• No results found

The CTLM Recommender Algorithm

4.4 RECOMMENDATION GENERATION

4.4.2 The CTLM Recommendation Approach

4.4.2.3 The CTLM Recommender Algorithm

This section discusses the novel item recommendation algorithm proposed above. The proposed CTLM recommendation approach is designed based on the relationships be- tween users and items according to a set of taxonomic categories (or concepts) that users interact with and that the items are associated with. We also describe the relevance of an item to a user as a kind of LM to understand how users, items and concepts are integrated to estimate the probability that a user is interested in an item. The proposed CTLM generates item recommendations by incorporating three essential components: item popularity npop(bk), the similarities between users and items based on concept

hierarchies cs(ua, bk) and the probability of the concepts used in concept hierarchies

P (ci).

To recommend a set of N items to a target user ua, we first form the item neighbour-

hood for the active user ua. Based on the proposed concept hierarchy model, both the user

and item are converted into the vector profiles for concepts relevant to both the active user −

ua and the item

− →

bk, which are then utilised to find a group of similar items with similar

concept hierarchy to the active user ua. The benefit of the proposed concept hierarchy

model is that it provides an effective method of neighbourhood formation when there are insufficient ratings data in the system or limited information about new users.

Additionally, the cosine similarity function is again utilised to calculate the degree to which an active user ua’s and item bk’s content match based on their profile vectors;

the top k-Nearest Neighbour items bkfor each active user uaare identified by computing

weight sim(−→ua,

− →

bk). Finally, concept similarity between user ua and item bk is denoted

by cs(ua, bk):

cs(ua, bk) = sim(−→ua,

− →

The k most similar items or the neighbourhood of the active user ua is denoted by

N eighbour(ua) = {bk|bk ∈ topK{cs(ua, bk)}}. Then, the user-item concept hierarchy

similarities are integrated with item popularity npop(bk) and the concept probability P (ci)

under the adaptation LM to predict how much interest active user ua will have in item

bk. Item popularity is the item commonly preferred by N eighbour(ua). The concept

probability P (ci) is the probability of concept ciin all relevant item hierarchy Hbk, which

is relevant to active user ua’s interests.

The recommendation of an item to the active user ua will be determined based on

the item recommendation function in Equation 4.17. Top-N item recommendation list is generated by ranking all items in descending order of the prediction score p score(ua, bk).

Formally, the prediction score for each active user is denoted by p score(ua, bk), as

calculated below:

p score(ua, bk) = α × npop(bk) + (1 − α)[β × cs(ua, bk)

+(1 − β) X

c∈Hua∩Hbk

P (ci)], (4.17)

where parameter α is an experimental coefficient that corresponds to item popularity npop(bk), and β is an experimental coefficient that corresponds to cs(ua, bk) and P (ci), 0

≥ α ≤ 1; these allow Mean Average Precision (MAP) measurement to perform at its best.

This CTLM algorithm is described based on the above definitions in subsection 4.4.2.1, it is designed according to a concept hierarchy that incorporates LM principles. In step 3, 0.5 is added to both the numerator and denominator to ensure that the denominator does not equal 0. At least in step 10, the prediction score p score is calculated for each item bkto make item recommendations. The CLTM recommender algorithm is shown in

Table 4.2: The CTLM Recommender Algorithm Algorithm 3: The CTLM recommender algorithm.

Inputs : a set of existing users ui ∈ U , a set of items bk ∈ B and each item concept

hierarchies Hbk, a new user or an active user uawith concept hierarchies Hua,

the explicit ratings Rik given by the existing users ui ∈ U to items bk, and

parameters α and β.

Output: a list of N items recommended for ua 1 for each item bk ∈ B do

//Calculate item popularity

pop(bk) =

X

ui∈U

Rik

end

2 let minpop= min{pop(bk)|bk ∈ B}, maxpop = max{pop(bk)|bk∈ B} 3 for each item bk ∈ B do

//Normalise item popularity

npop(bk) =

pop(bk) − minpop+ 0.5

maxpop− minpop+ 0.5

end

4 let−→uabe the concept profile vectors of ua 5 let−→bk be the concept profile vectors of bk 6 for each item bk ∈ B do

// Calculate the user-item concept hierarchy similarities to get the K nearest neighbours of the active user ua

cs(ua, bk) = − → ua· − → bk k−→uak × k − → bkk end

// Get top k nearest neighbour items of each active uawith the highest similarity score. 7 N eighbour(ua) ← {bk|bk∈ topK{(cs(ua, bk)}}

8 let C = {ci|ci ∈ Hbk, bk ∈ B}

9 for each concept ci ∈ C do

// Calculate the probability of concept P (ci) is for item bk, which is relevant to active

user ua P (ci) = |{bk ∈ B|c ∈ Hbk}| P bk∈B|{ci|ci ∈ Hbk}| end

10 for each item bk ∈ N eighbour(ua) do

//recommendation generation

// Generating a list of N items and order the items by the prediction score. p score(ua, bk) = α × npop(bk) + (1 − α)[β × cs(ua, bk)

+ (1 − β) × X

c∈Hua∩Hbk

P (ci)]

end

4.5

CHAPTER SUMMARY

This chapter discussed how to apply a user profile and an item representation based on a concept hierarchies to devise several approaches to making item recommendations. Specifically, it proposed two approaches: (1) the PopCs approach, which was developed using a linear combination of item popularity and user-item concept hierarchy similarities; and (2) the CTLM approach, which was designed as a combination of item popularity, user-item concept hierarchy similarities and concept probability under an adapted LM approach. The LM was applied to the second approach to facilitate an understanding of how likely item recommendations meet user needs. The efficiency of the proposed CTLM recommender algorithm based on the Big-O concept is presented in Chapter 5. This chapter also discussed how to form the neighbourhood of similar users and similar items for an active user and utilise them to make item recommendations. The experimental results and evaluations of the two recommenders’ performances are discussed in Chapter 5.

EXPERIMENTS AND RESULTS

The previous chapter presented the proposed concept hierarchies and recommendation approaches. It claimed that each approach will improve the performance of recommenda- tion accuracy and will also improve on some of the other techniques as well. This chapter will examine the evaluation of the proposed recommendation approaches, the research hypotheses, the datasets, the evaluation metrics, the baseline approaches involved in the comparative experiments, and discuss the application of the experimental results to the recommendation problems.

5.1

RESEARCH HYPOTHESES

The objective of the experiment was to show how the proposed user profiling and recom- mendation approaches can effectively improve the performance of recommender systems. To verify the effectiveness of the proposed approaches, the experiments were conducted and investigated in order to prove the following hypotheses:

• Hypothesis 1: The proposed concept hierarchy model based on taxonomy can effectively improve recommendation accuracy and solve the cold-start problem. • Hypothesis 2: The Concept Taxonomy with Language Model (CTLM) recommen-

dation approach can effectively improve recommendation accuracy. 106

• Hypothesis 3: Use of the concept taxonomy can effectively alleviate the cold-start problem.