• No results found

Efficiency of the CTLM Recommender Algorithm

5.3 EXPERIMENTAL RESULTS AND DISCUSSION

5.3.2 Results of the CTLM Recommendation Approach

5.3.2.8 Efficiency of the CTLM Recommender Algorithm

The efficiency of an algorithm is measured based on the time it takes for the algorithm to run as a function of the input size by using the Big-O notation concept. Big-O notation is the language used to articulate how long an algorithm takes to run. There are many factors affecting this time, such as the speed of the computer’s processor, memory, and so on. Consequently, it is difficult to estimate the exact runtime of an algorithm. Therefore, this thesis use Big-O notation to express how quickly the algorithm’s runtime grows. The input of this algorithm is m, which is denoted as the number of active users |U | in the test set and n is the number of the items |B| that is used in the training set. Based on the proposed CTLM recommender algorithm in subsection 4.4.2.3, the efficiency of the algorithm can be described as follows:

• step 1: O(n ∗ m) • step 2: O(n)

• step 3: O(n) • step 4: O(1) • step 5: O(1)

• step 6: using Algorithm 1 as shown in Chapter 3, section 3.5.3. For each item profile vector−→bk, there are three in average descriptors D to get the concept hierarchy H,

where p is the size of a descriptor. The total time complexity is O(p) + O(p2) + O(p) + O(p2) = O(p2). Therefore, the time complexity of Algorithm 1 is O(3 ∗ n ∗

p2) = O(np2)

• step 7: O(n log n) • step 8: O(np2)

• step 9: O(m ∗ n) • step 10: O(n) • step 11: O(n log n)

The total time complexity is O(n ∗ m) + O(n) + O(1) + O(np2) + O(n log n) = O(n ∗ m) + O(np2) + O(n log n). Therefore, the time complexity of the proposed CTLM recommender algorithm is O(n ∗ (m + p2+ log n)). Based on the experimental

results in subsection 5.3.2.2, we select the best baseline model ITB to compare with the proposed CTLM recommender algorithm. The proposed taxonomy-based ITB approach developed by Weng et al. and based on Ziegler’s topic weight method was adapted in order to generate the baseline model ITB. We generated the ITB’s algorithm according to the formulae and details provided in subsection 5.2.2. For more information on Ziegler’s taxonomy vector construction algorithm, please refer to [111, 121]. The time complexity of the baseline model ITB recommender algorithm is O(mp2) + O(n ∗ (n + p2+ log n)).

- The Runtime Comparison

Figure 5.7: The runtime comparison of CTLM and ITB

In addition, we conduct the efficiency comparison experiments. The proposed CTLM approach and the baseline model ITB were compared by their runtime on Dataset D1. They were run on a laptop with a 4 GB memory, Core i7 CPU. The runtime comparison of the proposed CTLM approach and the baseline model ITB are shown in Figure 5.7 according to the average seconds spent per recommendation. The results suggest that the efficiency of the baseline model ITB was slightly worse than the proposed CTLM. While the proposed CTLM approach took around 20.19 seconds, the ITB took around 26.72 seconds to process the same data. This is because the ITB is computation expensive; it needs to transform all the users and items into high dimension taxonomy vectors in order to compute the similarities between a user and user and between an item and user. ITB makes the prediction or recommendation by relying on these two similarities, therefore its efficiency might be affected by this circumstance. The results suggest that the proposed CTLM approach can be used for large-scale recommender systems.

5.4

CHAPTER SUMMARY

This chapter evaluated the effectiveness of using concept taxonomy hierarchies as a part of the recommendation approach. The proposed approaches were evaluated by using two real-world datasets on books and music that were collected from the Book-Crossing com- munity and Amazon’s website. This thesis proposed three recommendation approaches to improve the performance of recommendation making and alleviate the cold-start prob- lem. The experimental results were compared to state-of-the-art baseline approaches and other approaches involved in the experiments. The experimental results showed that the proposed PopCs and CTLM approach-based taxonomies can significantly improve the accuracy of the item recommendations. The proposed method for using the probability of concept and the user-item concept hierarchies’ similarities, along with the use of concept taxonomy, has yielded a significant improvement in the recommendation making in both normal and cold-start environments.

CONCLUSION AND FUTURE WORK

6.1

SUMMARY OF THE RESEARCH

The main objective of this thesis is to enhance approaches for making personalised recom- mendations. This thesis also provides an alternative way to alleviate the recommendation problems with cold-starts and the issue of users having uncertain information needs. This research explores and exploits the relationship between users and items, according to a set of taxonomic categories or concepts utilised by users and relating to items. The main idea is to understand how items are likely to generate what users want, according to their relationship in concept taxonomy. This thesis proposes the user profiling and item representation approaches based on the concept hierarchy model.

This thesis has proposed a new structure to use taxonomy information to describe uncertain knowledge regarding user information needs in a two-dimensional hierarchy. The concept hierarchy model has been proposed in order to acquire new users’ preferences when there is limited information about these users. The purpose was to solve the cold- start problem and users being confronted with uncertainties regarding their information needs. The new concept-weighting method, using rating data instead of considering the hierarchical relationship between each concept taxonomy and the frequency of each taxonomic concept in item taxonomic descriptors, was proposed to measure the weight of taxonomic concepts. Instead of representing user preferences by using rating data, the user preferences and the item descriptions can be represented via a user-concept hierarchy

and an item-concept hierarchy, respectively. This model was utilised to generate the active user profile and item representation based on concept hierarchy.

Furthermore, we have developed effective recommendation-making approaches to improve the quality of item recommendations. The combination of both the user profiles and the item representation, based on the concept hierarchy, were applied to recommendation- making approaches. In this thesis, we use the following factors for making recommenda- tions:

• Item popularity: the users’ opinion about items by means of ratings. The ratings a user assigns to the item correspond to the user’s overall preferences in regard to the items.

• The user-item concept hierarchy similarities: the similarities between user and item based on concept hierarchies.

• The concept probability: the probability of the concept in regard to an item that is relevant to the user’s concept preferences.

The linear item popularity combined with the user-item concept hierarchy similar- ities made up the PopCs recommendation approach. In addition, the adaptation of the language model incorporating the item popularity, the user and item concept hierarchy similarities and the probability of the concept were utilised to generate the CTLM recom- mendation approach.

The experimental results were evaluated on the two real world data sets collected from the BookCrossing community and Amazon website. This thesis proposes two rec- ommended approaches to improve the performance of the recommendation making as well as to reduce the new user problem. The experimental results were compared with the state-of-the-art baseline approaches and other experimental approaches.

effectively improve recommendation accuracy and quality. In addition, the proposed user- item concept hierarchy similarities and the probability of concept approaches based on the use of concept taxonomy outperformed the approach based on item popularity, and they can provide significant improvement in recommendation making in both normal and cold-start environments.