6.5 Temporal Decay User-Profile Evaluation
6.5.3 Decay Method versus Time-Slices evaluation
This subsection will compare the proposed decay method to the time-slices method. The first evaluation metric utilised to compare the methods is the MAP, which is the mean of precision values for all top-N recommendations. The MAP was previously
Figure 6.15: Evaluation of the Decay Method-F1-Measure
Figure 6.16: Evaluation Semantic Profiles-Precision
introduced in chapter 4 in Equation 7.4. A typical way to evaluate a prediction is to compute the deviation of the prediction from the true value. This is the basis for the Mean Average Error (MAE) described in Equation 6.3.
1 |K| X i, j∈K |ri, j− rˆi, j|= FP+ T N FP+ FN + T P + T N (6.3)
Figure 6.17: Evaluation of Semantic Profiles-Recall
Figure 6.18: Evaluation of the Semantic Profiles-F1-Measure
where K is the set of all user-item pairings (i,j) for which we have a predicted rating rˆ
i, j and a known rating ri, jwhich was not used to learn the recommendation model.
Figure 6.19 shows the comparison of different enrichment values using the two pro- posed method, timeslices and decay. As the figure shows, the decay method outper-
Figure 6.19: Comparing Time-Slice and Decay Methods-MAP
forms the timeslice method in all profiles with best one for the combined enriched spatial profile. The reason for better MAP is, increasing the length of the time-slice makes the data denser, leading to better precision. We also compared the two methods using the MAE metric, but using only the enriched combined profile in Figure 6.20. As figure shows, the decay method error are smaller than the time-slices, which means that rating prediction is better for the decay method. It is observed in the experiments that when increasing the length of time slots, more ground truth places is brought for each user at each time slice. With the number of recommendations (i.e., N) unchanged, poorer recall values are observed with increasing the length of time slot. Thus, the decay method shows poorer recall values because it adds up the profile along the time.
The two proposed methods, time-slice and decay, are used for top-N recommendation as illustrated previously. Each method has it characteristics. In time slice method, the recommended place are predicted based on the short term profile that appears only within the time-slice, while in the decay method, the recommended places are pre- dicted based on the whole historic profile of the user. Figure 6.21 illustrates an ex- ample of recommendation when using both methods. The figure shows the profile of a user during three months, January February and March. When we use both meth-
Figure 6.20: Comparing Time-Slice and Decay Methods-MAE Restaurant Coffee Shop Stadium Super Market Hotel Train Station Theatre Bar January February March Top-1 recommendation using time-slice Method
Top-1 recommendation using decay method
Figure 6.21: A Top-1 Recommendation Example using Time-Slice and Decay Methods.
ods for recommending places to this specific user, the time-slice method recommends the “theatre” because the recommender looks only at the user profile created during March. On the other hand, the decay method recommends the “café”’ because the recommender look at the user profile during the three months together. Thus,the time- slice method is considered as short-term recommendation, while the decay method is
considered as a long-term recommendation method. The long-term place recommend- ation is more realistic than the short term recommendation method because it captures routine activities and interests of the user. The short term recommendation method can be some-how misleading because it does not reflect the actual activities and interests of the user towards the places he visits.
6.6
Summary
This chapter began by describing the temporal geo-folksnomy model and arguing that it is important to study the temporal influence in user modelling. It went on to sug- gest two methods for representing the time influence. The first method extracts user preferences during different time slots. Then, it calculates user and place similarities in each time interval based on co-occurrences. In the second method, an exponential decay function is used to measure interest drifts. Then, the user’s location preferences are predicted using both methods at each time interval. The results of the location re- commendation for both methods were discussed. The experimental results show that the proposed decay methods beat all baselines, and improve the accuracy of location recommendation over the time-slice method.The next chapter describes anther method for evaluating the user profiles using user similarity evaluation, it also empathises the conclusions retrieved in this chapter.
Chapter 7
Evaluation of User Similarity
7.1
Introduction
With the growth of location-based social networks, there is a need to calculate the similarities between their users. User similarity has a substantial impact on users, communities, and service providers. In LSBNs, different factors affect users similarity including co-existence in the same place, common interests between users, and tem- poral dynamics, which monitors the users’ behaviour change over time. This Chapter studies and evaluate user similarity based on the users’ interests and common POIs and aims to answer Research Question 5: How can different user profiles be evaluated us- ing user similarity measures to assess their quality?. To answer this question, different user similarity measures are proposed and evaluated using a search process.
The rest of the chapter is organised as follows. Section 7.2 explains the different views of user similarity based on different criteria. In Section 7.3 introduces the evaluation method used to evaluate different user similarities. In Section 7.4, the experiment used to evaluate the approach is described and its results presented and discussed. The chapter concludes in Section 7.5.