Conversational recommendation: Theoretical model and complexity analysis
Introduction
System-generated recommendations have become a common feature of modern online services such as e-commerce sites, media streaming platforms, and social networks. In many cases, the suggestions made by the underlying recommender systems are personalized according to the user’s stated or assumed needs and preferences. In the most prominent applications of recommender systems, e.g., on Amazon.com, Netflix, or YouTube, user preferences are estimated based on past user behavior. However, there are also several application domains where no past interaction logs are available or where the user’s needs and preferences might differ each time the user interacts with the service. Consider, for example, someone seeking a recommendation for a restaurant this evening for a party of four, where the requirements include that the location is nearby, that the prices are modest, and that there is a vegetarian option. In such a situation, the user’s current needs and preferences have to be interactively acquired by the system to make a suitable recommendation.
The class of systems that support such interactions are called Conversational Recommender Systems (CRS). In these systems, the recommendation process consists of an interactive, multi-turn dialog, where the system’s goal is to learn about the user preferences to the extent that appropriate recommendations can be made. The corresponding preference elicitation process can be implemented in different ways, ranging from predefined fill-out forms to natural language interfaces—see [1] for an overview. In particular, for this latter class of interfaces, we observed substantial progress in terms of voice recognition and natural language understanding in recent years, leading to the development of voice-controlled devices like Apple’s Siri or Amazon’s Alexa, and continuously improved chat-bot systems.
In that context, a specific goal when designing a CRS is to minimize the effort for users by asking as few questions as possible, i.e.,to increase the efficiency of the dialog.
Today, research in the general area of recommender systems, and specifically area of CRS, is almost entirely empirical. Recently, Cena et al. [2] provide logical foundations for knowledge-based recommender systems, for which, a comprehensive formalization was not yet available. Regarding CRS, typical research designs are based on simulations or user studies, in which two or more interaction strategies are compared in one or two application domains based on real or synthetic datasets. The corresponding efficiency measures are, for example, the number of required user interactions or the perceived difficulty and effort of the recommendation dialogs, where the main assumption is that a lower number of required interactions leads to a better usable system.
Such empirical studies are certainly important and insightful. However, little is known about the theoretical aspects of the underlying interactive recommendation processes. Unfortunately, theoretical questions regarding, e.g., the computational complexity of determining a good or the best interaction strategy can not be answered without a formal characterization of the overall problem.
With this work, we address this research gap and provide a theoretical model of conversational recommendation. The model is designed in a domain-independent way and aims to cover a wide range of realistic application scenarios. A conversational recommendation process is modeled as a sequence of states, where state transitions correspond to common conversational moves [3] that can be found in the literature. Among the possible actions taken by the user, we admit utterances for expressing preferences on items or features or for relaxing or revising previously stated preferences.
Since our model is agnostic about the application domain and the algorithm that is used to select and rank the objects for recommendation—i.e.,the recommendation algorithm—it serves as a basis to analyze important theoretical properties of conversational recommendation processes.
The main contribution of this work is the study of the computational complexity for finding an efficient conversational strategy in terms of number of dialog turns.
From our study we found that:
- •
the problem of finding an efficient conversational strategy in terms of number of dialog turns is NP-hard, but in PSPACE1;
- •
some specific factors of the item catalog influence the complexity of the problem;
- •
for a special class of catalogs, the upper bound lowers to POLYLOGSPACE2.
From a practical perspective, our analysis leads to the observation that the efficiency of a conversation strategy is tied to the characteristics of the available item catalog. Observations from an empirical analysis on datasets based on MovieLens-1M support these theoretical considerations.
The paper is organized as follows. In the next section we discuss existing works and outline our research goals. We provide a model formalization in Section 3, and then introduce our theoretical results in Section 4. Section 5 is devoted to an experimental evaluation to support the outcomes of our empirical analyses. A summary and outlook of the presented results, together with the identification of future directions close the paper. In order to improve the paper readability, Table 1 and Table 2 collect acronyms and symbols used in the paper.
Section snippets
Previous Work and Research Goals
Dialog efficiency is one of the main dimensions in which CRS are evaluated in research literature [1]. Here, efficiency usually refers to the time or effort that a user needs to find a suitable item. The corresponding underlying assumption is that users will find a system more valuable if it requires less effort for them. Likewise, if a recommendation dialog takes too many steps, users might quit the conversation or, even worse, abandon the system as a whole.
Model Formalization
On a conceptual level, a CRS—as considered in our study—works as follows. The system main task in the interaction with the users is to elicit their preferences3 regarding item features and items. To that purpose, the system maintains information about these preference statements in a user model. These preference statements determine which of the items of a given catalog qualify to be recommended. Therefore, in our
Devising Optimal Strategies
In the development of CRS with system-driven conversations, a cornerstone is the method the CRS uses to decide how to proceed in the conversation; we call such a method a strategy. Our goal in this section is to study the problem of developing efficient conversation strategies for CRS. For our setting, a CRS strategy is an algorithm that suggests the system its following action in every possible state of a conversation, among the ones formalized above. As discussed in Section 2, the efficiency
A complexity-driven experiment with two protocols and two catalogs
Based on the results of the previous section, we devise an in vitro experiment that confirms the intuition in Corollary 1. More specifically, we demonstrate that the two alternatives in Algorithm 1—that is, entering or not Line 11—lead to different results in terms of the efficiency of a CRS, but depending on the characteristics of the catalog. We recall the two protocols below, then we make our hypotheses explicit, and explain in detail the setting of the experiment. Subsequently, we present
Summary and Outlook
While conversational approaches to recommendation have been explored since the late 1990s, we observe an increasing interest in these types of systems in recent years in particular due to recent developments in natural language processing. Independent of the interaction modality, efficiency in terms of required dialog turns has commonly been a main target of research. So far, however, questions of efficiency were almost exclusively investigated in an empirical manner. With this work, we
CRediT authorship contribution statement
Tommaso Di Noia: Conceptualization, Methodology, Writing - original draft, Writing - review & editing. Francesco Maria Donini: Conceptualization, Methodology, Formal analysis, Writing - original draft, Writing - review & editing, Supervision. Dietmar Jannach: Conceptualization, Methodology, Writing - original draft, Writing - review & editing. Fedelucio Narducci: Conceptualization, Methodology, Writing - original draft, Writing - review & editing. Claudio Pomo: Conceptualization, Methodology,
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgement
This research was partially supported by the following research projects: ERP 4.0, PASSEPARTOUT, Secure Safe Apulia, Servizi Locali 2.0, CTE Matera, San Lorenzo Defence (APQ6 – DTC – Centro di Eccellenza Codice CUP n.F85F21001090003).
References (50)
- et al.
Logical foundations of knowledge-based recommender systems: A unifying spectrum of alternatives
Information Sciences
(2021) - et al.
Self-supervised clarification question generation for ambiguous multi-turn conversation
Information Sciences
(2022) - et al.
Conversational recommender systems and natural language: A study through the converse framework
Decision Support Systems
(2020) - et al.
Concept abduction and contraction for semantic-based discovery of matches and negotiation spaces in an e-marketplace
Electron. Commer. Res. Appl.
(2005) - et al.
Constructing optimal binary decision trees is np-complete
Inf. Process. Lett.
(1976) - et al.
Totally optimal decision trees for boolean functions
Discrete Applied Mathematics
(2016) - et al.
Exact learning when irrelevant variables abound
Inf. Process. Lett.
(1999) - et al.
A note on succinct representations of graphs
Inf. Control.
(1986) Relationships between nondeterministic and deterministic tape complexities
J. Comput. Syst. Sci.
(1970)- et al.
A survey on conversational recommender systems
ACM Comput. Surv.
(2021)