14.2.1 Collaborative Filtering Recommender Systems

CF is the most popular technique in which intelligence/interest of the crowd (large group of people) is used for recommending items. A rated dataset id used for considering the others interest. Then, nearest neighbor algorithm is applied which returns the results that are comparable to given input [9]. To implement CF based systems various approaches are used like: item-based [34], user-based [3], matrix factorization, and model-based approaches [24].

14.2.2 Content-Based Recommender Systems (CB)

Here, the recommendations are made by exploiting information of available items like ingredients and nutrition’s and user profiles. CB based systems make recommendations with recorded content, like books [28], articles, and web-pages [30]. Machine learning algorithms [27] and information Retrieval [4] are the two main approaches used for development of these systems.

14.2.3 Knowledge-Based Recommender Systems

Knowledge-based systems (KBS) systems were designed to address the issues of classical approaches like ramp-up problems [6]. These types of systems are mainly used where availability of rated data is scarce like apartments and financial services or where user defines his/her own requirements explicitly (e.g., “the color of the car should be white”). To implement knowledge based recommendation systems, two main approaches are used i.e. constraint-based recommendation [15] and case-based recommendation [5]. Critiquing-based recommendation is an alternative of case-based system that is widely used to make recommendation of particular things, and it produces users’ opinion in terms of reviews/critiques so that system accuracy [14] can be enhanced. For the implementation of KB systems following steps are followed:

• Requirement specification: To specify the requirements, users must interact with recommender systems.
• Repair of inconsistent requirements: In case, if a recommender fails to provide a possible solution, it must suggest a set of corrective measures that should propose alternatives/options to user requirements [16].
• Result presentation: For a single input, a multiple recommendations can be generated by the same system. System should present them in a rank wise order [18].
• Explanation: For every recommendation, system should provide a brief explanation to understand the relevance of a particular item [17].

14.2.4 Hybrid Recommender Systems

Hybrid recommender systems (HRS) are designed by using two or more than two different techniques so that limitations of one technique can overcome by using the approach. Author [32] describes “A hybrid system combining techniques A and B tries to use the advantages of A to fix the disadvantages of B”. For example, collaborative filtering method performs worse when a new-item is given as input. While CB approach provides a valid result as it makes recommendation for new things depending upon existing description. Author [7] proposed many hybrid approaches by merging both collaborative filtering and content based, including, switching, mixed, weighted, cascade, feature augmentation, feature combination and meta-level.

14.2.5 Context Aware Approaches

Various exploratory data analysis has proved that context to a specific food is important in food recommendation. Thus gender, hobbies, location, food availability, and time are identified as important variables. Importance of every variable can be understood by identifying what is actually lacking with respect to context like color of the food with respect to cooking duration and effect of that on the nutritional content of that specified food.

14.2.6 Group-Based Methods

Normally people eat and go for dinner in a group. Usually these activities are done together with friends, families or colleagues as food has a social role too. According to the psychology, the food choices made by a user is influenced by the society to one belongs. If the food recommendation system in such social context is addressed by human group recommendation system will be a boon for online purchasing. In such systems, a list of items is produced, for a group of people rather than for an individual user. Despite the pervasiveness of shared food consumption experiences, group based food recommender systems research has been limited.

14.2.7 Different Types of Food Recommender Systems

As presented in [29] two kinds of food recommending systems are widely used. Earlier, first type (type 1) systems were mainly used to recommend healthier recipes or food items. The type 2 recommender system recommends items which are acknowledged by nutritionist. Type 3 makes recommendations by considering both type 1 and type 2 criteria to make a balance between the user liking and healthy life style. Above defined systems are mainly intended for particular users.

Type 1: Based on user preferences

To implement these systems, the first step is to learn and characterize the user taste. Many researchers have proposed various food recommendation systems based upon user preferences [10, 20, 36], and/or combine with other approaches to enhance the accuracy of system [11, 25].

Here, a simple food recommending system was presented [10] which made recommendations for single user. In this, TF–IDF (Term Frequency– Inverse Document Frequency) filtering method was used to create user profile and to calculate similarity between a food and profile of user. As a knowledge base, healthy and standard food database taken from the United States Department of Agriculture 3 (USDA) is used. Food items are rated by each user as relevant or not relevant. Final recommendation is made by computing the similarity value. If the computed value is greater than a fixed threshold, then food item is suggested, else it will be ignored.

Author [20] proposed an approach which uses a CB algorithm to provide the recommendation by exploiting relevant data of matching ingredients present in this recipe. The recommendation procedure follows the following steps:

1. Divide an unranked target recipe (rt) into its ingredients.
2. Allocate the ranking score for every ingredient in the target recipe (rt).
3. Calculate the ranking score of the user (ua) for the target recipe (rt) (i.e., pred(ua, rt )) based on the average of ranking scores of all ingredients ingr₁, ..., ingr_j included in this recipe.
4. High predicted ranking score things will be recommended to user.

Type 2: Based on nutritional needs of users

Due to bad eating habits and imbalanced nutritional food, people are suffered from obesity and other dietary-related diseases like hypertension, diabetes etc. Nutritionists or dieticians suggests regular exercises and give diet plans to their patients as a treatment and preventive measures. Thus, Nutritionists become burdened with too many patients to manually advice personalized diet plan for every user. In this situation, food recommending systems can act as a smart nutrition consultation system.

Consider an example where recommendation system provides recommendations based upon users health problems as well as nutritional needs. Suppose a user enters his personal information like age: 54, gender: female, occupation: office work, physical activities: walking (15 minutes per day), disease: cardiovascular. Based on this information, recommender system performs the following steps:

1. An energy table is referred to find the total calories (in kcal) a user needs in a day. The total calories required to every person are computed as per age, gender, medical history and PAL (Physical Activity Level) value. �In the above example, user works in an office and does very less physical activity (only 15 minutes per day for walking). By referring energy table according to age, gender, medical history and physical activity assume, calories requirement for user is 2,300 kcal.
2. Filter foods with the total calories ≤2,300 kcal/day.
3. Filtered foods must be ranked in the increasing order of fat as user suffered from heart disease; so less oily food will be recommended to him.

Type 3: Based upon preferences of users and nutrition’s requirements of users

If recommendation system considers only one parameter i.e. either user preference or nutritional needs, the system may provide sub-optimal results. For example, if system considers only preferences given by user, then bad ingestion behaviour would be encouraged. On the other hand, if only requirement of nutrition’s are referred then users will not be attracted towards the designed. Thus, taking together, nutritional needs and preferences of users will give the best solution as users get more interesting results.

Consider an example in which system provides recommendations by considering both preferences of users and nutrition’s requirements. Suppose user enters personal information as follows: Age: 54, Gender: female, Occupation: office work, worker, Physical activity: walking (15 min per day), disease: cardiovascular, Favorite ingredients: potato. The proposed model makes recommendations considering both elements preferred by the user and user-related information as follows:

1. Finding the total calories a user needs by referring to the energy table. As the user works in office and is a sedentary worker (only 15 min per day for walking), so calories requirement for user is 2,300 kcal.
2. Extracting food which contains less than or equal to 2,200 kcal of calories, and consider favorite ingredient “potato”.
3. Rank the extracted food in the increasing order of fat as user has vascular disease.

Author [12] suggested two ways to balance users’ preferences and nutritional needs to incorporate nutritional aspects into consideration.

• The first approach balances between the food liked by user and healthy foods. To implement such models, following steps are followed:
  1. A prediction algorithm is used which recommends favorite recipes with computed probability greater than a predefined threshold.
  2. For each recipe extracted by model calories and fat per gram present in it is calculated.
  3. Foods with minimum fat or calories per gram will be recommended to user.
• In the second approach, rather than recommending individual foods, complete diet plans are proposed to user, which are generated.

14.3.1 Item Profile Representation

In CB systems, features and attributes are used to describe the items. Different similarity metrics are used to compute the similarity between the feature vectors such as Euclidian or cosine similarity. Though, human decision of matching two things often gives different weights to different features. Other than this, document frequency is also commonly used criteria.

14.3.2 Information Retrieval

Information retrieval (IR) is a data search and collection technology which includes crawling, processing and indexing of content, and questioning for content. Web crawling is the process by which information is collected from the web, so that ranking can be performed. It supports a search engine. The aim of crawling is to quickly and efficiently collect as much suitable web pages as possible, along with the link structure that interconnects them.

Due to the high complexity of food domain, several challenges arise while making recommendations. For making suitable suggestions, maximum number of food items/ingredients need to be collected. As foods/ingredients are typically associated with each other in a recipe instead of being consumed independently [20], recipes dataset needs to collect. From the recipes datasets, food/ingredient words are extracted and all unnecessary words are discarded.

14.3.3 Word2vec

Machines are not good in dealing with raw data. In fact, it’s almost impossible for machines to deal with any other data except for numerical data. So, representing text in the form of vectors is the most important step in all NLP processes.

Word2vec embedding’s the most popular method introduced to the NLP community in 2013. This embedding’s proved to be state-of-the-art for tasks such as word analogies and word similarities. Word2vec embedding’s also able to accomplish tasks such as King—man, woman ~= Queen, which is considered an almost magical result.

There are two alternatives of a word2vec model II—Continuous Bag of Words and Skip-Gram model. In this implementation, Bag of Words model is used. In a document, the word occurrence is represented by bag-of-words model and widely used in applications like natural language processing (NLP) and information retrieval (IR). The model deals checks the presence of word without looking at the context. It consists of two things: a vocabulary of known words and a measure of the presence of known words. Features from the text can be extracted to model learning applications.

14.3.4 How are word2vec Embedding’s Obtained?

A word2vec model is a simple neural network (NN) model with a single hidden layer. The task of this model is to estimate the nearby words for each and every word in a sentence. However, our objective has nothing to with this task. All we want are the weights learned by the hidden layer of the model once the model is trained. These weights can then be used as the word embeddings. Let me give you an example to understand how a word2vec model works. Consider the sentence below:

“Quantum computing researcher’s teleport data inside a diamond.”

Let’s say the word “teleport” is our input word. It has a context window of size 2. This means we are considering only the 2 adjacent words on either side of the input word as the nearby words.

Note: The size of the context window is not fixed; it can be changed as per our requirement.

Now, the task is to pick the nearby words (words in the context window) one-by-one and find the probability of every word in the vocabulary of being the selected nearby word.

14.3.5 Obtaining word2vec Embeddings

Now, let’s say we have a bunch of sentences and we extract training samples from them in the same manner. We will end up with training data of considerable size. Suppose the number of unique words in this dataset is 5,000 and we wish to create word vectors of size 100 each. Then, with respect to the word2vec architecture given in Figure 14.2:

• V = 5000 (size of vocabulary)
• N = 100 (number of hidden units or length of word embeddings)

The inputs would be the one-hot-encoded vectors and the output layer would give the probability of being the nearby word for every word in the vocabulary. Once this model is trained, we can easily extract the learned weight matrix W_{V × N} and use it to extract the word vectors.

14.3.6 Dataset

For the implementation of recommender system, dataset has been collected from “eightportions.com”. It contains approximately 125,000 recipes which are collected by web scrapping various web sites. A recipe is defined in a systematic system using different components like title of the recipe, a list of ingredients and measurements, directions for preparation, and a photo of the resultant dish.

This dataset is mainly used for machine learning algorithms as each recipe comprises many components, each of which provides additional information about the recipe.

14.3.7 Web Scrapping For Food List

As we have lot of unnecessary words in the recipe like cups, tablespoons, number etc. Thus, these words must be removed and only words related to foods needs to be retained. For this, knowledge of food words is mandatory. So, a list of food words has been created by using web scrapping from various websites.

14.3.7.3 Final Data Frame with Dishes and Their Ingredients

Final data frame with dishes and their ingredients have been created using word2vec method.

14.3.7.3.1 Similarity Metrics

Results are computed using three different matrices: Cosine Similarity, Hamming Distance and Jaccard Distance.

14.3.7.3.2 Cosine Similarity

Cosine similarity: It computes the similarity between user and item. This method is best when we have high dimensional features especially in information retrieval and text mining. The range of this is between −1 and 1 and there are two approaches:

• Top-n approach: In this, best ‘n’ suggestions are selected and here n is decided by user.
• Rating scale approach: In this a threshold is set and all the recommendations above threshold are recommended.

The cosine of two non-zero vectors can be computed by using the Euclidean dot product formula:

(i)

Thus, cosine similarity, cos(θ) is computed for given two vectors of attributes, A and B, is as follows

(ii)

14.6.1 User Information Challenges

14.6.2 Recommendation Algorithms Challenges

Any algorithm requires the following information to calculate nutritional recommendations for user [29].

14.6.3 Challenges Concerning Changing Eating Behavior of Consumers

Presently, due to inappropriate eating habits, many people suffer from health problems [37]. For example, in relation with their level of physical activity, most people eat too much food and eventually become obese. While others limit extreme intake of nutrition, this leads to malnutrition. Therefore, recognizing the eating habits of consumers and persuading them to alter eating patterns in positive ways is one of the main functions of food recommender systems. This is a huge challenge for food recommendation system; however, as eating activity is influenced by many factors. In order to inspire consumers to follow healthy diet, food recommendation systems should incorporate health psychology theory. Using one simple change at a specific time, the first approach can be used until consumer behaviour becomes usual [32]. Further approach can be applied for food recommender system is to compare the ideal nutrient. Consumers can find the optimal diet structure from reputable data based on age and physical activity level (e.g., USDA, DACH) and then analyze the food they eat with what is prescribed in [37]. The comparison approach is also proposed to provide consumers with potential dietary changes in paper proposed [27].

14.6.4 Challenges Regarding Explanations and Visualizations

Visualizations and explanations have a vital role in recommendation systems as consumers’ confidence increases in the outcome of decisions [38]. Explanations are even more important in the healthy food domain as they not only increase confidence in recommenders but also encourage consumers to eat healthy foods and improve their eating habits. For the purpose of this, it makes sense that food recommendation systems explanations explain how adequacy results are obtained [11]. In addition, a specific explanation of food items such as a nutritional value table for a recipe must be involved in a manner that focuses the health of a particular consumer food.

1. 1. Aberg, J., Dealing with Malnutrition: A Meal Planning System for Elderly, in: AAAI Spring Symposium: Argumentation for Consumers of Healthcare, pp. 1–7, 2006.
2. 2. Aizawa, K., De Silva, G.C., Ogawa, M., Sato, Y., Food log by snapping and processing images, in: 2010 16th International Conference on Virtual Systems and Multimedia, IEEE, pp. 71–74, 2010.
3. 3. Asanov, D., Algorithms and methods in recommender systems, Berlin Institute of Technology, Berlin, Germany, 2011.
4. 4. Balabanović, M. and Shoham, Y., Fab: content-based, collaborative recommendation. Commun. ACM, 40, 3, 66–72, 1997.
5. 5. Bridge, D., Göker, M.H., McGinty, L., Smyth, B., Case-based recommender systems. Knowl. Eng. Rev., 20, 3, 315–320, 2005.
6. 6. Burke, R., Knowledge-based recommender systems, in: Encyclopedia of library and information systems, vol. 69, Marcel Dekker (Ed.), pp. 180–200, 2000.
7. 7. Burke, R., Hybrid recommender systems: Survey and experiments. User Model. User-ADAP, 12, 4, 331–370, 2002.
8. 8. Burke, R., Felfernig, A., Göker, M.H., Recommender systems: An overview. AI Mag., 32, 3, 13–18, 2011.
9. 9. Ekstrand, M.D., Riedl, J.T., Konstan, J.A., Collaborative filtering recommender systems. Found. Trends^® Hum.–Comput. Interact., 4, 2, 81–173, 2011.
10. 10. El-Dosuky, M.A., Rashad, M.Z., Hamza, T.T., El-Bassiouny, A.H., Food recommendation using ontology and heuristics, in: International Conference on Advanced Machine Learning Technologies and Applications, Springer, Berlin, Heidelberg, pp. 423–429, 2012.
11. 11. Elahi, M., Ge, M., Ricci, F., Fernández-Tobías, I., Berkovsky, S., David, M., Interaction design in a mobile food recommender system, in: CEUR Workshop Proceedings, CEUR-WS, 2015.
12. 12. Elsweiler, D., Harvey, M., Ludwig, B., Said, A., Bringing the “healthy” into Food Recommenders, in: DMRS, pp. 33–36, 2015.
13. 13. Evert, A.B., Boucher, J.L., Cypress, M., Dunbar, S.A., Franz, M.J., Mayer-Davis, E.J., Yancy, W.S., Nutrition therapy recommendations for the management of adults with diabetes. Diabetes Care, 37, Supplement 1, S120–S143, 2014.
14. 14. Felfernig, A., Biases in decision making, in: Proceedings of the First International Workshop on Decision Making and Recommender systems (DMRS2014), Bolzano, Italy, September 18–19, 2014, vol. 1278, CEUR Proceedings, pp. 32–37, 2014.
15. 15. Felfernig, A. and Burke, R., Constraint-based recommender systems: technologies and research issues, in: Proceedings of the 10th International Conference on Electronic Commerce, ACM, p. 3, 2008.
16. 16. Felfernig, A., Friedrich, G., Jannach, D., Zanker, M., Developing constraint-based recommenders, in: Recommender systems handbook, pp. 187–215, Springer, Boston, MA, 2011.
17. 17. Felfernig, A., Zehentner, C., Ninaus, G., Grabner, H., Maalej, W., Pagano, D., Reinfrank, F., Group decision support for requirements negotiation, in: International Conference on User Modeling, Adaptation, and Personalization, pp. 105–116, Springer, Berlin, Heidelberg, 2011.
18. 18. Felfernig, A., Hotz, L., Bagley, C., Tiihonen, J., Knowledge-based configuration: From research to business cases, Elsevier, Newnes, 2014.
19. 19. Felfernig, A., Stettinger, M., Ninaus, G., Jeran, M., Reiterer, S., Falkner, A.A., Tiihonen, J., Towards Open Configuration, in: Configuration Workshop, pp. 89–94, 2014.
20. 20. Freyne, J. and Berkovsky, S., Intelligent food planning: personalized recipe recommendation, in: Proceedings of the 15th International Conference on Intelligent User Interfaces, ACM, pp. 321–324, 2010.
21. 21. Ge, M., Elahi, M., Fernaández-Tobías, I., Ricci, F., Massimo, D., Using tags and latent factors in a food recommender system, in: Proceedings of the 5th International Conference on Digital Health 2015, ACM, pp. 105– 112, 2015.
22. 22. Hoxmeier, J.A. and DiCesare, C., System response time and user satisfaction: An experimental study of browser-based applications. AMCIS. 2000 Proc., 347, 140–145, 2000.
23. 23. Knowler, W.C., Barrett-Connor, E., Fowler, S.E., Hamman, R.F., Lachin, J.M., Walker, E.A., Nathan, D.M., Reduction in the incidence of type 2 diabetes with lifestyle intervention or metformin. N. Engl. J. Med., 346, 6, 393–403, 2002.
24. 24. Koren, Y., Bell, R., Volinsky, C., Matrix factorization techniques for recommender systems. Comput., 8, 30–37, 2009.
25. 25. Kuo, F.F., Li, C.T., Shan, M.K., Lee, S.Y., Intelligent menu planning: Recommending set of recipes by ingredients, in: Proceedings of the ACM Multimedia 2012 Workshop on Multimedia for Cooking and Eating Activities, pp. 1–6, ACM, 2012.
26. 26. LeFevre, M.L., Behavioral counseling to promote a healthful diet and physical activity for cardiovascular disease prevention in adults with cardiovascular risk factors: US Preventive Services Task Force Recommendation Statement. Ann. Intern. Med., 161, 8, 587–593, 2014.
27. 27. Mankoff, J., Hsieh, G., Hung, H.C., Lee, S., Nitao, E., Using low-cost sensing to support nutritional awareness, in: International Conference on Ubiquitous Computing, Springer, Berlin, Heidelberg, pp. 371–378, 2002.
28. 28. Mooney, R.J. and Roy, L., Content-based book recommending using learning for text categorization, in: Proceedings of the Fifth ACM Conference on Digital Libraries, ACM, pp. 195–204, 2000.
29. 29. Mika, S., Challenges for nutrition recommender systems, in: Proceedings of the 2nd Workshop on Context Aware Intel. Assistance, Berlin, Germany, pp. 25–33, 2011.
30. 30. Pazzani, M.J., Muramatsu, J., Billsus, D., Syskill & Webert: Identifying interesting web sites, in: AAAI/IAAI, vol. 1, pp. 54–61, 1996.
31. 31. Phanich, M., Pholkul, P., Phimoltares, S., Food recommendation system using clustering analysis for diabetic patients, in: 2010 International Conference on Information Science and Applications, IEEE, pp. 1–8, 2010.
32. 32. Ricci, F., Rokach, L., Shapira, B., Introduction to recommender systems handbook, in: Recommender systems handbook, pp. 1–35, Springer, Boston, MA, 2011.
33. 33. Runo, M., FooDroid: a food recommendation app for university canteens. Unpublished semester thesis, Swiss Federal Institute of Theology, Zurich, 2011.
34. 34. Sarwar, B.M., Karypis, G., Konstan, J.A., Riedl, J., Item-based collaborative filtering recommendation algorithms. Www, 1, 285–295, 2001.
35. 35. Robertson, A., Tirado, C., Lobstein, T., Knai, C., Jensen, J., Ferro-Luzzi, A., James, W., Food and Health in Europe: A New Basis for Action (European Series No 96), WHO, 2004.
36. 36. Svensson, M., Laaksolahti, J., Höök, K., Waern, A., A recipe based on-line food store, in: Proceedings of the 5th International Conference on Intelligent User Interfaces, ACM, pp. 260–263, 2000.
37. 37. Snooks, M.K., Health psychology: Biological, psychological, and sociocultural perspectives, Jones & Bartlett Publishers, 2009.
38. 38. Tintarev, N. and Masthoff, J., A survey of explanations in recommender systems, in: 2007 IEEE 23rd International Conference on Data Engineering Workshop, IEEE, pp. 801–810, 2007.
39. 39. Teng, C.Y., Lin, Y.R., Adamic, L.A., Recipe recommendation using ingredient networks, in: Proceedings of the 4th Annual ACM Web Science Conference, ACM, pp. 298–307, 2012.
40. 40. Ueta, T., Iwakami, M., Ito, T., A recipe recommendation system based on automatic nutrition information extraction, in: International Conference on Knowledge Science, Engineering and Management, Springer, Berlin, Heidelberg, pp. 79–90, 2011.
41. 41. Van Pinxteren, Y., Geleijnse, G., Kamsteeg, P., Deriving a recipe similarity measure for recommending healthful meals, in: Proceedings of the 16th International Conference on Intelligent User Interfaces, ACM, pp. 105–114, 2011.