User Behavior for Neural Network-based Web Search Results Filtering

This paper describes methodology and performance of an experimental research on filtering of web search results. Filtering was performed on the basis of predicted relevance of search results derived from users’ implicit feedback. The feedback was obtained from users’ web browsers and consisted of a set of browsing behavioral metrics, including reading time, clicks on links, mouse pointer and wheel movement patterns, bookmarking, sharing, copying, and whether the search was continued after the page was closed. A multi-layer neural network used to infer from the behaviors how much the user was interested in each filtered document. Neural network, therefore, performed deep learning without human supervision. Predicted relevance measure was compared to the explicit feedback. Obtained results of 89% correct relevance rating prediction suggest that selected set of metrics was successful in terms of correctly predict how relevant the web page was for the user involved in the study. More research is recommended for further advances of information filtering methods.

Depth, Connectedness, and Structural Relevance in the Object-Superiority Effect: Line Segments are Harder to See in Flatter Patterns

A briefly flashed line can be identified more accurately when it is part of certain types of pattern than in others (the ‘object-superiority effect’). Three experiments were designed to investigate what aspects of these patterns determine the facilitatory effect of context. Subjects identified which of four line segments was present in various briefly flashed figures. Other subjects rated the figures for three-dimensionality, connectedness, and ‘structural relevance’ of the target line. Little relationship was found between connectedness ratings and accuracy in the identification task, but accuracy was highly correlated with mean depth rating (accounting for 95% of the variance) and with mean structural-relevance rating (88%). Because of the high correlation ( r = 0·98) between these two judgments in the present experiments, and confounding with other stimulus variables in previously published studies, the relative importance of these two global attributes cannot yet be determined definitively (though there was some evidence that for these patterns depth judgments were primary and structural-relevance judgments derivative). A reexamination of pertinent research suggests that comparisons between well-matched stimuli (as in the object-superiority effect) are likely to be more robust and informative than comparisons between lines alone and in context (the ‘object-line effect’).