Learning in CCOBRA

CCOBRA is fundamentally based on the idea that models need to account for all cognitive information processing systems. For that matter, the evaluation scheme defined by CCOBRA’s model interface relies on two distinct phases: global pre-training and individual adaptation. The following paragraphs introduce these learning phases.

Modeling Problem

At the core of CCOBRA lies the problem of generating predictions for individual human reasoners. This problem is composed of the following algorithmic components:

  • Input: Reasoning task such as a single syllogism (“All A are B; All B are C; What follows?”).

  • Output: Generated prediction to the given task. In contrast to common statistical or data scientific analyses, the focus is not on predicting a distribution, but on generating a single response (e.g., “All A are C”).

  • Scope: In the end, CCOBRA models are not models for aggregated data accounting for a population of reasoners, but predictors for a specific individual.

Note

Models for the CCOBRA framework should be written with one specific individual in mind. The benchmark will handle model initialization and separation of individuals for the prediction and adaption phases by itself.

Phases of Learning

Since CCOBRA’s model interaction comprises two types of learning, it strays away from regular machine learning and pattern recognition problems where modeling is usually constrained to a single learning and a subsequent model application phase where parameters stay fixed. The following sections should give additional insight into the high-level ideas surrounding pre-training and adaption.

A Metaphor for CCOBRA’s Learning

To understand the learning procedure of CCOBRA, it is useful to consider an example. Let’s assume that there is a video platform which intends to provide its users with content suggestions they are likely to enjoy (predict).

If a new user creates an account the platform has no idea about the individual preferences but is still supposed to give useful suggestions. One way to acquire information about useful initial guesses is by considering the preferences of the population. By learning which content most other users enjoy, it becomes possible to identify trends which may also be of interest to unknown users (pre_train).

In some cases the system can utilize information about the new user that is available from other sources. For example, the user might link his account to an existing social-media account, which allows the system to retrieve some background information about the user, e.g., tags used by the user or videos that were liked by user. This information can be utilized to create an initial user profile (pre_person_background).

By interacting with the user and learning about which suggestions are aligned, a user profile is constructed which contains individual preferences and distinguishing features (adapt). The more the user interacts with the platform, the more knowledge can it amass. This information can then be integrated into the suggestion generation algorithm to provide suggestions ever increasing in quality.

Finally, assume that the video platform started without a system to recommend content and the user was already using the platform. In this case, the old interactions have to be taken into account as well (pre_train_person). This case is similar to the adaption, with the difference that the information is not presented one at a time, but as a batch of interactions instead. By default, CCOBRA implements pre_train_person by calling adapt for each recorded interaction of the user.

Pre-Training: Global Effects

Model pre-training in CCOBRA has the goal to leverage given information about the problem domain (training dataset) in order to extract an initial parameterization. This should allow for a sensible warm-start of the model and serve as the basis for fine-grained tuning to the individual. Pre-training is only performed in the beginning, once per participant. After the model is pre-trained, the predict-adapt-cycle begins.

Pre-Training: Person Background

CCOBRA can provide models with background information about a participant. This information is not part of the problem domain, but consist of behavioral data of the particpant in other domains. Models can utilize this information to adjust their initial parameterization.

Pre-Training: Person

In some scenarios, the warm-start of models should be improved even further. CCOBRA offers the possiblity to pass parts of the data to be predicted in advance. This way, models are given the opportunity to adjust their parameterization to the specific participant before the the predict-adapt-cycle begins.

Adaption: Individual Effects / User Profile

Adaption is the actual training of the model. It immediately follows a predict-call and contains the task information as well as the true response of the individual being modeled. It should be used to infer information about the specific reasoning processes of the individual being models. Adapt is called once per task, directly following the corresponding prediction.

Note

The benchmark specification decides wether or not the adaption and pre-training options mentioned above are available. Models for the CCOBRA framework should be written to utilize but not expect the additional information.