Dataset Prioritization - Ranking

The value of ranking deductions should be within the low to medium single figure range. We also recommend that the sum of all deductions should not exceed half of the Spread value, since otherwise the products will already with minor writing errors not be included in the search results. As deductions from multiple groups are cumulative, this can occur quite quickly. The use of too many rules is, therefore, not recommended.

Sometimes you may want to prioritise certain products over others. For instance, you may want to do this if your shop wishes to favour its own products over the comparable products from other manufacturers. The products from other manufacturers should not be excluded from the search results entirely. For example, these products should also be displayed if the product searched for is not available from your shop’s own brand. This is exactly where ranking comes into play.

As with Field Emphasis, Ranking rules take the form of a reduction in relevance. Thus 10% means “the record is 10% less relevant”, and not “the data record is 10% relevant”. It is not possible to increase the relevance. But reduction via a ranking rule is not relative like field emphasis, but rather absolute. 10% reduction actually means 10% less relevance. This way, a hit can be excluded from the search result if the spread is only 8%.
The configured record ranking value is deducted from the similarity value established by the search process. This influences the position of the products in the search result.

The record ranking is implemented by ranking rules. These rules allow you to select groups of records that are to be given the same ranking. They apply on the basis of the field content.

You can create any number of rules for each field. The rules are combined into groups. If several rules in one group could apply at the same time, only the first matching rule in the group is applied. This design allows you to combine semantically identical rules that query differing field values.

The following comparison operators are available for use in such comparative rules: is equal to, contains, is not equal to, does not contain, is empty, is not empty.

As mentioned at the outset, ranking is implemented by reducing the value of a record. This requires a different approach when creating the rules. Instead of creating a rule that increases the ranking of preferred records, you must create a rule that reduces the value of those records that are not preferred. There is a good reason as to why we adopted this inverse approach: There can be no products with a similarity of over 100%. If FACT-Finder finds 10 records with a similarity of 100%, and one record were to be improved by 1%, this improvement would have no effect (as there can be no 101% similarity). On the other hand, the reduction in value can ensure that 9 of these products are reduced in value by 1%. As a result, the product that you wish to prioritise, moves to the top.

Two rule groups are defined in the list above. Group A is intended to reduce the value of all products that do not belong to the shop’s own brand. Group B is intended to slightly reduce the value of all products that are considered to be accessories. This makes sense because generic searches, such as “iphone”, will then return the smartphone by Apple first, followed by its accessories. You can isolate accessories in three ways using this group. First, if the product is an accessory the “Accessories” field in the database will contain the checked value “true”. However, it may be that this field has been poorly maintained and does not contain “true” for all accessories. Therefore, the “Category” field is also checked for the word “Accessories” and the “Product name” field is checked for the word “for” (for example, “Headphones for iphone”).

If you like it more tidy, you can define multiple conditions for a depreciation within a single rule. For this, first create a rule and then click on Add condition. FACT-Finder will add a new row for an additional condition, which can be used just as the previous one. You can also set a condition which is different in a minor way. For example by using is equal to in the first rule and contains in the second one. Two or more conditions are connected with OR by FACT-Finder.

As mentioned above, only one rule from each group is applied, even if several match.

Rules from different groups are cumulative, however. Therefore, a product in the “Accessories” category and with the brand “Apple” – i.e. not the shop’s own brand – will be reduced in value by 3%.

Impact

Die Datensatzgewichtung beeinflusst die letztendliche Ähnlichkeit eines Datensatzes. Die Gewichtungen haben keine Auswirkung auf die Suchgeschwindigkeit.

Changing settings

You can modify ranking rules via the Ranking page in the Interface.

Alternative Ranking Possibilities

Additionally to the rather simple text comparison rules explained above, there are two more possibilities to influence the order of search results via the ranking. One option is raise dynamic values like revenue or sold amount as criteria. This is done by creating so called top seller rankings, where you don’t devalue a given top value and do a maximum devaluation for a given flop value. Values are calculated in between.

Such top seller rankings can use values from the feed as basis and also data from FACT-Finder tracking. Regardless of the data origin you can create rankings with such rules which not statically always return the same result order. And you can also bring the most successful products to the front – a good basis for conversion.

Additionally to number based rules, FACT-Finder also knows rules which work with dates. If you have many novelties and wish to place them at the front, then a time rank rule is ideal: In the feed, FACT-Finder receives the sale start date and the ranking rule calculates the space between the start date and the present date and devalues based on your settings. This way, new articles can e.g. receive a one week “safe space”.

Combining Ranking Rules

You will almost never only use a single ranking type, but rather two. With rules you can specifically influence the order of results without having to take a look at every search term. Which data should be emphasized in which way is usually determined in a ranking concept, because the positive and negative possibilities are equally extensive.

Recommendation

The value of ranking deductions should be in the low to medium single figure range. We recommend that the sum of all deductions should not exceed half of the Spread value. Otherwise the products with minor writing errors will not be included in the search results. As deductions from multiple groups are cumulative, this can occur quite quickly. The use of too many rules is, therefore, not recommended.