Predictive Text Input in a Mobile Shopping Assistant - CiteSeerX

8 downloads 0 Views 121KB Size Report
Feb 11, 2009 - interface design for integrating the predictive text input with ... perspective, mobile assistants can help to organize shopping visits and reduce ...
Predictive Text Input in a Mobile Shopping Assistant: Methods and Interface Design Petteri Nurmi, Andreas Forsblom, Patrik Flor´een Helsinki Institute for Information Technology HIIT Department of Computer Science, P.O. Box 68, FI-00014 University of Helsinki, Finland [email protected] ABSTRACT

The fundamental nature of grocery shopping makes it an interesting domain for intelligent mobile assistants. Even though the central role of shopping lists is widely recognized, relatively little attention has been paid to facilitating shopping list creation and management. In this paper we introduce a predictive text input technique that is based on association rules and item frequencies. We also describe an interface design for integrating the predictive text input with a web-based mobile shopping assistant. In a user study we compared two interfaces, one with text input support and one without. Our results indicate that, even though shopping list entries are typically short, our technique makes text input significantly faster, decreases typing error rates and increases overall user satisfaction. Author Keywords

user interface, adaptive, recommendations, usability ACM Classification Keywords

H.5.2 Information interfaces and presentation: User interfaces; H.3.5 Information storage and retrieval: On-line information systems INTRODUCTION

Grocery shopping is a fundamental activity that (almost) everyone performs on a frequent and regular basis. According to consumer studies, shopping lists play a central role in grocery shopping [1, 10, 13]. Among other things, shopping lists serve as memory aids and as a tool for budgeting. Moreover, studies on list usage have shown that the (vast) majority of consumers use, either written or mental shopping lists, at least for major shopping visits. Grocery shopping is a rather complex and often distressing process [5]. The inherent complexity, and the fundamen-

Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. IUI’09, February 8 - 11, 2009, Sanibel Island, Florida, USA. Copyright 2009 ACM 978-1-60558-331-0/09/02...$5.00.

Peter Peltonen, Petri Saarikko Helsinki Institute for Information Technology HIIT P.O. Box 9800, FI-02015 Helsinki University of Technology TKK, Finland [email protected] tal nature of grocery shopping make it an interesting domain for intelligent mobile assistants. From a consumer’s perspective, mobile assistants can help to organize shopping visits and reduce the time spent in the shop. On the other hand, from a retailer’s perspective, mobile assistants provide, among other things, a medium for targeted advertisement and shop layout optimization. Text input on mobile devices is often slow and cumbersome. This suggests that mobile shopping assistants should support techniques that facilitate creating shopping lists. In this paper we describe a predictive text input technique that uses association rules and item frequencies, as well as introduce an interface design for integrating the predictive text input with a web-based mobile shopping assistant. In a user study we compared two interfaces, one with text input support and one without. Our results indicate that, even though shopping list entries are typically short, our technique makes text input significantly faster, decreases typing error rates and increases overall user satisfaction.

RELATED WORK

There are two kinds of mobile shopping assistants: those for shopping malls [2, 4] and those for individual shops [3, 12]. Shopping assistants have been envisaged for navigation support (both finding shops in malls and products within shops), promotions, and many other functions [3, 7]. Assistants for individual shops tend to focus on product level information whereas mall assistants tend to focus on navigation and shop locating features. The scenarios and user studies described in [10] present central features for a shopping assistant and stress the design challenges following from providing information on a small screen and the need for the users to simultaneously handle groceries and the device. In earlier work, relatively little attention has been given to the creation of shopping lists. Either the shopping assistant presupposes a shopping list has been made [2], or provides a predefined list of items to choose from [12]. Ideas of using default items in the shopping lists and inclusion of items from recipes have been presented in [12]. Our idea is to support the creation of the shopping lists using predictive text input, based on predicting additional items for the shopping list.

PREDICTIVE TEXT INPUT

Various techniques have been proposed for facilitating mobile text entry; see [9]. One of the most popular techniques is to use language prediction. In language prediction, the idea is to model the input language and automatically suggest what the user might be entering based on what the user has inputted thus far. Language prediction tools have the potential to significantly facilitate text input, as long as the predictions are good [9]. Typically language prediction tools assume that the words that are entered are uncorrelated. However, in the grocery domain this assumption is seldom valid as users tend to purchase certain items together (e.g., ketchup and macaroni). Motivated by this, we designed a novel predictive text input functionality that takes item-wise correlations into account. An important aspect in the design of text entry studies is to ensure that the inputted text is representative of the target domain’s vocabulary [9]. Hence, in our case the inputs should be representative of items that users actually use in their shopping lists. We initially considered using real shopping lists, but this turned out to be infeasible for two reasons. First of all, our sample of real shopping lists (100) was too small for generating a useful dictionary. Secondly, the lists contained many non-words, colloquialisms and abbreviations, which could influence the user evaluations, especially if participants are unfamiliar with some of the words. In the end, we decided to use synthetic shopping lists. We implemented a shopping list generator for this purpose. Our shopping list generator randomly samples products from actual shopping data provided by a partner Finnish hypermarket. We limited the sampling to grocery items. The sampled products are then mapped into shopping list entries. First of all, we remove all package and unit sizes from the actual product name. Next we use a stop word list to remove words that correspond to popular manufacturers, and adjectives describing product properties. The remaining words are lemmatized, i.e., reduced to the grammatical base form. Finally, we calculate the TF-IDF (term frequency - inverse document frequency) score for each word in the product name, and return the word with the lowest TF-IDF score. The word with the lowest TF-IDF score has the lowest discriminative power, and hence it is likely to correspond to a generic term. As an example, {milk, coffee, tomato, potato, beer} is a shopping list generated by our shopping list generator. To ensure that the dictionary of the predictive text input functionality was representative of the lists that were used in the evaluation, we generated 10000 shopping lists with 10 items each. The text input tool is based on association rules and item frequencies. The predictions are based on association rules that are mined from our 10000 shopping lists. Our system contained 8029 association rules sorted by their confidence. When the user starts typing text, the user input is matched against associations rules that are triggered by items on the user’s shopping list. For example, assume the user has a shopping list with one item: ketchup. If the user now types

Figure 1. The interface for the predictive text input functionality.

m, the predictive tool could suggest macaroni before milk. If the rules do not trigger enough predictions, or the user does not have any items on his/her shopping list, frequent items are used instead. Currently we show six predictions at a time as this is the maximum amount we can easily fit on the screen of a Nokia E61i mobile phone, our target device. In order to interact with the predictive input tool, the user simply needs to start typing. The user input opens a suggestion box that shows the six best predictions; see Fig. 1. In order to add an item to the list, the user selects the appropriate entry from the suggestion list and clicks on the item. Naturally, the user can also type the entire product. If the user inputs space or backspace as the first character, this opens the suggestion box without any user input. In this case, the suggestion box shows default recommendations; see Fig. 1. The default recommendations are based on the same association rules as the predictions. EVALUATION SETUP

We evaluated the predictive text input functionality in a user study where users were asked to type a predefined shopping list with ten items that was generated with our shopping list generator. Users were given two different versions of the interface. One version, referred to as PREDICTIVE, used our predictive text input functionality, whereas the other, STANDARD, did not have any text input support. As the STANDARD version did not contain any text input support, it also did not show the suggestion box. As our goal was also to evaluate the mobile interface design, we did not consider standard, item frequence-based predictive input in the evaluation. Instead we have used offline analyses to compare the association rule predictions against item frequency based predictions; see the Summary and Discussion section. The interface that was presented first to the participants was randomized. In order to ensure that there are no differences in the distribution of word lengths across the two systems, we gave the participant the same shopping list for both versions. Regression analysis was used to verify that this did not cause biases to the evaluation; see next section. Be-

fore the user started inputting the shopping list, we demonstrated how the corresponding version of the interface functioned. After a participant finished inputting the shopping list with one of the versions, we presented him/her with a two-question questionnaire based on the IBM computer usability satisfaction questionnaire [8]: ”Overall, I’m satisfied with the ease of completing this task”; ”Overall, I’m satisfied with the amount of time it took to complete this task”. Some participants were asked to perform the study using single handed input whereas others were allowed to use both hands for input. Our motivation for considering separately one- and two-handed input was that this allows us to emulate home and in-store usage, the two most common contexts for using shopping list creation or management features [10]. At home, users are likely to be able to make time and place for the interaction, i.e., they can stop other activities while interacting with the mobile device. This suggests that two-handed input is well-suited for the home setting. On the other hand, inside the store the user must observe other persons, grab groceries and make time for the interaction. Field studies have suggested that one-handed interaction is an important design driver for situations where users must make time and place for interaction as it frees some of the user’s cognitive resources for other tasks [6, 11]. ANALYSIS AND RESULTS

The evaluation was performed as an intercept study at the premises of two university buildings. Participants were given candy for participating in the study. In total 30 participants participated. The majority of participants were between 18 and 25 years old (19 participants, 63%). The remaining participants were between 26 and 40 years old. Thirteen participants were females and the remaining seventeen were males. All participants reported that they used a mobile phone for writing at least 10 times a month, and the majority (20 participants, 67%) used the phone for writing at least 50 times a month. Eighteen participants had no previous experience with mobile web. Of the twelve participants who had previous experience, six indicated that they used mobile web very often, while the remaining participants used mobile web occasionally (three participants) or very rarely (three participants). Two participants had to be removed from the analysis. One participant was removed because a technical error caused the PREDICTIVE version to malfunction during the test. The second participant was removed because of a substantially high typing error rate with the STANDARD version1 . Of the remaining 28 participants, 13 started with PREDICTIVE (7 one-handed, 6 two-handed) and 15 with STANDARD (6 one-handed, 9 two-handed). The version that was used first was encoded as an explanatory variable and we used regression analysis to verify that the version that was used first did not influence the test results. 1 The error rate of the participant was ≈ 19%, whereas the error rate of other participants was between 0 and 5%

PREDICTIVE MED IQR 14.35 5.77 13.00 3.65 17.25 7.13

ALL 1-hand 2-hand

STANDARD MED IQR 10.34 4.10 9.62 2.82 10.81 4.28

Z-Score 3.34 ** 2.23 * 2.74 **

Table 1. Median and interquartile range of the task completion rates measured as words per minute (wpm). The row ALL contains a summary of all results, whereas the rows 1-hand and 2-hand contain the results for the input mode (one-handed or two-handed). The final column contains Z-scores from a Mann-Whitney U-Test and the corresponding significance levels (∗ = 0.05, ∗∗ = 0.01).

PREDICTIVE 2.46 1.18 0.71

MEAN STDV α

STANDARD 3.73 1.47 0.86

Table 2. Summary of the user satisfaction scores. The values are between 1 and 7, and lower values are better. The variable α indicates the reliability of the scale (Cronbach’s α).

As the unit of analysis, we consider the standardized measure words per minute (wpm), and refer in the following to this as the task completion time. Using wpm makes it possible to compare our results with other input methods; e.g., handwriting speeds are reported to be 15-25 wpm in text copying [9]. We obtain the wpm by scaling from characters per second, using the standard definition of one word corresponding to five characters, i.e., the speed one character per second is noted as 12 wpm. The task completion rates are summarized in Table 1. As the table indicates, participants achieved a significantly higher wpm when using the PREDICTIVE version. The difference was found significant with both input modes, though the difference was somewhat bigger with the two-handed input mode. Part of the slowness is caused by delays with the browser interaction. For example, there is a slight delay between adding an item to the shopping list and starting the input of the next item. We next analyzed the influence of background variables on the task completion rates. The analysis was performed using generalized linear models. After comparing different models we settled for an inverse Gaussian regression model. For STANDARD, the regression analysis indicated that gender (T = −2.65, p < .05) was a significant predictor of task completion rate. More specifically, males were somewhat faster than females (median 11.07 wpm vs 9.70 wpm). For PREDICTIVE, the input mode (T = −3.12, p < .01), i.e., whether the participant used one-handed or two-handed input, and frequency of text input (T = 3.34, p < .01) had an influence on the task completion rates. More specifically, persons using two-handed input were significantly faster than persons using one-handed input (median 17.25 wpm vs 13.00 wpm) and participants using mobile text entry over 50 times per month were around 40% faster (median 15.51 wpm vs 11.84 wpm).

As the next step of analysis, we analyzed the questionnaire responses of the users. The results of the questionnaire are summarized in Table 2. As the scale reliability was relatively high for both systems, we performed the analysis on scale level. In other words, we averaged the responses to the two questions and analyzed the averaged responses instead of performing question-by-question analysis. Oneway ANOVA indicated that the satisfaction scores were significantly different (F = 12.7, p < .001). Hence, users were more satisfied with the PREDICTIVE version. We also performed regression analysis on the scale scores. None of the background variables were found to have a significant influence on the satisfaction scores of either version.

3. C. Cumby, A. Fano, R. Ghani, and M. Krema. Building intelligent shopping assistants using individual customer models. In Proceedings of the 10th international conference on Intelligent user interfaces (IUI’05), pages 323 – 325. ACM, 2005.

We also separately analyzed the typing error rates. In total, participants made six errors when they used the PREDICTIVE version, and 33 errors when they used the STANDARD version. These values correspond to error rates of 0.3% and 1.4%. In other words, the PREDICTIVE version decreased the error rate by around 80%. We also verified that the difference was statistically significant using a Chisquared test (χ2 = 29.12, p < .05).

6. S. Kristoffersen and F. Ljungberg. Making place to make it work: Empirical exploration of HCI for mobile CSCW. In Proceedings of the International Conference on Supporting Group Work (Group’99), pages 276 – 285. ACM, 1999.

SUMMARY AND DISCUSSION

8. J. R. Lewis. IBM computer usability satisfaction questionnaires: psychometric evaluation and instructions. International Journal of Human-Computer Interactions, 7(1):57–78, 1995.

The predictive text input functionality increased the user’s input speed. On average, the increase was around 5 words per minute, somewhat more for persons using two-handed input than for persons using one-handed input. In addition, the text input support significantly reduced the error rates of the users (over 80%) and increased user satisfaction. Offline analysis regarding the usefulness of association rules for text input suggests that association rule suggestions are often selected (over 70% of the time). We have also compared the association rule predictions against item frequency based predictions. Our results suggest that the overall amount of input decreases by 2 − 3% (p < .001), and we are currently examining ways to further enhance the performance of the association rule predictions. Acknowledgments

The authors are grateful to their present and past colleagues in the project. This work was supported by the Finnish Funding Agency for Technology and Innovation TEKES, under the project Personalised Ubiservices in Public Spaces. The work was also supported in part by the ICT program of the European Community, under the PASCAL2 network of excellence, ICT-216886-PASCAL2. The publication only reflects the authors’ views. REFERENCES

1. L. G. Block and V. G. Morwitz. Shopping lists as an external memory aid for grocery shopping: Influences on list writing and list fulfillment. Journal of Consumer Psychology, 8(4):343–375, 1999. 2. T. Bohnenberger, O. Jacobs, A. Jameson, and I. Aslan. Decision-theoretic planning meets user requirements: Enhancements and studies of an intelligent shopping guide. In Proceedings of the 5th International Conference on Pervasive Computing (Pervasive), pages 279–296. Springer, 2005.

4. A. E. Fano. Shopper’s eye: using location-based filtering for a shopping agent in the physical world. In Proceedings of the second international conference on Autonomous agents (AGENTS), pages 416–421. ACM, 1998. 5. E. H. Fram and J. Axelrod. The distressed shopper. American Demographics, 12(10):44–45, 1990.

7. S. Kurkovsky and K. Harihar. Using ubiquitous computing in interactive mobile marketing. Personal and Ubiquitous Computing, 10(4):227–240, 2006.

9. I. S. MacKenzie and R. W. Soukoreff. Text Entry for Mobile Computing: Models and Methods, Theory and Practice. Human-Computer Interaction, 17(2):147–198, 2002. 10. E. Newcomb, T. Pashley, and J. Stasko. Mobile computing in the retail arena. In Proceedings of the SIGCHI Conference on Human Factors in Computing Systems (SIGCHI), pages 337–344. ACM, 2003. 11. A. Oulasvirta, S. Tamminen, V. Roto, and J. Kuorelahti. Interaction in 4-second bursts: The fragmented nature of attentional resources in mobile HCI. In Proceedings of the SIGCHI Conference on Human Factors in Computing Systems, pages 919–928. ACM Press, 2005. 12. S. Shekar, P. Nair, and A. S. Helal. iGrocer: a ubiquitous and pervasive smart grocery shopping system. In Proceedings of the ACM symposium on Applied computing (SAC), pages 645–652, 2003. 13. A. Thomas and R. Garland. Grocery shopping: list and non-list usage. Marketing Intelligence & Planning, 22:623 – 635, 2004.

Suggest Documents