• Tidak ada hasil yang ditemukan

Rating Conversion

Dalam dokumen submitted to the Department of Informatics (Halaman 43-47)

2.4 Multi-Criteria Recommender Systems and Rating Conversion

2.4.3 Rating Conversion

In order to provide an effective rating prediction, the memory-based CF approach rely crucially on the ratings from neighbors. However, exploiting those ratings to make a prediction for the other users directly might lead to a problem. This is because the habits or patterns on giving ratings among users vary due to their personal biases. For example, on the rating range of 1 to 10, Useru1might give rating score from 2 to 5 indicating ‘dislike’

to ‘like’, while Useru2, instead, gives the rating from 5 to 8 with the same intention. This means that ‘like’ for useru1 equals to ‘dislike’ to useru2. Therefore, using ratings from neighbors to predict the rating for an active user directly may not be practical.

In order to deal with the user personal biases in the ratings, many rating conversion techniques have been introduced in single criterion domain [8, 18, 43, 44, 56]. The main idea is to convert the ratings from the neighbors into the same scale as the active user, before utilizing them for a rating prediction. The most simplest approach that can be applied for converting the ratings is a normalization.

The normalization approach converts the user ratings into a specific range. Such range is usually between 0 and 1 where everyone’s ‘most like’ and ‘most dislike’ will be mapped to

score ‘1’ and ‘0’, respectively. Many normalization methods are proposed based on different assumptions, such as linear normalization, the Gaussian normalization, and the decoupling normalization.

Linear Normalization

This method maps ratings based on the maximum and minimum of personal user ratings. By using the linear function, the normalized rating valuernewua for the userua’s specific rating is computed as:

rnewu

a = ruolda −rua,min+1

rua,max−rua,min+1, (2.4.4) whereroldua is an original rating ofua,rua,max andrua,mindenote the maximum and minimum ratings useruahas rated, respectively. This normalization method maps ratings based only on maximum and minimum of the personal user ratings.

Gaussian Normalization

This method considers two factors that affect the variance of ratings among users with similar interests [43]. The first factor is a difference of a rating from the average ratings. This factor relates to the fact that some users are more tolerant and tend to give higher ratings than others.

Another factor is the difference of users rating scales. This comes from the fact that some users tend to assign items to a narrow range of ratings, whereas other users tend to assign items to a wide range. Combining these two factors, the ratings of each user are subtracted with his average and divided by the variance of his ratings, as expressed by:

rnewu

a = roldua −r¯ua

σua , (2.4.5)

where ¯rua andσua are an average and a standard deviation of user ratings, respectively.

Decoupling Normalization

This method converts a user rating on item into a probability for that item to be favored by the user [44]. When the ratingrua is going to be normalized, the probability is determined based on two factors. First, a ratio between two numbers: the number of items which was rated no more than valuerua by the useruaand the number of all items that the useruahas

2.4 Multi-Criteria Recommender Systems and Rating Conversion 29 rated. The high ratio means the ratingrua are likely to be favored by the user. The second factor is a ratio between the other two numbers: the number of items which was rated value rua by the useruaand the double number of all items that the user has rated. The low ratio means the ratingrua are likely to be favored by the user. Based on these two factors, a special formula; called halfway accumulative distribution was proposed as:

runewa = |{vj∈Iua|ra,j≤ruold

a }|

Iua −|{vj∈Iua|ra,j=ruold

a }|

2|Iua| , (2.4.6)

whereIua denotes the set of items to which useruahas rated.

Although the normalization techniques are able to convert a user’s ratings into the same range, the conversions are based only on the rating data of the only one user. This might lead to an inaccurate recommendation if there are two active users whose rating patterns are different but having the same neighbors. If the normalized ratings are used for the recommendations to these two active users, the results will be the same. For example, an active useruausually rates ‘0.4’ (normalized ratings) while another active userubusually rates ‘0.7’. If these two users share the same neighborucwhose rated the target item with

‘0.8’, they will receive the same predicted ratings of ‘0.8’. Although ‘0.8’ seems like no effect on user ub, it seems to be high value of rating for user ua since his usual rating is

‘0.4’. Thus, the better solution is to find the relationship between each pair of user ratings:

original user and target user, in order to convert neighbor ratings to individual active user ratings. The examples of such conversion techniques include linear mapping [8], Lathia’s rating conversion [56] and Warat’s rating conversion [18], which are explained further in Chapter 3.3.

Furthermore, the rating conversion techniques have been proposed only in the SC domain.

Such SC rating conversion techniques can be applied to MC ratings by converting ratings from each criterion independently. However, this could cause a scalability problem and consume a lot of resources. Moreover, usually there are implicit relation among the criteria ratings when user makes decision to select an item. For example, a user may choose a room that have high score on both service and location, while ignore its price. If each criterion rating is converted independently, such implicit relation could be lost.

Chapter 3

Related Work

3.1 Review-Based Recommendation Techniques

Dalam dokumen submitted to the Department of Informatics (Halaman 43-47)