In exploring the discriminative power of each feature on the performance of the proposed system, we find that features regarding a user’s behavior are robust for functional naming, and the environmental features mined from images or postings are effective for business naming. Figure 13(a) presents the residence-time distribution of several place types on weekdays. We found that the stay behaviors have strong features that can differentiate some types (e.g., residences and food places), but several types (e.g., colleges and professional offices) are confused, due to similar stay patterns. We further investigated the difference of distributions between places, as shown in Figure 13(b). The figure illustrates the Kullback-Leibler divergence [15] between the residence-time distributions. The results indicate that places of the same type have higher similarities compared to places of different types. We highlight the characteristics of the environmental features mined from images (i.e., GIST, SIFT, and OCR words). Figure 14 presents the distribution of attributes mined from images. We presented a few selected attributes since the semantic features from images were sparsely distributed over a large number of attributes. The result shows that clustered terms were observed differently on some types of places that can differentiate places. The example images in the same cluster inferred as food place are presented in Figure 15(a). The images contains cups or bowls with round shapes that derived similar terms. Similarly, the terms mined from images at the same place show consistent patterns since those images contains similar scene or local objects, as illustrated in Figure 15(b). In OCR words cases, the system observed a high frequency of OCR words in food places and most images captured at shopping and travel and transport contained written words, as illustrated in Figure 16. The words were mostly mined from images capturing store signs, menus, or a direction boards in a place. To filter out noisy results, we utilized the confidence score provided by the OCR engine. Figure 17(a) shows that false positives decrease as the threshold of confidence score increases. The result shows that even though the classifier produces noisy results, reliable results can be obtained using high threshold of confidence. The side-effect of this approach is a significant drop in true positives, but the system still extracts sufficient amount of results due to the large amounts of data collected by crowdsensing and crowdsourcing. We set the threshold of confidence as 300 to obtain high precision with small loss in recall, as illustrated in Figure 17(b). The low recall indicates that the classifier would require further investigation to obtain more classified features from noisy images. Table 3 indicates that major words on different types of places reflected the characteristics of places. Finally, we explored the relationship between the number of samples and the performance of the proposed system. Intuitively, as shown in Figure 18, the more data leads a more accurate result. We found that a larger number of visits induced better accuracy in functional naming, and a greater number of images is effective for business naming. Places with more than 100 visits showed 80% of accuracy, while more than 20 images derived 4.3±4.9 ranks in the list. The reason is that the number of matched features of the same places increases with the number of images. The visitations would increase as time goes by in daily life, but the collection of images requires the active user participation. Considering that 9.7% of places in SNS and 34% of places in crowdsensing contain image data, the reward for data collection is required to induce user participation.
In exploring the discriminative power of each feature on the performance of the proposed system, we find that features regarding a user’s behavior are robust for functional naming, and the environmental features mined from images or postings are effective for business naming. Figure 13(a) presents the residence-time distribution of several place types on weekdays. We found that the stay behaviors have strong features that can differentiate some types (e.g., residences and food places), but several types (e.g., colleges and professional offices) are confused, due to similar stay patterns. We further investigated the difference of distributions between places, as shown in Figure 13(b). The figure illustrates the Kullback-Leibler divergence [15] between the residence-time distributions. The results indicate that places of the same type have higher similarities compared to places of different types. We highlight the characteristics of the environmental features mined from images (i.e., GIST, SIFT, and OCR words). Figure 14 presents the distribution of attributes mined from images. We presented a few selected attributes since the semantic features from images were sparsely distributed over a large number of attributes. The result shows that clustered terms were observed differently on some types of places that can differentiate places. The example images in the same cluster inferred as food place are presented in Figure 15(a). The images contains cups or bowls with round shapes that derived similar terms. Similarly, the terms mined from images at the same place show consistent patterns since those images contains similar scene or local objects, as illustrated in Figure 15(b). In OCR words cases, the system observed a high frequency of OCR words in food places and most images captured at shopping and travel and transport contained written words, as illustrated in Figure 16. The words were mostly mined from images capturing store signs, menus, or a direction boards in a place. To filter out noisy results, we utilized the confidence score provided by the OCR engine. Figure 17(a) shows that false positives decrease as the threshold of confidence score increases. The result shows that even though the classifier produces noisy results, reliable results can be obtained using high threshold of confidence. The side-effect of this approach is a significant drop in true positives, but the system still extracts sufficient amount of results due to the large amounts of data collected by crowdsensing and crowdsourcing. We set the threshold of confidence as 300 to obtain high precision with small loss in recall, as illustrated in Figure 17(b). The low recall indicates that the classifier would require further investigation to obtain more classified features from noisy images. Table 3 indicates that major words on different types of places reflected the characteristics of places. Finally, we explored the relationship between the number of samples and the performance of the proposed system. Intuitively, as shown in Figure 18, the more data leads a more accurate result. We found that a larger number of visits induced better accuracy in functional naming, and a greater number of images is effective for business naming. Places with more than 100 visits showed 80% of accuracy, while more than 20 images derived 4.3±4.9 ranks in the list. The reason is that the number of matched features of the same places increases with the number of images. The visitations would increase as time goes by in daily life, but the collection of images requires the active user participation. Considering that 9.7% of places in SNS and 34% of places in crowdsensing contain image data, the reward for data collection is required to induce user participation.
การแปล กรุณารอสักครู่..
