Predicting change in time production -- A machine learning approach to time perception

TL;DR

The study addresses how to quantify and predict changes in subjective time production at the individual level within ecologically valid, naturalistic stimuli. It trains a logistic-regression model on data from 995 online participants watching driving-scene videos to predict the direction of change in production time, achieving $61\%$ accuracy and generalizing to a separate offline driving-simulator dataset with $66\%$ accuracy. The model’s output probabilities also encode information about the magnitude of change, and SHAP-based explanations link features to components of the attentional-gate theory, notably showing prior timing performance as a key driver and environmental cues as modulators of attentional gates. These findings offer a quantitative, interpretable bridge between cognitive timing theories and data-driven prediction, with potential applications in human-computer interaction systems that adapt to a user’s time perception, such as ChronoPilot.

Abstract

Time perception research has advanced significantly over the years. However, some areas remain largely unexplored. This study addresses two such under-explored areas in timing research: (1) A quantitative analysis of time perception at an individual level, and (2) Time perception in an ecological setting. In this context, we trained a machine learning model to predict the direction of change in an individual's time production. The model's training data was collected using an ecologically valid setup. We moved closer to an ecological setting by conducting an online experiment with 995 participants performing a time production task that used naturalistic videos (no audio) as stimuli. The model achieved an accuracy of 61%. This was 10 percentage points higher than the baseline models derived from cognitive theories of timing. The model performed equally well on new data from a second experiment, providing evidence of its generalization capabilities. The model's output analysis revealed that it also contained information about the magnitude of change in time production. The predictions were further analysed at both population and individual level. It was found that a participant's previous timing performance played a significant role in determining the direction of change in time production. By integrating attentional-gate theories from timing research with feature importance techniques from machine learning, we explained model predictions using cognitive theories of timing. The model and findings from this study have potential applications in systems involving human-computer interactions where understanding and predicting changes in user's time perception can enable better user experience and task performance.

Figures (12)

• Figure 1: Screenshots of the videos used as naturalistic stimuli in the main experiment. The leftmost image corresponds to low engagement level, the middle image to medium engagement level and the rightmost image to high engagement level.
• Figure 2: Main experiment design: One trial included a video followed by a questionnaire. The two trials were time production tasks, where participants had to stop the video after they thought 30 seconds had passed. In the questionnaire following each video, participants were asked about the video content (attention checks), self-evaluation of their timing performance and their visual and timing perception
• Figure 3: Prediction model: The machine learning model uses the selected features from the previous trial and the engagement level of the upcoming trial's video as inputs. Its output is a probability of decrease, which reflects the direction of change in production time. Additionally, these probabilities were found to contain information about the magnitude of change in production time (as detailed in the Results section). Using this information, we can approximately infer how long a participant might wait before stopping the video
• Figure 4: (a) The cognitive counter accumulates all or fewer ticks generated by the pacemaker depending on the attention-gate width. Arousal caused by a stimuli or the environment, affects the pacemaker by increasing clock speed. Similarly, attention to time determines the width of the attention-gate component and, contextual bias like regression to the mean influences reference memory. The decision phase involves, a comparator that waits until the accumulated ticks in cognitive counter are more than or equal to the reference memory ticks in order to take an action associated with marking the end of the target interval. (b) The white ovals depict the broad feature categories used by the machine learning model. These features can indicate a change in either reference memory or cognitive counter. We assume that self-evaluation of timing performance and prior timing performance capture information related to reference memory changes (contextual calibrations like regression to the mean and prior timing experience). While, environmental characteristics and participant sensitivity capture information about changes in cognitive counter (arousal or change in attention to time cause by stimuli or environment). For simplicity, we ignore any other factors that can cause a change within the two components of the attentional-gate model.
• Figure 5: Probability evaluation for magnitude and direction of change prediction: The graph shows the trained logistic regression model's predicted probability of decrease against $\Delta$T (difference in time produced between two trials) for all participants (706). The probability of decrease was obtained using LOOCV. These probabilities have been divided into three levels (indicated by the vertical grey dotted lines): low (probability $<$ 0.4), uncertain (0.4 $<$ probability $<$ 0.6) and high (probability $>$ 0.6). Similarly, $\Delta$T has been divided into three levels (indicated by the horizontal grey dotted lines): high decrease ($\Delta$T $<$ -5), small change (-5 $<$$\Delta$T $<$ 5) and high increase ($\Delta$T $>$ 5). The green *'s correspond to the cases when the predicted probability conveyed the correct information about magnitude of change level, the red X's correspond to cases when the predicted probability conveyed extremely incorrect information, and the black dots correspond to other slightly incorrect cases. Table \ref{['table:cm_magnitude']} shows the number of data points in each of the nine regions created by the grey dotted lines (both horizontal and vertical). The vertical and horizontal blue lines corresponds to probability of 0.5 and $\Delta$T of zero respectively. The four quadrants formed by these lines can be used as reference to evaluate the direction of change predictions. All points in the bottom-right and top-left correspond to correct direction of change prediction and all points to the top-right and bottom-left correspond to incorrect predictions.