Choice modeling using dot-product attention mechanism
- PDF / 1,265,911 Bytes
- 6 Pages / 595.276 x 790.866 pts Page_size
- 41 Downloads / 194 Views
ORIGINAL ARTICLE
Choice modeling using dot‑product attention mechanism Mofei Li1 · Yutaka Nakamura1 · Hiroshi Ishiguro1 Received: 13 April 2020 / Accepted: 21 August 2020 © International Society of Artificial Life and Robotics (ISAROB) 2020
Abstract Modelling the cognitive process is a challenging task. Contextual conditions and the scope of the options are critical factors that influence human decisions. We propose and formulate an attention-based network to model the various choices made by humans based on the various factors that predict the possible option in each scope. To evaluate our proposed method, we conducted a user choice experiment in which a user chose an option from among limited choices. Our results showed that our model successfully extracted the hidden context on the attention layer and even outperformed the chance level in terms of prediction accuracy. Keywords Attention network · Choice modeling · Hidden context
1 Introduction Many researchers have been studying models of humans to recognize the world around them. There are two main types of studies: (1) models for common sense models such as semantic networks (e.g., WordNet [1] and ConceptNet [2]) and (2) models using the individual subjectivity [3] such as human preferences and impressions. An artificial system capable of deeply communicating with a person would need to understand the individual subjectivity of the person. In this research, we focus on the choice made by a person. There are a number of methods for preference estimation; the collaborative filtering [4, 5] techniques are the typical framework for this purpose. In these methods, the estimation is based on the assumption that the preference of a person for an item is similar to the preference of similar persons for similar items. However, they are not a model of the choices made by a person, which may depend heavily on the situations encountered by that person. This work was presented in part at the 25th International Symposium on Artificial Life and Robotics (Beppu, Oita, January 22–24, 2020). * Yutaka Nakamura [email protected]‑u.ac.jp Mofei Li [email protected]‑u.ac.jp 1
In this paper, we designed a ‘choice task’ in which a person chooses one option from the given discrete options (scope) in a context, as shown in Fig. 1. Since not only the context but also the scope may affect the person’s choice [13], it is necessary to developed a mathematical model for a person’s making choices. Recently, the dot-product attention mechanism [6–8] was proposed and used in neural network frameworks to cope with contextual information. In this mechanism, the network output was determined based on three factors: query, keys, and value. The dot-product attention mechanism performs information processing similar to a dictionary which changes its own behavior according to the training of the network. That is, the connection between the index (key) and the query and the output are tuned to fit the training data. Since the original dot-product attention model is not suitable wh
Data Loading...
