We present a preliminary study on the problem of nding products similar to a product given as input, based solely on their attributes. We assume that we are given a set of products from a same category of a same on-line store, were each product is described in a catalog by a number of attributes (e.g., general characteristics, technical speci cations, etc.). This problem, which at a rst glance may be seen as straightforward or even mundane, is in fact challenging and intriguing. In fact, any automatic solution for it requires techniques for comparing tens of di erent atributes, whose semantics are often very technical and speci c (e.g., the shutter speed of a camera) and also requires dealing with hundreds of products in the category. To be generic, such a solution must also deal with several distinct product categories. In here, we describe and evaluate a similarity function we have proposed for comparing products based on their attributes. This function uses a number of attribute-speci c similarity functions, which are selected according to a class assigned to the attribute. The assignment of classes to attributes is carried out by a simple classi cation strategy, which we also describe and evaluate. Experiments we carried out to evaluate our proposed similarity function using data from real catalogs in ve distinct popular product categories have shown promising results.
Recommendation Systems are used by most e-commerce sites to suggest
products to their users and provide additional information to help customers to decide
which products to acquire [
A simple way of enabling content-based recommendation is, given a product, presenting to the user other products that are similar to it with respect to their attributes. This is useful, for instance, when costumers explicitly want to nd products with certain characteristics or when the seller wants to present to a customer products similar to a product she is interest in, e.g., for the sake of comparison or to provide alternatives to out-of-stock itens.
However, in typical e-commerce sites, looking for similar products may require the user to browse manually through a large number of pages and products. For instance, suppose a user is interested in a speci c camera, say, \Nikon S3500". Currently, if this user wants to nd alternative cameras that are similar to this model (i.e., having similar features), for the sake of comparing their prices, it is likely that she would have to browse over hundreds of other cameras in the catalog to nd them. On the other hand, if this camera is not in stock, it would be interesting to provide the user with similar alternative cameras in stock, without having her to look over the whole catalog.
Another interesting aspect of this kind of recommendation is that it enables suggesting products to the customers without relying on historical data. It means that the system can recommend products and provide buying options even if a costumer is new to the system or if the item is new to the catalog.
To nd whether two products are similar it is necessary to compare them. Products on e-commerce sites are often described by their attributes. It means that, to make a comparison between two products, it is necessary to compare their attributes. This can be unfeasible to be carried out manually by casual users on the Web.
For instance, in a certain e-commerce site, to verify whether the \Nikon S3500" camera is similar to another camera, say the \Sony W830", a user has the option of comparing the 26 atributes provided for the rst camera with the corresponding attributes of the second cameras. The lists of attributes available for each camera in this site are presented in Figure 1. Notice that the second camera has only 18 attributes. Also, notice that many attributes are di cult to be compared, unless the user is an expert in the eld.
In general, the same situation occurs in many other categories, that is, comparing products requires comparing tens of attributes, some of them with very speci c semantics.
In this paper we present a preliminary study on the problem of nding products similar to a given product. We assume that we are given a set of products from a same category of a same on-line store, along with their attributes. For instance, one of the datasets used in our experiments comprises a set of 489 camera models under the Cameras category of a real on-line store.
Speci cally, we describe and evaluate a generic similarity function we have proposed for comparing products based on their attributes. This function uses a number of attribute-speci c similarity functions, which are selected according to a class assigned to the attribute. The assignment of classes to attributes is carried out by a simple but e ective classi cation strategy, which we also describe and evaluate here.
An experimental evaluation we carried out and reported here has shown promising results. Our proposed similarity function showed to be accurate in nding similar products, achieving average F-1 values above 0.75 in 5 representative product categories we have tested. Also, our strategy for attribute classication has correctly classi ed most of the attributes from these categories.
Attribute Nikon S3500
Brand Nikon Type of Camera Compact Monitor/Display 2,7" LCD / TFT 230.000
Resolution 20,1 Internal Memory 25MB
Memory Cards Yes Compatible Memory Cards SD, SDHC and SDXC
Sensor Optical Zoom Digital Zoom
Lenses Shutter Speed Focus range
Opening Flash Modes Flash range Battery Type Video Features Scene modes
File Formats Built-in microphone
Tripod mount Menu Languages
Color Dimensions (HxWxD)
Weight
Sony W830 Although important and challenging, e ective methods for nding similar products are scarce both in industry and in the academy.
Kagie et. al. [
Our approach di ers from this in many aspects. First, in our approach the user does not need to specify an ideal product. In fact, this is avoided, since we consider that casual users in e-commerce sites are not willing to specify desired values for several attributes. Instead, we only require the user to select one product to be used for comparison. Second, besides categorical and numerical attributes, we consider two additional classes of atributes: multi-categorial and dimensional. We adopted these two additional attributes classes because they are very common in e-commerce products. Third, in our case there is no need to ask the user to provide the class of each attribute involved in the comparison. Fourth, while Kagie's work seems to focused on a single category, our approach was conceived to deal with many categories typically found in e-commerce sites. Fifth, we instead of using a 2D map with several products, our approach can produce, as output, a ranking of products in order of similarity. 3
Prior to the application of our similarity function, it is necessary to take each attribute found in the products of a given category we are interested in and assign each one to a single class of a simple attribute taxonomy comprising four classes, namely: Numerical, Categorical, Multicategorical and Dimensional.
This taxonomy was created based on previous work by Kagie et. al. [
Although a number of di erent approaches could have been used for this
task, we opted for using a simple strategy in which the values expected for the
attributes in a given class are described by a regular expression we call domain
descriptors. Domain descriptors are similar to the Data Frames used by Embley
et. al. in several methods (e.g., in [
The classi cation of a attribute is carried out as follows. Let Ai be an attribute that occurs for products p1;: : :;pm in a given category. For instance, attribute Scene Modes occurs in the description of many products in the Compact Cameras category. First, for all products pj (1 j p), we take the value vi;j for Ai occurring in pj .
Next, we perform several cleaning and standardization operations over set of values vi;j of Ai taken from products. These operations include duplicate values removal, white space and case normalization, among others. The result is a set of values a1;: : :;am which we call the occurrences of Ai. Notice that by doing so we assume that all values of Ai have the same semantics in all pj . For instance, we assume that the attribute Scene Modes has the same semantics in the description of all products in the Compact Cameras category.
Finally, we test each occurrence a1;: : :;am against each domain descriptor k (k=1;: : :;4) and associate atribute Ai with the atribute class Ck whose domain descriptor k recognizes the majority of its occurrences.
Although simple, this classi cation procedure is very e ective as we demonstrate in experiments we carried out and report later in this paper. 4
Based on the general coe cient similarity proposed by Gower [
Sij =
K K X mikmjksijk= max(X k=1 k=1
K mik; X k=1 mjk) (1)
In this equation, similarity scores sijk are computed for every atribute Ak that has value for both products pi and pj being compared. Also, mik (mjk) is 0 when the value for attribute Ak is missing for products pi (pj ) and 1 when it is not missing.
The speci c functions used for computing the similarity score sijk depend on the class of the attribute Ak. Recall from Section 3 that this class was already de ned. For each one of the four atribute classes we de ned an appropriate similarity function.
For the Numerical class, the similarity function is de ned as the absolute di erence between the values of the attribute in the two products, as shown in Equation 2. implying that objects having the same value get a similarity score of 1 and 0 otherwise.
N sijk = 1
jvik vjkj max (vik; vjk); C sijk = 1(vik = vjk) where vik and vjk are, respectively, the values of the attribute k for products pi and pj .
For the Categorical class, the similarity function is de ned as (2) (3) siMjk = jvik \ vjkj
jvik [ vjkj In this case vik and vjk denote the sets of individual categorical values composing the actual values. For instance, vik would be fAuto, On, O , Slow Syncro, : : : g for the attribute Flash Modes in the camera Sony W830 of Figure 1.
For the Dimensional class, the similarity function is the normalized euclidean distance over the dimension values, as described in Equation 5.
For the Multicategorical class, the similarity function is computed using the
Jaccard coe cient [
D sijk = 1
D [X((vidk)0 d=1 (vjdk)0 )2] 21 (4) (5) (6) where D is the number of dimensions found in the values of the attribute and vidk is the value for dimension d in vik (the same applies to vjdk). For computing this function, each dimension is mean-centered and normalized using (vxdk)0 = ((vxdk) d)= d d and d are, respectively, the mean and the standard deviation of the set of values of dimension p in all values of atribute k, for the products in the category.
As a nal comment, it is worth noting that the general coe cient similarity
proposed by Gower [
In this section we report the results of experiments we performed to evaluate the attribute classi cation strategy presented in Section 3, and the similarity function described in Section 4. 5.1
For the experiments, we have used ve datasets provided by Neemu1, a company that develops search and recommendation technology for major e-commerce sites in Brazil. These datasets comprise ve di erent popular product categories, namely: Cameras, Camcorders, Laptops, Smartphones and TVs. The product 1 http://www.neemu.com descriptions available in these datasets often provide many attributes that are not related to the product characteristics themselves. For instance, attributes related to the packing of the products such as, packing dimension, package contents, etc., are very common. Thus, we disregarded these attributes in our experiments. In addition, we removed all atributes that are not found in at least 20% of the products in a given category. By doing so, we tried to increased the percentage of attributes that can be e ectively compared to calculate the similarity between products.
Table 1 compares the number of attributes originally available in each dataset and the nal number of attributes we considered in each category. Notice that, even though many attributes were removed, still the number of attributes considered is large to be handled manually by humans. This table also presents the number of distinct products available in each dataset.
In Table 2, we present the number of attributes in each of the classes of our taxonomy. This classi cation was carried out manually to be used as a golden standard. Notice that the large majority of the attributes are categorical. This trend was observed in all categories. Also, a single dimensional attribute was available in each category,. In Table 3, we summarize the results obtained with our attribute classi cation strategy. For this, we used the well known Precision, Recall and F-1 metrics. In this table, each line corresponds to the results obtained with attributes of a distinct classe, namely, \NUM" (Numerical), \CAT" (Categorical), \MCA" (Multicategorical) and \DIM" (Dimensionall).
As it can be notice, our strategy has correctly classi ed most of the attributes from all categories we tested. We obtained perfect classi cation in many cases and F-1 values equal or above 0.8 were obtained in all cases but one. This case is the Multicategorical class in the Laptops category, which has a single attribute (see Table 2), and our classi cation strategy missed it. In many cases, the value of some attributes eventually presented noise our cleaning operation was unable to identify and x Nevertheless, we believe the small number of failures does not compromise the e ectiveness of our strategy and, as will see next, did not harm the overall results of our method. 5.3
Evaluating the e ectiveness of the similarity measure we described in Section 4 proved to be a challenge by itself. Indeed, carrying out a thorough evaluation to obtain values of Precision, Recall and F-1 would require to compare hundreds of products, examining the values of tens of attributes, some of them very technical. Thus, we opted to evaluate our proposed similarity measure in a task close to its intended application. This task consists in taking a product given as input, using the similarity measure to compare this product with all others in the same category, and verifying if the k products deemed as the most similar are indeed similar to the input product, according to a human-based evaluation. The results are reported in terms of the precision considering these top-k answers, a metric often known as P@k. In our case we used k = 5, which is reasonable in terms of recommender systems.
For each of the ve product categories, we randomly selected 10 products, which we refer to as query products, and, for each of them, we examine the 5 most similar products in the same category according to our similarity measure. Thus, a total of 250 pairs of products were manually evaluated. The results are presented in Figure 2.
In Figure 2, each graph corresponds to a product category and shows the P@5 values resulting from each of the 10 query products, along with the average of the ten values. Our similarity measure led to P@5=1 in 22 out of the 50 query products. Only in 8 cases, the P@5 values were below 0.5. In all categories, the average of P@5 values was around 0.75. An average above 0.8 was observed for the TVs category. Notice that the very low P@5 values obtained for some queries (e.g., 0 for query 1 in Smartphones or 1 for query 9 in Cancorders) does not necessarily implies that our similarity measure failed. For instance, it might happen that the query product has very few or none similar product in the catalog. In this case, our function just gave a low similarity score, but no similar products would appear among the top-5 answers. To solve this, a threshold on similarity score could be applied. However, there is no obvious way of imposing this threshold. Thus, we leave this study for future work.
In this paper we presented a preliminary study on the problem of nding products similar to a product given as input. This problem, although important for e-commerce sites, has been ill addressed so far both in the industry and in the academy. We described and evaluated a similarity function we have proposed for comparing products based on their attributes. Our function is generic in the sense that it deals di erent types of attributes occurring in products from distinct categories. Prior to its application, the function requires that each attribute has been classi ed into to a class that determines an speci c similarity function that handles this attribute. We demonstrate that this classi cation can be carry out by a simple but highly e ective strategy we proposed, which relies of regular expressions. Experiments we have performed with our similarity function with datasests with real products, revealed that it is accurate in nding similar products, achieving average F-1 values above 0.75 in 5 representative product categories.
Our plans for future work address two main aspects. First, we are working
on improving the e ectiveness of our function by considering that di erent
attributes may have di erent degrees of importante for users when comparing two
products of a given category. Thus, we are investiganting ways for capturing this
knowledge from the user and using it to improve our function. For this, we have
been working on machine learning techniques, which require training from user
data. Thus, the second aspect we are currently addressing is on how to obtain
training data without requiring users to label instances speci cally for this
problem. Another interesting future work we plan to address is considering additional
similarity functions for attributes. For instance, in the case of categorical data
it is worth investigating the metrics studied in [