17.5 Graphical Shopping Interface
The idea of map based visualization can be combined with recommender systems
[39]. The most direct way for doing this is by just representing recommendations
of the system in a map instead of a list. This approach is taken in the graphical recommender
system introduced in [23]. Another system introduced in that paper, the
graphical shopping interface (GSI), implements a more interactive recommendation
process for users who do not have a clear idea about what they are looking for and
need to shape their preferences. The GSI implements an approach which in recommender
system literature is called recommendation by proposing [45] or inspiration
seeking [40]. The idea is to let the user navigate through the complete product catalog
in steps, where at each step a set of products is represented in a map. In this
map, the user can select a product and then a new set of products, generally more
similar to the selected product, is produced and visualized by MDS in a similar way
as is done by the MDS based product catalog map introduced in Section 17.3.1.
In [23], three different types of the GSI were proposed, the random system, the
clustering system, and the hierarchical system. Since the random system performed
best in a simulation study in that paper, we only consider the random system here.
In this system, each time a small set of products is randomly selected to be shown
to the user out of a larger set. This larger set contains products that are similar to a
product selected by the user. First, the GSI needs to be initialized. We refer to this
initialization as iteration t = 0. In this initialization, the larger set of products Dt is
set to be the complete product catalog, that is, D0 = D. Out of set D0, p products
are selected at random (without replacement). These products form together the
smaller set D∗0. Then, dissimilarity matrix Δ∗0 is computed using the adapted Gower
coefficient introduced in Section 17.3.1 and D∗0 . Finally, a map Z0 in which these
randomly selected products are mapped is created using MDS and shown to the user.
The iterative process starts when the user selects one of the shown products we
denote by x∗t . Then, the dissimilarities between x∗t and all other products in D are computed. To create Dt , we select the max(p−1,αt I −1) products that are most
similar to x∗t . The parameter α, where 0 < α ≤ 1, determines how much the size
of Dt is decreased each iteration. Thereafter, the process is almost identical to the
steps taken in the initialization. We create a small set D∗t consisting of x∗t and p−1
products randomly selected from Dt and compute a dissimilarity matrix Δ ∗t based
on these products. This matrix is the input for the MDS algorithm returning the new
map Zt .
The parameter α determines how large the influence of the selections of the user
are on the complete process. When α = 1, this influence is very small, since each
time a completely random selection is shown to the user except for the product
selected by the user in the last iteration. Whenα is lower, this influence is higher, but
the variance in Dt also decreases more quickly. The random system is summarized
in Figure 17.5.
Results (
Thai) 1:
[Copy]Copied!
17.5 Graphical Shopping InterfaceThe idea of map based visualization can be combined with recommender systems[39]. The most direct way for doing this is by just representing recommendationsof the system in a map instead of a list. This approach is taken in the graphical recommendersystem introduced in [23]. Another system introduced in that paper, thegraphical shopping interface (GSI), implements a more interactive recommendationprocess for users who do not have a clear idea about what they are looking for andneed to shape their preferences. The GSI implements an approach which in recommendersystem literature is called recommendation by proposing [45] or inspirationseeking [40]. The idea is to let the user navigate through the complete product catalogin steps, where at each step a set of products is represented in a map. In thismap, the user can select a product and then a new set of products, generally moresimilar to the selected product, is produced and visualized by MDS in a similar wayas is done by the MDS based product catalog map introduced in Section 17.3.1.In [23], three different types of the GSI were proposed, the random system, theclustering system, and the hierarchical system. Since the random system performedbest in a simulation study in that paper, we only consider the random system here.In this system, each time a small set of products is randomly selected to be shownto the user out of a larger set. This larger set contains products that are similar to aproduct selected by the user. First, the GSI needs to be initialized. We refer to thisinitialization as iteration t = 0. In this initialization, the larger set of products Dt isset to be the complete product catalog, that is, D0 = D. Out of set D0, p productsare selected at random (without replacement). These products form together thesmaller set D∗0. Then, dissimilarity matrix Δ∗0 is computed using the adapted Gowercoefficient introduced in Section 17.3.1 and D∗0 . Finally, a map Z0 in which theserandomly selected products are mapped is created using MDS and shown to the user.The iterative process starts when the user selects one of the shown products wedenote by x∗t . Then, the dissimilarities between x∗t and all other products in D are computed. To create Dt , we select the max(p−1,αt I −1) products that are mostsimilar to x∗t . The parameter α, where 0 < α ≤ 1, determines how much the sizeof Dt is decreased each iteration. Thereafter, the process is almost identical to thesteps taken in the initialization. We create a small set D∗t consisting of x∗t and p−1products randomly selected from Dt and compute a dissimilarity matrix Δ ∗t basedon these products. This matrix is the input for the MDS algorithm returning the newmap Zt .The parameter α determines how large the influence of the selections of the userอยู่กระบวนการเสร็จสมบูรณ์ เมื่อα = 1 อิทธิพลนี้เป็นขนาดเล็กมาก เนื่องจากแต่ละเวลาที่แสดงขึ้นให้ผู้ใช้ยกเว้นผลิตภัณฑ์เลือกสุ่มอย่างสมบูรณ์เลือกผู้ใช้ในการเกิดซ้ำล่าสุด Whenα ล่าง อิทธิพลนี้จะสูงขึ้น แต่ผลต่างใน Dt ยังลดลงได้เร็วขึ้น สรุประบบสุ่มในรูป 17.5
Being translated, please wait..
