This paper presents IntelliMeal, a case-based reasoning (CBR) system for recommending recipes. The main focus of the system is customizing recipes to a given user query re-using the domain knowledge of a CBR system within adaptation rules. In this work we implement a CBR system that works with a limited case base (21 recipes) and increase the amount of recipe recommendations by adapting these recipes using addition, creation, substitution, suitability and title name customization rules.
This paper presents IntelliMeal1, a case-based recipe recommendation system
addressing the open challenge of the 2017 Computer Cooking Contest (CCC).
Since its initialization in 2008, the competition has been running almost every
year with minor adjustments. Several research groups have contributed to the
CCC over the years using information retrieval, information extraction and
semantic technologies along with Case-Based Reasoning when developing recipe
recommendation systems. The researchers have contributed with various
approaches to the task. Four of the more in uential systems are Taaable [
IntelliMeal is a knowledge engineering heavy system utilizing Case-Based
Reasoning (CBR)[
The case base consists of twenty-one sandwich recipe cases, hierarchical
taxonomies and a set of adaptation rules. The taxonomies are separated into nine
attributes or ingredient categories. Each taxonomy de nes similarities between
a restricted set of ingredients that belong to the given ingredient category. The
taxonomies are extended versions of the taxonomies used in CookIIS [
The paper is structured as follows: Section 2 explains an adapted version of
the CBR cycle [
IntelliMeal implements all steps of the CBR cycle, but it includes a second retrieval phase after an ephemeral case base including temporarily adapted cases as been created. This allows IntelliMeal to assess the similarity of cases that have been modi ed based on desired and undesired ingredients speci ed in the query.
Figure 1 shows an overview of our implemented version of the cycle. Step 1 The problem presented is the user query, which consists of desired and undesired ingredients. As gure 1 illustrates, the query is split in two: One undesired query containing the undesired ingredients and one desired query containing the desired ingredient.
Step 2 Our modi ed version of the CBR cycle involves two retrieval steps. The rst process is the case base retrieval. It involves retrieving the cases from the case base with the highest similarity score to the user query. Hence, cases with the best starting point to end up in successful recipe recommendations. As the retrieval method employed in IntelliMeal is an important feature of the system, it is explained more detailed in section 2.1. The retrieved cases are further copied. The retrieved original cases are kept for later use while considering the copied versions in the reuse step.
Step 3 The reuse step is the most comprehensive step of the cycle. It is also the main focus of IntelliMeal. Hence, it is explained more detailed in section 2.3. The goal is to customize cases (i.e. recipes) so that they better t the user query. However, with restrictions to avoid distasteful recipe results. As gure 1 illustrates, domain knowledge, rules, and the queries are used in the adaptation process. The result from the reuse step is adapted versions of the cases in the case base, further referred to as adapted instances. Step 4 The strategies used to set up an ephemeral case base mainly concerns using the undesired query to discard cases that are not satisfactorily. As gure 1 illustrates, both original cases from the case base as well as the adapted instances resulting from the reuse step takes part in the setup. The result is an ephemeral case base containing a selection of both types.
Step 5 The ephemeral case base retrieval involves comparing the cases in the ephemeral case base to the desired query. The result is a mixture of original cases and adapted instances, together with their resulting similarity score. Step 6 The revision step involves getting feedback on recommendations. More speci cally, the user gets the opportunity to con rm that adapted recipe recommendations are tasty.
Step 7 Only when an adapted recipe is con rmed tasty, the recipe instance will be added as a case to the case base together with the original cases. This refers to the retain step of the cycle. 2.1
Our retrieval method compares one query instance to all cases in a case base. The method considers one case at a time. It iterates through every attribute in the query and ignores attributes that are unde ned. The rest of the attributes are considered valid and takes part in the similarity calculation.
For each valid attribute, the similarity between the query attribute and the corresponding case attribute is calculated by using the taxonomies. Then, they are weighted with a con gured attribute weight. When calculating the total similarity between the query and a case, the attribute similarities are summed up and divided by the number of valid attributes. The retrieval method returns all cases ordered by their retrieved similarity score. 2.2
A rule engine was created for this speci c system as a set of adaptation rules. There are two rule formats: Simple Rule and Substitution Rule.
If one rule were to speci cally target only one ingredient, many rules would have to be written. Therefore, the rule engine is also able to consider all children of the ingredient. The children are fetched from the taxonomies. This enables one single rule to apply to hundreds of ingredients. For example, consider a rule saying any type of meat can substitute for any type of sh. These ingredients have 191 and 74 ingredients, respectively. Hence, this one rule will form 14 134 various combinations.
However, the functionality to ignore the children of a speci c ingredients was also implemented. This can be used by writing a * after the ingredient. With this, a rule containing meat* and sh* would only form one combination.
Rule requirements refer to the ingredients in the recipe that have to be present for the rule to be valid. There can be zero or as many requirements as desired for a rule to re. One requirement is satis ed if the recipe considered contains either the stated ingredient requirement or one of its children. A rule is valid when all requirements are satis ed.
Simple rules are the most basic rule type used in IntelliMeal. Based on given set of conditions (requirements req) an action is taken. For this rule type, all requirements have to be satis ed, before the rule can be applied. req1; :::; reqn ! ingredient (1)
This rule type is used to create addition rules, deletion rules, suitable rules and title rules.
Substitution rules are more complex since they take both requirements and the ingredient to be substituted into account. Rules described by equation (2) takes the user requirement together with an ingredient to be substituted and the action describes which new ingredient can be included given a conditions, e.g. the existence or non-existence of other ingredients.
req1; :::; reqn + ingredientold ! ingredientnew + condition
An example for this rule is if the recipe contains tuna, any type of undesired supplements can be substituted out in favor for mayonnaise.
Furthermore, equation (3) shows a bidirectional rule that allows forward and backward substitutions: req1; :::; reqn + ingredientold ! ingredientnew 2.3
In general, the reuse step involves three steps illustrated in gure 2: 1) Adaptation with the undesired query. 2) adaptation with the desired query and 3) suitable adaptation. The overall goal with the two rst adaptation processes is to customize the retrieved recipes to better t the user query. Suitable adaptation considers the modi cations applied in the two rst processes, examining whether more substitutions are necessary with the goal of making the recipe as a whole successful. (2) (3)
Adaptation with the undesired query involves getting rid of undesired ingredients, while the goal with adaptation with the desired query is to add desired ingredients to recipes. These two adaptation processes are very similar. The di erence is their opposite goals. Both processes loops through undesired/desired ingredients, respectively. For each ingredient, three substitution methods are considered: 1) Simple deletion/addition, respectively, 2) substitutions, and 3) similarity substitutions.
Deletion and addition rules are written in the simple rule format. When given requirements are satis ed, an ingredient may be deleted from/added to the recipe. Adaptation with the undesired query is carried out by deletion rules, while the adaptation with the desired query uses the addition rules.
Substitution rules are written in the substitution rule format. When some requirements are satis ed, a modi cation may be done to a recipe ingredient. That means, one ingredient is removed from the recipe and one ingredient is added. For adaptation with the undesired query, an undesired ingredient is removed and a substitution alternative is found. For adaptation with the desired query, a desired ingredient is added and a substitution o er within the recipe is found.
Similarity substitutions are the last type of substitutions considered. These substitutions are based on taxonomies. The system fetches ingredients that are similar to the ingredient that is considered. For adaptation with the desired query, the goal is to nd an ingredient that is similar within the recipe, then remove this and add the desired ingredient. For adaptation with the undesired query, a similar alternative is added, and the undesired ingredient is removed. However, the substitution only goes through if an adaptation threshold is satis ed. A threshold is set for each ingredient category. It de nes how similar the substitution alternative and the considered ingredient needs to be for the substitution to go through.
After the undesired and desired adaptation process, the style of the recipe may have changed. The idea with suitable adaptation is to make the new ingredients t the recipe better. The process starts by iterating all modi cations of the recipe instance. For every new ingredient, the system checks whether a so-called suitable rule applies. Suitable rules are of the type Simple Rule, which means that there may be requirements for the rule to be valid. Both addition rules, substitution rules, and similarity substitutions are considered in the process. If several suitable ingredients are found, a random one among them is chosen.
Suitable adaptation also involves modifying titles. When ingredients are replaced by new ingredients in a recipe, the recipe title may be out of context. Two types of adaptation methods were implemented to rename the title to better t the new recipe. First, the system considers pre-de ned rules. For the record, the rules are called title rules and are of the type Simple Rule. The rule may have requirements to be valid. If no speci c rules on ingredients apply, a di erent approach is considered. The method focuses on the previous title, and apply to titles containing ingredients. The system aims to match the content of the title with the substitutions that are carried out. If substituted ingredients are found in the title, the same substitutions are carried out in the title. 3
Several evaluation methods were assessed to conduct the system's evaluation and three evaluation goals were set: 1) evaluate the calculated similarity score, and hence, the order of the suggested recipes, 2) show that adapted recipes better match the user query, and 3) evaluate whether the user can distinguish the computer created recipes from the human created recipes.
For the rst goal, we compared the system's ranking of recipes with ten test subjects ranking of recipes, for three given queries. The di erence was measured by calculating the di erence between each recipes ranking by the system and its ranking by the test subject. Per query, this resulted in a score from 0 to 25, where 0 represents equal ranking, and 25 represents the opposite order between the system and the test subject's ranking. On average, the result of this process showed that the discrepancy in ranking between the test subject's and the system was 5.52 out of 25. Also, the test subject's and the system shared 3.57 out of 5 recipes as their top 5.
For the second goal, a three step process were conducted: 1) doing a set of queries with the adaptation process turned o and note the similarity scores achieved for the top ve results, 2) doing the same set of queries with the adaptation process turned on, and 3) compare the similarity scores. The evaluation results showed that the average similarity score increases for all the test queries. Figure 3 illustrates the evaluation results. In the gure, the yellow bars show the average similarity score for the top ve results with no adaptation, while the green bars show the average increment for the top ve results with adaptation on. On average, the mean similarity score for the top ve recipes suggested per query increased with 0.32.
For the third goal, an online quiz was implemented2. The quiz displayed one recipe at a time for the user, and the user was to guess whether the recipe was human or computer created. The quiz ran for seven days (168 hours) and gathered in total 3414 responses distributed over 42 recipes. Figure 4 shows a confusion matrix of the quiz feedback. To clarify, true refers to an adapted recipe. Results showed that people guessed that a human had created the computer adapted recipes in 53.43% of the cases. Also, people recognized the original cases as created by a computer in 50.09% of the cases. This result reveals that most users are not able to distinguish the recipes from one another. 4
Earlier participants in the CCC have chosen various approaches for their systems. All four systems presented involve a hierarchical taxonomy. Some systems generate substitution rules from their taxonomy or cooking communities, while JaDaWeb has a table of ingredients that can substitute each other across categories. However, none explicitly de ne removal, addition or suitable rules like 2 www.intellimeal.no/botornot implemented in IntelliMeal. The rules and the similarity substitutions complement each other's weaknesses. Together, the components result in comprehensive adaptation of recipes. The rules employed in IntelliMeal have the ability to serve both speci c and general purposes. Incorporating taxonomies into the rule engine has enabled this. Also, rules may be speci ed for ingredients across attributes. The result is more radical modi cations of recipes, where the general style of a recipe may change completely.
In conclusion, IntelliMeal, extends the traditional CBR cycle by adding a second retrieval from an ephemeral case base populated with cases from a multilayer rule engine. The adaptation mechanism gets creative in the way that it exploits domain knowledge de ned in taxonomies to adjust similarity and to expand recipe case base. Therewith, we were able to achieve overall satisfying results with a limited case base. The measurable outcome of this project presented in section 3 is exceedingly satisfying. Adaptation of cases increased similarity scores for a given user query in all test cases, and humans had di culties distinguish computer and human created recipes from one another.