Selecting Ordering Policy and Items Classification Based on Canonical Correlation and Cluster Analysis
- Author: Nagasawa Keisuke, Irohara Takashi, Matoba Yosuke, Liu Shuling
- Organization: Nagasawa Keisuke; Irohara Takashi; Matoba Yosuke; Liu Shuling
- Publish: Industrial Engineering and Management Systems Volume 11, Issue2, p134~141, 30 June 2012
It is difficult to find an appropriate ordering policy for a many types of items. One of the reasons for this difficulty is that each item has a different demand trend. We will classify items by shipment trend and then decide the ordering policy for each item category. In this study, we indicate that categorizing items from their statistical characteristics leads to an ordering policy suitable for that category. We analyze the ordering policy and shipment trend and propose a new method for selecting the ordering policy which is based on finding the strongest relation between the classification of the items and the ordering policy. In our numerical experiment, from actual shipment data of about 5,000 items over the past year, we calculated many statistics that represent the trend of each item. Next, we applied the canonical correlation analysis between the evaluations of ordering policies and the various statistics. Furthermore, we applied the cluster analysis on the statistics concerning the performance of ordering policies. Finally, we separate items into several categories and show that the appropriate ordering policies are different for each category.
Inventory Management , Ordering Policy , Multivariable Analysis , Canonical Correlation Analysis , Cluster Analysis
Inventory management has been studied by many researchers and many practitioners. Tiacci and Saetta (2009) considered the relationship between forecasting and ordering policy with carrier capacity. Flores
et al. (1992) provided a matrix-based methodology formulticriteria ABC classification. Ramanathan (2006) proposed a weighted linear optimization model and classified inventory items by multiple criteria. Ernst and Cohen (1990) used cluster analysis to group similar items. Xaio et al. (2011) considered the importance of inventory items based on a loss rule. Berling and Marklund (2006) used linear regression technique to obtain approximate values of the induced backorder cost in a one-warehouse multiple- retailer system.
Many researchers started studying the lost-sales model. In the case of stock-out, Gruen
et al. (2002) revealed that only 15% of the customers who observe a stock out wait for the item to be on the shelves again. The other 85% of the customers decide to buy a different product, visit another store, or do not buy any product at all. According to Zipkin (2008), the cost deviations can run up to 30% when the lost-sales system is approximated by a backorder model. Janakiraman et al. (2007) compared the performance of optimal replenish-ment policies in lost-sales and backorder models. Huh et al. (2009) compared the performance of base-stock policies in lost-sales and backorder models. Levi et al. (2008) proposed a dual-balancing policy, in which the risk of ordering and holding are balanced. The authors proved that the expected total cost of this policy is at most twice the expected cost of the optimal policy. Bijvank and Vis (2011) classified the models based on the characteristics of the inventory system and they reviewed their proposed replenishment policies. Van Donselaar and Broekmeulen (2011) showed that the assumption which the lost sales system can simply be approximated by a backordering system if the target fill rate is at least 95%, which may lead to serious approximate errors.
Forecasting is never separate from inventory management. Liao and Chang (2010) investigated the effects of five demand forecasting methods, two inventory policies, and three lead times on the total inventory cost of a 3-echelon serial supply chain system. Tanaka
et al. (2012) proposed a decision support system that measures demand risks through a sales forecasting method, especially for new products. The system enables us to make the demand uncertainty controllable and gives the proper volume and timing guidance for daily reproduction of products.
One expects that if a single ordering policy is applied to many different items with different shipping trends, the inventory would be a mix of items which has frequent shortages and those having excess inventory. On the other hand, even in using multiple ordering policies, it is not easy to find an appropriate policy for many types of items, particularly when each has a different shipping trend. In addition, given a particular shipping trend, no method or criteria has been specifically developed for choosing the ordering policy which would be applied to these items. For these reasons, the proper relationship between different item trends and the appropriate ordering policies is still not clear.
Therefore, we propose a new algorithm for finding a reasonable ordering policy using the relationships between shipping trends. In particular, our proposed algorithm classifies items according to shipping trends in order to find a suitable ordering policy and in addition reveals the shipping trends which are most desirable. In this case, shipping trends refer to shipping statistics related to the criteria of the ordering policy, which are calculated from past shipping data.
In section 2, we introduce the proposed model. The model development includes the converting of actual shipping data to data matrices for analysis. The details of the experiment and results of analysis are reported in section 3.
This study considers the inventory management problem for items having different shipping trends. Here, the ordering policyis assumed to perform differently according to item trends. To achieve efficient inventory management, we would like to select a suitable ordering policy for each item. Hence, we want to know which item trends are suitable for determining the ordering policy, or to find out a reasonable method for grouping which items would be suitable and which is not for a given ordering policy.
This study assumes that there has not yet been revealed a difference in item shipping trends and also what the suitable ordering policy should be. For deciding an ordering policy, the ideal case would be when the suitable ordering policy could be determined using only a single statistic. However, such a statistic is not yet known. Thus, many researchers have sought to find such a statistic based on theoretical considerations. Similarly, practitioners are also struggling to find such a trend.
In this study, we calculate and use shipping statistics and evaluation values of ordering policies. A general algorithm is presented in the following section.
In the following steps, we repeatedly use multivariate data matrices, as shown in Table 2. Before we proceed to explaining the algorithm flowchart, we will first explain the premise behind multivariate data matrices.
Xdenote an n × vmultivariate data matrix, which represents the shipping statistics. There are nitems and vshipping criteria. xijgives the value for the jth shipping criteria of the ith item. For example, in Table 1, the daily shipping average is shipping criteria j, xij= ( di1 + di2 + ㆍㆍㆍ dim)/m. Let xj denotean n× 1 multivariate data matrix. These are the values of all items for the jth criteria.
Ydenote the multivariate data matrix, which represents evaluation values of ordering policies. Yis similar to the previously mentioned X. For example, yizpgives the value for the zth evaluation criteria of the pthordering policy of the ithitem. So, yzp denotes the n× 1 variable matrix of the zth evaluation criteria of the pthordering policy. Then, Yz denotesthe n × qvariable matrix ofthe zth evaluation criteria.
For analysis, since there are differences in the scale of numbers, we should standardize the evaluation value of each evaluation criteria and item in order to comparethe ordering policies. If
yikpis the value for the kthevaluation criteria of the pthordering policy of the ithitem, then let y’ikpbe the standardized evaluation value for this same combination, where the y’ikpare calculated using Eq. (1) and take valuesin [0, 1].
In order to evaluate multiple evaluation values, we apply a weighted linear mixed evaluate criteria. For example, if the criteria under evaluation are criterion
e, criterion f, … and the corresponding weights are we, wf, …, then ylpi, the value of the weighted linear mixed evaluation of criterion l, ordering policy p, and item i, is calculated by yilp= weyiep+ wfyifp+ … . Similarly, ylp = weyep+ wfyfp+ …. In addition, in order to find differences in ordering policy tends, we applied another weighted linear mixed evaluate criteria. Notably, the second evalu-ation value was calculated from the difference between evaluation values. For example, to analyze the trend difference for criteria cbetween policy fand policy g, naming the evaluation criterion d, it is calculated yd = ycf - ycg.
All the notations above, shown in Tables 1 and 2, are used in the sections that follow. The two large sets of variables, X and Y, are standard to, for example, the canonical correlation analysis and cluster analysis.
2.2.1 Canonical correlation analysis
The simplest known measure of relationships is the simple correlation coefficient between two variables. When interested in the relation between the set of variables
Xand a variable y, the multiple linear regression analysis can be used. However, in this study, we are interested in the correlation between the set of variables Xand the set of variables Y, that is, each object for comparison has multiple variables. Canonical correlation analysis is one method for determining the relationship between sets of variables. According to Basu and Mandal( 2010), this was initially developed by Hotelling (1936).
The canonical correlation analysis focuses on the correlation between the
ithlinear combination of the set of variables X(say Ui) and the ithlinear combination of the set of variables Y(say Vi). It determines the coefficients of the linear combination which make the correlation between Ui and Vi the highest possible, whereas the correlation between both Ui and Vi and each of Ui-1, …, U1, Vi-1, …, V1 is zero (i.e., uncorrelatedness of other combinations is a constraint on the optimization). The pairs of linear combinations are called the canonical variables, and their correlation coefficients are called the canonical correlation coefficients. In addition, the correlations between a linear combination of the set of variables and the set of variables themselves are called canonical loadings. The canonical correlation analysis is also useful for determining how many dimensions are needed to account for the relationship.
In the present study, we apply the canonical correlation analysis to the set of shipping statistics
Xand the evaluation values Y. In addition, we expect to find relationships between the sets of linear combinations of variables. This would allow us to sort shipping statistics into those which are significant and those which are not.
2.2.2 Cluster analysis
The cluster analysis is an exploratory technique. Cluster analysis techniques themselves can be broadly grouped into hierarchical clustering and non-hierarchical clustering. We here apply hierarchical clustering using the Ward method, which is a frequently used method.
In this study, we separate items into some number of clusters from xs, the shipping statistics, according to their relationship to
Yi, the evaluation value of the criteria iwhich the practitioner is focused on. Then we search for trends in shipping statistics and evaluation values of ordering policies.
Now, we calculate shipping statistics from actual shipping data and generate evaluation values of ordering policies from the simulation results. Then, we apply canonical correlation analysis to the set of shipping statistics and the evaluation values of the ordering policies. From these results, we can obtain the shipping statistics which exhibit a relatively strong relation with evaluation values of ordering policies. Clustering the items by these shipping statistics, we can group items according to a particular trend in the evaluation values of ordering policies.
In this study, we used SPSS ver. 19.0 (IBM Co., New York, NY, USA), a data mining and statistical analysis software package, for the canonical correlation analysis and cluster analysis, as well as the multiple linear regression analysis. The experimental environment is an Intel® Core™ 2 Duo CPU, E8400 at 3.00 GHz, 4.00 GB RAM; we did not need long computational times for our algorithm.
We used two ordering policies for this study: the so-called
(s, Q)policy and the (R, s, S)policy. There are two reasons why we choose these two ordering policies. First, these policies are the most commonly used ordering policies for inventory control. Second, the company that provided actual shipment data for us set the parameters and applied these ordering policies. So, we focused on these policies.
(s, Q)policy, when the inventory position declines to or below the reorder point s, a batch quantity of size Qis ordered. In this case, we never go beyond the position s + Q.
(R, s, S)policy, at the end of each time period of length R, if the inventory position declines to or below the reorder point s, order an amount equal to the difference between the order-up-to level Sand the current inventory position. In the special case in which s= S- 1 , we call this the (R, S)policy.
In this study, we use
(s, Q)and (R, S)policies. In particular, s, Q, and Sof the ithitem in the tthdate period are denoted by si,t, Qi,t, and Si,t, which are calculated as follows: Qi,t: ordering quantity of item ion the tth day si,t: reorder point for item ion the tth day Si,t: order-up-to level of item ion the tth day k: safety coefficient di,t: shipments of item ion the tth day R: periodic review term LT: lead time SP: sampling period I: the set of items T: the set of days C: the set of days that the parameter is recalculated (C ⊆ T ) D: the complement of Cwith respect to T
If practitioners would like to manage items precisely, it is necessary to shift ordering policy by some reason, like demand forecasting or seasonality. But, because we would like to find the relation between the shipping statistics values and the evaluation values of ordering policies, we did not shift ordering policies in the example applied.
In numerical and simulation experiments, we used the actual data of a distribution company’s shipments for the previous year. These items are categorized broadly into 11 varieties, for example, beer, Japanese sake, foods, and spices. In these categories, there are no obvious characteristic trends. So we selected about 5,000 items to be analyzed from about 7,000 items total. The other 2,000 items were less frequently shipped, and in smaller amounts. We applied our algorithm to the selected 5,000 items, those considered analyzable.
(s, Q)and (R, S)policies from the shipping data, which were used to calculate shipping statistics. We set common parameters between the set of (s, Q)and (R, S)policies, as described below. Safety coefficient kwas set to 1.96, sampling period SPwas set to 2 months, and lead time LT, the time between when the items are ordered and when they are delivered, was set to 2 days.
We compared the performance between five ordering policies: the
(s, Q)policy and the (R, S)policies for Requal to 1, 2, 3, or 4 week. For simulation, we used the first 2 months of shipping data for calculating initial values of si,t, Qi,t, and Si,t. For the (R, S)policies, Si,twere recalculated periodically, where as si,tand Qi,twere fixed at the initial calculation values.
From the results of the five ordering policy simulations, we define
Y πas the set of evaluation values, where Y πis calculated as Y π= [ y π, (s, Q), y π, (R= 1week, S), …, y π, ( R= 4weeks, S)]. y π, pis the set of linear combinations for ordering policy p, of the standardized evaluation of shortage time, y β, p, and the standardized evaluation of total daily stock, y γ,p. We define y π,pas follows:
We take the weights of shortage time,
wβ, as 0.5, and the weights of total daily stock, wγ, as 0.5.
Similarly, we define
Yδ as the difference between evaluation values for criteria πof policy fand another policy using yδ g = y πf - y πg. If yiδ g is positive, then ordering policy gis more suitable than ffor item i. In the present study, we compare the evaluation value of criteria
(s, Q)and (R, S). Thus, Yδ are calculated as follows:
We then apply canonical correlation analysis to the set of shipping statistics
Xand the set of evaluation criteria, using either Y πor Y δ. Next, cluster analysis is applied for all items using the shipping statistics, which are with respect to the ordering policy.
The results of the canonical correlation analysis are listed in Table 3. We apply the canonical analysis to the sets (
σ, π) and ( σ, δ), Here, σ is the set of shipping statistics and πis the set of evaluation values. The set δis the set of differences in evaluation value. This notation is also used in Tables 4-6.
Conducting the analysis shown in Table 3, we obtained the five canonical variables, ？π1, ？ π 2, ？π3 , ？ π 4 and ？π 5 from the canonical analysis of the sets
(σ, π )and four canonical variables, ？δ 1, ？δ 1, ？δ 2, ？δ 3 and ？ δ4 from the canonical analysis of the sets ( σ , δ). We extracted as canonical variables only those for which the canonical correlation coefficient was at least 0.20, the value at which variables can be said to have a weak relation in terms of one group being affected by the other groups. Here, in addition to the canonical correlation coefficients, we must consider the absolute value of canonical loadings. If these are large in absolute value, then the statistic indicates a strong effect on the other groups. For example, in Table 3, if ？π 1 in Xare listed in order of the strength of their relation to Y, then the canonical loadings are σ 6 = 0.751, σ 13 = -0.695 , σ 11 …. So, σ 6, which is the average of between the date of shipping (BDS), BDSAvg is the most significant statistic for the first canonical variable, ？π 1 of Y. And, if ？ π 1 in Yare listed in order of their relation to X, these are π 5 = -0.732, = -0.558 , …. So, π 5, which is evaluation value (EV; R = 4 week, S), EV of (R, S)policy when R= 4 weeks, and π1, which is EV (R = 1 week, S), EV of (R, S)policy when R = 1 week, were strongly affected by the first canonical variable, ?π1 of X. Thus, either π 5 or π 1 and σ 6, or possibly σ 13, are strongly related.
The results of correlation analysis are listed in Table 4. In the Table 4, (σ6, π 2) represents the correlation of the average of BDS and EV of ordering policy
(R, S)for R= 1 week, which is 0.04. It can seem from Table 3 that there is no relation between the shipping statistics and ordering policy.
The last result is interesting in light of a comparison of Tables 3 and 4. Table 4 gives simple correlations of shipping statistics and evaluation values of ordering policies. By comparing these Tables, it is possible to say that no individual variables and evaluation values of an ordering policy clearly relate to each other. But individual groups of statistics and groups of evaluation values do. A comparison of Tables 3 and 5 reveals similar aspects, where Table 5 shows the results of the multiple linear regression analysis. For example, in Table 5, all entries of the σ13 and σ11 rows indicate a lack of a strong relation, whereas Table 3 does indicate a strong relation.
We, therefore, conjecture that searching for relationships between the set of shipping statistics and the ordering policies would produce results different than the other analyses.
By using cluster analysis, we separate all items into two groups, also called clusters. In Table 6, the averages of each statistic of each separated group, columns χα1 and χα2, and the averages of each statistic of all items, column χαall, are shown. In the cluster analysis, we use the Ward method and calculate distance of each itemfrom the statistics σ6, σ11, and σ13, those for which canonical correlation analysis results showed a relationship between shipping statistics and evaluation values of ordering policies in the canonical variable ?π1, the first canonical variable of canonical analysis to the sets (σ, π). The signs " + " and "?" to the right of the values for columns χα1 and χα2 are relative to χαall. For example, the value of (σ6, χα1) is higher than (σ6, χαall). Thus, the sign to the right, separated by a dashed, is " ." From Table 6, we can observe that there are trends within each cluster.
From the results shown in Table 3 and the σ rows in Table 6, it is possible to make predictions. For example, from looking at the ?π1 column of Table 3, let us consider σ6 , σ11 , and σ13 , which were the top three in terms of the strength of their relation in
Xto Yof column ?π1 In Table 3, the values of those σ’s , (σ6, ?π1), (σ11, ?π1), and (σ13, ?π1), are greater than zero, less than zero, and less than zero, respectively; thus, (σ6, ?π1), (σ11, ?π1), and (σ13, ?π1) correspond to ( , -, -). As another example, in the χα1 column of Table 6, the signs of (σ6, χα1 ), (σ11, χα1 ), and (σ13, χα1 ) are (+, -, -). Therefore, we can predict that (π1, χα1 ) and (π5, χα1 ) in Table 6 will follow the same tends as that represented in the ?π1 column of Table 3, that is, they will be less than the average.
In a similar way, in the ？π2 column of Table 3, let us consider σ8, σ12, and σ14, which were the top three in terms of the strength of their relation in
Xto Yof column ?π2 In Table 3, all three values were less than zero. In the χα2 column of Table 6, (σ8, χα2 ), (σ12, χα2 ), and (σ 14, χα2 ) correspond to (-, -, -). From these signs, we can predict that (π2, χα2 ) and (π3, χα2 ), and possibly (π4, χα2 ) and (π1, χα2 ) in Table 6, will follow the same trend as that represented in the ?π2 column of Table 3, which means they will be less than the average.
Next, in the ？π1 column of Table 3, (π6, ?π1 ), (π11, ?π1 ) and (π13, ?π1 ) have signs ( + , -, -); in the χα1 column of Table 6, (π6, χα1 ), (π11, χα1 ), and (π13, χα1 ) have the same signs, ( + , -, -). Thus, as we would predict, (π1, χα1 ), (π2, χα1 ), and (π3, χα1 ) in Table 6 would follow the same trend as that shown in the ?π1 column of Table 3, meaning they would be less than the average.
For the case of comparing the evaluation value of a policy, if the average performance of ordering policies were not significantly different and there are complementarities, such as π1 and π3, then a difference in evaluation values of ordering policies implies which ordering policy should be applied. For example, from Table 6, applying the original policy of the evaluation value π1, which is a
(s, Q)policy, for group χα1 , and applying the origin policy of the evaluation value π3 for group χα2 is superior to using either one alone.
In this paper, our focus was on an approach to how to categorize items for a suitable ordering policy based on shipping statistics according to an evaluation of the ordering policy.
The practical point of view for real business of our proposed method is that practitioners can find, easily, significant shipping statistics values for selecting an ordering policy and can select an ordering policy for each item based on those significant statistics values efficiently. In the phase of finding significant shipping statistics values, we can look at the relation between the set of shipping statistics and the set of evaluation values of ordering policies.
In the numerical experiments, actual shipping data were used for calculating shipping data and simulating ordering policies. To evaluate the ordering policy, we considered the simulation results of
(s, Q)and (R, S)policies for each item. Then, we applied canonical correlation analysis between the set of shipping statistics and the set of evaluation values of inventory policies.
Using the results, we showed the effectiveness of grouping items by shipping statistics. Because groups, which are separated by cluster analysis, have different trends in evaluation values, it is possible to see how to select an ordering policy from shipping statistics. For example, by selecting ordering policies appropriately, we can reduce both shortage time and total daily stock.
Furthermore, we conjecture that the results of applying our method would vary according to the demand forecasting method, demand uncertainty, parameter settings for ordering policies, evaluation criteria settings and criteria of shifting ordering policy. By analyzing the interactions of these factors, we would like to extend the model beyond inventories to consider the global optimization of the performance of the entire supply chain management system.
[Figure 1.] Flowchart of proposed algorithm.
[Table 1.] Shipping data
[Table 2.] General variable of each item
[Table 3.] Canonical loadings (shipping statisticsevaluation value, differences between policy evaluation values)
[Table 4.] Correlation between shipping statistics and evaluation value of ordering policy
[Table 5.] Standard partial regression coefficients of multiple linear regression analysis between the set of shipping statistics and each evaluation value
[Table 6.] Statistical average of all items and two groups separated by the ward method (cluster analysis)