneous in their product attributes. Without knowing who are among the competitors during consumer shortlisting, priming or promotions emphasizing product attributes similar to the shortlisted items will unlikely be effective. Therefore, using the right priming keywords that effectively differentiate one's product from the competition is crucial after product shortlisting.

The predominant approaches to study priming effects have been simulation and lab-controlled experiments. Despite useful in delineating construct relationships, opportunities for generalizability and scalability for individualization could be limited [9]. As Jones & Chen indicated, the final shortlist is not always a static list of selections [6]. It could expand or contract as a consumer gains more experience, has more information regarding the product or revises their heuristics. The static results produced from simulation and lab studies could be too rigid for practical use [3] and costly to become scalable longitudinally.

To accomplish shortlist identification as it expands and contract, identification of primed keywords, and distilling of consumers' long-term preferences from a series of micro-increments of hotel selection [8], one would need a framework that breaks the technical and managerial boundaries by modeling long-term user preference, identifying shortlist formation and items within, and extracting primed keywords for shortlisted items from their product reviews. To the best of our knowledge, this is an area that has profound marketing opportunities, but has received very little attention in scholarly works. Therefore, this study is designed to take an initial look at these issues by answering the following key research questions:

RQ1

How to build a framework that allows for the identification and monitoring of product shortlisting BnBs?

RQ2

How to capture long-term user preference with deep learning and NLP techniques to recommend BnBs?

RQ1 is a theoretical advancement by extending the boundary of the priming theory into a product shortlisting phase that is internal to individual consumers. We proposed a framework that uses natural language processing (NLP) to extract consumer shortlisting longitudinally throughout user sessions across time and builds deep learning models to capture consumer heuristics during the product selection and purchase decisions. The framework produces three interesting deliverables for practitioners: (a) A recommendation system built based on the learned consumer heuristics to recommend B&B options, (b) identification of shortlists, and (b) keywords relevant to user heuristics extracted from repeat visits for concurrent and subsequent priming. Because these keywords are from the B&B materials that consumers repeated visited, they create a stronger appeal than the primed keywords solely defined by marketers. Not only does this approach allow for subsequent priming to be performed after shortlisting, it also provides opportunities for automating heuristics extraction to tackle the issue of shortlist not being static over time [6]. To answer RQ2, we propose a novel multi-layer solution to optimize deep learning and NLP techniques to capture consumer preferences from the BnB pages they visited and dimensions of the product reviews associated with the visited BnBs. We also present technical challenges of using deep learning techniques in the context of BnB booking and possible solutions.

In section 2, we highlight the current literature for developing a two-stage product evaluation process, and the current development of NLP and recommendation systems. Section 3 outlines our approach to empirically validate the two-stage process and proposes empirical analyses. Section 4 reports findings from the empirical validation. Discussion and conclusion are presented in section 5 before our contributions are outlined.