Determining the Weights of the Hybrid Recommendation Scheme

The hybrid recommendation scheme is based on the hybrid weighting ratios wM

and wP (wP=1-wM) of the mobile phone and product preference clusters. Hybrid recommendation becomes pure preference-based recommendation when wM equals zero, and pure MPF-based recommendation when wM equals one.

The top-N recommendations are divided into two segments. One segment is from the top-1 to the top-10 recommendations and the other is from the top-11 to the top-20 recommendations. We choose the top-5 and top-15 recommendations to represent the first and second segments respectively. The quality of the top-5 and top-15 hybrid recommendations with different MPF weights (wM) is shown in Figure 6. The best recommendation quality for the top-5 and top-15 occurs when wM =0.9 and wM =0.6 respectively. We use these weights as the hybrid weighting ratios of the hybrid recommendation scheme in Section 3.3.5.

Hybrid recommendations

0.1 0.12 0.14 0.16 0.18 0.2 0.22

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 wM

F1-metric

top-5 top-15

Figure 6 The weighting ratio wM of the hybrid recommendations 3.3.5 Evaluation of MPF-Preference Hybrid Recommendation Methods

We compare two proposed methods, namely, MPF-based and Hybrid MPF-Preference methods, with the other two methods, preference-based and CF methods. MPF-based method cluster users into groups based on users’ mobile phone features and recommend products according to the association rules and most frequent items extracted from user groups. Preference-based method makes recommendations based on user groups that are clustered according to the users’

product preferences. Hybrid MPF-Preference recommendations are generated by a combination of the MPF-based and preference-based recommendation schemes with the hybrid weighting ratio described in Section 3.1.3. The hybrid weighting ratio described in Section 3.3.4 is set at wM =0.9 for the first top-N segment (top1-10) and wM =0.6 for the second top-N segment (top11-20). The CF method is a typical k-NN CF method that recommends the top-N most frequently occurring products of the k-nearest neighbors (similar users). Because the average number of users in the product clusters is 232.5(=930/4), we choose k=200 as the number of nearest neighbors. Table 7 presents the precision, recall and F1-metric evaluation of k-NN CF, Preference-based, MPF-based and Hybrid MPF-Preference methods.

The F1 values of all methods are low, since the user-item matrix of our experiment data is very sparse. Although the F1 values of our proposed methods are still low, our methods can achieve better improvement over conventional methods. For example,

as listed in Table 7, the average F1-metric of the MPF-based method is 11% better than the preference-based method. Furthermore, the average F1-metric of the hybrid MPF-Preference method, which combined MPF-based and preference-based methods, is 33% better than the preference-based method. The F1-mertic of the hybrid MPF-Preference, MPF-based, Preference-based and k-NN CF methods are shown in Figure 7.

Table 7 Evaluation of k-NN CF, Preference-based, MPF-based and hybrid methods

k-NN CF Preference-based MPF-based Hybrid MPF-Preference TopN

Precision Recall F1 Precision Recall F1 Precision Recall F1 Precision Recall F1 2 0.015 0.004 0.006 0.153 0.085 0.092 0.161 0.088 0.099 0.176 0.100 0.110 4 0.026 0.017 0.017 0.104 0.113 0.089 0.122 0.128 0.106 0.140 0.157 0.125 6 0.036 0.055 0.035 0.080 0.124 0.080 0.096 0.156 0.100 0.113 0.186 0.118 8 0.039 0.098 0.045 0.072 0.146 0.081 0.079 0.165 0.091 0.092 0.195 0.106 10 0.035 0.107 0.044 0.063 0.156 0.076 0.067 0.172 0.082 0.081 0.212 0.100 12 0.030 0.109 0.040 0.057 0.165 0.072 0.058 0.178 0.075 0.072 0.221 0.094 14 0.027 0.111 0.036 0.051 0.171 0.066 0.051 0.180 0.069 0.064 0.227 0.087 16 0.023 0.112 0.033 0.046 0.174 0.062 0.045 0.181 0.063 0.058 0.236 0.081 18 0.021 0.112 0.030 0.042 0.179 0.059 0.043 0.197 0.062 0.053 0.244 0.077 20 0.019 0.112 0.028 0.040 0.187 0.057 0.041 0.210 0.061 0.049 0.248 0.073 Avg. 0.027 0.084 0.032 0.071 0.150 0.073 0.076 0.165 0.081 0.090 0.203 0.097

As shown in Fig 7, the recommendation quality of all the methods declines after the top-4 recommendations, as the number of recommended products increases.

Recall that association rule-based recommendations are based on the items users browsed previously. There are only a few recommended products because the average number of previously browsed products was 3.87. Therefore, the most frequent item recommendations are used to support the association rule recommendations if the number of recommended products is not sufficient. However, most frequent item-based recommendations are not better than association rule-based recommendations, so the recommendation quality deteriorates after the top-4 recommendations.

0 0.02 0.04 0.06 0.08 0.1 0.12 0.14

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 Top-N

F1-metric

Hybrid MPF-Preference MPF-based

Preference-based k-NN CF

Figure 7 Evaluation of the hybrid, MPF-based, preference-based and k-NN CF methods

3.4 Discussions

Figure 8 shows that the mobile phone cluster 0 (camera phones with java, video, Bluetooth, card slot and flash light functions) achieves the best recommendation quality, followed by cluster 1 (simple phones with java and video functions), and cluster 2 (feature phones with java, video, Bluetooth and card slot functions). Among all the phone types, camera phones with the flash light feature yield the best recommendation quality. The owners of camera phones like to browse for digital cameras and travel products because they like to travel and take photographs. We also evaluate the effect of the hybrid method on the recommendation quality of MPF-based clusters. Figure 8 shows that the recommendation quality of the hybrid method (hybrid0 – 2) is better than that of the MPF-based method (mphone0 – 2) for each MPF-based cluster. In other words, the effect of combining MPF-based recommendations with preference-based recommendations is positive.

0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 Top-N

F1-metric

hybrid0 mphone0 hybrid1 mphone1 hybrid2 mphone2

Figure 8 Effect of the hybrid method on the recommendation quality for mobile phone clusters

From Figure 9, we observe that product cluster 0 (lingerie, pants and skincare products) achieves the best recommendation quality in terms of product preferences, followed by product cluster 2 (hotels, travel coupons, food and domestic travel), product cluster 1 (mobile phones, digital cameras, cordless phones and notebooks), and product cluster 3 (skincare, mp3, cosmetics and consumer products). Users who prefer lingerie and underwear products receive better quality recommendations than users who prefer other products. We also evaluate the effect of the hybrid method on the recommendation quality of preference-based clusters. Figure 9 shows that the recommendation quality of the hybrid method (hybrid0 – 3) is better than that of the preference-based method (product0 – 3) for each preference-based cluster. Hence, combining the preference-based method with the MPF-based method can improve the quality of recommendations.

0 0.05 0.1 0.15 0.2 0.25 0.3 0.35

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 Top-N

F1-metric

hybrid0 product0 hybrid2 product2 hybrid1 product1 hybrid3 product3

Figure 9 Effect of the hybrid method on the recommendation quality for product clusters

Chapter 4. Hybrid Multiple Channels-based (HMC) Method

4.1 Hybrid Multiple Channels-based (HMC) Method

In this section, we describe the proposed recommendation method based on the hybrid multiple channels, which are composed of mobile, television, catalog, and Web channels, as shown in Figure 10. Users of the multiple channels are divided into RFM groups to find heavy users based on their recency (R), frequency (F) and monetary (M) values; then these heavy users are divided into preference groups based on their product category preferences to provide recommendations for the new channel users. First, we use the K-means clustering method to cluster existing channel users into RFM groups based on the Euclidean distance of R, F, and M values and compare the average R, F, and M values of the clusters to the average R, F, and M values of all users. Heavy users who were selected by the clusters of the lower R values but the higher F and M values, provides more transaction instances that could be used to find more similar users for the new channel. Second, we use the K-means clustering method to cluster the heavy users of each channel into preference groups based on users’ similarity which is measured by Pearson’s correlation coefficient of users’ product category preferences. Heavy users in the preference group could find more similar users for the new channel, which could solve the sparsity problem of the new channel and derive more association rules to improve the recommendation quality. For every target mobile channel user, similar users are selected from the clusters of mobile, television, catalog, and Web channels based on product category preferences. The system then finds the association rules of products and product categories as well as the most frequent items of the similar users of each channel. The association rules and most frequent items of the hybrid multiple channels are determined from the rules and items of multiple channels using the weighted sum of the associated confidence scores and frequent counts with different hybrid weights of wM, wT, wC, and wW. The hybrid weights are the relative importance of the multiple channels to the mobile channel, which are determined by the best recommendation quality of the recommendation engine based on the preliminary analytical data, which will be described in Section 4.1.2. Finally, the method recommends products based on the association rules and most frequent items by using the hybrid weights (wM, wT, wC, wW).

Figure 10 An overview of the proposed recommendation scheme

4.1.1 User Selection and Clustering of the Existing Channels

Heavy users are valuable customers who spend a large amount of money to purchase products frequently and recently in a channel. Figure 11 shows the selection of the heavy users with lower recency (R), higher frequency (F) and monetary (M) values. First, we calculate the R, F, and M values of each user in a channel. Second, we cluster users into groups by the K-means clustering method based on the

Euclidean distance of R, F, and M values, and compare the average R, F, and M values of clusters to the average R, F, and M values of all users in a channel. Finally, the clusters of heavy users in a channel are selected by the lower recency (R) values but the higher frequency (F) and monetary (M) values.

Figure 11 Selection of heavy users of a channel

After selected the existing channel heavy users to represent consumption behavior of users of a channel, the mobile, television, catalog, and Web channel users are clustered by the K-means clustering method into groups based on users’ similarity which is measured by Pearson’s correlation coefficient of the user–product category rating matrix as shown in Table 8.

Table 8 User–product category rating matrix

User ID Cosmetics Perfumes Skincare Pants Shoes Toys Shirts Notebooks …

1 1 0 1 1 1 0 0 1 …

2 0 1 1 0 1 1 0 0 …

3 1 0 0 1 0 0 1 1 …

4 0 1 1 1 0 1 1 0 …

… … … … … … … … … …

4.1.2 The Recommendation Engine

The proposed hybrid multiple-channel method derives recommendations based on the association-rule and most-frequent items approaches. For each group of users, two kinds of association rules are extracted, namely, product-level association rules and category-level association rules. The former are extracted from the product transactions; and the latter are extracted from category-level transactions, which are derived by replacing the products in product transactions with their respective categories. The recommendation engine is comprised of three components: the product association rules (X_H^PRi →Y_H^PRi) component, the product category association rules (X_H^CRj →Y_H^CRj)component, and the most frequent items (Y_H^Mf) component, as shown in Fig. 12. In the figure, H represents either M, T, C, or W, which denote the mobile, television, catalog and Web channels respectively.

In the multiple channel approach, let X_H^PRi →Y_H^PRi,H∈{M,T,C,W} be the product-level association rules extracted from the product transactions of a group of channel users, comprised of mobile, television, catalog, and Web channel users; and let their associated confidence scores be cf_M^PRi,cf_T^PRi,cf_C^PRi, and cf_W^PRi respectively.

In addition, let X_u represent the previous set of products that the target user u browsed in the mobile channel; and let Y_u^AR be the set of candidate products generated from the union of Y_H^PRi −X_u according to all the association rules

i catalog, and Web channels respectively.

Let Y_H^Mf,H∈{ , , , }M T C W denote the set of most frequent items derived from the user groups of target user u in multiple channels. The frequency count of an item v for a user group Ug is equal to the number of users in Ug that had browsed/purchased item v. Let f_{v M}^Mf_, ,f_{v T}^Mf_, ,f_{v C}^Mf_, , and f_{v W}^Mf_, represent the frequency counts of an item v in

Mf

YH , respectively. Let Y_u^Mf be the set of candidate products generated from the union of Y_H^Mf −X_u. The products in Y_u^Mf are ranked according to the weighted sum of their frequency counts calculated as Eq. (11).

Mf

Let }X_H^CRj →Y_H^CRj,H∈{M,T,C,W be the category-level association rules extracted from the category-level transactions of a group of channel users, comprised of mobile, television, catalog, and Web channels; and let their associated confidence scores be cf_M^CRj,cf_T^CRj,cf_C^CRj, and cf_W^CRj respectively. In addition, let X_u^C represent the set of product categories that the target user u browsed previously from the mobile channel; and let Y_u^C be the set of candidate product categories generated from the union of Y_H^CRj according to all the category-level association rules

j the weighted sum of their confidence scores (Eq. 12).

j Let Y_u^CMf denote the set of most frequent candidate items derived from the candidate product categories Y_u^C and most frequent candidate items Y_u^Mf . We note that Y_u^Mf is derived from the user groups of target user u in multiple channels.

CMf

Yu is the set of items in Y_u^Mf that also belong to the candidate categories in Y_u^C. Each item v in Y_u^CMf is associated with a pair of (cf^Ck ,f_v^Mf ), where cf^Ck is the associated confidence score of v’s category Ck derived using Eq. (12), and f_v^Mf is the frequency count of item v calculated using Eq. (11). The product items in Y_u^CMf are ranked as follows. The items with the highest frequency counts in each category of

C

Yu are selected first and ranked according to their associated confidence scores.

Then, the items with the highest frequency counts among the remaining items in each category are selected and ranked according to their associated confidence scores. The process repeats to select and rank items in Y_u^CMf by recommending most frequent items from diverse candidate categories.

We compare the number of candidate products |Y_u^AR| and the top-N recommendations. Note that Y_u^AR is the set of candidate products generated from the product-level association rules. If the number of candidate products |Y_u^AR| is higher than the number of top-N recommendations (|Y_u^AR|≥N) , the system will recommend the top-N products from Y_u^AR. If the number of candidate products

|

|Y_u^AR is less than the number of top-N recommendations (|Y_u^AR|<N) , but

AR CMf

u u

| Y UY | is larger than the number of top-N recommendations

(| Y_u^ARUY_u^CMf | N≥ ), the system will recommend |Y_u^AR| products from Y_u^AR. The remaining |N−|Y_u^AR products for recommendation are selected from Y_u^CMf . Note that Y_u^CMf is the set of most frequent product items belonging to the associated

product categories in Y_u^C .

If | Y_u^ARUY_u^CMf | is less than the number of top-N recommendations

(| Y_u^ARUY_u^CMf | N< ), the remaining N | Y− _u^ARUY_u^CMf | products for recommendation are selected from Y_u^Mf - (Y_u^ARUY_u^CMf), which is the set of most frequent items that the target user u has not browsed in the mobile channel and are not in Y_u^ARUY_u^CMf. The products are ranked according to the weighted sum of the frequency counts of the products.

Figure 12 The recommendation engine

4.2 Experimental Setup and Datasets

The multichannel company is a home shopping company which has owned the television, catalog and Web channels in Taiwan. Because of the rapid development of 3G mobile network, the company would develop the new mobile channel. The television channel is a sale channel of the home shopping company. The products are introduced in television channel and people can purchase products by a toll-free telephone.

The mobile channel is an on-line experimental mobile shopping website which tried to find the consumption behaviors of the new mobile channel users. Users could access the mobile website by their own mobile phones via 2G, 3G, 3.5G and Wi-Fi networks. Data for the mobile channel and the existing channels were collected from the mobile website and CRM system of a retailer from October 2006 to January 2007, which contained information of about 1,692 users who own 184 different models of the mobile phones and offered 1,416 products which are included in 194 product categories. The product categories which are frequently browsed are mobile phones, lingerie, digital cameras, skincare, MP3 players, watches, living products, cosmetics, cordless phones and travel coupons. The products offered by the mobile channel were also provided in the other three channels.

The dataset was divided up as follows: 80% was used for training and 20% for testing. The training set was also used as the dataset in the preliminary analytical experiment. Specifically, 55% of the data set was used to derive recommendation rules and 25% was used as a preliminary analytical dataset to determine the hybrid weights assigned to mobile, television, catalog, and Web channels based on the quality of the recommendations. There were 1,353 users in the training dataset and 339 users in the test dataset.

The consumption behaviors of the applications in e-commerce are different, so the datasets are different. The support and confidence of the association rules are set to retrieve the interesting patterns in datasets. Based on the characteristics of our dataset, the minimum support and confidence of the association rules were set at 0.004 and 0.4 to find the interesting rules, which were both higher than the study by Cooley et al. [12] but lower than the study by Cho et al. [9].

4.3 Experimental Results

4.3.1 Heavy Users’ Selection of the Existing Channels

The groups of heavy users were determined by comparing the group average of RFM values to the total average of RFM values in a channel. The heavy user groups are those groups with group average R smaller than and group average FM larger than the total average RFM in each channel. First, the R, F, and M values of every user of television, catalog, and Web channels were calculated. Users of a channel were clustered into groups. By comparing the group average to the total average in a channel, the group average may be larger () or smaller () than the total average.

Because each R, F, and M value of a group can have two alternative values, larger () or smaller () than the total average, we cluster users based on three R, F, and M values into 8 groups (2 × 2 × 2). Second, the heavy user groups were checked (9) in Table 9 due to their average R were smaller () but F, M were larger () than the total average in each channel. The clustering results are considered significant (p<0.05) based on R, F and M variable differences for television, catalog and Web channels.

For example, the clusters of heavy users in the television channel are clusters 4 and 5 because their average R were smaller () than the total average R, their average F were larger () than the total average F, and their average M were larger () than the total average M in the television channel. Similarly, based on the selection criteria, the clusters of heavy users in the catalog channel are clusters 3 and 6, and the clusters of heavy users in the Web channel are clusters 2 and 7 as checked in Table 9.

Table 9 R, F, and M values of users in each channel by clusters

Channel Television* Catalog* Web*

Cluster ID Users R F M Users R F M Users R F M 0 1,156 80 2 3,932 132 54 2 3,951 26 82 2 2,677

1 4,844 40 4 10,366 187 40 3 5,990 235 40 3 5,174

2 562 93 2 3,059 61 63 2 2,917 9 216 16 14 38,158

3 2,013 69 3 4,969 9 83 19 7 23,019 155 52 3 4,260

4 9 5,694 32 5 16,831 21 68 2 2,566 321 28 4 9,822

5 9 4,534 23 9 40,772 101 60 2 3,331 262 37 3 7,610

6 4,032 47 3 7,556 9178 32 4 9,101 67 61 2 3,069

7 2,765 60 3 6,287 140 47 3 4,577 9 325 22 6 16,012

Total 25,600 44 4 15,431 903 44 3 7,067 1,607 33 5 12,909

* Significant at the 0.05 level

We note that there are exactly two heavy user groups for each channel based on the selection criteria. For other dataset, there may exist more than two heavy user groups. The final selection result is shown in Table 10. In our study, we are interested in the major consumption behaviors which are contributed by heavy users in channels. Users are selected due to their heavy consumption behaviors in channels.

Table 10 Clusters of heavy users selected in each channel

Television Catalog Web Cluster ID Users Cluster ID Users Cluster ID Users

4 5,694 3 83 2 216

5 4,534 6 178 7 325

(4,5) 10,228 (3,6) 261 (2,7) 541

4.3.2 Determining Channel Weights for the Hybrid Recommendation Scheme The hybrid multiple channel recommendation scheme is based on the hybrid weighting ratios of mobile (wM), television (wT), catalog (wC), and Web (wW) channels (wM + wT + wC + wW = 100%). The derivation of these weights is as follows. First, the dataset is divided into 80% training dataset and 20% testing dataset.

The training dataset trains a model to evaluate the testing dataset. In the 80% training dataset, 55% is used to derive the association rules and 25% is used as the preliminary analytical data to derive the weights. Second, these weights are determined by the best recommendation quality of the recommendation engine based on the preliminary analytical data. Because the average number of browsed products is 3.87 in the mobile channel, we choose the top four recommendations to determine the hybrid weights of multiple channels. We systematically adjust the values of channel weights in increments of 1%. The qualities of the top four hybrid recommendations according to different hybrid weight combinations (wM, wT, wC, wW) are shown in Figure 13. The best recommendation quality F1-metric of 0.1573 for the top four recommendations occurs when (wM, wT, wC, wW) = (60%, 1%, 33%, 6%). We use these weights as the hybrid weighting ratios of the hybrid

The training dataset trains a model to evaluate the testing dataset. In the 80% training dataset, 55% is used to derive the association rules and 25% is used as the preliminary analytical data to derive the weights. Second, these weights are determined by the best recommendation quality of the recommendation engine based on the preliminary analytical data. Because the average number of browsed products is 3.87 in the mobile channel, we choose the top four recommendations to determine the hybrid weights of multiple channels. We systematically adjust the values of channel weights in increments of 1%. The qualities of the top four hybrid recommendations according to different hybrid weight combinations (wM, wT, wC, wW) are shown in Figure 13. The best recommendation quality F1-metric of 0.1573 for the top four recommendations occurs when (wM, wT, wC, wW) = (60%, 1%, 33%, 6%). We use these weights as the hybrid weighting ratios of the hybrid