=Paper=
{{Paper
|id=None
|storemode=property
|title= Intrapersonal Retrospective Recommendation: Lifestyle Change Recommendations Using Stable Patterns of Personal Behavior
|pdfUrl=https://ceur-ws.org/Vol-891/LIFESTYLE2012_paper3.pdf
|volume=Vol-891
}}
== Intrapersonal Retrospective Recommendation: Lifestyle Change Recommendations Using Stable Patterns of Personal Behavior==
Intrapersonal Retrospective Recommendation: Lifestyle Change Recommendations Using Stable Patterns of Personal Behavior Robert G. Farrell, Catalina M. Danis, Sreeram Ramakrishnan, Wendy A. Kellogg IBM, T J Watson Research Center 19 Skyline Drive Hawthorne, NY 10532 {robfarr, danis, sramakr, wkellogg}@us.ibm.com ABSTRACT regimen comprised of flexibility, strengthening, and difficult to Leading a healthy lifestyle can prevent or delay medical maintain over time as motivation and commitment wax and wane. conditions, elevate mood, improve energy, stabilize sleep, and Thus, there is an opportunity for recommender systems (Ricci, have other positive effects. Recommender systems are one Rokach, and Shapira, 2010.) to suggesting incremental changes to possible technology to support making lifestyle changes. one’s routines that collectively can bring about a lifestyle change. Recommender systems often use ratings of other users to make recommendations, but this approach may be problematic for making lifestyle change recommendations because of the large This paper introduces Intrapersonal Retrospective variations in human behavior . This paper proposes Intrapersonal Recommendation (IRR) as a promising method of generating Retrospective Recommendation as a new method for generating lifestyle change recommendations. The key idea behind this lifestyle change recommendations that uses only personal history. approach is that recommendations can be based on what behaviors We explain the benefits and drawbacks of this approach and worked and did not work for the individual in the past. Stable suggest some future directions. patterns of behavior within a prior time period may be more predictive of an individuals’ future behavior than the common Categories and Subject Descriptors behavior patterns of other users. Hence, behavioral patterns in H5.m. Information interfaces and presentation (e.g., HCI): periods of success at lifestyle change or maintenance that are not Miscellaneous. being followed can be recommended when the user is facing a similar goal but not succeeding. Similarly, the system can General Terms recommend cessation of behavior patterns that are found in prior Design, Experimentation, Human Factors. periods of failure at lifestyle change or maintenance as long as a similar goal is being pursued. Keywords The remainder of the paper is structured as follows: the next Lifestyle change, habit, retrospective, recommendation, section introduces the problem of lifestyle change recommender system recommendation; section three explains some of the problems with using collaborative filtering in this domain; section four introduces the IRR method; section five reviews some related 1. INTRODUCTION work; section six provides a real-world example; and section Changing long-term behavior can be a challenging task for seven provides a generalized algorithm. We wrap up with a anyone. Bad habits can be entrenched and good habits difficult to discussion and some conclusions. establish, transition and sustain. In this work, we focus on lifestyle changes motivated by maintaining or improving personal health. 2. THE PROBLEM For example, changing a diet in order to achieve weight loss, or We define lifestyle as the pattern of behavior choices an individual supporting the individual in evolving a sustainable exercise makes during a period of time. Mobile phones, sensors, and other devices are making it increasingly possible to collect fine-grained behavioral data about individuals, often with little work on the part of the user. In addition, tools1 such as Lose It!, DailyBurn™ Paper presented at the Workshop on Recommendation Technologies for FitDay™, and MyNetDairy™ allow their users to track their food Lifestyle Change 2012, in conjunction with the 6th ACM conference on Recommender Systems. Copyright © 2012 for the individual papers is held by the papers' authors. This volume is published and copyrighted by 1 LoseIt! is an unregistered trademark of FitNow, Inc., its editors. DailyBurn™ is a registered trademark of Daily Burn, FitDay is a registered service mark of Internet Brands, Inc., MyNetDiary Lifestyle@RecSys’12, September 13, 2012, Dublin, Ireland is a registered trademark of 4Technologies Corporation. and/or exercise manually on a web site or smart phone. This different calorie targets and this impacts food choices and tracking makes it possible for users to maintain a long-term fine- exercise regimens. grained history of their lifestyle-related choices. In addition, 3. Varied Contexts: Ratings across individuals in real-world meters, scales, and so on allow individuals to track measurements situations, whether explicit or implicit, may diverge because such as weight, waist size, and number of calories above or below of the varied contexts in which ratings were collected. For some target budget or level to determine progress against their example, one user may rate pizza high and another low goals over time. because of the quality of the different pizza parlors they We have been developing a system to analyze a person’s tracking frequent. data and display information useful for making and maintaining a 4. Distinctiveness: The diversity of ratings may be exacerbated healthy lifestyle. Alone, these data may be too detailed for people by the fact that, for many people, lifestyle is, by definition, to distinguish meaningful patterns. Recommendations have a aimed at being distinctive. People may seek out items that significant role to play in helping users dynamically make are unique or significantly different than others. For example, intelligent choices that help achieve their goals. In our system, one person might search out exotic foods and another each individual has a set of lifestyle goals in the form of unusual places to exercise, thus reducing their similarity. constraints to satisfy over an interval of time. For example, Bob wants to do aerobic exercise three times per week. Aerobic While many of these issues can be addressed by having a larger exercise is an activity, each time that Bob performs that activity is data set, a technique is needed that is not sensitive to the sparse a behavior and repeated behaviors are a behavioral pattern (i.e., item space, diverse user characteristics and goals, varied contexts, Bob aerobic exercise three times per week). Given this and distinctiveness. formulation, the recommendation problem is to suggest one or more activities, either individually or in a sequence, for a user, 4. INTRAPERSONAL RETROSPECTIVE given their history of behaviors by finding stable behavioral RECOMMENDATIONS patterns. An alternate technique for generating recommendations is to use In the examples in this paper, we use the term item to refer to a the user’s own history of choices. We call this Intrapersonal food (an eating activity) or exercise (a physical activity). This Retrospective Recommendation (IRR). IRR is based on the could be at various level of specificity. For example, the item observation that most people are highly stable (i.e., change slowly could be “coffee” or “coffee with crème and sugar” or the item over time) and highly distinctive (i.e., different that others) in their could be “running” or “running two miles in 10 minutes”. While it lifestyle-related choices. When a set of outcome measures are is possible for users to explicitly rate items, tracking data tracked (e.g., weight), IRR can return users to stable patterns that indicating that a user consumed a food or performed an exercise is worked in the past (i.e., where measures were consistent with their itself an implicit rating. We consider the count of the number of goals) or avoid stable patterns that did not work in the past (i.e., times the user performed the behavior, total amount of the food or where measures were short of goals). We call a stable pattern of time exercising, and total calories burned or consumed to be part lifestyle-related choices over a time period a prior self. In the of the implicit rating of an item. lifestyle change domain, for example, an individual may have the prior self who ran at least twice a week and who consistently ate 3. CHALLENGES WITH USER-USER high calorie desserts and the prior self who did not exercise. A key insight is that prior selves can serve the role of “similar users” SIMILARITY APPROACHES for the purposes of recommendation. Collaborative filtering (Resnick et. al., 1994) is one of the most successful approaches to generating recommendations. It uses the To illustrate this idea, suppose Rachel has toast and coffee every known ratings of a group of users to make predictions about the morning and is maintaining her weight. One day she starts unknown ratings of other users. The prediction accuracy of buttering her toast every day. If Rachel’s peers have coffee and collaborative filtering depends on the similarity of the ratings orange juice for breakfast, then collaborative filtering would from the group of users. There are several reasons to believe that recommend orange juice to Rachel. IRR would instead LCRSes based on collaborative filtering may have relatively low recommend she use less butter or butter her toast less often, since prediction accuracy: during her stable period she did not butter her toast. 1. Sparse Item Space: The items in LCRSes may be selected The time dimension plays an important role in retrospective by users from a very large item space or even constructed by recommendations. For example, the “stage” of the individual with users, thus reducing the probability of two individuals rating respect to his or her goal can influence what is recommended. the same item. For example, one user may like Vietnamese Early on, simply limiting quantity might lead to successful frog’s legs but it might not ever appear on the menus of other adherence and weight loss. Later, once the individual has lost users. While comparing items at a higher level of abstraction significant weight but not yet achieved his goal, it might be may increase the item overlap, recommendations of highly necessary to also change the types of foods that are recommended. abstract items may be less useful. Even if two users rate We anticipate that IRR will not be as sensitive to some of the single items, impact on measurements may vary. For problems we outlined with user-user similarity approaches. First, example, how foods are prepared makes a large difference in the item space may be smaller because individuals may explore their calorie count. only a small number of items; trying a new item is often risky or 2. Diverse User Characteristics and Goals: There may be breaks an established habit. Second, given that individuals are large individual differences in lifestyles due to differences in relatively stable in their preferences over time, user diversity may individual characteristics (age, height, gender, weight, etc.) be lower with prior selves because individual characteristics and and goals (lose 100 pounds vs. lose 5 pounds.) For example, goals will have not changed significantly. The context of item males and females at different weights typically have quite ratings within individuals will tend to be consistent over time, improving the stability of IRR. Finally, distinctiveness is not an data to compare with other approaches, we have used IRR to issue since all comparisons are intrapersonal. generate recommendations at various points in time for each of Retrospective recommendation has a number of other advantages. these users. User A provided feedback on the suitability of the First, the lifestyle change data needed can be made private, thus recommendations. circumventing the need for data from other users that would be After three months of steadily losing weight at a rate of 3.3 required for other approaches, such as social recommending. pounds per month, User A hit a plateau i.e., his weight leveled off. Second, IRR may be more transparent (Herlocker, 2004) since the During the 4th month, User A’s net calories reached 13,500 behaviors being recommended will be familiar. Finally, calories, whereas during the previous 3 months, User A’s average recommendations from one’s own history are easier to trust than net calories had leveled off at 10,000 calories per month. Thus, we recommendations from others. can differentiate the three months of making steady progress toward the weight loss goal and the month of making little 5. RELATED WORK progress toward the goal. Other researchers have investigated how to make lifestyle change We analyzed User A’s logs over the 4 month period to find each recommendations. Van Pinxteren, Geleijnse, and Kamsteeg food and exercise frequency and cumulative calories. During the (2011) created a recipe recommender system that suggested fourth month, (period two) User A had many of the same patterns healthy alternatives to commonly selected recipes by using a of behavior as in the first three months (period one). He ate recipe similarity measure. Luo, Tang, and Thomas (2010) created Coffee, Juice and Milk at breakfast, chocolate squares and nuts as a system to recommend home nursing activities and home medical a snack, and field greens and salad dressing at lunch. However, in products. Hammer, Kim, and André (2010) describe a rule-based period two User A started drinking beer and had increased his recommender system for diabetes patients that balance short-term consumption of chocolate, wine, kung pao chicken, quiche, pizza, user preferences with long-term medical prescriptions. Sami, bananas, brown rice, and jelly. Together, these differences Nagatomi, Terabe, and Hashimoto (2008) designed a system to accounted for the majority of the calorie change. User A also at recommend leisure-time physical activities and identified the tortilla chips in month 3 but then stopped in month 4, probably a problem of varied contexts across individuals, primarily the issue positive development. However, if tortilla chips were substituting that people prefer different places to exercise. Wiesner and Pfeifer for French fries, for example, then it might be better to go back to (2010) developed a semantic distance metric for health concepts tortilla chips. While we cannot be sure that these patterns had and used it to make personalized health recommendations from an become new habits, it might still be useful to recommend changes electronic health record. None of these systems make use of the early. user’s history. Given the patterns, we generated recommendations by suggesting There has been recent interest in temporal variation, such as decreasing frequency or portion size for food and increasing adjusting recommendations to situations where the end users frequency, intensity, or time for exercise. User A then sorted the interests “drift” over time (Cao, Chen, Xiong, 2009). Koren and recommendations into three categories of suitability: “follow” Bell (2011) show how the predictive accuracy of matrix (likely to take the suggestion), “consider” (like the suggestion but factorization models can be improved using temporal information. unlikely to follow it), and ignore (don’t like the suggestion and Tanaka, Hori, and Yamamoto (2010) developed LifeLog, a won’t follow it). Here is how he sorted the recommendations: recommender system that captures a history of offline and on-line • Follow – reduce quiche and chocolate; Web activities and recommends information on Web sites to help users “enjoy waves of information again”. This system uses prior • Consider – stop beer, reduce wine and pizza; stable patterns from personal history, but does not make • Ignore – reduce kung pao chicken, bananas, and brown rice. recommendations based on success or failure relative to personal goals and it does not make lifestyle change recommendations. User A reported that eating chocolate squares was a behavior acquired during the weight loss, so it was relatively easy to moderate. The wine and pizza were more entrenched habits. In 6. AN EXAMPLE future work we would like to address the problem of entrenched We have collected log data from several individuals who are habits and strong preferences against recommended items. trying to lose weight. Here is a typical log for food and exercise: 7. AN ALGORITHM FOR GENERATING Date Item Time Number Units Calories RETROSPECTIVE RECOMMENDATIONS 3/19/12 Coffee Breakfast 32 oz 9 Given our analysis of log data, we designed the following general algorithm for IRR: 3/19/12 Melon Lunch 1 cup 61 1. Find Periods of Success and Failure: Establish the 3/19/12 Running Exercise 50 minutes 535 individual’s goals as a Boolean combination of measures to Each line indicates a particular instance of an item or type of maintain, increase, or decrease (e.g., maintain weight while behavior that happened at a date and time. The item can be a type reducing fat intake by 10%)Calculate periods of consistent of food and an amount (number and units) or a type of exercise goal achievement (e.g., maintaining weight) or failure over with an amount of time. The calories for each item are listed. time using historical data (e.g., weight 190 and fat intake of On June 1st, we collected log data from three users (A, B, and C) only 25mg/day average). A period of 1 week is used as a who logged their food and exercise daily. User A lost 20 pounds minimum length period since physical changes are difficult and logged for 6 months, User B lost 9 pounds over 7 months, and to measure in smaller periods. User C lost 55 pounds over 36 months. While all three users were 2. Find Stable Patterns: Identify stable patterns as repeated undergoing significant weight loss, their trajectories included items within each period contributing most and least to the plateaus and periods of weight gain. While we do not have enough goal (i.e., net calories). For example, eating a diet snack an average of 4 times per month while reducing fat intake or system will be low, since it only covers the part of the item space running on the treadmill 12 times a month while losing that the user has already explored. weight. Any item in the log more than once is used, but the Some of these problems may be solved with a hybrid solution. top N=20 items are selected according to their contribution Ideal profiles of foods and exercise could be stored for various toward the goal (i.e., highest net calories.) weight ranges and weight loss targets and used in lieu of historical 3. Find Potential Changes in Stable Patterns: Compute 3 data. In this case, retrospective recommendation could categories of changes across the “current period” and prior recommend the foods and exercises from the ideal profile needed periods: to establish new patterns, transition from old patterns to new a. Cessation – Patterns that existed but no longer patterns, or .maintain existing patterns. appear in the current period (e.g., no more toast and skim milk at breakfast). For this we use the 9. CONCLUSION Lifestyle change is a challenging domain for recommender immediately prior period. systems. People are often purposefully distinctive in their b. Formation – Patterns that emerged in the current lifestyles. The complexity of the real-world activities of eating period (e.g., chocolate croissant and bacon start and exercising makes it difficult to find similarities between users. appearing at breakfast) We introduced Intrapersonal Retrospective Recommendation as c. Substitution – Patterns that existed but have been an alternative recommendation method that uses an individual’s modified (e.g., toast has butter, a cup of skim milk own history of goal achievement to identify behavior patterns to instead of 4 oz) re-establish, transition, or sustain. Next, we hope to evaluate this approach on a larger data set and integrate the system into an 4. Determine If Potential Changes in Stable Patterns Will overall personal health solution. Contribute or Detract from Goals: Label stable patterns of change as contributing to or detracting from the user’s goals. For example, toast and skim milk at breakfast add 180 10. ACKNOWLEDGMENTS calories but with only 5mg of fat. On average, other breakfast Thanks to Jim Christensen for his early involvement in the choices of 180 calories had 10 mg of fat. Therefore, toast and medical adherence project and Sen Hirano for his feedback on this skim milk would be labeled as contributing to the goal. For paper. formation, the proportion of contribution to goals is used (e.g., chocolate croissant had a disproportionate contribution 11. REFERENCES to both fat and calories.) For substitution, the difference can [1] Cao, H., Chen, E., Yang,J., and Xiong., H. 2009. Enhancing be evaluated as a formation (e.g., adding butter had a recommender systems under volatile user interest drifts. In disproportionate contribution to fat). The results are ordered Proceedings of the 18th ACM conference on Information and by net calories added or subtracted. knowledge management (CIKM '09). ACM, New York, NY, 5. Recommend Changes With The Largest Impact: USA, 1257-1266. Recommendations are generated to a) decrease the frequency [2] Hammer, S., Kim, J. and André, E. 2010. MED-StyleR: and/or portion size of high calorie foods; b) increase the METABO diabetes-lifestyle recommender. In Proceedings of frequency of low calorie foods; or c) increase the frequency, the fourth ACM conference on Recommender systems intensity, or time for exercise. These changes are based on (RecSys '10). ACM, New York, NY, USA, 285-288. the net calories saved. For example, if the pattern of eating [3] Herlocker, J.L., Konstan, J., Terveen, L.G. and Riedl, J.T. hamburgers started in period 4 and it contributes significantly 2004. Evaluating collaborative filtering recommender to the net calories then recommend reducing the number systems. ACM Trans. Inf. Syst. 22, 1 (January 2004), 5-53. and/or size of hamburgers and offer. [4] Koren, Y. and Bell, R.M. 2011. Advances in collaborative 6. Tie Changes to Particular Times and Places filtering. Recommender Systems Handbook, pages 145-186. Recommendations can be associated with the appropriate time of day (i.e., meal.) or place (i.e., a restaurant). [5] Luo, G., Tang, C. and Thomas, S.B. 2010. Intelligent Recommendations can be offered repeatedly to establish new personal health record: experience and open issues. In habits. Proceedings of the 1st ACM International Health Informatics Symposium (IHI '10), Tiffany Veinot (Ed.). ACM, New 8. DISCUSSION York, NY, USA, 326-335. IRR appears to generate useful recommendations in some cases, [6] Resnick, P., Iacovou,, N., Suchak, M., Bergstrom, P., and but we have not done a formal evaluation. There are significant Riedl., J. 1994. GroupLens: an open architecture for challenges with using IRR, many of which were identified in collaborative filtering of netnews. In Proceedings of the 1994 (Herlocker, 2004). First, it can suffer from the cold start problem. ACM conference on Computer supported cooperative work There may be no tracking data at the start, particularly since users (CSCW '94). ACM, New York, NY, USA, 175-186. may start using the service to make a lifestyle change without having already performed tracking. The system can wait for [7] Ricci, F.. Rokach, F. and Shapira, B. Introduction to enough tracking data to be available before making a Recommender Systems Handbook, Recommender Systems recommendation, but presumably users may simply be tracking an Handbook, Springer, 2011, pp. 1-35 initial state that is far from their goals. Second, in its most basic [8] Sami, A., Nagatomi, R., Terabe, M., and Hashimoto, K. 2008 form IRR guarantees that items will never be novel, since they are Design of A Physical Activity Recommendation System. In selected from the user’s own history. Users may be dissatisfied Proceedings of the IADIS European Conference on Data with a system that repeatedly recommends known items. Third, Mining (Amsterdam, The Netherlands 22-27 July 2008). DM the coverage of an intrapersonal retrospective recommender ’08. [9] Tanaja, K., Hori, K., and Yamamoto, M. 2010. Development Informatics Symposium (IHI '10), Tiffany Veinot (Ed.). of a Recommender System based on Extending Contexts of ACM, New York, NY, USA, 410-414. Content and Personal History. Journal of Emerging [11] van Pinxteren, Y., Geleijnse, G. and Kamsteeg, P. 2011. Technologies in Web Intelligence, Vol 2, No 3 (2010), 197- Deriving a recipe similarity measure for recommending 203, Aug 2010. healthful meals. In Proceedings of the 16th international [10] Wiesner, M. and Pfeifer, D. 2010. Adapting recommender conference on Intelligent user interfaces (IUI '11). ACM, systems to the requirements of personal health record New York, NY, USA, 105-11 systems. In Proceedings of the 1st ACM International Health